diff options
| author | firecoperana <xuqiaowei1124@gmail.com> | 2025-07-02 01:49:42 -0500 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2025-07-02 08:49:42 +0200 |
| commit | d5cd99f9c8f652452c8c5229b6ad95d3e80ef0ca (patch) | |
| tree | 94ed84ec2af95e57bcb079da85884f08341701ee | |
| parent | bce7697d64dc09d52dec468b7ed69c768967b8b6 (diff) | |
Merge vulkan code from mainline up to commit of 6/28/2025 (#563)
* Merge vulkan code from mainline up to commit of 6/28/2025
* Vulkan Optimizations and Fixes (#8959)
* Optimize Vulkan REPEAT performance
* Use Vulkan GLSL fused multiply-add instruction where possible
* Add GGML_VULKAN_PERF option to output performance data per operator
* Rework and fix Vulkan descriptor set and descriptor pool handling
* Fix float32 concat f16 shader validation error
* Add Vulkan GROUP_NORM eps parameter
* Fix validation error with transfer queue memory barrier flags
* Remove trailing whitespaces
vulkan : do not use tensor->extra (#9407)
* vulkan : do not use tensor->extra
This patch allows using the Vulkan backend with the RPC backend as
tensor->extra is no longer used.
Ref: #8536
* Adapt GGML_VULKAN_CHECK_RESULTS to extra removal (#2)
---------
Co-authored-by: 0cc4m <picard12@live.de>
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan : fix build (#0)
ggml-ci
Improve Vulkan shader build system (#9239)
* Improve Vulkan shader builds system
- Add dependency to vulkan-shaders-gen to rebuild shaders when changing the shader compilation utility.
- Add option to generate debug info for Vulkan shaders to provide shader source to Vulkan shader profiling tools
* remove not required self dependency
ggml : fix build break for the vulkan-debug (#9265)
- windows build : Ok.
- linux build : Ok.
Signed-off-by: Changyeon Kim <cyzero.kim@samsung.com>
vulkan: correctly report support for OP_CONT (ggml/946)
test-backend-ops fails because ggml_cont aborts
when invoked passing an unsupported type.
This commit makes ggml_cont tests pass
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
vulkan: add dryrun support to sin and cos ops (ggml/947)
sin and cos failed test-backend-ops because they
tried to dereference a context pointer that is null
on dry runs.
This commit prevents that segfault.
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
# Conflicts:
# ggml/src/ggml-vulkan.cpp
Overlap cmdbuffer creation and cmdbuffer execution in Vulkan backend by submitting smaller cmdbuffers early. (#9118)
* Overlap cmdbuffer creation and cmdbuffer execution in Vulkan backend by submitting smaller cmdbuffers early.
* fix compile issues
* Fix issues where the last submit wasn't executed or handled properly.
* remove trailing whitespace
* Repair GGML_VULKAN_CHECK_RESULTS
* Increase submit counter only if actual work has been submitted and increase submit count to 100.
* Fix some nodes are not checked with GGML_VULKAN_CHECK_RESULTS enabled.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
Enable use to the rebar feature to upload buffers to the device. (#9251)
vulkan : argsort barriers must be under uniform control flow (ggml/951)
a return before a barrier (that happens only in some threads in
a workgroup) leads to UB.
While the old code actually works on some devices,
it fails on some others (i.e. "smaller" GPUs).
BTW, I think it would be better to set specialization constants
when the graph is built, in that way the local workgroup
could be sized appropriately.
But it would take a lot of work.
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
vulkan : fix build for GGML_VULKAN_RUN_TESTS, add TFLOPS to log (ggml/961)
vulkan : multithread pipeline creation (ggml/963)
vulkan : mul_mat: fix UB with small warps (ggml/952)
When the device's warp size is less than 16,
it is possible for loadstride_a (mul_mm.comp:114)
and loadstride_b (mul_mm.comp:115) to be set to 0.
Because they are calculated as: the workgroup size,
multiplied by LOAD_VEC_* (which can be 1) and divided by 16.
And the workgroup size is set to be the same as the
warp/subgroup size.
The loadstride_* variables are used as increments in the
loops that populate the buffers used for the multiplication.
When they are 0 they cause an infinite loop.
But infinite loops without side-effects are UB and the
values of loadstride_* are known at compile time.
So, the compiler quietly optimizes all the loops away.
As a consequence, the buffers are not populated and
the multiplication result is just a matrix with all elements
set to 0.
We prevent the UB by making sure that the workgroup size
will never be less than 16, even if our device has a
smaller warp size (e.g. 8).
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
vulkan : retry allocation with fallback flags (whisper/2451)
Co-authored-by: Samuel Morris <samuel.morris@artlist.io>
vulkan : improve ggml_vk_create_buffer error handling (#9898)
vulkan: Fix newly added tests for permuted mul_mat and 1D im2col (#10226)
vulkan: Throttle the number of shader compiles during the build step. (#10222)
Fixes #9582
Spawning too many concurrent copies of glslc leads to "Failed to create pipes"
errors on Linux. This change applies the same throttling we use for
multithreaded pipeline creation.
# Conflicts:
# ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
vulkan: Optimize contiguous copies (#10254)
* tests: Fix memory bandwidth calculation for perf tests
Add a flops calculation for flash attention.
Add one GGML_OP_CPY perf test.
* vulkan: Optimize contiguous copies
Add a variant of the copy shader for when the tensors are contiguous. Avoid
the complex addressing calculations, and do four elements per invocation
to hide some other overhead.
Apply similar changes to the scale shader, since scale is always contiguous.
Add a "progress bar" for shader compiles.
# Conflicts:
# tests/test-backend-ops.cpp
vulkan: Use macros to make the mat mul pipeline creation more concise (#10259)
Also add vk_matmul_pipeline2 to hold f16/f32 accumulator versions of a
pipeline. This isn't really used yet.
vulkan: Optimize binary ops (#10270)
Reuse the index calculations across all of src0/src1/dst. Add a shader
variant for when src0/src1 are the same dimensions and additional modulus
for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that
have a fast path when the calculation isn't needed or can be done more
cheaply.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/acc.comp
ggml : vulkan logs (whisper/2547)
vulkan: Optimize some mat-vec mul quant shaders (#10296)
Compute two result elements per workgroup (for Q{4,5}_{0,1}). This reuses
the B loads across the rows and also reuses some addressing calculations.
This required manually partially unrolling the loop, since the compiler
is less willing to unroll outer loops.
Add bounds-checking on the last iteration of the loop. I think this was at
least partly broken before.
Optimize the Q4_K shader to vectorize most loads and reduce the number of
bit twiddling instructions.
Vulkan: Fix device info output format specifiers (#10366)
* Vulkan: Fix device info output format specifiers
* Vulkan: Use zu printf specifier for size_t instead of ld
vulkan: remove use of null initializer (#10372)
Seems like this isn't working for vulkan-over-metal when the array is sized
by a spec constant. Maybe a spirv-cross limitation?
vulkan: Optimize soft_max (#10301)
* vulkan: Optimize soft_max
Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.
Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.
Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.
* vulkan: Further soft_max optimizations
Restore the workgroup size of 512 case, use it for >1024.
Use unrollable loops for more iteration counts.
vulkan: further optimize mul_mat_vec using larger loads (#10387)
* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.
Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.
Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.
* vulkan: Add GLSL structure aliases for quant types to allow larger loads
In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.
* vulkan: use larger loads in q5_k and q6_k shaders.
Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.
* vulkan: use larger K step per iteration in mul_mat_vec.
Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.
The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.
Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.
vulkan: copy iq4_nl LUT into shared memory (#10409)
vulkan: predicate max operation in soft_max shaders/soft_max (#10437)
Fixes #10434
vulkan: Fix a vulkan-shaders-gen arugment parsing error (#10484)
The vulkan-shaders-gen was not parsing the --no-clean argument correctly.
Because the previous code was parsing the arguments which have a value only
and the --no-clean argument does not have a value, it was not being parsed
correctly. This commit can now correctly parse arguments that don't have values.
vulkan: fix group_norm (#10496)
Fix bad calculation of the end of the range. Add a backend test that
covers the bad case (taken from stable diffusion).
Fixes https://github.com/leejet/stable-diffusion.cpp/issues/439.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: optimize Q2_K and Q3_K mul_mat_vec (#10459)
vulkan: skip integer div/mod in get_offsets for batch_idx==0 (#10506)
vulkan: further optimize q5_k mul_mat_vec (#10479)
vulkan: Handle GPUs with less shared memory (#10468)
There have been reports of failure to compile on systems with <= 32KB
of shared memory (e.g. #10037). This change makes the large tile size
fall back to a smaller size if necessary, and makes mul_mat_id fall
back to CPU if there's only 16KB of shared memory.
vulkan: define all quant data structures in types.comp (#10440)
vulkan: get the first command buffer submitted sooner (#10499)
This is an incremental improvement over #9118 to get work to the GPU a bit
sooner. The first part is to start with a smaller number of nodes before
the first submit, and ramp it up to the current 100 nodes/submit. The
second part is to reduce the dryrun overhead for all the nodes that just
need to request descriptor space.
With these changes I get around 1-2% speedup on RTX 4070 combined with my
old Haswell-era CPU.
vulkan: Dynamic subgroup size support for Q6_K mat_vec (#10536)
* subgroup 64 version with subgroup add. 15% faster
scalable version
tested for subgroup sizes 16-128
* check for subgroup multiple of 16 and greater than 16
* subgroup sizes are always a power of 2 (https://github.com/KhronosGroup/GLSL/issues/45)
* force 16 sequential threads per block
* make 16 subgroup size a constant
vulkan: optimize and reenable split_k (#10637)
Use vector loads when possible in mul_mat_split_k_reduce. Use split_k
when there aren't enough workgroups to fill the shaders.
vulkan: Implement "fast divide" (mul+shift) for unary ops like copy (#10642)
vulkan: Add VK_NV_cooperative_matrix2 support for mul_mat and flash attention (#10206)
# Conflicts:
# ggml/src/vulkan-shaders/dequant_funcs_cm2.comp
# ggml/src/vulkan-shaders/flash_attn_cm2.comp
# ggml/src/vulkan-shaders/mul_mm_cm2.comp
Vulkan: VK_KHR_cooperative_matrix support to speed up prompt processing (#10597)
* Vulkan: Implement VK_KHR_cooperative_matrix support in the matrix matrix multiplication shader
* Improve performance with better q4_k and q5_k dequant and store unrolling
* Add Vulkan MUL_MAT and MUL_MAT_ID accumulator precision selection
* Rework mulmat shader selection and compilation logic, avoid compiling shaders that won't get used by device
* Vulkan: Implement accumulator switch for specific mul mat mat shaders
* Vulkan: Unroll more loops for more mul mat mat performance
* Vulkan: Add VK_AMD_shader_core_properties2 support to read Compute Unit count for split_k logic
* Disable coopmat support on AMD proprietary driver
* Remove redundant checks
* Add environment variable GGML_VK_DISABLE_COOPMAT to disable VK_KHR_cooperative_matrix support
* Fix rebase typo
* Fix coopmat2 MUL_MAT_ID pipeline selection
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: compile a test shader in cmake to check for coopmat2 support (#10713)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/ggml-vulkan/CMakeLists.txt
# ggml/src/vulkan-shaders/test_coopmat2_support.comp
Vulkan: fix NaN in tanh.comp with AMD proprietary driver on Windows (#10723)
* Vulkan: fix NaN in tanh.comp
* Faster NaN-free tanh
vulkan: fix compile warnings (#10731)
vulkan: disable spirv-opt for coopmat shaders (#10763)
There are some bugs in the 1.3.296 SDK, so disable this. It isn't strictly
necessary anyway.
Add missing dependency on vulkan-shaders-gen, so shaders get recompiled when it
changes.
Fix coopmat support reporting when glslc doesn't support NV_coopmat2.
vulkan: dynamic subgroup size for the remaining k quants (#10745)
* q5_k
q4_k
q3_k
q2_k
q6_k multi row example
* revert as multi row isnt faster for k quants
vulkan: request round-to-even for fp16 in im2col/rope_head (#10767)
Vulkan doesn't mandate a specific rounding mode, but the shader_float_controls
feature allows rounding mode to be requested if the implementation supports it.
Vulkan: Add VK_EXT_subgroup_size_control support to ensure full subgroups for coopmats (#10721)
* Vulkan: Add VK_EXT_subgroup_size_control support to ensure full subgroups for coopmats
* Fix subgroup size control extension support check
Add accf32 and accf16 checks for coopmats
* Also disable coopmats on amdvlk
Vulkan: Use improved q4_k and q5_k dequant code in dequant shaders (#10798)
vulkan: small mul_mat_vec optimizations (#10665)
* double the number of rows per workgroup
* Update ggml-vulkan.cpp
* Vulkan: Add VK_EXT_subgroup_size_control support to ensure full subgroups for coopmats
* only increase the number of rows for amd and subgroup size 64
* fix missing NUM_ROWS for mul_mat_vec_iq4_nl_f16_f32, untested
* use subgroup min and max to check for gcn (requires https://github.com/ggerganov/llama.cpp/pull/10721)
* manual merge ggml-vulkan.cpp
* set min and max subgroup size in any case
* Also double the number of rows for Intel GPUs
Change Debug print name
add GGML_ROPE_TYPE_MROPE
rwkv6: add wkv6 support for Vulkan backend (#10829)
* rwkv_wkv6 vulkan shader
* RWKV_WKV6 Vulkan op tests passed
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
* Apply code format changes
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
* add [[unroll]] and remove unnecessary conditions
* add uma support
* fix erros in EditorConfig Checker
---------
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
Co-authored-by: Molly Sophia <mollysophia379@gmail.com>
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/wkv6.comp
vulkan: bugfixes for small subgroup size systems + llvmpipe test (#10809)
* ensure mul mat shaders work on systems with subgroup size less than 32
more fixes
add test
* only s_warptile_mmq needs to be run with 32 threads or more
# Conflicts:
# .github/workflows/build.yml
vulkan : fix soft_max.comp division by zero (whisper/2633)
This change prevents a division by zero error when p.KY is 0.
vulkan: optimize coopmat2 dequant functions (#10855)
Change the code to do 16b loads when possible and extract the appropriate
component late, so the code is effectively decoding a pair of elements and
then selecting one. This can allow more commoning to happen in the compiler
when neighboring elements are loaded.
vulkan: build fixes for 32b (#10927)
* vulkan: build fixes for 32b
Should fix #10923
* vulkan: initialize some buffer/offset variables
examples, ggml : fix GCC compiler warnings (#10983)
Warning types fixed (observed under MSYS2 GCC 14.2.0):
* format '%ld' expects argument of type 'long int', but argument has type 'size_t'
* llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp:81:46: warning: missing initializer for member '_STARTUPINFOA::lpDesktop' [-Wmissing-field-initializers] (emitted for all struct field except first)
# Conflicts:
# examples/export-lora/export-lora.cpp
vulkan: multi-row k quants (#10846)
* multi row k quant shaders!
* better row selection
* more row choices
* readjust row selection
* rm_kq=2 by default
vulkan: Use push constant offset to handle misaligned descriptors (#10987)
vulkan: im2col and matmul optimizations for stable diffusion (#10942)
* tests: Add im2col perf tests
* vulkan: optimize im2col, more elements per thread
* vulkan: increase small tile size for NV_coopmat2
* vulkan: change im2col to 512 elements per workgroup
vulkan: optimize mul_mat for small values of N (#10991)
Make the mul_mat_vec shaders support N>1 (as a spec constant, NUM_COLS) where
the batch_strides are overloaded to hold the row strides. Put the loads from the
B matrix in the innermost loop because it should cache better.
Share some code for reducing the result values to memory in mul_mat_vec_base.
# Conflicts:
# tests/test-backend-ops.cpp
fix: Vulkan shader gen binary path (#11037)
Vulkan: Add device-specific blacklist for coopmat for the AMD proprietary driver (#11074)
* Vulkan: Add device-specific blacklist for coopmat for the AMD proprietary driver
* Add (TM) to AMD name check
fix lora print
Disable GL_KHR_cooperative_matrix Vulkan extension if not available. (#11117)
* Disable GL_KHR_cooperative_matrix Vulkan extension if not available.
* Perform Vulkan extensions checks in a more sensible order
* Remove unnecessary #ifdef directive
# Conflicts:
# ggml/src/vulkan-shaders/test_coopmat_support.comp
llama: add support for QRWKV6 model architecture (#11001)
Vulkan: Fix float16 use on devices without float16 support + fix subgroup_size_control validation error (#11161)
* Vulkan: Remove float16 use in shaders
* Fix validation error about subgroup_size_control extension
fix: ggml: fix vulkan-shaders-gen build (#10448)
* fix: ggml: fix vulkan-shaders-gen build
The vulkan-shaders-gen target was not being built correctly
in case of cross-compilation.
Other outputs need to be built for the cross compile target,
but vulkan-shaders-gen needs to be built for the host.
* refactor: ggml: Improve vulkan-shaders-gen toolchain setup
- Add GGML_SHADERS_GEN_TOOLCHAIN CMake option.
- Auto-detect host toolchain if not set.
* refactor: ggml: Improve vulkan-shaders-gen toolchain setup
Use configure_file to generate host_toolchain.cmake from template
* fix: ggml: Fix compile error
Fix compile error not finding vulkan-shaders-gen
* fix: vulkan-shaders-gen build and path handling
Fix build issues with vulkan-shaders-gen:
- Add target dependency for correct build order
- Use CMAKE_HOST_SYSTEM_NAME for executable suffix
- Fix MSVC output directory in host toolchain
- Normalize path handling for cross-compilation
* fix: improve host compiler detection in vulkan shader build
Improve host compiler detection for vulkan shader generation:
- Add NO_CMAKE_FIND_ROOT_PATH to all compiler searches
- Consolidate compiler detection logic
- Fix Windows-specific MSVC detection
- Ensure correct compiler search in cross-compilation
* refactor: Simplify CMake function for detecting host compiler
Simplified the CMake function to improve the process of detecting the host compiler.
* fix: Remove unnecessary Vulkan library linkage in CMakeLists.txt
Since `vulkan-shader-gen.cpp` only requires the `glslc` executable
and not the Vulkan headers or libraries, CMakeLists.txt needs to
be corrected.
(See: ecc93d0558fc3ecb8a5af69d2ece02fae4710ade)
* refactor: Rename host_toolchain.cmake.in
- Rename host_toolchain.cmake.in to cmake/host-toolchain.cmake.in
* refactor: GGML_VULKAN_SHADERS_GEN_TOOLCHAIN
Rename the macro GGML_SHADERS_GEN_TOOLCHAIN to GGML_VULKAN_SHADERS_GEN_TOOLCHAIN
# Conflicts:
# ggml/src/ggml-vulkan/CMakeLists.txt
vulkan: scale caching for k quants + misc fixes (#11081)
* q6_k scale caching
* 16 bit unpack
* q4_k test (slow)
* revert it
* q3_k
* q2_k
* little stuff
* try precalculating products of a and q2_k scales
* Revert "try precalculating products of a and q2_k scales"
This reverts commit 65110b81f23f66331a50c6e889a7c1ab9470a86b.
* unpack should be u16, add vim swap to gitignore (about time)
* better q4_k scales
* q5_k
* better q6_k with separate paths for all threads and partial threads in use, plus some more optimizations
* q2_k better dequant
* q3_k optimizations
* q3_k use hmask simd from cpu avx version
* make the caches happy
* q3_k separate out calculation
* q2_k separate out
* little stuff
* use calc_superblock everywhere
* q2_k optimize scale calculation
* more barriers
vulkan: optimize coopmat2 q2_k dequant function (#11130)
vulkan: optimize coopmat2 q4_k/q5_k dequant functions. (#11206)
Do masking on whole dwords, fetch all scales at once.
vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl (#11166)
* vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl
Shaders are based on cpy.cu.
* vulkan: support copy from q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl to f32
* ggml: copy q->f32 assumes some contiguity in the destination
# Conflicts:
# ggml/src/ggml-cpu/ggml-cpu.c
# ggml/src/vulkan-shaders/copy_from_quant.comp
# ggml/src/vulkan-shaders/copy_to_quant.comp
vulkan: fix coopmat2 flash attention for non-contiguous inputs (#11281)
Add code similar to mul_mm_cm2 to force alignment of strides, to avoid
a performance regression.
Add noncontiguous FA tests in test-backend-ops.
Fixes #11268.
# Conflicts:
# tests/test-backend-ops.cpp
vulkan: fix coopmat2 validation failures (#11284)
mul mat and flash attention shaders were loading f32 types directly into
A/B matrices, which happens to work but is technically invalid usage.
For FA, we can load it as an Accumulator matrix and convert and this
is not in the inner loop and is cheap enough. For mul mat, it's more
efficient to do this conversion in a separate pass and have the input(s)
be f16.
coopmat2 requires SPIR-V 1.6 (related using to LocalSizeId). LocalSizeId
requires maintenance4 be enabled, and SPIR-V 1.6 requires Vulkan 1.3.
vulkan: fix diag_mask_inf (#11323)
With robustbufferaccess disabled, this shader was showing OOB stores. There
is a bounds check in the code, but the workgrouop dimensions were reversed vs
CUDA and it was running the wrong number of threads. So fix the workgroup
dimensions and disable robustness for this pipeline.
vulkan: sort shaders for more deterministic binary (#11315)
Fixes #11306.
Vulkan-run-test: fix mmq_wg_denoms (#11343)
There should be a copy-and-paste error here.
*mmq_wg_denoms should be used together with *warptile_mmq, instead of
wg_denoms.
vulkan: compile shaders on-demand (#11406)
Reduce first-run startup time and memory consumption.
Should fix #11339.
vulkan: Catch pipeline creation failure and print an error message (#11436)
* vulkan: Catch pipeline creation failure and print an error message
Also, fix some warnings from my on-demand compile change.
* vulkan: fix pipeline creation logging
vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)
* vulkan: initial support for IQ3_S
* vulkan: initial support for IQ3_XXS
* vulkan: initial support for IQ2_XXS
* vulkan: initial support for IQ2_XS
* vulkan: optimize Q3_K by removing branches
* vulkan: implement dequantize variants for coopmat2
* vulkan: initial support for IQ2_S
* vulkan: vertically realign code
* port failing dequant callbacks from mul_mm
* Fix array length mismatches
* vulkan: avoid using workgroup size before it is referenced
* tests: increase timeout for Vulkan llvmpipe backend
---------
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
# Conflicts:
# ggml/src/vulkan-shaders/dequant_iq2_s.comp
# ggml/src/vulkan-shaders/dequant_iq2_xs.comp
# ggml/src/vulkan-shaders/dequant_iq2_xxs.comp
# ggml/src/vulkan-shaders/dequant_iq3_s.comp
# ggml/src/vulkan-shaders/dequant_iq3_xxs.comp
CUDA: non-contiguous (RMS) norm support (#11659)
vulkan: use smaller combined allocations to avoid fragmentation (#11551)
# Conflicts:
# ggml/src/ggml-alloc.c
vulkan: initial support for IQ4_XS quantization (#11501)
# Conflicts:
# ggml/src/vulkan-shaders/dequant_iq4_xs.comp
vulkan: optimize coopmat2 iq2/iq3 callbacks (#11521)
* vulkan: optimize coopmat2 iq2/iq3 callbacks
* build: trigger CI on GLSL compute shader changes
vulkan: print shared memory size (#11719)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: account for lookup tables when checking shared memory size (#11502)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: add environment variable GGML_VK_PREFER_HOST_MEMORY to avoid VRAM allocation (#11592)
vulkan: linux builds + small subgroup size fixes (#11767)
* mm subgroup size
* upload vulkan x86 builds
vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)
* vulkan: initial support for IQ1_S and IQ1_M quantizations
* vulkan: define MMV kernels for IQ1 quantizations
* devops: increase timeout of Vulkan tests again
* vulkan: simplify ifdef for init_iq_shmem
# Conflicts:
# ggml/src/vulkan-shaders/dequant_iq1_m.comp
# ggml/src/vulkan-shaders/dequant_iq1_s.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq1_m.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq1_s.comp
vulkan: support multi/vision rope, and noncontiguous rope (#11902)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/rope_multi.comp
# ggml/src/vulkan-shaders/rope_vision.comp
vulkan: implement several ops relevant for ggml_opt (#11769)
* vulkan: support memset_tensor
* vulkan: support GGML_OP_SUM
* vulkan: implement GGML_OP_ARGMAX
* vulkan: implement GGML_OP_SUB
* vulkan: implement GGML_OP_COUNT_EQUAL
* vulkan: implement GGML_OP_OPT_STEP_ADAMW
* vulkan: fix check_results RWKV_WKV6 crash and memory leaks
* vulkan: implement GGML_OP_REPEAT_BACK
* tests: remove invalid test-backend-ops REPEAT_BACK tests
* vulkan: fix COUNT_EQUAL memset using a fillBuffer command
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/argmax.comp
# ggml/src/vulkan-shaders/count_equal.comp
# ggml/src/vulkan-shaders/opt_step_adamw.comp
# ggml/src/vulkan-shaders/repeat_back.comp
# ggml/src/vulkan-shaders/sub.comp
# tests/test-backend-ops.cpp
vulkan: implement more backpropagation operators (#11914)
* vulkan: implement GGML_OP_ROPE_BACK
* vulkan: implement GGML_OP_RMS_NORM_BACK
* vulkan: implement GGML_OP_SILU_BACK
* vulkan: implement GGML_OP_SOFTMAX_BACK
# Conflicts:
# ggml/src/vulkan-shaders/rms_norm_back.comp
# ggml/src/vulkan-shaders/silu_back.comp
# ggml/src/vulkan-shaders/soft_max_back.comp
Add memset tensor in all backend interface
SYCL: implement memset ggml backend buffer interface (#12580)
* SYCL: implement memset ggml backend buffer interface
* use GGML_ABORT macro
* Do not wait for all queues to finish for memset operation
# Conflicts:
# ggml/src/ggml-sycl.cpp
add OP sigmoid (#12056)
Co-authored-by: Judd <foldl@boxvest.com>
# Conflicts:
# ggml/src/vulkan-shaders/sigmoid.comp
vulkan: fix assertion when qy_needs_dequant (#12068)
Looks like a copy/paste bug from qx_needs_dequant.
vulkan: improve im2col (#11826)
* vulkan: improve im2col performance
vulkan: matmul dequantization improvements (#12015)
* faster dequant for old quants
* dont use unpack for iq4_nl
* vec2 unpack for q8
vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)
* vulkan: implement specialized MMV kernels for IQ2 quantizations
* vulkan: add MMV kernels for IQ3 quants
* vulkan: Increase MMV batch size and unroll IQ LUT setup
* vulkan: fix init_iq_shmem for WG sizes larger than tables
* vulkan: common batch size for all I-quants
# Conflicts:
# ggml/src/vulkan-shaders/mul_mat_vec_iq2_s.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq2_xs.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq2_xxs.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq3_s.comp
# ggml/src/vulkan-shaders/mul_mat_vec_iq3_xxs.comp
cuda/vulkan: specify fp32-only support for some operations in supports_op (ggml/1129)
ggml-ci
# Conflicts:
# ggml/src/ggml-cuda.cu
# tests/test-backend-ops.cpp
mat vec double buffer (#12188)
vulkan: fix bug in coopmat1 mul_mat_id (#12316)
* tests: run mul_mat_id with a larger N
* vulkan: fix bug in coopmat1 mul_mat_id
Update build.yml for Windows Vulkan builder to use Vulkan 1.4.304 SDK for VK_NV_cooperative_matrix2 support (#12301)
vulkan: Adjust coopmat2 tile sizes and selection heuristic (#12258)
vulkan: Pad N dimension of B matrix for coopmat2 perf, to avoid bounds checking (#12273)
* vulkan: Pad N dimension of B matrix for coopmat2 perf, to avoid bounds checking
vulkan: use fp32 in coopmat2 q4_k dequant function (#12309)
vulkan: subgroup size tuning (#12087)
* vulkan: subgroup size test
* Vulkan: Add device architecture enum and logic to recognize AMD generations
* vulkan: use new architecture logic to specify subgroup size
* Initial vulkan subgroup size tuning for RDNA3
* vulkan: commonize RDNA subgroup tuning
* vulkan: override subgroup size if required_subgroup_size = 0
* vulkan: disable warp 32 for RDNA3
* vulkan: fine tuned RDNA1 subgroup sizes
* vulkan: adjusted subgroup size map
* vulkan: fixed RDNA2 subgroup map
---------
Co-authored-by: 0cc4m <picard12@live.de>
vulkan: Add N/2 and N/4 optimized paths in coopmat2 shader (#12312)
ggml-vulkan: remove unused find_program(glslc) (#12416)
It's already found by FindVulkan.cmake in the parent CMakeLists
Vulkan: Default to 1GB allocations instead of 4GB to avoid fragmentation and driver issues (#12434)
vulkan: Submit once enough matmul work has been recorded (#12406)
I've been seeing significantly worse performance for tg with flash attention
enabled vs disabled, and it seems to be related to the submit heuristic.
Change the heuristic to check how many bytes worth of weight matrix are
used and flush every 100MB, and ramp up after the first few submits.
This seems to resolve the issue, and also increases perf for non-FA a bit.
vulkan: optimize iq1 coopmat2 dequant functions (#12427)
vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)
Vulkan: RTE rounding for cpy to quant (#12480)
* Vulkan: RTE rounding for cpy to quant
Co-Authored-By: Jeff Bolz <jbolz@nvidia.com>
* remove trailing whitespace
* avoid duplicating pipeline_cpy_f32_quant
* fix copypasting issue
* remove duplicated code
---------
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
vulkan: Optimize mul_mat_vec p021 and nc shaders (#12505)
* tests: add mul_mat perf/functional tests for p021/nc vulkan shaders
* vulkan: Optimize mul_mat_vec p021 and nc shaders.
These shaders are used in attention calculations, and when the KV cache grows
large they start to dominate the run time. For the nc shader (which is called
with large 'k' dimension), use unrolling and vector loads. For the p021 shader
(which is called with large 'm' and small 'k' dimensions), take advantage of
grouped query attention to reuse loads from the A matrix for the whole group,
and reduce the number of workgroups (too much overhead from tiny dispatches).
Using subgroupAdd in the p021 shader also helps, use that conditionally.
# Conflicts:
# tests/test-backend-ops.cpp
vulkan: fix mul_mat_vec failure in backend tests (#12529)
The OOB calculation could be wrong if the last iteration was during one of
the unrolled loops. Adjust the unrolling counts to avoid this. Add a couple
new backend tests that hit this failure on NVIDIA GPUs.
vulkan: fix coopmat shader generation when cross-compiling (#12272)
* vulkan: fix coopmat shader generation when cross-compiling
Previously the status of coopmat{,2} support isn't passed to the
vulkan-shaders-gen project building on the host, which leads to build
failure because of the cross-compiling code expecting coopmat{,2}
shaders that didn't get generated.
Fix this by passing the coopmat{,2} support status to vulkan-shaders
subproject.
Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
* Only call coop-mat shaders once
* Fix whitespace
---------
Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
Co-authored-by: bandoti <141645996+bandoti@users.noreply.github.com>
cmake: improve Vulkan cooperative matrix support checks (whisper/2966)
Co-authored-by: Sandro Hanea <me@sandro.rocks>
cmake : fix whitespace (#0)
Vulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)
* Vulkan: Add DP4A MMQ and Q8_1 quantization shader
* Add q4_0 x q8_1 matrix matrix multiplication support
* Vulkan: Add int8 coopmat MMQ support
* Vulkan: Add q4_1, q5_0 and q5_1 quants, improve integer dot code
* Add GL_EXT_integer_dot_product check
* Remove ggml changes, fix mmq pipeline picker
* Remove ggml changes, restore Intel coopmat behaviour
* Fix glsl compile attempt when integer vec dot is not supported
* Remove redundant code, use non-saturating integer dot, enable all matmul sizes for mmq
* Remove redundant comment
* Fix integer dot check
* Fix compile issue with unsupported int dot glslc
* Update Windows build Vulkan SDK version
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/mul_mmq.comp
# ggml/src/vulkan-shaders/mul_mmq_funcs.comp
# ggml/src/vulkan-shaders/quantize_q8_1.comp
# ggml/src/vulkan-shaders/test_integer_dot_support.comp
vulkan: fix build when glslc doesn't support coopmat (#12683)
Vulkan: Fix mmq int dot float cache size (#12722)
vulkan: Implement grouped query attention in the coopmat2 FA shader (#12559)
When adjacent batches of Q share the same batches of K/V, batch them into
the same workgroup. For example, when:
dst(128,32,1,1) = FA(q(128,1,32,1), k(128,16640,8,1), v(128,16640,8,1))
previously we would run 32 workgroups computing 1 result each, now we will
run 8 workgroups computing 4 results each.
This doesn't directly translate to better performance (at least when you have
>=32 SMs), but in a subsequent change I'll enable split_k which will scale much
better with 4x fewer workgroups.
cmake: remove caching from vulkan coopmat checks (#12719)
vulkan: Implement split_k for coopmat2 flash attention. (#12627)
When using group query attention, we have one workgroup per KV batch and this
can be very few workgroups (e.g. just 8 in some models). Enable split_k to
spread the work across SMs. This helps a lot when the KV cache is large.
# Conflicts:
# ggml/src/vulkan-shaders/flash_attn_split_k_reduce.comp
vulkan: Fix missing cmake logic for dot product extension (#12721)
vulkan: set cmake minimum and project name in vulkan-shaders (#12744)
vulkan: Hybrid waitForFences/getFenceStatus to reduce fence latency (#12630)
There seems to be a bubble waking up from waitForFences, which costs a few
percent performance and also increased variance in performance. This change
inserts an "almost_ready" fence when the graph is about 80% complete and we
waitForFences for the almost_ready fence and then spin (with _mm_pauses) waiting
for the final fence to be signaled.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
cmake: fix ggml-shaders-gen compiler paths containing spaces (#12747)
fixes error for compiler paths with spaces
Vulkan: Tune Vulkan mmq int dot shader for performance (#12767)
vulkan: Use unclamped loads for flash attention mask (#12720)
nem1 must be a multiple of GGML_KQ_MASK_PAD, and GGML_KQ_MASK_PAD is a multiple
of the number of rows in the matrix. The KV dim is a multiple of the number of
columns for the aligned shader.
vulkan: fix NaN issue in flash attention shader (#12776)
Use -FLT_MAX/2 rather than -inf as the initial value for computing the maximum.
vulkan: Use fp16 for the flash attention P*V multiplication (#12783)
This is consistent with the ggml-cuda behavior and the mul_mat fallback.
vulkan: In coopmat2 mmq, load q4_k/q5_k scales through shared memory (#12833)
q4_k and q5_k had a lot of redundant global loads where the same 16B of
scale information is repeatedly loaded and decoded during each loop iteration.
This change restructures the loops to more explicitly iterate over whole
blocks in the outer loop (with unrolled inner loop) and to copy/decode the
scale data into shared memory once at the start of each outer loop. The copy
is pipelined so the scale load from global memory is relatively cheap.
This improves q4_k/q5_k model prompt processing performance by around 5-7%.
I briefly tried applying this to q6_k and q4_0, and it didn't help for q6_k
and hurt for q4_0.
The big "else" path in mul_mm_cm2.comp that had all the clamped/unclamped
variants isn't used as often as it originally was (e.g. due to the padded_N
change), so I trimmed it down to offset some of the new complexity of the
semi-manual loop unrolling.
vulkan: use aligned loads for flash attention mask (#12853)
Rewrite the stride logic for the mask tensor in the FA shader to force the
stride to be aligned, to allow using more efficient loads.
vulkan: enable coopmat2 FA gqa and split_k optimizations more often (#12931)
The grouped query attention optmization doesn't require a power of two ratio,
the only thing relying on it was the modulo operation written as bitwise &.
split_k need not depend on gqa_ratio - enable it any time there's only one
workgroup in the X dimension. The shader gets the split index from the x coord,
and multiple workgroups in the X dimension (pre-split) indicates a larger
FA operation that wouldn't need splitting.
vulkan: support noncontiguous rms_norm (#13031)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: matmul gcn tuning (#13016)
* tune matmul for gcn
* this one is more power efficient
* Update ggml/src/ggml-vulkan/ggml-vulkan.cpp
Co-authored-by: 0cc4m <picard12@live.de>
* disable this tune for the proprietary driver
---------
Co-authored-by: 0cc4m <picard12@live.de>
vulkan: use uint array index to avoid glslang bug (#13193)
vulkan: Handle src1 batch dimension in non-contiguous mat-vec-mul shader (#13191)
* vulkan: Handle src1 batch dimension in non-contiguous mat-vec-mul shader
vulkan: Add bfloat16 support (#12554)
* vulkan: Add bfloat16 support
This adds bfloat16 matrix multiply support based on VK_KHR_shader_bfloat16.
The extension is required for coopmat multiply support, but matrix-vector
multiply trivially promotes bf16 to fp32 and doesn't require the extension.
The copy/get_rows shaders also don't require the extension.
It's probably possible to fall back to non-coopmat and promote to fp32 when
the extension isn't supported, but this change doesn't do that.
The coopmat support also requires a glslc that supports the extension, which
currently requires a custom build.
* vulkan: Support bf16 tensors without the bf16 extension or coopmat support
Compile a variant of the scalar mul_mm shader that will promote the bf16
values to float, and use that when either the bf16 extension or the coopmat
extensions aren't available.
* vulkan: bfloat16 fixes (really works without bfloat16 support now)
* vulkan: fix spirv-val failure and reenable -O
# Conflicts:
# ggml/src/vulkan-shaders/test_bfloat16_support.comp
vulkan: Additional type support for unary, binary, and copy (#13266)
Support f16->f32 copy.
Support f16->f16 and f32->f32 unary ops.
Support all combinations of f16/f32 for src0/src1/dst for add/sub/mul/div.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: Allow up to 4096 elements for mul_mat_id row_ids (#13326)
This assert fired running Qwen_Qwen3-30B-A3B-Q2_K.gguf:
GGML_ASSERT(nei0 * nei1 <= 3072);
The tensor is 8 x 512. Increase this array size to accommodate.
vulkan: scalar flash attention implementation (#13324)
* vulkan: scalar flash attention implementation
* vulkan: always use fp32 for scalar flash attention
* vulkan: use vector loads in scalar flash attention shader
* vulkan: remove PV matrix, helps with register usage
* vulkan: reduce register usage in scalar FA, but perf may be slightly worse
* vulkan: load each Q value once. optimize O reduction. more tuning
* vulkan: support q4_0/q8_0 KV in scalar FA
* CI: increase timeout to accommodate newly-supported tests
* vulkan: for scalar FA, select between 1 and 8 rows
* vulkan: avoid using Float16 capability in scalar FA
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/flash_attn.comp
vulkan: workaround FA compile failures on macos (#13517)
vulkan: KHR_coopmat flash attention (#13506)
This shader uses coopmat1 to do the Q*K^T multiply. The P*V multiply is more
difficult for various reasons so I haven't done it. Performance for this
shader is around 2.5x better than for the scalar shader when doing prompt
processing. Some of the benefit may be from other optimizations like staging
through shared memory, or splitting by rows.
# Conflicts:
# ggml/src/vulkan-shaders/flash_attn_cm1.comp
cmake: simplify vulkan shader test logic (#13263)
vulkan: use scalar FA rather than coopmat2 when N==1 (#13554)
Add pipeline_acc_f32
vulkan: move common FA code to flash_attn_base.comp (#13556)
* vulkan: move common FA code to flash_attn_base.comp
* vulkan: move common FA index/stride setup code to flash_attn_base.comp
* build fix
# Conflicts:
# ggml/src/vulkan-shaders/flash_attn_base.comp
cmake: use the current build config for vulkan-shaders-gen (#13595)
* fix: use the current build config for `vulkan-shaders-gen`
* fix: only pass a valid build type to `--config`
Vulkan: Add f32 accumulator support to quantized mul mat to fix GLM4 32B incoherence (#13607)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: fix warnings (#13626)
* small fixes
* remove ifdef
use LOG_WARN to replace `std::cerr` (#13657)
vulkan: Disable coopmat/coopmat2/bfloat extensions if glslc doesn't support it (#13696)
vulkan: support CPY from any type to itself (#13695)
Reuse the f16/f32 copy shaders, and just scale the number of elements
according to the type size.
add GGML_LOG_WARN
vulkan: mark IM2COL as supporting non-contig (#13783)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: use timestamp queries for GGML_VULKAN_PERF (#13817)
Also change it to be controlled by an env var rather than cmake flag
vulkan : Remove unexpected ; (ggml/1253)
vulkan: fix warnings in perf logger querypool code (#13937)
ggml-vulkan: adds support for op CONV_TRANSPOSE_1D (#13813)
* * ggml-vulkan: adds op CONV_TRANSPOSE_1D
* test-backend-ops: adds more spohisticated tests for CONV_TRANSPOSE_1D
* Missing barrier added to shader.
Number of additional tests reduced to 108.
* * Fixes typo in variable name.
* Removes extra whitespaces.
* Adds int64->int32 casts to prevent possible warnings.
* Problem size reduced in tests to pass tests with llvmpipe.
* supports_op condition moved from unintended position
# Conflicts:
# ggml/src/ggml-vulkan.cpp
# ggml/src/vulkan-shaders/conv_transpose_1d.comp
vulkan: Enable VK_KHR_cooperative_matrix extension for Intel Xe2 GPUs (#14001)
* allowing B580 and U9-288V
* experimenting code to detect Xe2
* allowing coopmat only for Xe2 GPUs
* fixed comment wording
* fixed comment wording
* removed unnecessary driver check
Vulkan: Don't default to CPU device (like llvmpipe), even if no other device is available, to allow fallback to CPU backend (#14099)
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: force device 0 in CI (#14106)
Add GGML_LOG_INFO
vulkan: Track descriptor pools/sets per-context (#14109)
Use the same descriptor set layout for all pipelines (MAX_PARAMETER_COUNT == 8)
and move it to the vk_device. Move all the descriptor pool and set tracking to
the context - none of it is specific to pipelines anymore. It has a single vector
of pools and vector of sets, and a single counter to track requests and a single
counter to track use.
vulkan: Better thread-safety for command pools/buffers (#14116)
This change moves the command pool/buffer tracking into a vk_command_pool
structure. There are two instances per context (for compute+transfer) and
two instances per device for operations that don't go through a context.
This should prevent separate contexts from stomping on each other.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: mutex around vkQueueSubmit (#14127)
This fixes the remaining crash in test-thread-safety on my system.
cmake: clean up external project logic for vulkan-shaders-gen (#14179)
* Remove install step for vulkan-shaders-gen
* Add install step to normalize msvc with make
* Regenerate modified shaders at build-time
# Conflicts:
# .github/workflows/build.yml
cmake: remove shader-gen step-targets from ggml-vulkan (#14226)
* Remove step-targets from vulkan-shaders-gen
* Unset DESTDIR when building vulkan-shaders-gen
Vulkan: Set device max size for host memory to avoid OOM warning and fallback to CPU buffer (#14249)
Add support for VK_EXT_debug_utils to add labels to Vulkan objects. (#13792)
* Add support for VK_EXT_debug_utils to add labels to Vulkan objects. In step 1 compute pipelines are getting labeled.
* remove #ifdef for debug utils and add queue marker.
# Conflicts:
# ggml/src/ggml-vulkan.cpp
vulkan: update windows SDK in CI (#14334)
vulkan: update windows SDK in release.yml (#14344)
# Conflicts:
# .github/workflows/release.yml
cmake: regen vulkan shaders when shaders-gen sources change (#14398)
* Add shaders-gen sources as target deps
vulkan: Fix GGML_VULKAN_SHADER_DEBUG_INFO (#14427)
This setting needs to be passed through to vulkan-shaders-gen
vulkan: lock accesses of pinned_memory vector (#14333)
vulkan: handle noncontig in the final case of ggml_vk_get_cpy_pipeline (#14378)
Fix cuda build error
test
* remove new cpu backend and yml files
* remove new op and GGML_ROPE_TYPE_NEOX
* fix build error
* change cmake file to add matrix operation
* remove coopmat2 check in flash attention
* print gpu info for vulkan
* disable fuse to recover vulkan performance
---------
Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: firecoperana <firecoperana>
116 files changed, 14566 insertions, 3282 deletions
@@ -17,6 +17,7 @@ *.metallib *.o *.so +*.swp *.tmp # IDE / OS @@ -774,6 +774,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG MK_CPPFLAGS += -DGGML_VULKAN_MEMORY_DEBUG endif +ifdef GGML_VULKAN_PERF + MK_CPPFLAGS += -DGGML_VULKAN_PERF +endif + ifdef GGML_VULKAN_VALIDATE MK_CPPFLAGS += -DGGML_VULKAN_VALIDATE endif diff --git a/examples/cvector-generator/mean.hpp b/examples/cvector-generator/mean.hpp index 16be5ce3..4eeac1ee 100644 --- a/examples/cvector-generator/mean.hpp +++ b/examples/cvector-generator/mean.hpp @@ -15,7 +15,7 @@ static void run( for (size_t il = 0; il < v_input.size(); ++il) { // prepare output vector struct ggml_tensor * ctrl_out = v_output[il]; - ggml_format_name(ctrl_out, "direction.%ld", il+1); + ggml_format_name(ctrl_out, "direction.%zu", il+1); // calculate mean vector struct ggml_tensor * t_layer = v_input[il]; diff --git a/examples/cvector-generator/pca.hpp b/examples/cvector-generator/pca.hpp index 6ec3141a..85be07dd 100644 --- a/examples/cvector-generator/pca.hpp +++ b/examples/cvector-generator/pca.hpp @@ -312,7 +312,7 @@ static void run_pca( // prepare output vector struct ggml_tensor * ctrl_out = v_output[il]; - ggml_format_name(ctrl_out, "direction.%ld", il+1); + ggml_format_name(ctrl_out, "direction.%zu", il+1); // run power_iteration params.i_layer = il; diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 3176d6e2..1aaf281e 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -251,8 +251,8 @@ struct lora_merge_ctx { fout.write((const char *)data.data(), data.size()); } - printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged); - printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size()); + printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged); + printf("%s : wrote %zu tensors to output file\n", __func__, base_to_out_tensors.size()); } void copy_tensor(struct ggml_tensor * base) { @@ -334,7 +334,7 @@ struct lora_merge_ctx { const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale; delta = ggml_scale(ctx0, delta, scale); cur = ggml_add(ctx0, delta, cur); - printf("%s : + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type)); + printf("%s : + merging from adapter[%zu] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type)); printf("%s : input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]); } cur = ggml_cast(ctx0, cur, out->type); diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 5c891105..4dbd8d34 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -143,6 +143,7 @@ option(GGML_VULKAN "ggml: use Vulkan" option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF) option(GGML_VULKAN_DEBUG "ggml: enable Vulkan debug output" OFF) option(GGML_VULKAN_MEMORY_DEBUG "ggml: enable Vulkan memory debug output" OFF) +option(GGML_VULKAN_SHADER_DEBUG_INFO "ggml: enable Vulkan shader debug info" OFF) option(GGML_VULKAN_VALIDATE "ggml: enable Vulkan validation" OFF) option(GGML_VULKAN_RUN_TESTS "ggml: run Vulkan tests" OFF) option(GGML_KOMPUTE "ggml: use Kompute" OFF) @@ -160,6 +161,9 @@ option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl ca set (GGML_SYCL_TARGET "INTEL" CACHE STRING "ggml: sycl target device") +# toolchain for vulkan-shaders-gen +set (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen") + # extra artifacts option(GGML_BUILD_TESTS "ggml: build tests" ${GGML_STANDALONE}) option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE}) diff --git a/ggml/include/ggml-vulkan.h b/ggml/include/ggml-vulkan.h index af661c2d..e3e62864 100644 --- a/ggml/include/ggml-vulkan.h +++ b/ggml/include/ggml-vulkan.h @@ -10,7 +10,7 @@ extern "C" { #define GGML_VK_NAME "Vulkan" #define GGML_VK_MAX_DEVICES 16 -GGML_API void ggml_vk_instance_init(void); +GGML_API GGML_CALL void ggml_vk_instance_init(void); // backend API GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num); diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 0a14ba57..67551eb2 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -628,7 +628,6 @@ extern "C" { GGML_OP_WIN_UNPART, GGML_OP_GET_REL_POS, GGML_OP_ADD_REL_POS, - GGML_OP_UNARY, GGML_OP_MAP_UNARY, @@ -644,7 +643,6 @@ extern "C" { GGML_OP_CROSS_ENTROPY_LOSS, GGML_OP_CROSS_ENTROPY_LOSS_BACK, - GGML_OP_COUNT, }; @@ -1130,6 +1128,7 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + // if a is the same shape as b, and a is not parameter, return a // otherwise, return a new tensor: repeat(a) to fit in b GGML_API struct ggml_tensor * ggml_repeat( diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 4f181841..4f9428f9 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -643,12 +643,85 @@ if (GGML_RPC) set(GGML_SOURCES_RPC ggml-rpc.cpp) endif() +find_package(Vulkan COMPONENTS glslc REQUIRED) + +function(detect_host_compiler) + if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows") + find_program(HOST_C_COMPILER NAMES cl gcc clang NO_CMAKE_FIND_ROOT_PATH) + find_program(HOST_CXX_COMPILER NAMES cl g++ clang++ NO_CMAKE_FIND_ROOT_PATH) + else() + find_program(HOST_C_COMPILER NAMES gcc clang NO_CMAKE_FIND_ROOT_PATH) + find_program(HOST_CXX_COMPILER NAMES g++ clang++ NO_CMAKE_FIND_ROOT_PATH) + endif() + set(HOST_C_COMPILER "${HOST_C_COMPILER}" PARENT_SCOPE) + set(HOST_CXX_COMPILER "${HOST_CXX_COMPILER}" PARENT_SCOPE) +endfunction() + +# Function to test shader extension support +# Parameters: +# EXTENSION_NAME - Name of the extension to test (e.g., "GL_EXT_integer_dot_product") +# TEST_SHADER_FILE - Path to the test shader file +# RESULT_VARIABLE - Name of the variable to set (ON/OFF) based on test result +function(test_shader_extension_support EXTENSION_NAME TEST_SHADER_FILE RESULT_VARIABLE) + execute_process( + COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${TEST_SHADER_FILE}" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error + ) + + if (${glslc_error} MATCHES ".*extension not supported: ${EXTENSION_NAME}.*") + message(STATUS "${EXTENSION_NAME} not supported by glslc") + set(${RESULT_VARIABLE} OFF PARENT_SCOPE) + else() + message(STATUS "${EXTENSION_NAME} supported by glslc") + set(${RESULT_VARIABLE} ON PARENT_SCOPE) + add_compile_definitions(${RESULT_VARIABLE}) + + # Ensure the extension support is forwarded to vulkan-shaders-gen + list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -D${RESULT_VARIABLE}=ON) + set(VULKAN_SHADER_GEN_CMAKE_ARGS "${VULKAN_SHADER_GEN_CMAKE_ARGS}" PARENT_SCOPE) + endif() +endfunction() + if (GGML_VULKAN) find_package(Vulkan COMPONENTS glslc REQUIRED) if (Vulkan_FOUND) message(STATUS "Vulkan found") + # add vulkan test + + set(VULKAN_SHADER_GEN_CMAKE_ARGS "") + + # Test all shader extensions + test_shader_extension_support( + "GL_KHR_cooperative_matrix" + "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp" + "GGML_VULKAN_COOPMAT_GLSLC_SUPPORT" + ) + + test_shader_extension_support( + "GL_NV_cooperative_matrix2" + "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp" + "GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT" + ) + + test_shader_extension_support( + "GL_EXT_integer_dot_product" + "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp" + "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT" + ) + + test_shader_extension_support( + "GL_EXT_bfloat16" + "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp" + "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT" + ) + + + + + # end vulkan test list(APPEND GGML_CDEF_PUBLIC GGML_USE_VULKAN) # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build @@ -669,6 +742,14 @@ if (GGML_VULKAN) add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG) endif() + if (GGML_VULKAN_SHADER_DEBUG_INFO) + add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO) + endif() + + if (GGML_VULKAN_PERF) + add_compile_definitions(GGML_VULKAN_PERF) + endif() + if (GGML_VULKAN_VALIDATE) add_compile_definitions(GGML_VULKAN_VALIDATE) endif() @@ -677,31 +758,79 @@ if (GGML_VULKAN) add_compile_definitions(GGML_VULKAN_RUN_TESTS) endif() - add_subdirectory(vulkan-shaders) - - set (_ggml_vk_genshaders_cmd vulkan-shaders-gen) - set (_ggml_vk_header ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp) - set (_ggml_vk_source ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp) - set (_ggml_vk_input_dir ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders) - set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv) - - file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp") - - add_custom_command( - OUTPUT ${_ggml_vk_header} - ${_ggml_vk_source} - - COMMAND ${_ggml_vk_genshaders_cmd} - --glslc ${Vulkan_GLSLC_EXECUTABLE} - --input-dir ${_ggml_vk_input_dir} - --output-dir ${_ggml_vk_output_dir} - --target-hpp ${_ggml_vk_header} - --target-cpp ${_ggml_vk_source} - --no-clean - - DEPENDS ${_ggml_vk_shader_deps} - COMMENT "Generate vulkan shaders" - ) + # Set up toolchain for host compilation whether cross-compiling or not + if (CMAKE_CROSSCOMPILING) + if (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN) + set(HOST_CMAKE_TOOLCHAIN_FILE ${GGML_VULKAN_SHADERS_GEN_TOOLCHAIN}) + else() + detect_host_compiler() + if (NOT HOST_C_COMPILER OR NOT HOST_CXX_COMPILER) + message(FATAL_ERROR "Host compiler not found") + else() + message(STATUS "Host compiler: ${HOST_C_COMPILER} ${HOST_CXX_COMPILER}") + endif() + configure_file(${CMAKE_CURRENT_SOURCE_DIR}/cmake/host-toolchain.cmake.in ${CMAKE_BINARY_DIR}/host-toolchain.cmake @ONLY) + set(HOST_CMAKE_TOOLCHAIN_FILE ${CMAKE_BINARY_DIR}/host-toolchain.cmake) + endif() + else() + # For non-cross-compiling, use empty toolchain (use host compiler) + set(HOST_CMAKE_TOOLCHAIN_FILE "") + endif() + + include(ExternalProject) + + if (CMAKE_CROSSCOMPILING) + list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE}) + message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}") + endif() + + ExternalProject_Add( + vulkan-shaders-gen + SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/$<CONFIG> + -DCMAKE_INSTALL_BINDIR=. + -DCMAKE_BUILD_TYPE=$<CONFIG> + ${VULKAN_SHADER_GEN_CMAKE_ARGS} + + BUILD_COMMAND ${CMAKE_COMMAND} --build . --config $<CONFIG> + + # NOTE: When DESTDIR is set using Makefile generators and + # "make install" triggers the build step, vulkan-shaders-gen + # would be installed into the DESTDIR prefix, so it is unset + # to ensure that does not happen. + + INSTALL_COMMAND ${CMAKE_COMMAND} -E env --unset=DESTDIR + ${CMAKE_COMMAND} --install . --config $<CONFIG> + ) + + set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>) + set (_ggml_vk_genshaders_dir "${CMAKE_BINARY_DIR}/$<CONFIG>") + set (_ggml_vk_genshaders_cmd "${_ggml_vk_genshaders_dir}/vulkan-shaders-gen${_ggml_vk_host_suffix}") + set (_ggml_vk_header "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp") + set (_ggml_vk_source "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp") + set (_ggml_vk_input_dir "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders") + set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv") + + file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp") + + add_custom_command( + OUTPUT ${_ggml_vk_header} + ${_ggml_vk_source} + + COMMAND ${_ggml_vk_genshaders_cmd} + --glslc ${Vulkan_GLSLC_EXECUTABLE} + --input-dir ${_ggml_vk_input_dir} + --output-dir ${_ggml_vk_output_dir} + --target-hpp ${_ggml_vk_header} + --target-cpp ${_ggml_vk_source} + --no-clean + + DEPENDS ${_ggml_vk_shader_files} + ${_ggml_vk_shader_gens_sources} + vulkan-shaders-gen + + COMMENT "Generate vulkan shaders" + ) set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header}) set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source}) diff --git a/ggml/src/cmake/host-toolchain.cmake.in b/ggml/src/cmake/host-toolchain.cmake.in new file mode 100644 index 00000000..2d8a8569 --- /dev/null +++ b/ggml/src/cmake/host-toolchain.cmake.in @@ -0,0 +1,15 @@ +set(CMAKE_BUILD_TYPE Release) +set(CMAKE_C_FLAGS -O2) +set(CMAKE_CXX_FLAGS -O2) +set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER) +set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER) +set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER) +set(CMAKE_C_COMPILER "@HOST_C_COMPILER@") +set(CMAKE_CXX_COMPILER "@HOST_CXX_COMPILER@") +set(CMAKE_RUNTIME_OUTPUT_DIRECTORY @CMAKE_RUNTIME_OUTPUT_DIRECTORY@) + +if("@CMAKE_C_COMPILER_ID@" STREQUAL "MSVC") + foreach(CONFIG IN ITEMS DEBUG RELEASE MINSIZEREL RELWITHDEBINFO) + set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}) + endforeach() +endif() diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c index 3f2d2023..67c1ba18 100644 --- a/ggml/src/ggml-alloc.c +++ b/ggml/src/ggml-alloc.c @@ -987,9 +987,9 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte if (this_size > max_size) { fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n", - __func__, t->name, - ggml_backend_buft_name(buft), - this_size, max_size); + __func__, t->name, + ggml_backend_buft_name(buft), + this_size, max_size); for (size_t i = 0; i < n_buffers; i++) { ggml_backend_buffer_free(buffers[i]); } diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h index 36ca3708..e08bde83 100644 --- a/ggml/src/ggml-backend-impl.h +++ b/ggml/src/ggml-backend-impl.h @@ -42,6 +42,7 @@ extern "C" { void (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer); void * (*GGML_CALL get_base) (ggml_backend_buffer_t buffer); void (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + void (*GGML_CALL memset_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor* tensor, uint8_t value, size_t offset, size_t size); void (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); void (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); bool (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c index 5903c89c..28850e0d 100644 --- a/ggml/src/ggml-backend.c +++ b/ggml/src/ggml-backend.c @@ -266,6 +266,21 @@ GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * #endif } +GGML_CALL void ggml_backend_tensor_memset(struct ggml_tensor* tensor, uint8_t value, size_t offset, size_t size) { + ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; + + if (size == 0) { + return; + } + + GGML_ASSERT(buf != NULL && "tensor buffer not set"); + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); + GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not implemented by backend buffer"); + + buf->iface.memset_tensor(buf, tensor, value, offset, size); +} + void ggml_backend_synchronize(ggml_backend_t backend) { if (backend->iface.synchronize == NULL) { return; @@ -593,6 +608,12 @@ GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t free(buffer->context); } +static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor* tensor, uint8_t value, size_t offset, size_t size) { + memset((char*)tensor->data + offset, value, size); + + GGML_UNUSED(buffer); +} + GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { memcpy((char *)tensor->data + offset, data, size); @@ -624,6 +645,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = { /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, /* .get_base = */ ggml_backend_cpu_buffer_get_base, /* .init_tensor = */ NULL, // no initialization required + /* .memset_tensor = */ ggml_backend_cpu_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor, @@ -637,6 +659,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed /* .get_base = */ ggml_backend_cpu_buffer_get_base, /* .init_tensor = */ NULL, // no initialization required + /* .memset_tensor = */ ggml_backend_cpu_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor, @@ -993,6 +1016,7 @@ static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface( /* .free_buffer = */ ggml_backend_multi_buffer_free_buffer, /* .get_base = */ NULL, /* .init_tensor = */ NULL, + /* .memset_tensor = */ NULL, /* .set_tensor = */ NULL, /* .get_tensor = */ NULL, /* .cpy_tensor = */ NULL, diff --git a/ggml/src/ggml-cann.cpp b/ggml/src/ggml-cann.cpp index 06930ba2..a2d64da5 100644 --- a/ggml/src/ggml-cann.cpp +++ b/ggml/src/ggml-cann.cpp @@ -1036,6 +1036,7 @@ static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = { /* .free_buffer = */ ggml_backend_cann_buffer_free_buffer, /* .get_base = */ ggml_backend_cann_buffer_get_base, /* .init_tensor = */ ggml_backend_cann_buffer_init_tensor, + /* .memset_tensor = */ NULL, /* .set_tensor = */ ggml_backend_cann_buffer_set_tensor, /* .get_tensor = */ ggml_backend_cann_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_cann_buffer_cpy_tensor, diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index da3dc334..ae7a55c6 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -500,6 +500,14 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t } } +GGML_CALL static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; + + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + offset, value, size, cudaStreamPerThread)); + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); +} + GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; @@ -551,6 +559,7 @@ static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = { /* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer, /* .get_base = */ ggml_backend_cuda_buffer_get_base, /* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_cuda_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor, /* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor, @@ -867,6 +876,7 @@ static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { /* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer, /* .get_base = */ ggml_backend_cuda_split_buffer_get_base, /* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor, + /* .memset_tensor = */ NULL, /* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor, /* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor, /* .cpy_tensor = */ NULL, @@ -3566,17 +3576,24 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons } return false; } break; + case GGML_OP_SILU_BACK: + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; + break; + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + return true; + case GGML_OP_RMS_NORM_BACK: + return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0; + break; case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: - case GGML_OP_NORM: case GGML_OP_ADD: case GGML_OP_MULTI_ADD: case GGML_OP_MUL: case GGML_OP_DIV: - case GGML_OP_RMS_NORM: case GGML_OP_FUSED_RMS_NORM: case GGML_OP_SCALE: case GGML_OP_SOFTCAP: diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute.cpp index 41ac63fa..d0395ff9 100644 --- a/ggml/src/ggml-kompute.cpp +++ b/ggml/src/ggml-kompute.cpp @@ -1872,6 +1872,7 @@ static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = { /* .free_buffer = */ ggml_backend_kompute_buffer_free_buffer, /* .get_base = */ ggml_backend_kompute_buffer_get_base, /* .init_tensor = */ NULL, + /* .memset_tensor = */ NULL, /* .set_tensor = */ ggml_backend_kompute_buffer_set_tensor, /* .get_tensor = */ ggml_backend_kompute_buffer_get_tensor, /* .cpy_tensor = */ NULL, diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m index 4cd44cb9..ff68ee19 100644 --- a/ggml/src/ggml-metal.m +++ b/ggml/src/ggml-metal.m @@ -4068,6 +4068,12 @@ GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t return ctx->all_data; } +GGML_CALL void ggml_backend_metal_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + memset((char *)tensor->data + offset, value, size); + + GGML_UNUSED(buffer); +} + GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { memcpy((char *)tensor->data + offset, data, size); @@ -4101,6 +4107,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = { /* .free_buffer = */ ggml_backend_metal_buffer_free_buffer, /* .get_base = */ ggml_backend_metal_buffer_get_base, /* .init_tensor = */ NULL, + /* .memset_tensor = */ ggml_backend_metal_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_metal_buffer_set_tensor, /* .get_tensor = */ ggml_backend_metal_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_metal_buffer_cpy_tensor, diff --git a/ggml/src/ggml-rpc.cpp b/ggml/src/ggml-rpc.cpp index 3b5c8f46..4c44ae54 100644 --- a/ggml/src/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc.cpp @@ -631,6 +631,7 @@ static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = { /* .free_buffer = */ ggml_backend_rpc_buffer_free_buffer, /* .get_base = */ ggml_backend_rpc_buffer_get_base, /* .init_tensor = */ ggml_backend_rpc_buffer_init_tensor, + /* .memset_tensor =*/ NULL, /* .set_tensor = */ ggml_backend_rpc_buffer_set_tensor, /* .get_tensor = */ ggml_backend_rpc_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_rpc_buffer_cpy_tensor, diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index d8eb86c2..e0f8536a 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -37,6 +37,7 @@ #include "ggml-backend-impl.h" #include "ggml-sycl/backend.hpp" +#include "ggml-sycl/common.hpp" #include "ggml-sycl/presets.hpp" bool ggml_sycl_loaded(void); @@ -83,12 +84,64 @@ static __dpct_inline__ float op_add(const float a, const float b) { return a + b; } -static __dpct_inline__ float op_mul(const float a, const float b) { - return a * b; +static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, + size_t offset, size_t size) { + GGML_SYCL_DEBUG(" [SYCL] call %s\n", __func__); + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + SYCL_CHECK(ggml_sycl_set_device(ctx->device)); + auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue()); + if (size == 0) { + return; // Nothing to do + } + if (tensor->data == nullptr) { + GGML_ABORT("Error: Tensor data pointer is null.\n"); + } + void * target_ptr = static_cast<char *>(tensor->data) + offset; + SYCL_CHECK(CHECK_TRY_ERROR((*stream).memset(target_ptr, value, size))); + SYCL_CHECK(CHECK_TRY_ERROR((*stream).wait())); +} + +static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) { + GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__); + if (buffer == nullptr) { + return; + } + + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + + if (ctx != nullptr) { + for (ggml_tensor_extra_gpu * extra : ctx->tensor_extras) { + release_extra_gpu(extra); + } + ctx->tensor_extras.clear(); // reset the tensor_extras vector + } } -static __dpct_inline__ float op_div(const float a, const float b) { - return a / b; +static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { + /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, + /* .get_base = */ ggml_backend_sycl_buffer_get_base, + /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor, + /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, + /* .clear = */ ggml_backend_sycl_buffer_clear, + /* .reset = */ ggml_backend_sycl_buffer_reset, +}; + +// sycl buffer type +struct ggml_backend_sycl_buffer_type_context { + int device; + std::string name; + + // each buffer type has its own stream + queue_ptr stream = nullptr; +}; + +static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; + + return ctx->name.c_str(); } template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t> @@ -4407,7 +4460,9 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, /* .get_base = */ ggml_backend_sycl_buffer_get_base, /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, + /* .memset_tensor =*/ NULL, /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, /* .clear = */ ggml_backend_sycl_buffer_clear, @@ -4810,6 +4865,7 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = { /* .free_buffer = */ ggml_backend_sycl_split_buffer_free_buffer, /* .get_base = */ ggml_backend_sycl_split_buffer_get_base, /* .init_tensor = */ ggml_backend_sycl_split_buffer_init_tensor, + /* .memset_tensor =*/ NULL, /* .set_tensor = */ ggml_backend_sycl_split_buffer_set_tensor, /* .get_tensor = */ ggml_backend_sycl_split_buffer_get_tensor, /* .cpy_tensor = */ NULL, diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index 86732837..c0bdfb7b 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -1,7 +1,8 @@ #include "ggml-vulkan.h" #include <vulkan/vulkan_core.h> -#ifdef GGML_VULKAN_RUN_TESTS +#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_CHECK_RESULTS) #include <chrono> +#include "ggml-cpu.h" #endif #include <vulkan/vulkan.hpp> @@ -17,37 +18,72 @@ #include <memory> #include <limits> #include <map> +#include <unordered_map> #include <memory> #include <mutex> +#include <future> +#include <thread> + +#if defined(_MSC_VER) +# define NOMINMAX 1 +# include <windows.h> +# define YIELD() YieldProcessor() +#elif defined(__clang__) || defined(__GNUC__) +# if defined(__x86_64__) ||defined(__i386__) +# include <immintrin.h> +# define YIELD() _mm_pause() +# elif defined(__arm__) || defined(__aarch64__) +# if defined(__clang__) +# include <arm_acle.h> +# define YIELD() __yield() +# else +# define YIELD() asm volatile("yield") +# endif +# endif +#endif + +#if !defined(YIELD) +#define YIELD() +#endif -#include "ggml.h" +#include "ggml-impl.h" #include "ggml-backend-impl.h" #include "ggml-vulkan-shaders.hpp" -#define VK_API_VERSION VK_API_VERSION_1_2 +// remove this once it's more widely available in the SDK +#if !defined(VK_KHR_shader_bfloat16) + +#define VK_KHR_shader_bfloat16 1 +#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION 1 +#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME "VK_KHR_shader_bfloat16" +#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000) +#define VK_COMPONENT_TYPE_BFLOAT16_KHR ((VkComponentTypeKHR)1000141000) + +typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR { + VkStructureType sType; + void* pNext; + VkBool32 shaderBFloat16Type; + VkBool32 shaderBFloat16DotProduct; + VkBool32 shaderBFloat16CooperativeMatrix; +} VkPhysicalDeviceShaderBfloat16FeaturesKHR; +#endif +#define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) +static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; } #define VK_VENDOR_ID_AMD 0x1002 #define VK_VENDOR_ID_APPLE 0x106b #define VK_VENDOR_ID_INTEL 0x8086 #define VK_VENDOR_ID_NVIDIA 0x10de -#define VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN 0 -#define VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI 1 -#define VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE 2 +#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256 #define GGML_VK_MAX_NODES 8192 #define MAX_VK_BUFFERS 256 -#ifndef K_QUANTS_PER_ITERATION -#define K_QUANTS_PER_ITERATION 1 -#else -static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); -#endif - #define VK_CHECK(err, msg) \ do { \ vk::Result err_ = (err); \ @@ -63,32 +99,32 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA #else #define VK_LOG_DEBUG(msg) ((void) 0) #endif // GGML_VULKAN_DEBUG +#define GGML_DEBUG 0 +#if (GGML_DEBUG >= 1) +#define GGML_LOG_DEBUG(...) printf(__VA_ARGS__) +#else +#define GGML_LOG_DEBUG(...) +#endif +#define GGML_LOG_WARN(...) printf(__VA_ARGS__) +#define GGML_LOG_INFO(...) printf(__VA_ARGS__) struct ggml_backend_vk_context; -struct vk_queue { - uint32_t queue_family_index; - vk::Queue queue; - vk::CommandPool pool; - uint32_t cmd_buffer_idx; - std::vector<vk::CommandBuffer> cmd_buffers; - - vk::PipelineStageFlags stage_flags; -}; +#define MAX_PARAMETER_COUNT 8 struct vk_pipeline_struct { std::string name; vk::ShaderModule shader_module; - vk::DescriptorSetLayout dsl; - std::vector<vk::DescriptorPool> descriptor_pools; - std::vector<vk::DescriptorSet> descriptor_sets; - uint32_t descriptor_set_idx; vk::PipelineLayout layout; vk::Pipeline pipeline; uint32_t push_constant_size; uint32_t parameter_count; std::array<uint32_t, 3> wg_denoms; uint32_t align; + // set to true to request the pipeline is compiled after the dryrun + bool needed {}; + // set to true when the shader has been compiled + bool compiled {}; }; typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline; @@ -103,6 +139,15 @@ struct vk_matmul_pipeline_struct { typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline; +struct vk_matmul_pipeline2 { + vk_matmul_pipeline2() { + f16acc = std::make_shared<vk_matmul_pipeline_struct>(); + f32acc = std::make_shared<vk_matmul_pipeline_struct>(); + } + vk_matmul_pipeline f32acc; + vk_matmul_pipeline f16acc; +}; + struct vk_device_struct; typedef std::shared_ptr<vk_device_struct> vk_device; typedef std::weak_ptr<vk_device_struct> vk_device_ref; @@ -116,11 +161,50 @@ struct ggml_backend_vk_buffer_type_context { vk_device device; }; -GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft); -GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor); +struct vk_queue; + +// Stores command pool/buffers. There's an instance of this +// for each (context,queue) pair and for each (device,queue) pair. +struct vk_command_pool { + void init(vk_device& device, vk_queue *q_); + void destroy(vk::Device& device); + + vk::CommandPool pool; + uint32_t cmd_buffer_idx; + std::vector<vk::CommandBuffer> cmd_buffers; + + vk_queue *q; +}; + +// Prevent simultaneous submissions to the same queue. +// This could be per vk_queue if we stopped having two vk_queue structures +// sharing the same vk::Queue. +static std::mutex queue_mutex; + +struct vk_queue { + uint32_t queue_family_index; + vk::Queue queue; + + vk_command_pool cmd_pool; + + vk::PipelineStageFlags stage_flags; + + bool transfer_only; + + // copy everything except the cmd_pool + void copyFrom(vk_queue &other) { + queue_family_index = other.queue_family_index; + queue = other.queue; + stage_flags = other.stage_flags; + transfer_only = other.transfer_only; + } +}; + +static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft); +static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size); +static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft); +static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft); +static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor); static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = { /* .get_name = */ ggml_backend_vk_buffer_type_name, /* .alloc_buffer = */ ggml_backend_vk_buffer_type_alloc_buffer, @@ -133,79 +217,276 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = { #ifdef GGML_VULKAN_MEMORY_DEBUG class vk_memory_logger; #endif +class vk_perf_logger; static void ggml_vk_destroy_buffer(vk_buffer& buf); +static constexpr uint32_t mul_mat_vec_max_cols = 8; +static constexpr uint32_t p021_max_gqa_ratio = 8; + +enum vk_device_architecture { + OTHER, + AMD_GCN, + AMD_RDNA1, + AMD_RDNA2, + AMD_RDNA3, + INTEL_XE2, +}; + +static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) { + vk::PhysicalDeviceProperties props = device.getProperties(); + + if (props.vendorID == VK_VENDOR_ID_AMD) { + const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties(); + + bool amd_shader_core_properties = false; + bool integer_dot_product = false; + bool subgroup_size_control = false; + + for (const auto& properties : ext_props) { + if (strcmp("VK_AMD_shader_core_properties", properties.extensionName) == 0) { + amd_shader_core_properties = true; + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) { + integer_dot_product = true; + } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) { + subgroup_size_control = true; + } + } + + if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) { + return vk_device_architecture::OTHER; + } + + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props; + vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props; + + props2.pNext = &shader_core_props_amd; + shader_core_props_amd.pNext = &integer_dot_props; + integer_dot_props.pNext = &subgroup_size_control_props; + + device.getProperties2(&props2); + + if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) { + return vk_device_architecture::AMD_GCN; + } + if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) { + // RDNA + if (shader_core_props_amd.wavefrontsPerSimd == 20) { + return vk_device_architecture::AMD_RDNA1; + } + if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) { + return vk_device_architecture::AMD_RDNA3; + } + return vk_device_architecture::AMD_RDNA2; + } + } else if (props.vendorID == VK_VENDOR_ID_INTEL) { + const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties(); + + bool subgroup_size_control = false; + + for (const auto& properties : ext_props) { + if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) { + subgroup_size_control = true; + } + } + + if (!subgroup_size_control) { + return vk_device_architecture::OTHER; + } + + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props; + + props2.pNext = &subgroup_size_control_props; + device.getProperties2(&props2); + + if (subgroup_size_control_props.minSubgroupSize == 16) { + // Xe2 architecture uses SIMD16 while previous Xe and Gen architecture uses SIMD8. + // Minimum subgroup size matches the SIMD width so we distinguish architecture by checking this value. + // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html + // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html + return vk_device_architecture::INTEL_XE2; + } + } + return vk_device_architecture::OTHER; +} + struct vk_device_struct { - std::mutex mutex; + std::recursive_mutex mutex; vk::PhysicalDevice physical_device; vk::PhysicalDeviceProperties properties; std::string name; uint64_t max_memory_allocation_size; + uint64_t suballocation_block_size; bool fp16; + bool pipeline_robustness; vk::Device device; uint32_t vendor_id; + vk::DriverId driver_id; + vk_device_architecture architecture; vk_queue compute_queue; vk_queue transfer_queue; bool single_queue; - uint32_t descriptor_set_mode; uint32_t subgroup_size; + uint32_t shader_core_count; bool uma; + bool prefer_host_memory; + bool float_controls_rte_fp16; + bool subgroup_add; + bool subgroup_shuffle; + + bool integer_dot_product; + + bool subgroup_size_control; + uint32_t subgroup_min_size; + uint32_t subgroup_max_size; + bool subgroup_require_full_support; + + bool coopmat_support; + bool coopmat_acc_f32_support {}; + bool coopmat_acc_f16_support {}; + bool coopmat_bf16_support {}; + bool coopmat_support_16x16x16_f16acc {}; + bool coopmat_support_16x16x16_f32acc {}; + bool coopmat1_fa_support {}; + uint32_t coopmat_m; + uint32_t coopmat_n; + uint32_t coopmat_k; + + bool coopmat_int_support; + uint32_t coopmat_int_m; + uint32_t coopmat_int_n; + uint32_t coopmat_int_k; + + bool coopmat2; size_t idx; - vk_matmul_pipeline pipeline_matmul_f32; - vk_matmul_pipeline pipeline_matmul_f32_f16; - vk_matmul_pipeline pipeline_matmul_f16; - vk_matmul_pipeline pipeline_matmul_f16_f32; - vk_pipeline pipeline_matmul_split_k_reduce; + bool mul_mat_l[GGML_TYPE_COUNT]; + bool mul_mat_m[GGML_TYPE_COUNT]; + bool mul_mat_s[GGML_TYPE_COUNT]; + bool mul_mat_id_l[GGML_TYPE_COUNT]; + bool mul_mat_id_m[GGML_TYPE_COUNT]; + bool mul_mat_id_s[GGML_TYPE_COUNT]; - vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT]; + // set to true to indicate that some shaders need to be compiled after the dryrun + bool need_compiles {}; + + vk::DescriptorSetLayout dsl; - vk_matmul_pipeline pipeline_matmul_id_f32; - vk_matmul_pipeline pipeline_matmul_id_f16; - vk_matmul_pipeline pipeline_matmul_id_f16_f32; + vk_matmul_pipeline pipeline_matmul_f32 {}; + vk_matmul_pipeline pipeline_matmul_f32_f16 {}; + vk_matmul_pipeline pipeline_matmul_bf16 {}; + vk_matmul_pipeline2 pipeline_matmul_f16; + vk_matmul_pipeline2 pipeline_matmul_f16_f32; - vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT]; + + vk_matmul_pipeline pipeline_matmul_id_f32 {}; + vk_matmul_pipeline pipeline_matmul_id_bf16 {}; + vk_matmul_pipeline2 pipeline_matmul_id_f16; + vk_matmul_pipeline2 pipeline_matmul_id_f16_f32; + + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT]; + + vk_pipeline pipeline_matmul_split_k_reduce; + vk_pipeline pipeline_quantize_q8_1; vk_pipeline pipeline_dequant[GGML_TYPE_COUNT]; - vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT]; - vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT]; + vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols]; + vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols]; vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT]; - vk_pipeline pipeline_mul_mat_vec_p021_f16_f32; + vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio]; vk_pipeline pipeline_mul_mat_vec_nc_f16_f32; vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT]; vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT]; - vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16; - vk_pipeline pipeline_mul_f32; - vk_pipeline pipeline_div_f32; + vk_pipeline pipeline_acc_f32; + + // [src0 0=fp32,1=fp16][src1 0=fp32,1=fp16][dst 0=fp32,1=fp16] + vk_pipeline pipeline_add[2][2][2]; + vk_pipeline pipeline_add_norepeat[2][2][2]; + vk_pipeline pipeline_sub[2][2][2]; + vk_pipeline pipeline_sub_norepeat[2][2][2]; + vk_pipeline pipeline_mul[2][2][2]; + vk_pipeline pipeline_mul_norepeat[2][2][2]; + vk_pipeline pipeline_div[2][2][2]; + vk_pipeline pipeline_div_norepeat[2][2][2]; + vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32; vk_pipeline pipeline_upscale_f32; vk_pipeline pipeline_scale_f32; vk_pipeline pipeline_sqr_f32; vk_pipeline pipeline_clamp_f32; vk_pipeline pipeline_pad_f32; - vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; + vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32; + vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f16_f32, pipeline_cpy_f32_bf16; + vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16; + vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; + vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_norm_f32; vk_pipeline pipeline_group_norm_f32; vk_pipeline pipeline_rms_norm_f32; - vk_pipeline pipeline_gelu_f32; - vk_pipeline pipeline_gelu_quick_f32; - vk_pipeline pipeline_silu_f32; - vk_pipeline pipeline_relu_f32; + vk_pipeline pipeline_rms_norm_back_f32; + + // [src/dst 0=fp32,1=fp16] + vk_pipeline pipeline_gelu[2]; + vk_pipeline pipeline_gelu_quick[2]; + vk_pipeline pipeline_silu[2]; + vk_pipeline pipeline_relu[2]; + vk_pipeline pipeline_tanh[2]; + vk_pipeline pipeline_sigmoid[2]; + vk_pipeline pipeline_leaky_relu_f32; - vk_pipeline pipeline_tanh_f32; + vk_pipeline pipeline_silu_back_f32; vk_pipeline pipeline_diag_mask_inf_f32; vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16; + vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512; + vk_pipeline pipeline_soft_max_back_f32; vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16; vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16; + vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16; + vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16; vk_pipeline pipeline_argsort_f32; vk_pipeline pipeline_sum_rows_f32; + vk_pipeline pipeline_argmax_f32; + vk_pipeline pipeline_count_equal_i32; vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16; vk_pipeline pipeline_timestep_embedding_f32; - - std::vector<vk_pipeline_ref> pipelines; + vk_pipeline pipeline_conv_transpose_1d_f32; + vk_pipeline pipeline_pool2d_f32; + vk_pipeline pipeline_rwkv_wkv6_f32; + vk_pipeline pipeline_opt_step_adamw_f32; + + // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned} + vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2]; + + vk_pipeline pipeline_flash_attn_f32_f16_D64_cm1[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D80_cm1[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D96_cm1[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D112_cm1[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D128_cm1[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D256_cm1[GGML_TYPE_COUNT][2][2][2]; + + vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2]; + + vk_pipeline pipeline_flash_attn_split_k_reduce; + + std::unordered_map<std::string, vk_pipeline_ref> pipelines; std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory; @@ -218,6 +499,11 @@ struct vk_device_struct { std::unique_ptr<vk_memory_logger> memory_logger; #endif + // for GGML_VK_PERF_LOGGER + std::unique_ptr<vk_perf_logger> perf_logger; + vk::QueryPool query_pool; + int32_t num_queries; + ~vk_device_struct() { VK_LOG_DEBUG("destroy device " << name); @@ -225,25 +511,39 @@ struct vk_device_struct { ggml_vk_destroy_buffer(sync_staging); - device.destroyCommandPool(compute_queue.pool); - if (!single_queue) { - device.destroyCommandPool(transfer_queue.pool); - } + compute_queue.cmd_pool.destroy(device); + transfer_queue.cmd_pool.destroy(device); for (auto& pipeline : pipelines) { - if (pipeline.expired()) { + if (pipeline.second.expired()) { continue; } - vk_pipeline pl = pipeline.lock(); + vk_pipeline pl = pipeline.second.lock(); ggml_vk_destroy_pipeline(device, pl); } pipelines.clear(); + device.destroyDescriptorSetLayout(dsl); + device.destroy(); } }; +void vk_command_pool::init(vk_device& device, vk_queue *q_) { + cmd_buffer_idx = 0; + q = q_; + + vk::CommandPoolCreateInfo command_pool_create_info(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), q->queue_family_index); + pool = device->device.createCommandPool(command_pool_create_info); +} + +void vk_command_pool::destroy(vk::Device& device) { + device.destroyCommandPool(pool); + pool = nullptr; + cmd_buffers.clear(); +} + struct vk_buffer_struct { vk::Buffer buffer = VK_NULL_HANDLE; vk::DeviceMemory device_memory = VK_NULL_HANDLE; @@ -293,6 +593,7 @@ struct vk_mat_mat_push_constants { uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t k_split; uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3; + uint32_t padded_N; }; struct vk_mat_vec_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -305,6 +606,7 @@ struct vk_mat_mat_id_push_constants { uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11; + uint32_t padded_N; }; struct vk_mat_vec_id_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -312,6 +614,47 @@ struct vk_mat_vec_id_push_constants { uint32_t nei0; uint32_t ne11; }; +struct vk_flash_attn_push_constants { + uint32_t N; + uint32_t KV; + + uint32_t ne1; + uint32_t ne2; + uint32_t ne3; + + uint32_t neq2; + uint32_t neq3; + uint32_t nek2; + uint32_t nek3; + uint32_t nev2; + uint32_t nev3; + uint32_t nem1; + + uint32_t nb01; + uint32_t nb02; + uint32_t nb03; + uint32_t nb11; + uint32_t nb12; + uint32_t nb13; + uint32_t nb21; + uint32_t nb22; + uint32_t nb23; + uint32_t nb31; + + float scale; + float max_bias; + float logit_softcap; + + uint32_t mask; + uint32_t n_head_log2; + float m0; + float m1; + + uint32_t gqa_ratio; + uint32_t split_kv; + uint32_t k_num; +}; + struct vk_op_push_constants { uint32_t KX; uint32_t KY; @@ -323,16 +666,55 @@ struct vk_op_unary_push_constants { uint32_t ne; uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13; - uint32_t d_offset; + uint32_t misalign_offsets; float param1; float param2; + uint32_t ne0_012mp; uint32_t ne0_012L; + uint32_t ne0_01mp; uint32_t ne0_01L; + uint32_t ne0_0mp; uint32_t ne0_0L; + uint32_t ne1_012mp; uint32_t ne1_012L; + uint32_t ne1_01mp; uint32_t ne1_01L; + uint32_t ne1_0mp; uint32_t ne1_0L; }; +static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128"); + +// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1. +// Precompute mp (m' in the paper) and L such that division +// can be computed using a multiply (high 32b of 64b result) +// and a shift: +// +// n/d = (mulhi(n, mp) + n) >> L; +static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L) +{ + // compute L = ceil(log2(d)); + L = 0; + while (L < 32 && (uint32_t{1} << L) < d) { + L++; + } + + mp = (uint32_t)((uint64_t{1} << 32) * ((uint64_t{1} << L) - d) / d + 1); +} + +template <typename T> void init_pushconst_fastdiv(T &p) { + GGML_UNUSED(p); + static_assert(!std::is_const<T>::value, "unexpected type"); +} + +template <> void init_pushconst_fastdiv(vk_op_unary_push_constants &p) { + // Compute magic values to divide by these six numbers. + init_fastdiv_values(p.ne02*p.ne01*p.ne00, p.ne0_012mp, p.ne0_012L); + init_fastdiv_values(p.ne01*p.ne00, p.ne0_01mp, p.ne0_01L); + init_fastdiv_values(p.ne00, p.ne0_0mp, p.ne0_0L); + init_fastdiv_values(p.ne12*p.ne11*p.ne10, p.ne1_012mp, p.ne1_012L); + init_fastdiv_values(p.ne11*p.ne10, p.ne1_01mp, p.ne1_01L); + init_fastdiv_values(p.ne10, p.ne1_0mp, p.ne1_0L); +} struct vk_op_binary_push_constants { uint32_t ne; uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13; uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23; - uint32_t d_offset; + uint32_t misalign_offsets; float param1; float param2; int32_t param3; }; @@ -353,6 +735,11 @@ struct vk_op_rope_push_constants { float corr_dims[2]; float theta_scale; uint32_t has_ff; + uint32_t ne02; + uint32_t s1; + uint32_t s2; + int32_t sections[4]; + uint32_t is_back; }; struct vk_op_soft_max_push_constants { @@ -363,6 +750,7 @@ struct vk_op_soft_max_push_constants { float m0; float m1; uint32_t n_head_log2; + uint32_t nrows_x; }; struct vk_op_argsort_push_constants { @@ -390,22 +778,75 @@ struct vk_op_timestep_embedding_push_constants { uint32_t max_period; }; -// Allow pre-recording command buffers -struct vk_staging_memcpy { - vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {} +struct vk_op_conv_transpose_1d_push_constants { + uint32_t Cout; + uint32_t Cin; + uint32_t K; + uint32_t L; + uint32_t KL; - void * dst; - const void * src; - size_t n; + uint32_t nb01; + uint32_t nb02; + uint32_t nb11; + uint32_t nb1; + + int32_t s0; +}; + +struct vk_op_pool2d_push_constants { + uint32_t IW; uint32_t IH; + uint32_t OW; uint32_t OH; + uint32_t OC; + uint32_t pelements; + uint32_t op; + int32_t k0; int32_t k1; + int32_t s0; int32_t s1; + int32_t p0; int32_t p1; +}; + +struct vk_op_rwkv_wkv6_push_constants { + uint32_t B; + uint32_t T; + uint32_t C; + uint32_t H; +}; + + +struct vk_op_conv2d_dw_push_constants { + uint32_t ne; + uint32_t batches; + uint32_t channels; + uint32_t dst_w; + uint32_t dst_h; + uint32_t src_w; + uint32_t src_h; + uint32_t knl_w; + uint32_t knl_h; + int32_t stride_x; + int32_t stride_y; + int32_t pad_x; + int32_t pad_y; + int32_t dilation_x; + int32_t dilation_y; }; + struct vk_op_upscale_push_constants { - uint32_t ne; uint32_t d_offset; + uint32_t ne; uint32_t a_offset; uint32_t d_offset; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; float sf0; float sf1; float sf2; float sf3; }; +// Allow pre-recording command buffers +struct vk_staging_memcpy { + vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {} + + void * dst; + const void * src; + size_t n; +}; + struct vk_context_struct { vk_submission * s; std::vector<vk_sequence> seqs; @@ -415,21 +856,11 @@ struct vk_context_struct { std::vector<vk_staging_memcpy> in_memcpys; std::vector<vk_staging_memcpy> out_memcpys; - vk_queue * q; + vk_command_pool * p {}; }; typedef std::shared_ptr<vk_context_struct> vk_context; typedef std::weak_ptr<vk_context_struct> vk_context_ref; -struct ggml_tensor_extra_gpu { - vk_buffer_ref buffer_gpu; - uint64_t offset; - - void reset() { - buffer_gpu.reset(); - offset = 0; - } -}; - struct ggml_vk_garbage_collector { std::vector<vk_semaphore> tl_semaphores; std::vector<vk_semaphore> semaphores; @@ -479,6 +910,45 @@ private: #define VK_LOG_MEMORY(msg) ((void) 0) #endif // GGML_VULKAN_MEMORY_DEBUG +class vk_perf_logger { +public: + void print_timings() { + std::cerr << "----------------\nVulkan Timings:" << std::endl; + for (const auto& t : timings) { + uint64_t total = 0; + for (const auto& time : t.second) { + total += time; + } + std::cerr << t.first << ": " << t.second.size() << " x " << (total / t.second.size() / 1000.0) << " us" << std::endl; + } + + timings.clear(); + } + + void log_timing(const ggml_tensor * node, uint64_t time) { + if (node->op == GGML_OP_UNARY) { + timings[ggml_unary_op_name(ggml_get_unary_op(node))].push_back(time); + return; + } + if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) { + const uint64_t m = node->src[0]->ne[1]; + const uint64_t n = node->src[1]->ne[1]; + const uint64_t k = node->src[1]->ne[0]; + std::string name = ggml_op_name(node->op); + if (n == 1) { + name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k); + } else { + name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k); + } + timings[name].push_back(time); + return; + } + timings[ggml_op_name(node->op)].push_back(time); + } +private: + std::map<std::string, std::vector<uint64_t>> timings; +}; + struct ggml_backend_vk_context { std::string name; @@ -488,10 +958,8 @@ struct ggml_backend_vk_context { ggml_vk_garbage_collector gc; size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k; vk_buffer prealloc_x, prealloc_y, prealloc_split_k; - vk::Fence fence; - vk_buffer staging; - size_t staging_size; - size_t staging_offset; + vk::Fence fence, almost_ready_fence; + bool almost_ready_fence_pending {}; vk_buffer buffer_pool[MAX_VK_BUFFERS]; @@ -499,6 +967,39 @@ struct ggml_backend_vk_context { vk_context_ref transfer_ctx; std::vector<vk_context_ref> tensor_ctxs; + + std::vector<vk::DescriptorPool> descriptor_pools; + std::vector<vk::DescriptorSet> descriptor_sets; + uint32_t descriptor_set_idx {}; + uint32_t pipeline_descriptor_set_requirements {}; + + vk_command_pool compute_cmd_pool; + vk_command_pool transfer_cmd_pool; +}; + +static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT + +static uint64_t vk_tensor_offset(const ggml_tensor * tensor) { + if (tensor->view_src) { + return (uint8_t *) tensor->view_src->data - (uint8_t *) vk_ptr_base; + } + return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base; +} + +struct ggml_backend_vk_buffer_context { + vk_device_ref device; + vk_buffer dev_buffer; + std::string name; + + ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) : + device(device), + dev_buffer(dev_buffer), + name(name) { + } + + ~ggml_backend_vk_buffer_context() { + ggml_vk_destroy_buffer(dev_buffer); + } }; #ifdef GGML_VULKAN_MEMORY_DEBUG @@ -537,6 +1038,14 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) { struct vk_instance_t { vk::Instance instance; + bool debug_utils_support = false; // VK_EXT_debug_utils enabled + PFN_vkSetDebugUtilsObjectNameEXT pfn_vkSetDebugUtilsObjectNameEXT = {}; + PFN_vkQueueBeginDebugUtilsLabelEXT pfn_vkQueueBeginDebugUtilsLabelEXT = {}; + PFN_vkQueueEndDebugUtilsLabelEXT pfn_vkQueueEndDebugUtilsLabelEXT = {}; + PFN_vkCmdBeginDebugUtilsLabelEXT pfn_vkCmdBeginDebugUtilsLabelEXT = {}; + PFN_vkCmdEndDebugUtilsLabelEXT pfn_vkCmdEndDebugUtilsLabelEXT = {}; + PFN_vkCmdInsertDebugUtilsLabelEXT pfn_vkCmdInsertDebugUtilsLabelEXT = {}; + std::vector<size_t> device_indices; vk_device devices[GGML_VK_MAX_DEVICES]; }; @@ -544,6 +1053,8 @@ struct vk_instance_t { static bool vk_instance_initialized = false; static vk_instance_t vk_instance; +static bool vk_perf_logger_enabled = false; + #ifdef GGML_VULKAN_CHECK_RESULTS static size_t vk_skip_checks; static size_t vk_output_tensor; @@ -557,77 +1068,64 @@ typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend); -static void ggml_vk_create_pipeline(vk_device& device, vk_pipeline& pipeline, const std::string& name, size_t spv_size, const void* spv_data, const std::string& entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) { - VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")"); - GGML_ASSERT(parameter_count > 0); - GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT +// Wait for ctx->fence to be signaled. +static void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx) { + // Use waitForFences while most of the graph executes. Hopefully the CPU can sleep + // during this wait. + if (ctx->almost_ready_fence_pending) { + VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence"); + ctx->device->device.resetFences({ ctx->almost_ready_fence }); + ctx->almost_ready_fence_pending = false; + } + + // Spin (w/pause) waiting for the graph to finish executing. + vk::Result result; + while ((result = ctx->device->device.getFenceStatus(ctx->fence)) != vk::Result::eSuccess) { + if (result != vk::Result::eNotReady) { + fprintf(stderr, "ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__); + exit(1); + } + for (uint32_t i = 0; i < 100; ++i) { + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + } + } + ctx->device->device.resetFences({ ctx->fence }); +} - std::lock_guard<std::mutex> guard(device->mutex); +// variables to track number of compiles in progress +static uint32_t compile_count = 0; +static std::mutex compile_count_mutex; +static std::condition_variable compile_count_cond; - pipeline = std::make_shared<vk_pipeline_struct>(); - pipeline->name = name; - pipeline->parameter_count = parameter_count; - pipeline->push_constant_size = push_constant_size; - pipeline->wg_denoms = wg_denoms; - pipeline->align = align; +static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, size_t spv_size, const void* spv_data, const std::string entrypoint, + uint32_t parameter_count, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, + bool disable_robustness, bool require_full_subgroups, uint32_t required_subgroup_size) { + VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << pipeline->name << ", " << entrypoint << ", " << parameter_count << + ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << + disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")"); + GGML_ASSERT(parameter_count > 0); + GGML_ASSERT(parameter_count <= MAX_PARAMETER_COUNT); + GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data)); pipeline->shader_module = device->device.createShaderModule(shader_module_create_info); - std::vector<vk::DescriptorSetLayoutBinding> dsl_binding; - std::vector<vk::DescriptorBindingFlags> dsl_binding_flags; - for (uint32_t i = 0; i < parameter_count; i++) { - dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute}); - dsl_binding_flags.push_back({}); - } - - vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags }; - vk::PushConstantRange pcr( vk::ShaderStageFlagBits::eCompute, 0, pipeline->push_constant_size ); - vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info( - {}, - dsl_binding); - descriptor_set_layout_create_info.setPNext(&dslbfci); - pipeline->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info); - - // Check if device supports multiple descriptors per pool - if (device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN) { - const uint32_t alloc_count = 2; - - // Try allocating multiple sets from one pool - // This fails on AMD for some reason, so add a fall back to allocating one pool per set - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); - vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, alloc_count, descriptor_pool_size); - vk::DescriptorPool pool = device->device.createDescriptorPool(descriptor_pool_create_info); - - std::vector<vk::DescriptorSetLayout> layouts(alloc_count); - for (uint32_t i = 0; i < alloc_count; i++) { - layouts[i] = pipeline->dsl; - } - try { - vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pool, alloc_count, layouts.data()); - std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info); - } catch(vk::OutOfPoolMemoryError const&) { - device->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE; - } - - device->device.destroyDescriptorPool(pool); - } - - if (device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); - vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 128, descriptor_pool_size); - pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info)); - } - - pipeline->descriptor_set_idx = 0; - - vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr); + vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), device->dsl, pcr); pipeline->layout = device->device.createPipelineLayout(pipeline_layout_create_info); std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size()); @@ -645,32 +1143,71 @@ static void ggml_vk_create_pipeline(vk_device& device, vk_pipeline& pipeline, co specialization_constants.data() ); + vk::PipelineShaderStageCreateFlags pipeline_shader_stage_create_flags{}; + + if (device->subgroup_require_full_support && require_full_subgroups) { + pipeline_shader_stage_create_flags |= vk::PipelineShaderStageCreateFlagBits::eRequireFullSubgroupsEXT; + } + vk::PipelineShaderStageCreateInfo pipeline_shader_create_info( - vk::PipelineShaderStageCreateFlags(), + pipeline_shader_stage_create_flags, vk::ShaderStageFlagBits::eCompute, pipeline->shader_module, entrypoint.c_str(), &specialization_info); + + vk::PipelineShaderStageRequiredSubgroupSizeCreateInfoEXT pipeline_shader_stage_required_subgroup_size_create_info; + pipeline_shader_stage_required_subgroup_size_create_info.requiredSubgroupSize = required_subgroup_size; + if (device->subgroup_size_control && required_subgroup_size > 0) { + GGML_ASSERT(device->subgroup_min_size <= required_subgroup_size && required_subgroup_size <= device->subgroup_max_size); + pipeline_shader_create_info.setPNext(&pipeline_shader_stage_required_subgroup_size_create_info); + } + vk::ComputePipelineCreateInfo compute_pipeline_create_info( - vk::PipelineCreateFlags(), + vk::PipelineCreateFlags{}, pipeline_shader_create_info, pipeline->layout); - pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; - device->pipelines.push_back(pipeline); -} + vk::PipelineRobustnessCreateInfoEXT rci; -static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) { - VK_LOG_DEBUG("ggml_pipeline_destroy_pipeline(" << pipeline->name << ")"); - for (auto& pool : pipeline->descriptor_pools) { - device.destroyDescriptorPool(pool); + if (device->pipeline_robustness && disable_robustness) { + rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + compute_pipeline_create_info.setPNext(&rci); + } + + try { + pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; + } catch (const vk::SystemError& e) { + std::cerr << "ggml_vulkan: Compute pipeline creation failed for " << pipeline->name << std::endl; + std::cerr << "ggml_vulkan: " << e.what() << std::endl; + throw e; } - pipeline->descriptor_pools.clear(); - pipeline->descriptor_sets.clear(); - pipeline->descriptor_set_idx = 0; + pipeline->compiled = true; - device.destroyDescriptorSetLayout(pipeline->dsl); + if (vk_instance.debug_utils_support) { + vk::DebugUtilsObjectNameInfoEXT duoni; + duoni.objectType = vk::ObjectType::ePipeline; + duoni.pObjectName = pipeline->name.c_str(); + duoni.objectHandle = reinterpret_cast<uint64_t>(static_cast<VkPipeline_T*>(pipeline->pipeline)); + vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast<VkDebugUtilsObjectNameInfoEXT &>(duoni)); + } + { + std::lock_guard<std::recursive_mutex> guard(device->mutex); + device->pipelines.insert({ pipeline->name, pipeline }); + } + + { + std::lock_guard<std::mutex> guard(compile_count_mutex); + assert(compile_count > 0); + compile_count--; + } + compile_count_cond.notify_all(); +} + +static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) { + VK_LOG_DEBUG("ggml_pipeline_destroy_pipeline(" << pipeline->name << ")"); device.destroyPipelineLayout(pipeline->layout); device.destroyShaderModule(pipeline->shader_module); @@ -678,76 +1215,78 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) device.destroyPipeline(pipeline->pipeline); } -static void ggml_pipeline_allocate_descriptor_sets(vk_device& device, vk_pipeline& pipeline, uint32_t n) { - VK_LOG_DEBUG("ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")"); - if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) { +static void ggml_pipeline_request_descriptor_sets(ggml_backend_vk_context *ctx, vk_pipeline& pipeline, uint32_t n) { + VK_LOG_DEBUG("ggml_pipeline_request_descriptor_sets(" << pipeline->name << ", " << n << ")"); + ctx->pipeline_descriptor_set_requirements += n; + if (!pipeline->compiled) { + pipeline->needed = true; + ctx->device->need_compiles = true; + } +} + +static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx) { + + if (ctx->descriptor_sets.size() >= ctx->pipeline_descriptor_set_requirements) { // Enough descriptors are available return; } - std::lock_guard<std::mutex> guard(device->mutex); + vk_device& device = ctx->device; - if (device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { - const uint32_t alloc_count = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size(); + uint32_t to_alloc = ctx->pipeline_descriptor_set_requirements - ctx->descriptor_sets.size(); + uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - ctx->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE; + uint32_t pool_idx = ctx->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE; + + while (to_alloc > 0) { + const uint32_t alloc_count = std::min(pool_remaining, to_alloc); + to_alloc -= alloc_count; + pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE; + + if (pool_idx >= ctx->descriptor_pools.size()) { + vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, MAX_PARAMETER_COUNT * VK_DEVICE_DESCRIPTOR_POOL_SIZE); + vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size); + ctx->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info)); + } std::vector<vk::DescriptorSetLayout> layouts(alloc_count); for (uint32_t i = 0; i < alloc_count; i++) { - layouts[i] = pipeline->dsl; + layouts[i] = device->dsl; } - vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[0], alloc_count, layouts.data()); + vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(ctx->descriptor_pools[pool_idx], alloc_count, layouts.data()); std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info); - pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end()); - } else { - for (uint32_t i = pipeline->descriptor_sets.size(); i < pipeline->descriptor_set_idx + n; i++) { - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); - vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 1, descriptor_pool_size); - pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info)); + ctx->descriptor_sets.insert(ctx->descriptor_sets.end(), sets.begin(), sets.end()); - vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[i], 1, &pipeline->dsl); - std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info); - pipeline->descriptor_sets.push_back(sets[0]); - } + pool_idx++; } } -static void ggml_pipeline_cleanup(vk_pipeline& pipeline) { - VK_LOG_DEBUG("ggml_pipeline_cleanup(" << pipeline->name << ")"); - pipeline->descriptor_set_idx = 0; -} - -static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_queue& q) { +static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_command_pool& p) { VK_LOG_DEBUG("ggml_vk_create_cmd_buffer()"); - std::lock_guard<std::mutex> guard(device->mutex); - if (q.cmd_buffers.size() > q.cmd_buffer_idx) { + if (p.cmd_buffers.size() > p.cmd_buffer_idx) { // Reuse command buffer - return q.cmd_buffers[q.cmd_buffer_idx++]; + return p.cmd_buffers[p.cmd_buffer_idx++]; } vk::CommandBufferAllocateInfo command_buffer_alloc_info( - q.pool, + p.pool, vk::CommandBufferLevel::ePrimary, 1); const std::vector<vk::CommandBuffer> cmd_buffers = device->device.allocateCommandBuffers(command_buffer_alloc_info); auto buf = cmd_buffers.front(); - q.cmd_buffers.push_back(buf); - q.cmd_buffer_idx++; + p.cmd_buffers.push_back(buf); + p.cmd_buffer_idx++; return buf; } -static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) { - VK_LOG_DEBUG("ggml_vk_create_submission()"); - vk_submission s; - s.buffer = ggml_vk_create_cmd_buffer(device, q); - s.wait_semaphores = std::move(wait_semaphores); - s.signal_semaphores = std::move(signal_semaphores); - return s; -} - static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) { if (ctx->seqs.empty()) { + if (fence) { + std::lock_guard<std::mutex> guard(queue_mutex); + ctx->p->q->queue.submit({}, fence); + } return; } VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")"); @@ -785,7 +1324,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) { tl_signal_vals.push_back({}); tl_signal_semaphores.push_back({}); for (size_t i = 0; i < submission.wait_semaphores.size(); i++) { - stage_flags[idx].push_back(ctx->q->stage_flags); + stage_flags[idx].push_back(ctx->p->q->stage_flags); tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value); tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s); } @@ -815,7 +1354,8 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) { } } - ctx->q->queue.submit(submit_infos, fence); + std::lock_guard<std::mutex> guard(queue_mutex); + ctx->p->q->queue.submit(submit_infos, fence); ctx->seqs.clear(); } @@ -866,34 +1406,32 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope abort(); } -static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags) { +static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only) { VK_LOG_DEBUG("ggml_vk_create_queue()"); - std::lock_guard<std::mutex> guard(device->mutex); + std::lock_guard<std::recursive_mutex> guard(device->mutex); q.queue_family_index = queue_family_index; + q.transfer_only = transfer_only; - vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index); - q.pool = device->device.createCommandPool(command_pool_create_info_compute); - - q.cmd_buffer_idx = 0; + q.cmd_pool.init(device, &q); q.queue = device->device.getQueue(queue_family_index, queue_index); q.stage_flags = stage_flags; } -static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) { +static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_command_pool& p) { vk_context result = std::make_shared<vk_context_struct>(); VK_LOG_DEBUG("ggml_vk_create_context(" << result << ")"); ctx->gc.contexts.emplace_back(result); - result->q = &q; + result->p = &p; return result; } -static vk_context ggml_vk_create_temporary_context(vk_queue& q) { +static vk_context ggml_vk_create_temporary_context(vk_command_pool& p) { vk_context result = std::make_shared<vk_context_struct>(); VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")"); - result->q = &q; + result->p = &p; return result; } @@ -926,15 +1464,29 @@ static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) { return ctx->gc.events[ctx->event_idx++]; } -static void ggml_vk_queue_cleanup(vk_device& device, vk_queue& q) { - VK_LOG_DEBUG("ggml_vk_queue_cleanup()"); - std::lock_guard<std::mutex> guard(device->mutex); +static void ggml_vk_command_pool_cleanup(vk_device& device, vk_command_pool& p) { + VK_LOG_DEBUG("ggml_vk_command_pool_cleanup()"); // Requires command buffers to be done - device->device.resetCommandPool(q.pool); - q.cmd_buffer_idx = 0; + device->device.resetCommandPool(p.pool); + p.cmd_buffer_idx = 0; +} + +static void ggml_vk_queue_command_pools_cleanup(vk_device& device) { + VK_LOG_DEBUG("ggml_vk_queue_command_pools_cleanup()"); + + // Arbitrary frequency to cleanup/reuse command buffers + static constexpr uint32_t cleanup_frequency = 10; + + if (device->compute_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) { + ggml_vk_command_pool_cleanup(device, device->compute_queue.cmd_pool); + } + if (device->transfer_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) { + ggml_vk_command_pool_cleanup(device, device->transfer_queue.cmd_pool); + } } + static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) { for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) { vk::MemoryType memory_type = mem_props->memoryTypes[i]; @@ -953,8 +1505,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit"); } - std::lock_guard<std::mutex> guard(device->mutex); - vk_buffer buf = std::make_shared<vk_buffer_struct>(); if (size == 0) { @@ -962,7 +1512,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor return buf; } - buf->size = size; vk::BufferCreateInfo buffer_create_info{ vk::BufferCreateFlags(), size, @@ -990,17 +1539,29 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor if (memory_type_index == UINT32_MAX) { device->device.destroyBuffer(buf->buffer); - buf->size = 0; throw vk::OutOfDeviceMemoryError("No suitable memory type found"); } try { buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index }); } catch (const vk::SystemError& e) { - // Out of Host/Device memory, clean up buffer - device->device.destroyBuffer(buf->buffer); - buf->size = 0; - throw e; + if (buf->memory_property_flags != fallback_flags) { + // Try again with fallback flags + memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags); + buf->memory_property_flags = fallback_flags; + + try { + buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index }); + } + catch (const vk::SystemError& e) { + device->device.destroyBuffer(buf->buffer); + throw e; + } + } else { + // Out of Host/Device memory, clean up buffer + device->device.destroyBuffer(buf->buffer); + throw e; + } } buf->ptr = nullptr; @@ -1011,6 +1572,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0); buf->device = device; + buf->size = size; #ifdef GGML_VULKAN_MEMORY_DEBUG device->memory_logger->log_allocation(buf, size); @@ -1032,11 +1594,14 @@ static vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk: static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) { vk_buffer buf; try { - if (device->uma) { + if (device->prefer_host_memory) { + buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal); + } else if (device->uma) { // Fall back to host memory type buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); } else { - buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal); + // use rebar if available, otherwise fallback to device only visible memory + buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal); } } catch (const vk::SystemError& e) { std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl; @@ -1067,13 +1632,16 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) { static void ggml_vk_sync_buffers(vk_context& ctx) { VK_LOG_DEBUG("ggml_vk_sync_buffers()"); + + const bool transfer_queue = ctx->p->q->transfer_only; + ctx->s->buffer.pipelineBarrier( - ctx->q->stage_flags, - ctx->q->stage_flags, + ctx->p->q->stage_flags, + ctx->p->q->stage_flags, {}, { { - {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite}, - {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite} + { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) }, + { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) } } }, {}, {} @@ -1088,515 +1656,927 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events ctx->s->buffer.waitEvents( events, - ctx->q->stage_flags, - ctx->q->stage_flags, + ctx->p->q->stage_flags, + ctx->p->q->stage_flags, {}, {}, {} ); } +enum FaCodePath { + FA_SCALAR, + FA_COOPMAT1, + FA_COOPMAT2, +}; + +// number of rows/cols for flash attention shader +static constexpr uint32_t flash_attention_num_small_rows = 32; +static constexpr uint32_t scalar_flash_attention_num_small_rows = 1; +static constexpr uint32_t scalar_flash_attention_num_large_rows = 8; + +// The FA coopmat1 shader assumes 16x16x16 matrix multiply support. +// 128 threads split into four subgroups, each subgroup does 1/4 +// of the Bc dimension. +static constexpr uint32_t coopmat1_flash_attention_num_large_rows = 16; +static constexpr uint32_t scalar_flash_attention_Bc = 64; +static constexpr uint32_t scalar_flash_attention_workgroup_size = 128; + +static uint32_t get_fa_num_small_rows(FaCodePath path) { + if (path == FA_COOPMAT2) { + return flash_attention_num_small_rows; + } else { + return scalar_flash_attention_num_small_rows; + } +} + +static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) { + GGML_UNUSED(clamp); + + if (path == FA_SCALAR) { + if (small_rows) { + return {scalar_flash_attention_num_small_rows, 64}; + } else { + return {scalar_flash_attention_num_large_rows, 32}; + } + } + + if (path == FA_COOPMAT1) { + if (small_rows) { + return {scalar_flash_attention_num_small_rows, scalar_flash_attention_Bc}; + } else { + return {coopmat1_flash_attention_num_large_rows, scalar_flash_attention_Bc}; + } + } + + // small rows, large cols + if (small_rows) { + return {get_fa_num_small_rows(FA_COOPMAT2), 32}; + } + + // small cols to reduce register count + if (ggml_is_quantized(type) || D == 256) { + return {64, 32}; + } + return {64, 64}; +} + +static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) { + + uint32_t lut_size = 0; + switch (src0_type) { + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + lut_size = 2*2048; + break; + case GGML_TYPE_IQ2_XXS: + lut_size = 8*256; + break; + case GGML_TYPE_IQ2_XS: + lut_size = 8*512; + break; + case GGML_TYPE_IQ2_S: + lut_size = 8*1024; + break; + case GGML_TYPE_IQ3_XXS: + lut_size = 4*256; + break; + case GGML_TYPE_IQ3_S: + lut_size = 4*512; + break; + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + lut_size = 4*16; + break; + default: + break; + } + + // Needs to be kept up to date on shader changes + const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1; + const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float); + const uint32_t warps = warptile[0] / warptile[10]; + + const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size; + const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0; + const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0; + + const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size; + const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize; + + VK_LOG_DEBUG("ggml_vk_matmul_shmem_support(warptile=(" << warptile[0] << "," << warptile[1] << "," << warptile[2] << "), " + "mul_mat_id=" << mul_mat_id << ", src0_type=" << ggml_type_name(src0_type) << ", supported=" << supported); + + return supported; +} + +struct GpuPipelineConfig { + // GPU architecture identifier. + // Example: vk_device_architecture::AMD_GCN + vk_device_architecture arch; + + // Mapping of pipeline names to their specific subgroup sizes. + // Example: {"soft_max_f32", 64} + std::unordered_map<std::string, uint32_t> pipelines; + + // Default subgroup size for this GPU. + // Defaults to 0 if not explicitly provided. + uint32_t default_subgroup_size = 0; +}; + +// Pipeline configuration for RDNA1 GPUs. +static const std::unordered_map<std::string, uint32_t> rdna1_pipelines = { + {"soft_max", 64}, {"im2col", 64}, + {"argmax", 64}, {"mul_mat_vec", 64}, + {"mul_mat_vec_f16", 32}, {"mul_mat_vec_f32_f16", 32} +}; + +// Pipeline configuration for RDNA2 GPUs. +static const std::unordered_map<std::string, uint32_t> rdna2_pipelines = { + {"soft_max", 64}, {"im2col", 64}, +}; + +static constexpr uint32_t RDNA_DEFAULT_SUBGROUP_SIZE = 32; + +// Define configurations for different GPUs. +static std::vector<GpuPipelineConfig> gpu_pipeline_configs = { + { + vk_device_architecture::AMD_RDNA1, + { + rdna1_pipelines, + }, + RDNA_DEFAULT_SUBGROUP_SIZE + }, + { + vk_device_architecture::AMD_RDNA2, + { + rdna2_pipelines, + }, + RDNA_DEFAULT_SUBGROUP_SIZE + }, +}; + +static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_device_architecture &arch) { + for (const auto &config : gpu_pipeline_configs) { + if (config.arch == arch) { + auto pipIt = config.pipelines.find(pipeline_name); + if (pipIt != config.pipelines.end()) { + return pipIt->second; + } + std::vector<std::pair<std::string, uint32_t>> sorted_pipelines(config.pipelines.begin(), config.pipelines.end()); + std::sort(sorted_pipelines.begin(), sorted_pipelines.end(), + [](const auto &a, const auto &b) { return a.first.size() > b.first.size(); }); + for (const auto &entry : sorted_pipelines) { + if (pipeline_name.find(entry.first) != std::string::npos) { + return entry.second; + } + } + return config.default_subgroup_size; + } + } + return 0; // If no matching configuration is found +} + static void ggml_vk_load_shaders(vk_device& device) { VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")"); + // some shaders have a minimum subgroup size + const uint32_t subgroup_size_8 = std::max(device->subgroup_size, 8u); + const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u); + const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u); + // mulmat - std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; - std::initializer_list<uint32_t> warptile_m = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - std::initializer_list<uint32_t> warptile_s = { device->subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; - - std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; - std::initializer_list<uint32_t> warptile_mmq_m = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - std::initializer_list<uint32_t> warptile_mmq_s = { device->subgroup_size, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; - - std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 }; - std::array<uint32_t, 3> m_wg_denoms = { 64, 64, 1 }; - std::array<uint32_t, 3> s_wg_denoms = { 32, 32, 1 }; - - uint32_t l_align = 128; - uint32_t m_align = 64; - uint32_t s_align = 32; - - device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>(); - - device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_matmul_id_f16 = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>(); + std::vector<uint32_t> l_warptile, m_warptile, s_warptile, + l_warptile_mmq, m_warptile_mmq, s_warptile_mmq, + l_warptile_mmq_int, m_warptile_mmq_int, s_warptile_mmq_int, + l_warptile_mmq_k, m_warptile_mmq_k, s_warptile_mmq_k, + l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid; + std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms, + l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms, + l_mmq_wg_denoms_k, m_mmq_wg_denoms_k, s_mmq_wg_denoms_k, + l_mmqid_wg_denoms, m_mmqid_wg_denoms, s_mmqid_wg_denoms; + + uint32_t l_align, m_align, s_align; + if (device->coopmat2) { + // spec constants and tile sizes for non-quant matmul/matmul_id + l_warptile = { 256, 128, 256, 64, 1 }; + m_warptile = { 256, 128, 128, 64, 0 }; + s_warptile = { 128, 64, 64, 64, 0 }; + l_wg_denoms = {128, 256, 1 }; + m_wg_denoms = {128, 128, 1 }; + s_wg_denoms = { 64, 64, 1 }; + + // spec constants and tile sizes for quant matmul (non-Qi_K) + l_warptile_mmq = { 256, 128, 256, 64, 1 }; + m_warptile_mmq = { 256, 128, 128, 64, 1 }; + s_warptile_mmq = { 256, 32, 64, 128, 0 }; + l_mmq_wg_denoms = { 128, 256, 1 }; + m_mmq_wg_denoms = { 128, 128, 1 }; + s_mmq_wg_denoms = { 32, 64, 1 }; + + // spec constants and tile sizes for quant matmul (Qi_K) + l_warptile_mmq_k = { 256, 64, 128, 64, 1 }; + m_warptile_mmq_k = { 256, 32, 64, 64, 0 }; + s_warptile_mmq_k = { 256, 32, 32, 128, 0 }; + l_mmq_wg_denoms_k = { 64, 128, 1 }; + m_mmq_wg_denoms_k = { 32, 64, 1 }; + s_mmq_wg_denoms_k = { 32, 32, 1 }; + + // spec constants and tile sizes for quant matmul_id + l_warptile_mmqid = { 256, 128, 64, 16, 0 }; + m_warptile_mmqid = { 256, 128, 64, 16, 0 }; + s_warptile_mmqid = { 256, 128, 64, 16, 0 }; + l_mmqid_wg_denoms = { 128, 64, 1 }; + m_mmqid_wg_denoms = { 128, 64, 1 }; + s_mmqid_wg_denoms = { 128, 64, 1 }; + + l_align = 128; + m_align = 64; + s_align = 32; + } else { + // Matrix cores require different warp group sizes + const uint32_t tm_l = device->coopmat_support ? device->coopmat_m : 4; + const uint32_t tm_m = device->coopmat_support ? device->coopmat_m : 4; + const uint32_t tm_s = device->coopmat_support ? device->coopmat_m : 2; + const uint32_t tn_l = device->coopmat_support ? device->coopmat_n : 4; + const uint32_t tn_m = device->coopmat_support ? device->coopmat_n : 2; + const uint32_t tn_s = device->coopmat_support ? device->coopmat_n : 2; + const uint32_t tk_l = device->coopmat_support ? device->coopmat_k : 1; + const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1; + const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1; + + l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; + m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; + + l_warptile_mmq = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; + m_warptile_mmq = { 128, 64, 64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; + s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; + + l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; + m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 }; + s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32, 32, 2, 2, 1, 1, subgroup_size_8 }; + + // chip specific tuning + if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) { + m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 }; + } + + l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 }; + m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 }; + s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 }; + l_align = 128; + m_align = 64; + s_align = 32; + + for (uint32_t i = 0; i < GGML_TYPE_COUNT; ++i) { + ggml_type t = (ggml_type)i; + // Disable medium and large matrix multiplication if not enough shared memory is available + // Check mmq warptiles as the largest configuration + // Throw an error if not enough for any matrix multiplication is available + if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, false, t)) { + std::cerr << "ggml_vulkan: Error: Shared memory size too small for matrix multiplication." << std::endl; + throw std::runtime_error("Shared memory size too small for matrix multiplication."); + } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, false, t)) { + device->mul_mat_m[i] = false; + device->mul_mat_l[i] = false; + } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, false, t)) { + device->mul_mat_l[i] = false; + } + + // Disable mul_mat_id if not enough shared memory is available + if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, true, t)) { + device->mul_mat_id_s[i] = false; + device->mul_mat_id_m[i] = false; + device->mul_mat_id_l[i] = false; + } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, true, t)) { + device->mul_mat_id_m[i] = false; + device->mul_mat_id_l[i] = false; + } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, true, t)) { + device->mul_mat_id_l[i] = false; + } + } + } + + if (!device->pipeline_matmul_f32) { + device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>(); + } + if (!device->pipeline_matmul_f32_f16) { + device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>(); + } + if (!device->pipeline_matmul_id_f32) { + device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>(); + } + if (!device->pipeline_matmul_bf16) { + device->pipeline_matmul_bf16 = std::make_shared<vk_matmul_pipeline_struct>(); + } + if (!device->pipeline_matmul_id_bf16) { + device->pipeline_matmul_id_bf16 = std::make_shared<vk_matmul_pipeline_struct>(); + } + + std::vector<std::future<void>> compiles; + auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, + uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, + uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) { + + if (!require_full_subgroups && required_subgroup_size == 0) { + required_subgroup_size = get_subgroup_size(name, device->architecture); + } + + if (!pipeline) { + pipeline = std::make_shared<vk_pipeline_struct>(); + pipeline->name = name; + pipeline->parameter_count = parameter_count; + pipeline->push_constant_size = push_constant_size; + pipeline->wg_denoms = wg_denoms; + pipeline->align = align; + } + + if (!pipeline->needed || pipeline->compiled) { + return; + } + { + // wait until fewer than N compiles are in progress + uint32_t N = std::max(1u, std::thread::hardware_concurrency()); + std::unique_lock<std::mutex> guard(compile_count_mutex); + while (compile_count >= N) { + compile_count_cond.wait(guard); + } + compile_count++; + } + compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint, + parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size)); + }; + + auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> { + return {fa_rows_cols(path, D, clamp, type, small_rows)[0], 1, 1}; + }; + + auto const &fa_spec_constants = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> { + // For large number of rows, 128 invocations seems to work best. + // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we + // can't use 256 for D==80. + // For scalar, use 128 (arbitrary) + uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1) + ? scalar_flash_attention_workgroup_size + : ((small_rows && (D % 32) == 0) ? 256 : 128); + auto rows_cols = fa_rows_cols(path, D, clamp, type, small_rows); + + // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it. + // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader. + const uint32_t D_lsb = D ^ (D & (D-1)); + uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4); + + // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads + GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0); + return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split}; + }; + +#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, D) \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true), fa_rows_cols(FAPATH,D,0,TYPE,true)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true), fa_rows_cols(FAPATH,D,0,TYPE,true)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + +#define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128) \ + CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256) + + CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, ) + CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, ) + CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_SCALAR, ) +#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + if (device->coopmat1_fa_support) { + CREATE_FA(GGML_TYPE_F16, f16, FA_COOPMAT1, _cm1) + CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_COOPMAT1, _cm1) + CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_COOPMAT1, _cm1) + } +#endif +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (device->coopmat2) { + CREATE_FA(GGML_TYPE_F16, f16, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_Q4_1, q4_1, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_Q5_0, q5_0, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_Q5_1, q5_1, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_COOPMAT2, _cm2) + CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_COOPMAT2, _cm2) + } +#endif +#undef CREATE_FA2 +#undef CREATE_FA + +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (device->coopmat2) { + + // Create 6 variants, {s,m,l}x{unaligned,aligned} +#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ + + // Create 2 variants, {f16,f32} accumulator +#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ + CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ + CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ + + CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3) + } +#endif + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0], matmul_q4_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1], matmul_q4_1_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0], matmul_q5_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_1], matmul_q5_1_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0], matmul_q8_0_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K], matmul_q2_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K], matmul_q3_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K], matmul_q4_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K], matmul_q5_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K], matmul_q6_k_f16, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_S], matmul_iq1_s_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_M], matmul_iq1_m_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XS], matmul_iq2_xs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_S], matmul_iq2_s_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_S], matmul_iq3_s_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_XS], matmul_iq4_xs_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL], matmul_iq4_nl_f16, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + + CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + } +#endif + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) +#undef CREATE_MM +#undef CREATE_MM2 + } else +#endif // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) +#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + if (device->coopmat_support) { + // Create 6 variants, {s,m,l}x{unaligned,aligned} +#define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true); \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true); \ + + // Create 2 variants, {f16,f32} accumulator +#define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->coopmat_acc_f16_support) { \ + CREATE_MM(TYPE, PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + } \ + if (device->coopmat_acc_f32_support) { \ + CREATE_MM(TYPE, PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + } \ + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ) + } +#endif + + if (device->coopmat_acc_f16_support) { + CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0], matmul_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1], matmul_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0], matmul_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K], matmul_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S], matmul_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M], matmul_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS], matmul_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S], matmul_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S], matmul_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS], matmul_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL], matmul_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + } else { + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f32acc, matmul_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f32acc, matmul_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f32acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc, matmul_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc, matmul_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f32acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc, matmul_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc, matmul_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + } + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + } +#endif + if (device->coopmat_acc_f16_support) { + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + } else { + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + } +#undef CREATE_MM2 +#undef CREATE_MM + } else +#endif // defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) if (device->fp16) { - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); + // Create 6 variants, {s,m,l}x{unaligned,aligned} +#define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ + +#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) { \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->l, #NAMELC "_f16acc_l", NAMELC ## _f16acc_len, NAMELC ## _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->l, #NAMELC "_l", NAMELC ## _len, NAMELC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + } \ + if (device->mul_mat ## ID ## _m[TYPE]) { \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->m, #NAMELC "_f16acc_m", NAMELC ## _f16acc_len, NAMELC ## _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->m, #NAMELC "_m", NAMELC ## _len, NAMELC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + } \ + if (device->mul_mat ## ID ## _s[TYPE]) { \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->s, #NAMELC "_f16acc_s", NAMELC ## _f16acc_len, NAMELC ## _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->s, #NAMELC "_s", NAMELC ## _len, NAMELC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + } \ + + // Create 2 variants, {f16,f32} accumulator +#define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + CREATE_MM(TYPE, PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + CREATE_MM(TYPE, PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + + CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0], matmul_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1], matmul_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0], matmul_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K], matmul_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K], matmul_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S], matmul_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M], matmul_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS], matmul_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS], matmul_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S], matmul_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS], matmul_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S], matmul_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS], matmul_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL], matmul_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (device->integer_dot_product) { + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1], matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0], matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1], matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + } +#endif + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); +#undef CREATE_MM2 +#undef CREATE_MMQ +#undef CREATE_MM } else { - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align); + // Create 6 variants, {s,m,l}x{unaligned,aligned} +#define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ + +#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC "_l", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC "_m", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC "_s", NAMELC ## _fp32_len, NAMELC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f32acc, matmul_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f32acc, matmul_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f32acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc, matmul_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc, matmul_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f32acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc, matmul_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc, matmul_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (device->integer_dot_product) { + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + } +#endif + + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + } + // reusing CREATE_MM from the fp32 path + if ((device->coopmat2 || device->coopmat_support) +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + && !device->coopmat_bf16_support +#endif + ) { + // use scalar tile sizes + l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; + m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 }; + + l_wg_denoms = {128, 128, 1 }; + m_wg_denoms = { 64, 64, 1 }; + s_wg_denoms = { 32, 32, 1 }; + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); } +#undef CREATE_MM // mul mat vec - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32", mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32", mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32", mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32", mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + + // the number of rows computed per shader depends on GPU model and quant + uint32_t rm_stdq = 1; + uint32_t rm_kq = 2; + if (device->vendor_id == VK_VENDOR_ID_AMD) { + if (device->architecture == AMD_GCN) { + rm_stdq = 2; + rm_kq = 4; + } + } else if (device->vendor_id == VK_VENDOR_ID_INTEL) + rm_stdq = 2; + uint32_t rm_iq = 2 * rm_kq; + + for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) { + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1), mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1), mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq1_s_f32_f32_len, mul_mat_vec_iq1_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq1_m_f32_f32_len, mul_mat_vec_iq1_m_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f32_f32_len, mul_mat_vec_iq2_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f32_f32_len, mul_mat_vec_iq2_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f32_f32_len, mul_mat_vec_iq3_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f32_f32_len, mul_mat_vec_iq4_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1), mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1), mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq1_s_f16_f32_len, mul_mat_vec_iq1_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq1_m_f16_f32_len, mul_mat_vec_iq1_m_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f16_f32_len, mul_mat_vec_iq2_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f16_f32_len, mul_mat_vec_iq2_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f16_f32_len, mul_mat_vec_iq3_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f16_f32_len, mul_mat_vec_iq4_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + } + + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S], "mul_mat_vec_id_iq1_s_f32", mul_mat_vec_id_iq1_s_f32_len, mul_mat_vec_id_iq1_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M], "mul_mat_vec_id_iq1_m_f32", mul_mat_vec_id_iq1_m_f32_len, mul_mat_vec_id_iq1_m_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS], "mul_mat_vec_id_iq2_xs_f32", mul_mat_vec_id_iq2_xs_f32_len, mul_mat_vec_id_iq2_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S], "mul_mat_vec_id_iq2_s_f32", mul_mat_vec_id_iq2_s_f32_len, mul_mat_vec_id_iq2_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S], "mul_mat_vec_id_iq3_s_f32", mul_mat_vec_id_iq3_s_f32_len, mul_mat_vec_id_iq3_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS], "mul_mat_vec_id_iq4_xs_f32", mul_mat_vec_id_iq4_xs_f32_len, mul_mat_vec_id_iq4_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); // dequant shaders ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); @@ -1610,45 +2590,131 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_k", dequant_q6_k_len, dequant_q6_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL], "dequant_iq4_nl", dequant_iq4_nl_len, dequant_iq4_nl_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ1_S], "dequant_iq1_s", dequant_iq1_s_len, dequant_iq1_s_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ1_M], "dequant_iq1_m", dequant_iq1_m_len, dequant_iq1_m_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XXS], "dequant_iq2_xxs", dequant_iq2_xxs_len, dequant_iq2_xxs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XS], "dequant_iq2_xs", dequant_iq2_xs_len, dequant_iq2_xs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_S], "dequant_iq2_s", dequant_iq2_s_len, dequant_iq2_s_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ3_XXS], "dequant_iq3_xxs", dequant_iq3_xxs_len, dequant_iq3_xxs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ3_S], "dequant_iq3_s", dequant_iq3_s_len, dequant_iq3_s_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_XS], "dequant_iq4_xs", dequant_iq4_xs_len, dequant_iq4_xs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL], "dequant_iq4_nl", dequant_iq4_nl_len, dequant_iq4_nl_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); // get_rows ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32", get_rows_f32_len, get_rows_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl", get_rows_iq4_nl_len, get_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ1_S], "get_rows_iq1_s", get_rows_iq1_s_len, get_rows_iq1_s_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ1_M], "get_rows_iq1_m", get_rows_iq1_m_len, get_rows_iq1_m_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs", get_rows_iq2_xxs_len, get_rows_iq2_xxs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XS], "get_rows_iq2_xs", get_rows_iq2_xs_len, get_rows_iq2_xs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_S], "get_rows_iq2_s", get_rows_iq2_s_len, get_rows_iq2_s_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ3_XXS], "get_rows_iq3_xxs", get_rows_iq3_xxs_len, get_rows_iq3_xxs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ3_S], "get_rows_iq3_s", get_rows_iq3_s_len, get_rows_iq3_s_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_XS], "get_rows_iq4_xs", get_rows_iq4_xs_len, get_rows_iq4_xs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl", get_rows_iq4_nl_len, get_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32", get_rows_f32_f32_len, get_rows_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); - - ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1); - - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ1_S], "get_rows_iq1_s_f32", get_rows_iq1_s_f32_len, get_rows_iq1_s_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ1_M], "get_rows_iq1_m_f32", get_rows_iq1_m_f32_len, get_rows_iq1_m_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs_f32", get_rows_iq2_xxs_f32_len, get_rows_iq2_xxs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XS], "get_rows_iq2_xs_f32", get_rows_iq2_xs_f32_len, get_rows_iq2_xs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_S], "get_rows_iq2_s_f32", get_rows_iq2_s_f32_len, get_rows_iq2_s_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ3_XXS], "get_rows_iq3_xxs_f32", get_rows_iq3_xxs_f32_len, get_rows_iq3_xxs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ3_S], "get_rows_iq3_s_f32", get_rows_iq3_s_f32_len, get_rows_iq3_s_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_XS], "get_rows_iq4_xs_f32", get_rows_iq4_xs_f32_len, get_rows_iq4_xs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 2, 3 * sizeof(uint32_t), {1, 1, 1}, {}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1); + + for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) { + if (device->subgroup_add && device->subgroup_require_full_support) { + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); + } else { + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); + } + } + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f32, "cpy_f16_f32", cpy_f16_f32_len, cpy_f16_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f32, "contig_cpy_f16_f32", contig_cpy_f16_f32_len, contig_cpy_f16_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + + if (device->float_controls_rte_fp16) { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } else { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } + + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_0], "cpy_q5_0_f32", cpy_q5_0_f32_len, cpy_q5_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_1], "cpy_q5_1_f32", cpy_q5_1_f32_len, cpy_q5_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + + auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { + std::string s; + s += std::string(src0_f16 ? "_f16" : "_f32"); + s += std::string(src1_f16 ? "_f16" : "_f32"); + s += std::string(dst_f16 ? "_f16" : "_f32"); + return s; + }; + +#define CREATE_BINARY(name, namemod, spec) \ + for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \ + #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d], name ## _data[s0][s1][d], \ + "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); + CREATE_BINARY(add, , {0}) + CREATE_BINARY(add, _norepeat, {1}) + CREATE_BINARY(sub, , {0}) + CREATE_BINARY(sub, _norepeat, {1}) + CREATE_BINARY(mul, , {0}) + CREATE_BINARY(mul, _norepeat, {1}) + CREATE_BINARY(div, , {0}) + CREATE_BINARY(div, _norepeat, {1}) +#undef CREATE_BINARY - ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); @@ -1664,34 +2730,82 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_repeat_back_f32, "repeat_back_f32", repeat_back_f32_len, repeat_back_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + +#define CREATE_UNARY(name) \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + + CREATE_UNARY(gelu) + CREATE_UNARY(gelu_quick) + CREATE_UNARY(silu) + CREATE_UNARY(relu) + CREATE_UNARY(tanh) + CREATE_UNARY(sigmoid) +#undef CREATE_UNARY + ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + + if (device->float_controls_rte_fp16) { + ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + } else { + ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + } ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1); + + ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); + if (device->float_controls_rte_fp16) { + ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); + } else { + ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); + } ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_conv_transpose_1d_f32, "conv_transpose_1d_f32", conv_transpose_1d_f32_len, conv_transpose_1d_f32_data, "main", 3, sizeof(vk_op_conv_transpose_1d_push_constants), {1, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + + for (auto &c : compiles) { + c.wait(); + } + device->need_compiles = false; } +static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch); + static vk_device ggml_vk_get_device(size_t idx) { VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")"); @@ -1703,6 +2817,9 @@ static vk_device ggml_vk_get_device(size_t idx) { #ifdef GGML_VULKAN_MEMORY_DEBUG device->memory_logger = std::unique_ptr<vk_memory_logger>(new vk_memory_logger()); #endif + if (vk_perf_logger_enabled) { + device->perf_logger = std::unique_ptr<vk_perf_logger>(new vk_perf_logger()); + } size_t dev_num = vk_instance.device_indices[idx]; @@ -1716,12 +2833,60 @@ static vk_device ggml_vk_get_device(size_t idx) { device->physical_device = physical_devices[dev_num]; const std::vector<vk::ExtensionProperties> ext_props = device->physical_device.enumerateDeviceExtensionProperties(); + device->architecture = get_device_architecture(device->physical_device); + + const char* GGML_VK_PREFER_HOST_MEMORY = getenv("GGML_VK_PREFER_HOST_MEMORY"); + device->prefer_host_memory = GGML_VK_PREFER_HOST_MEMORY != nullptr; + + bool fp16_storage = false; + bool fp16_compute = false; bool maintenance4_support = false; + bool sm_builtins = false; + bool amd_shader_core_properties2 = false; + bool pipeline_robustness = false; + bool coopmat2_support = false; + device->coopmat_support = false; + device->integer_dot_product = false; + bool bfloat16_support = false; - // Check if maintenance4 is supported for (const auto& properties : ext_props) { if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { maintenance4_support = true; + } else if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { + fp16_storage = true; + } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { + fp16_compute = true; + } else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) { + sm_builtins = true; + } else if (strcmp("VK_AMD_shader_core_properties2", properties.extensionName) == 0) { + amd_shader_core_properties2 = true; + } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) { + pipeline_robustness = true; + } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) { + device->subgroup_size_control = true; +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT")) { + device->coopmat_support = true; + device->coopmat_m = 0; + device->coopmat_n = 0; + device->coopmat_k = 0; +#endif +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT2")) { + coopmat2_support = true; +#endif +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { + device->integer_dot_product = true; +#endif +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_BFLOAT16")) { + bfloat16_support = true; +#endif } } @@ -1729,44 +2894,104 @@ static vk_device ggml_vk_get_device(size_t idx) { vk::PhysicalDeviceMaintenance3Properties props3; vk::PhysicalDeviceMaintenance4Properties props4; vk::PhysicalDeviceSubgroupProperties subgroup_props; + vk::PhysicalDeviceDriverProperties driver_props; + vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props; + vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props; + vk::PhysicalDeviceVulkan11Properties vk11_props; + vk::PhysicalDeviceVulkan12Properties vk12_props; + vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR shader_integer_dot_product_props; + props2.pNext = &props3; props3.pNext = &subgroup_props; + subgroup_props.pNext = &driver_props; + driver_props.pNext = &vk11_props; + vk11_props.pNext = &vk12_props; + + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props; + if (maintenance4_support) { - subgroup_props.pNext = &props4; + last_struct->pNext = (VkBaseOutStructure *)&props4; + last_struct = (VkBaseOutStructure *)&props4; } + if (sm_builtins) { + last_struct->pNext = (VkBaseOutStructure *)&sm_props; + last_struct = (VkBaseOutStructure *)&sm_props; + } + if (amd_shader_core_properties2) { + last_struct->pNext = (VkBaseOutStructure *)&amd_shader_core_properties2_props; + last_struct = (VkBaseOutStructure *)&amd_shader_core_properties2_props; + } + if (device->subgroup_size_control) { + last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_props; + last_struct = (VkBaseOutStructure *)&subgroup_size_control_props; + } + +#if defined(VK_NV_cooperative_matrix2) + vk::PhysicalDeviceCooperativeMatrix2PropertiesNV coopmat2_props; + if (coopmat2_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat2_props; + last_struct = (VkBaseOutStructure *)&coopmat2_props; + } +#endif + + if (device->integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_props; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_props; + } + device->physical_device.getProperties2(&props2); device->properties = props2.properties; + device->vendor_id = device->properties.vendorID; + device->driver_id = driver_props.driverID; const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE"); if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) { - device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE); + device->max_memory_allocation_size = std::stoul(GGML_VK_FORCE_MAX_ALLOCATION_SIZE); } else if (maintenance4_support) { device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize); } else { device->max_memory_allocation_size = props3.maxMemoryAllocationSize; } - device->vendor_id = device->properties.vendorID; + const char* GGML_VK_SUBALLOCATION_BLOCK_SIZE = getenv("GGML_VK_SUBALLOCATION_BLOCK_SIZE"); + + if (GGML_VK_SUBALLOCATION_BLOCK_SIZE != nullptr) { + device->suballocation_block_size = std::stoul(GGML_VK_SUBALLOCATION_BLOCK_SIZE); + } else { + // Limit batching of allocations to 1GB by default to avoid fragmentation issues + device->suballocation_block_size = 1024*1024*1024; + } + device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size); + device->subgroup_size = subgroup_props.subgroupSize; device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; + if (sm_builtins) { + device->shader_core_count = sm_props.shaderSMCount; + } else if (amd_shader_core_properties2) { + device->shader_core_count = amd_shader_core_properties2_props.activeComputeUnitCount; + } else { + device->shader_core_count = 0; + } + device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16; - bool fp16_storage = false; - bool fp16_compute = false; + device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && + (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic); - for (const auto& properties : ext_props) { - if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { - fp16_storage = true; - } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { - fp16_compute = true; - } - } + device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && + (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle); - const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); - const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; + const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr; device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; + if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props, device->architecture)) { + device->coopmat_support = false; + } + + device->integer_dot_product = device->integer_dot_product && shader_integer_dot_product_props.integerDotProduct4x8BitPackedSignedAccelerated; + std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties(); // Try to find a non-graphics compute queue and transfer-focused queues @@ -1804,10 +3029,185 @@ static vk_device ggml_vk_get_device(size_t idx) { vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; vk11_features.pNext = &vk12_features; + last_struct = (VkBaseOutStructure *)&vk12_features; + + VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features; + pl_robustness_features.pNext = nullptr; + pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT; + pl_robustness_features.pipelineRobustness = VK_FALSE; + + if (pipeline_robustness) { + last_struct->pNext = (VkBaseOutStructure *)&pl_robustness_features; + last_struct = (VkBaseOutStructure *)&pl_robustness_features; + device_extensions.push_back("VK_EXT_pipeline_robustness"); + } + + VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_size_control_features; + subgroup_size_control_features.pNext = nullptr; + subgroup_size_control_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT; + subgroup_size_control_features.computeFullSubgroups = false; + subgroup_size_control_features.subgroupSizeControl = false; + + if (device->subgroup_size_control) { + last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_features; + last_struct = (VkBaseOutStructure *)&subgroup_size_control_features; + } + +#if defined(VK_KHR_cooperative_matrix) + VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; + coopmat_features.pNext = nullptr; + coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; + coopmat_features.cooperativeMatrix = VK_FALSE; + + if (device->coopmat_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; + last_struct = (VkBaseOutStructure *)&coopmat_features; + } +#endif + +#if defined(VK_NV_cooperative_matrix2) + VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {}; + coopmat2_features.pNext = nullptr; + coopmat2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_2_FEATURES_NV; + if (coopmat2_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat2_features; + last_struct = (VkBaseOutStructure *)&coopmat2_features; + device_extensions.push_back("VK_NV_cooperative_matrix2"); + } +#endif + +#if defined(VK_KHR_shader_bfloat16) + VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {}; + bfloat16_features.pNext = nullptr; + bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR; + if (bfloat16_support) { + last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features; + last_struct = (VkBaseOutStructure *)&bfloat16_features; + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif + + VkPhysicalDeviceMaintenance4Features maint4_features {}; + maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES; + if (maintenance4_support) { + last_struct->pNext = (VkBaseOutStructure *)&maint4_features; + last_struct = (VkBaseOutStructure *)&maint4_features; + device_extensions.push_back("VK_KHR_maintenance4"); + } + + VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR shader_integer_dot_product_features {}; + shader_integer_dot_product_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_FEATURES_KHR; + if (device->integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_features; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_features; + device_extensions.push_back("VK_KHR_shader_integer_dot_product"); + } + vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2); device->fp16 = device->fp16 && vk12_features.shaderFloat16; + device->pipeline_robustness = pl_robustness_features.pipelineRobustness; + + if (device->subgroup_size_control) { + device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize; + device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize; + device_extensions.push_back("VK_EXT_subgroup_size_control"); + } + + device->subgroup_size_control = device->subgroup_size_control && + (subgroup_size_control_props.requiredSubgroupSizeStages & vk::ShaderStageFlagBits::eCompute) && + subgroup_size_control_features.subgroupSizeControl; + + if (device->subgroup_size_control) { + device->subgroup_require_full_support = subgroup_size_control_features.computeFullSubgroups; + } + +#if defined(VK_KHR_cooperative_matrix) + device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix; + + // coopmat1 fa shader currently assumes 32 invocations per subgroup + device->coopmat1_fa_support = device->coopmat_support && device->subgroup_require_full_support && + device->subgroup_size_control && device->subgroup_min_size <= 32 && + device->subgroup_max_size >= 32; +#endif + + if (coopmat2_support) { +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (coopmat2_features.cooperativeMatrixWorkgroupScope && + coopmat2_features.cooperativeMatrixFlexibleDimensions && + coopmat2_features.cooperativeMatrixReductions && + coopmat2_features.cooperativeMatrixConversions && + coopmat2_features.cooperativeMatrixPerElementOperations && + coopmat2_features.cooperativeMatrixTensorAddressing && + coopmat2_features.cooperativeMatrixBlockLoads && + vk12_features.bufferDeviceAddress) { + + std::vector<VkCooperativeMatrixFlexibleDimensionsPropertiesNV> flexible_dimensions; + uint32_t count = 0; + + PFN_vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV + _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV = + (PFN_vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV) + vk_instance.instance.getProcAddr("vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV"); + + _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV(device->physical_device, &count, nullptr); + + VkCooperativeMatrixFlexibleDimensionsPropertiesNV empty_prop {}; + empty_prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_FLEXIBLE_DIMENSIONS_PROPERTIES_NV; + flexible_dimensions.resize(count, empty_prop); + + _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV(device->physical_device, &count, flexible_dimensions.data()); + + bool found_fp16_128 = false, + found_fp16_256 = false, + found_fp32_128 = false, + found_fp32_256 = false; + // need to support fp16*fp16 with fp16/fp32 accumulator, for workgroupsize 128 + // with 32x16x16 and 256 with 32x32x16. + for (auto &prop : flexible_dimensions) { + if (prop.saturatingAccumulation == VK_FALSE && + prop.scope == VK_SCOPE_WORKGROUP_KHR && + prop.AType == VK_COMPONENT_TYPE_FLOAT16_KHR && + prop.BType == VK_COMPONENT_TYPE_FLOAT16_KHR) { + + if (prop.workgroupInvocations == 128 && + prop.MGranularity <= 32 && + prop.NGranularity <= 16 && + prop.KGranularity <= 16) { + if (prop.CType == VK_COMPONENT_TYPE_FLOAT16_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT16_KHR) { + found_fp16_128 = true; + } + if (prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR) { + found_fp32_128 = true; + } + } + if (prop.workgroupInvocations == 256 && + prop.MGranularity <= 32 && + prop.NGranularity <= 32 && + prop.KGranularity <= 16) { + if (prop.CType == VK_COMPONENT_TYPE_FLOAT16_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT16_KHR) { + found_fp16_256 = true; + } + if (prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR) { + found_fp32_256 = true; + } + } + } + } + if (found_fp16_128 && found_fp16_256 && + found_fp32_128 && found_fp32_256 && + coopmat2_props.cooperativeMatrixFlexibleDimensionsMaxDimension >= 512) { + device->coopmat2 = true; + } + } +#endif + } + if (!vk11_features.storageBuffer16BitAccess) { std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl; throw std::runtime_error("Unsupported device"); @@ -1822,7 +3222,120 @@ static vk_device ggml_vk_get_device(size_t idx) { if (device->fp16) { device_extensions.push_back("VK_KHR_shader_float16_int8"); } - device->name = device->properties.deviceName.data(); + +#if defined(VK_KHR_cooperative_matrix) + if (device->coopmat_support) { + // Query supported shapes + std::vector<VkCooperativeMatrixPropertiesKHR> cm_props; + + PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR = + (PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR)vkGetInstanceProcAddr(vk_instance.instance, "vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR"); + + uint32_t cm_props_num; + + pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, nullptr); + + cm_props.resize(cm_props_num); + + for (auto& prop : cm_props) { + prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR; + } + + pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, cm_props.data()); + + VK_LOG_DEBUG("ggml_vulkan: Cooperative Matrix Shapes: " << cm_props.size()); + + for (auto& prop : cm_props) { + VK_LOG_DEBUG("ggml_vulkan: M: " << prop.MSize << " N: " << prop.NSize << " K: " << prop.KSize << " A: " << vk::to_string((vk::ComponentTypeKHR)prop.AType) << " B: " << vk::to_string((vk::ComponentTypeKHR)prop.BType) << " C: " << vk::to_string((vk::ComponentTypeKHR)prop.CType) << " Result: " << vk::to_string((vk::ComponentTypeKHR)prop.ResultType) << " saturatingAccumulation: " << prop.saturatingAccumulation << " scope: " << vk::to_string((vk::ScopeKHR)prop.scope)); + + if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eFloat16 && + (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eFloat16 && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup + ) { + if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat32 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat32) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_acc_f32_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_acc_f32_support = true; + } + if (prop.MSize == 16 && prop.NSize == 16 && prop.KSize == 16) { + device->coopmat_support_16x16x16_f32acc = true; + } + } else if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat16 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat16) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_acc_f16_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_acc_f16_support = true; + } + if (prop.MSize == 16 && prop.NSize == 16 && prop.KSize == 16) { + device->coopmat_support_16x16x16_f16acc = true; + } + } + } else if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eSint8 && + (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eSint8 && + (vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eSint32 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eSint32 && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup && + device->coopmat_int_m == 0 + ) { + device->coopmat_int_support = true; + device->coopmat_int_m = prop.MSize; + device->coopmat_int_n = prop.NSize; + device->coopmat_int_k = prop.KSize; + } +#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup + ) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_bf16_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_bf16_support = true; + } + } +#endif + } + + if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) { + // No suitable matmul mode found + GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n"); + device->coopmat_support = false; + } + if (getenv("GGML_VK_DISABLE_BFLOAT16")) { + device->coopmat_bf16_support = false; + } + } + + if (device->coopmat_support) { + device_extensions.push_back("VK_KHR_cooperative_matrix"); + } +#if defined(VK_KHR_shader_bfloat16) + if (device->coopmat_bf16_support) { + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif +#endif + device->name = GGML_VK_NAME + std::to_string(idx); device_create_info = { vk::DeviceCreateFlags(), @@ -1833,20 +3346,68 @@ static vk_device ggml_vk_get_device(size_t idx) { device_create_info.setPNext(&device_features2); device->device = device->physical_device.createDevice(device_create_info); - device->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN; - // Queues - ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }); + ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false); // Shaders + // Disable matmul tile sizes early if performance low or not supported + for (uint32_t i = 0; i < GGML_TYPE_COUNT; ++i) { + switch (device->vendor_id) { +#ifndef GGML_VULKAN_RUN_TESTS + case VK_VENDOR_ID_AMD: + case VK_VENDOR_ID_INTEL: + device->mul_mat_l[i] = false; + device->mul_mat_m[i] = true; + device->mul_mat_s[i] = true; + device->mul_mat_id_l[i] = false; + device->mul_mat_id_m[i] = true; + device->mul_mat_id_s[i] = true; + break; + case VK_VENDOR_ID_APPLE: + device->mul_mat_l[i] = false; + device->mul_mat_m[i] = true; + device->mul_mat_s[i] = false; + device->mul_mat_id_l[i] = false; + device->mul_mat_id_m[i] = true; + device->mul_mat_id_s[i] = false; + break; +#endif + default: + device->mul_mat_l[i] = true; + device->mul_mat_m[i] = true; + device->mul_mat_s[i] = true; + device->mul_mat_id_l[i] = true; + device->mul_mat_id_m[i] = true; + device->mul_mat_id_s[i] = true; + break; + } + } + + + std::vector<vk::DescriptorSetLayoutBinding> dsl_binding; + std::vector<vk::DescriptorBindingFlags> dsl_binding_flags; + for (uint32_t i = 0; i < MAX_PARAMETER_COUNT; i++) { + dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute}); + dsl_binding_flags.push_back({}); + } + + vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags }; + + vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info( + {}, + dsl_binding); + descriptor_set_layout_create_info.setPNext(&dslbfci); + device->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info); + ggml_vk_load_shaders(device); if (!device->single_queue) { const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0; - ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }); + ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true); } else { // TODO: Use pointer or reference to avoid copy - device->transfer_queue = device->compute_queue; + device->transfer_queue.copyFrom(device->compute_queue); + device->transfer_queue.cmd_pool.init(device, &device->transfer_queue); } device->buffer_type = { @@ -1864,7 +3425,6 @@ static vk_device ggml_vk_get_device(size_t idx) { return vk_instance.devices[idx]; } - static void ggml_vk_print_gpu_info(size_t idx) { GGML_ASSERT(idx < vk_instance.device_indices.size()); size_t dev_num = vk_instance.device_indices[idx]; @@ -1881,40 +3441,64 @@ static void ggml_vk_print_gpu_info(size_t idx) { vk::PhysicalDevice physical_device = devices[dev_num]; std::vector<vk::ExtensionProperties> ext_props = physical_device.enumerateDeviceExtensionProperties(); - vk::PhysicalDeviceProperties2 props2; - vk::PhysicalDeviceMaintenance3Properties props3; - vk::PhysicalDeviceSubgroupProperties subgroup_props; - vk::PhysicalDeviceDriverProperties driver_props; - props2.pNext = &props3; - props3.pNext = &subgroup_props; - subgroup_props.pNext = &driver_props; - physical_device.getProperties2(&props2); - - const size_t subgroup_size = subgroup_props.subgroupSize; - const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; - bool fp16_storage = false; bool fp16_compute = false; + bool coopmat_support = false; + bool coopmat2_support = false; + bool integer_dot_product = false; for (auto properties : ext_props) { if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { fp16_storage = true; } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { fp16_compute = true; +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT")) { + coopmat_support = true; +#endif +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT2")) { + coopmat2_support = true; +#endif +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { + integer_dot_product = true; +#endif } } + const vk_device_architecture device_architecture = get_device_architecture(physical_device); + const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; bool fp16 = !force_disable_f16 && fp16_storage && fp16_compute; - vk::PhysicalDeviceFeatures device_features = physical_device.getFeatures(); + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceMaintenance3Properties props3; + vk::PhysicalDeviceSubgroupProperties subgroup_props; + vk::PhysicalDeviceDriverProperties driver_props; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR shader_integer_dot_product_props; + props2.pNext = &props3; + props3.pNext = &subgroup_props; + subgroup_props.pNext = &driver_props; + + // Pointer to the last chain element + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&driver_props; + + if (integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_props; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_props; + } + + physical_device.getProperties2(&props2); VkPhysicalDeviceFeatures2 device_features2; device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2; device_features2.pNext = nullptr; - device_features2.features = (VkPhysicalDeviceFeatures)device_features; VkPhysicalDeviceVulkan11Features vk11_features; vk11_features.pNext = nullptr; @@ -1926,22 +3510,64 @@ static void ggml_vk_print_gpu_info(size_t idx) { vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; vk11_features.pNext = &vk12_features; + // Pointer to the last chain element + last_struct = (VkBaseOutStructure *)&vk12_features; + +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; + coopmat_features.pNext = nullptr; + coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; + coopmat_features.cooperativeMatrix = VK_FALSE; + + if (coopmat_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; + last_struct = (VkBaseOutStructure *)&coopmat_features; + } +#endif + + VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR shader_integer_dot_product_features {}; + shader_integer_dot_product_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_FEATURES_KHR; + if (integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_features; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_features; + } + vkGetPhysicalDeviceFeatures2(physical_device, &device_features2); fp16 = fp16 && vk12_features.shaderFloat16; + uint32_t default_subgroup_size = get_subgroup_size("", device_architecture); + const size_t subgroup_size = (default_subgroup_size != 0) ? default_subgroup_size : subgroup_props.subgroupSize; + const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; + + integer_dot_product = integer_dot_product + && shader_integer_dot_product_props.integerDotProduct4x8BitPackedSignedAccelerated + && shader_integer_dot_product_features.shaderIntegerDotProduct; + + coopmat_support = coopmat_support +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + && coopmat_features.cooperativeMatrix +#endif + && ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture); + + std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none"; + std::string device_name = props2.properties.deviceName.data(); - std::cerr << GGML_VK_NAME << idx << ": " << device_name << " (" << driver_props.driverName << ") | uma: " << uma << " | fp16: " << fp16 << " | warp size: " << subgroup_size << std::endl; + GGML_LOG_INFO("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n", + idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, + props2.properties.limits.maxComputeSharedMemorySize, integer_dot_product, matrix_cores.c_str()); if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) { - std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl; + GGML_LOG_INFO("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n"); } } static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions); static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions); -void ggml_vk_instance_init() { +static bool ggml_vk_instance_debug_utils_ext_available(const std::vector<vk::ExtensionProperties> & instance_extensions); + +GGML_CALL void ggml_vk_instance_init() { if (vk_instance_initialized) { return; } @@ -1949,14 +3575,21 @@ void ggml_vk_instance_init() { vk_instance_initialized = true; - vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION }; + uint32_t api_version = vk::enumerateInstanceVersion(); + + if (api_version < VK_API_VERSION_1_2) { + std::cerr << "ggml_vulkan: Error: Vulkan 1.2 required." << std::endl; + GGML_ABORT("fatal error"); + } + + vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, api_version }; const std::vector<vk::ExtensionProperties> instance_extensions = vk::enumerateInstanceExtensionProperties(); const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions); #ifdef __APPLE__ const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions); #endif - + const bool debug_utils_ext = ggml_vk_instance_debug_utils_ext_available(instance_extensions) && getenv("GGML_VK_DEBUG_MARKERS") != nullptr; std::vector<const char*> layers; if (validation_ext) { @@ -1971,6 +3604,9 @@ void ggml_vk_instance_init() { extensions.push_back("VK_KHR_portability_enumeration"); } #endif + if (debug_utils_ext) { + extensions.push_back("VK_EXT_debug_utils"); + } vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions); #ifdef __APPLE__ if (portability_enumeration_ext) { @@ -1989,16 +3625,29 @@ void ggml_vk_instance_init() { }; validation_features.setPNext(nullptr); instance_create_info.setPNext(&validation_features); - - std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl; + GGML_LOG_DEBUG("ggml_vulkan: Validation layers enabled\n"); } vk_instance.instance = vk::createInstance(instance_create_info); + if (debug_utils_ext) { + vk_instance.debug_utils_support = true; + vk_instance.pfn_vkSetDebugUtilsObjectNameEXT = (PFN_vkSetDebugUtilsObjectNameEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkSetDebugUtilsObjectNameEXT"); + vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT = (PFN_vkQueueBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueBeginDebugUtilsLabelEXT"); + vk_instance.pfn_vkQueueEndDebugUtilsLabelEXT = (PFN_vkQueueEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueEndDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT = (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT"); + vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT"); + + } + size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size(); + vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr; // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan char * devices_env = getenv("GGML_VK_VISIBLE_DEVICES"); if (devices_env != nullptr) { + size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size(); + std::string devices(devices_env); std::replace(devices.begin(), devices.end(), ',', ' '); @@ -2014,9 +3663,9 @@ void ggml_vk_instance_init() { } else { std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices(); - // Make sure at least one device exists + // If no vulkan devices are found, return early if (devices.empty()) { - std::cerr << "ggml_vulkan: Error: No devices found." << std::endl; + GGML_LOG_INFO("ggml_vulkan: No devices found.\n"); GGML_ABORT("fatal error"); } @@ -2099,13 +3748,23 @@ void ggml_vk_instance_init() { } } - // If no dedicated GPUs found, fall back to GPU 0 + // If no dedicated GPUs found, fall back to the first non-CPU device. + // If only CPU devices are available, return without devices. if (vk_instance.device_indices.empty()) { - vk_instance.device_indices.push_back(0); + for (size_t i = 0; i < devices.size(); i++) { + if (devices[i].getProperties().deviceType != vk::PhysicalDeviceType::eCpu) { + vk_instance.device_indices.push_back(i); + break; + } + } } - } - std::cerr << "ggml_vulkan: Found " << vk_instance.device_indices.size() << " Vulkan devices:" << std::endl; + if (vk_instance.device_indices.empty()) { + GGML_LOG_INFO("ggml_vulkan: No devices found.\n"); + return; + } + } + GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size()); for (size_t i = 0; i < vk_instance.device_indices.size(); i++) { ggml_vk_print_gpu_info(i); @@ -2129,9 +3788,10 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { ctx->prealloc_size_split_k = 0; ctx->fence = ctx->device->device.createFence({}); + ctx->almost_ready_fence = ctx->device->device.createFence({}); - ctx->staging_size = 0; - ctx->staging_offset = 0; + ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue); + ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue); #ifdef GGML_VULKAN_CHECK_RESULTS const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS"); @@ -2155,6 +3815,14 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: @@ -2164,22 +3832,45 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type return ctx->device->pipeline_dequant[type]; } -static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) { - VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); +static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) { + VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ", " << prec << ")"); if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_f32; } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { return ctx->device->pipeline_matmul_f32_f16; } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { - return ctx->device->pipeline_matmul_f16_f32; + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_bf16; } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { - return ctx->device->pipeline_matmul_f16; + if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_f16_f32.f16acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_f16.f16acc; + } + } else { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_f16_f32.f32acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_f16.f32acc; + } + } + + // MMQ + if (src1_type == GGML_TYPE_Q8_1) { + vk_matmul_pipeline pipelines = (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc; + + if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) { + return nullptr; + } + + return pipelines; } - if (src1_type != GGML_TYPE_F32) { + if (src1_type != GGML_TYPE_F32 && !ctx->device->coopmat2) { return nullptr; } @@ -2194,22 +3885,39 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: return nullptr; } - return ctx->device->pipeline_dequant_mul_mat_mat[src0_type]; + if (ctx->device->coopmat2) { + assert(src1_type == GGML_TYPE_F16); + return prec == GGML_PREC_DEFAULT ? ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f32acc; + } + if (ctx->device->coopmat_support) { + return (ctx->device->fp16 && ctx->device->coopmat_acc_f16_support && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc; + } + return (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc; } -static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) { +static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) { VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()"); GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16); + GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols); switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -2220,28 +3928,48 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: return nullptr; } - return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type]; + return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1]; } -static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) { +static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) { VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_id_pipeline()"); if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f32; } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { - return ctx->device->pipeline_matmul_id_f16_f32; + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_id_bf16; } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { - return ctx->device->pipeline_matmul_id_f16; + if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_id_f16_f32.f16acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_id_f16.f16acc; + } + } else { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_id_f16_f32.f32acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_id_f16.f32acc; + } } - GGML_ASSERT(src1_type == GGML_TYPE_F32); + GGML_ASSERT(src1_type == GGML_TYPE_F32 || (ctx->device->coopmat2 && src1_type == GGML_TYPE_F16)); switch (src0_type) { case GGML_TYPE_Q4_0: @@ -2254,13 +3982,21 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: return nullptr; } - return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type]; + return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc; } static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) { @@ -2270,6 +4006,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -2280,6 +4017,14 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: @@ -2368,6 +4113,7 @@ static void * ggml_vk_host_malloc(vk_device& device, size_t size) { return nullptr; } + std::lock_guard<std::recursive_mutex> guard(device->mutex); device->pinned_memory.push_back(std::make_tuple(buf->ptr, size, buf)); return buf->ptr; @@ -2378,6 +4124,8 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) { return; } VK_LOG_MEMORY("ggml_vk_host_free(" << ptr << ")"); + std::lock_guard<std::recursive_mutex> guard(device->mutex); + vk_buffer buf; size_t index; for (size_t i = 0; i < device->pinned_memory.size(); i++) { @@ -2400,6 +4148,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) { } static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) { + std::lock_guard<std::recursive_mutex> guard(device->mutex); buf = nullptr; buf_offset = 0; for (size_t i = 0; i < device->pinned_memory.size(); i++) { @@ -2413,9 +4162,9 @@ static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf } } -static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bool one_time = true) { +static vk_submission ggml_vk_begin_submission(vk_device& device, vk_command_pool& p, bool one_time = true) { vk_submission s; - s.buffer = ggml_vk_create_cmd_buffer(device, q); + s.buffer = ggml_vk_create_cmd_buffer(device, p); if (one_time) { s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit }); } else { @@ -2425,21 +4174,19 @@ static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bo return s; } - - static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) { const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]); const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]); const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]); VK_LOG_DEBUG("ggml_vk_dispatch_pipeline(" << pipeline->name << ", {"; for (auto& buffer : descriptor_buffer_infos) { - std::cerr << "(" << buffer << ", " << buffer.offset << ", " << buffer.size << "), "; + std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.range << "), "; } std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))"); - GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size()); - GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count); + GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size()); + GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT); - vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++]; + vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++]; vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() }; ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {}); @@ -2476,7 +4223,7 @@ static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) { ggml_vk_ctx_end(subctx); } - subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->q) }); + subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->p) }); subctx->s = subctx->seqs[subctx->seqs.size() - 1].data(); } @@ -2512,8 +4259,8 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont GGML_ABORT("fatal error"); } // Check if src is pinned memory - vk_buffer buf; - size_t buf_offset; + vk_buffer buf = nullptr; + size_t buf_offset = 0; ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset); const uint64_t ne0 = tensor->ne[0]; @@ -2565,39 +4312,31 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont return; } - // Staging buffer required - vk_buffer staging = ctx->staging; - size_t staging_offset = ctx->staging_offset; - const size_t copy_size = ts*ne/bs; - if (ctx->staging->size < ctx->staging_offset + copy_size) { - if (sync_staging) { - // Create temporary larger buffer - ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size); - - staging = ctx->device->sync_staging; - staging_offset = 0; - } else { - GGML_ABORT("fatal error"); - } + if (!sync_staging) { + GGML_ABORT("Asynchronous write to non-pinned memory not supported"); } - VkBufferCopy buf_copy{ staging_offset, offset, copy_size }; + // Staging buffer required + vk_buffer& staging = ctx->device->sync_staging; + const uint64_t copy_size = ts*ne/bs; + ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size); + VkBufferCopy buf_copy{ 0, offset, copy_size }; ggml_vk_sync_buffers(subctx); - vkCmdCopyBuffer(subctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy); + vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy); for (uint64_t i3 = 0; i3 < ne3; i3++) { for (uint64_t i2 = 0; i2 < ne2; i2++) { // Find longest contiguous slice if (ne1*nb1 == dstnb2) { - deferred_memcpy((uint8_t *)staging->ptr + staging_offset + i3*dstnb3 + i2*dstnb2, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2, dstnb2, &subctx->in_memcpys); + deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2, dstnb2, &subctx->in_memcpys); } else { for (uint64_t i1 = 0; i1 < ne1; i1++) { if (ne0*nb0/bs == dstnb1) { - deferred_memcpy((uint8_t *)staging->ptr + staging_offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2 + i1*nb1, dstnb1, &subctx->in_memcpys); + deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2 + i1*nb1, dstnb1, &subctx->in_memcpys); } else { const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1; - const uint64_t d_off = staging_offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1; + const uint64_t d_off = i3*dstnb3 + i2*dstnb2 + i1*dstnb1; for (uint64_t i0 = 0; i0 < ne0; i0++) { deferred_memcpy((uint8_t *)staging->ptr + d_off + i0*dstnb0, (const uint8_t *) tensor->data + s_off + i0*nb0, dstnb0, &subctx->in_memcpys); } @@ -2608,7 +4347,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont } } -static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) { +static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, bool sync_staging = false) { VK_LOG_DEBUG("ggml_vk_buffer_write_2d_async(" << width << ", " << height << ")"); // Buffer is already mapped if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) { @@ -2617,7 +4356,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz } // Check if src is pinned memory vk_buffer buf = nullptr; - size_t buf_offset; + size_t buf_offset = 0; ggml_vk_host_get(dst->device, src, buf, buf_offset); if (buf != nullptr) { @@ -2643,39 +4382,36 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz } VK_LOG_DEBUG("STAGING"); + if (!sync_staging) { + GGML_ABORT("Asynchronous write to non-pinned memory not supported"); + } + // Staging buffer required const size_t copy_size = width*height; - if (staging_buffer == nullptr || staging_buffer->size < staging_offset + copy_size) { - if (sync_staging) { - ggml_vk_ensure_sync_staging_buffer(dst->device, copy_size); + ggml_vk_ensure_sync_staging_buffer(dst->device, copy_size); - staging_buffer = dst->device->sync_staging; - staging_offset = 0; - } else { - GGML_ABORT("fatal error"); - } - } + vk_buffer& staging_buffer = dst->device->sync_staging; VkBufferCopy buf_copy = { - staging_offset, + 0, offset, copy_size}; ggml_vk_sync_buffers(subctx); - vkCmdCopyBuffer(subctx->s->buffer, staging_buffer->buffer, dst->buffer, 1, &buf_copy); + vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy); if (width == spitch) { - deferred_memcpy((uint8_t *)staging_buffer->ptr + staging_offset, src, width * height, &subctx->in_memcpys); + deferred_memcpy((uint8_t *)staging_buffer->ptr, src, width * height, &subctx->in_memcpys); } else { for (size_t i = 0; i < height; i++) { - deferred_memcpy((uint8_t *)staging_buffer->ptr + staging_offset + i * width, (const uint8_t *) src + i * spitch, width, &subctx->in_memcpys); + deferred_memcpy((uint8_t *)staging_buffer->ptr + i * width, (const uint8_t *) src + i * spitch, width, &subctx->in_memcpys); } } } -static void ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) { +static void ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, bool sync_staging = false) { VK_LOG_DEBUG("ggml_vk_buffer_write_async(" << size << ")"); - return ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, size, size, 1, staging_buffer, staging_offset, sync_staging); + return ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, size, size, 1, sync_staging); } static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height) { @@ -2688,9 +4424,11 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width); } } else { - vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue); + std::lock_guard<std::recursive_mutex> guard(dst->device->mutex); + + vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool); ggml_vk_ctx_begin(dst->device, subctx); - ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, nullptr, 0, true); + ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, true); ggml_vk_ctx_end(subctx); for (auto& cpy : subctx->in_memcpys) { @@ -2700,6 +4438,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * ggml_vk_submit(subctx, dst->device->fence); VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences"); dst->device->device.resetFences({ dst->device->fence }); + ggml_vk_queue_command_pools_cleanup(dst->device); } } @@ -2708,7 +4447,7 @@ static void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src ggml_vk_buffer_write_2d(dst, offset, src, 0, size, 1); } -static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) { +static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false) { VK_LOG_DEBUG("ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")"); GGML_ASSERT(width > 0); GGML_ASSERT(height > 0); @@ -2718,7 +4457,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size // Check if dst is pinned memory vk_buffer buf = nullptr; - size_t buf_offset; + size_t buf_offset = 0; ggml_vk_host_get(src->device, dst, buf, buf_offset); std::vector<vk::BufferCopy> slices(1); @@ -2745,18 +4484,15 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size } VK_LOG_DEBUG("STAGING"); + if (!sync_staging) { + GGML_ABORT("Asynchronous read from non-pinned memory not supported"); + } + // Fall back to staging buffer const size_t copy_size = dpitch * height; - if (staging_buffer == nullptr || staging_buffer->size < staging_offset + copy_size) { - if (sync_staging) { - // Create temporary larger buffer - ggml_vk_ensure_sync_staging_buffer(src->device, copy_size); + ggml_vk_ensure_sync_staging_buffer(src->device, copy_size); - staging_buffer = src->device->sync_staging; - } else { - GGML_ABORT("fatal error"); - } - } + vk_buffer& staging_buffer = src->device->sync_staging; ggml_vk_sync_buffers(subctx); subctx->s->buffer.copyBuffer(src->buffer, staging_buffer->buffer, slices); @@ -2764,25 +4500,32 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys); } -static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, vk_buffer staging_buffer, size_t staging_offset, bool sync_staging = false) { - return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, staging_buffer, staging_offset, sync_staging); +static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, bool sync_staging = false) { + return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, sync_staging); } static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) { VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")"); - if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) { + + // If the device is not an UMA device the memory is host-accessible through rebar. While writing + // through PCIe is sufficient fast reading back data from PCIe is slower than going through + // the HW device to host copy path. + if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) { GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent); memcpy(dst, (uint8_t *) src->ptr + offset, size); } else { - vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue); + std::lock_guard<std::recursive_mutex> guard(src->device->mutex); + + vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool); ggml_vk_ctx_begin(src->device, subctx); - ggml_vk_buffer_read_async(subctx, src, offset, dst, size, nullptr, 0, true); + ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true); ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, src->device->fence); VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences"); src->device->device.resetFences({ src->device->fence }); + ggml_vk_queue_command_pools_cleanup(src->device); for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -2797,20 +4540,22 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds VkBufferCopy bc{ src_offset, dst_offset, size }; - vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc); + vkCmdCopyBuffer(ctx->s->buffer, (VkBuffer)src->buffer, (VkBuffer)dst->buffer, 1, &bc); } static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) { if (src->device == dst->device) { + std::lock_guard<std::recursive_mutex> guard(src->device->mutex); VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")"); // Copy within the device - vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue); + vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool); ggml_vk_ctx_begin(src->device, subctx); ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size); ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, src->device->fence); VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences"); src->device->device.resetFences({ src->device->fence }); + ggml_vk_queue_command_pools_cleanup(src->device); } else { VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")"); // Copy device to device @@ -2826,10 +4571,17 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr } } +static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size) { + VK_LOG_DEBUG("ggml_vk_buffer_memset_async(" << offset << ", " << c << ", " << size << ")"); + + ctx->s->buffer.fillBuffer(dst->buffer, offset, size, c); +} + static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) { VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")"); - vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue); + std::lock_guard<std::recursive_mutex> guard(dst->device->mutex); + vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool); ggml_vk_ctx_begin(dst->device, subctx); subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c); ggml_vk_ctx_end(subctx); @@ -2837,65 +4589,67 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz ggml_vk_submit(subctx, dst->device->fence); VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences"); dst->device->device.resetFences({ dst->device->fence }); + ggml_vk_queue_command_pools_cleanup(dst->device); } -static uint32_t ggml_vk_guess_split_k(int m, int n, int k) { +static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) { VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")"); - // if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) { - // return 4; - // } - return 1; - - GGML_UNUSED(m); GGML_UNUSED(n); GGML_UNUSED(k); -} - -static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { - if (m <= 32 || n <= 32) { - return aligned ? mmp->a_s : mmp->s; + uint32_t split_k = 1; + if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) { + // If k is 'large' and the SMs will fill less than halfway, use split_k. + uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]); + uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]); + if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) { + split_k = ctx->device->shader_core_count / (m_tiles * n_tiles); + // Clamp to 2 or 4 + split_k = std::min(split_k, 4u); + if (split_k == 3) { + split_k = 2; + } + if (ctx->device->coopmat2) { + // coopmat2 shader expects splits to be aligned to 256 + while (split_k > 1 && ((k / split_k) % 256) != 0) { + split_k /= 2; + } + } + } } - return aligned ? mmp->a_m : mmp->m; - - GGML_UNUSED(ctx); -} - -static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { - return aligned ? mmp->a_m : mmp->m; - GGML_UNUSED(ctx); + return split_k; } -static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { - return aligned ? mmp->a_s : mmp->s; - - GGML_UNUSED(ctx); -} +static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type, ggml_type src1_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); -static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")"); - switch (ctx->device->vendor_id) { - case VK_VENDOR_ID_AMD: - return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned); - case VK_VENDOR_ID_APPLE: - return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned); - case VK_VENDOR_ID_INTEL: - return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned); - default: - break; + if (ctx->device->coopmat2) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) { + return aligned ? mmp->a_l : mmp->l; + } + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_s[src0_type]) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_s : mmp->s; } - if (m <= 32 || n <= 32) { + if ((ctx->device->mul_mat_s[src0_type] && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_l[src0_type])) { return aligned ? mmp->a_s : mmp->s; } - if (m <= 64 || n <= 64) { + if ((ctx->device->mul_mat_m[src0_type] && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_l[src0_type]) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_l : mmp->l; + + GGML_UNUSED(src1_type); } -static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")"); - return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true)->align; +static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type src0_type, ggml_type src1_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); + return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true, src0_type, src1_type)->align; } static void ggml_vk_matmul( @@ -2903,18 +4657,19 @@ static void ggml_vk_matmul( vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) { - VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")"); + uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, + uint32_t padded_n) { + VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")"); ggml_vk_sync_buffers(subctx); if (split_k == 1) { - const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch }); return; } GGML_ASSERT(batch_stride_d == m * n); - const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n }; // Make sure enough workgroups get assigned for split k to work ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch }); ggml_vk_sync_buffers(subctx); @@ -2922,19 +4677,51 @@ static void ggml_vk_matmul( ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 }); } +static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_id_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); + + if (ctx->device->coopmat2) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_id_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) { + return aligned ? mmp->a_l : mmp->l; + } + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_id_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_id_s[src0_type]) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_s : mmp->s; + } + + if ((ctx->device->mul_mat_id_s[src0_type] && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_l[src0_type])) { + return aligned ? mmp->a_s : mmp->s; + } + if ((ctx->device->mul_mat_id_m[src0_type] && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_id_l[src0_type]) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_l : mmp->l; +} + +static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type src0_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ", " << ggml_type_name(src0_type) << ")"); + return ggml_vk_guess_matmul_id_pipeline(ctx, mmp, m, n, true, src0_type)->align; +} + static void ggml_vk_matmul_id( ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline, vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) { + uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11, + uint32_t padded_n) { VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " << "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " << "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " << "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")"); ggml_vk_sync_buffers(subctx); const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, - nei0, nei1, nbi1, ne11 }; + nei0, nei1, nbi1, ne11, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as }); } @@ -2945,18 +4732,96 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } -static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) { - if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) { - return ctx->device->pipeline_cpy_f32_f32; +static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to) { + + // Choose "contiguous copy" shader if src/dst are contiguous + bool contig = ggml_is_contiguous(src) && (!dst || ggml_is_contiguous(dst)); + + if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_f32; + } else { + return ctx->device->pipeline_cpy_f32_f32; + } + } + if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_f16; + } else { + return ctx->device->pipeline_cpy_f32_f16; + } + } + if (src->type == GGML_TYPE_F16 && to == GGML_TYPE_F16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f16_f16; + } else { + return ctx->device->pipeline_cpy_f16_f16; + } } - if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) { - return ctx->device->pipeline_cpy_f32_f16; + if (src->type == GGML_TYPE_F16 && to == GGML_TYPE_F32) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f16_f32; + } else { + return ctx->device->pipeline_cpy_f16_f32; + } } - if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) { - return ctx->device->pipeline_cpy_f16_f16; + if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_bf16; + } else { + return ctx->device->pipeline_cpy_f32_bf16; + } + } + if (src->type == GGML_TYPE_F32) { + switch (to) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ4_NL: + return ctx->device->pipeline_cpy_f32_quant[to]; + default: + break; + } + } + + if (to == GGML_TYPE_F32) { + switch (src->type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ4_NL: + return ctx->device->pipeline_cpy_quant_f32[src->type]; + default: + break; + } } - std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl; + if (src->type == to) { + // Copy two or four bytes at a time, depending on block size. + // For quantized types, we scale by block size/type size. But + // this path is also used for bf16->bf16 for example, where the + // type size must be exactly 2 or 4. + GGML_ASSERT(ggml_is_quantized(to) || ggml_type_size(src->type) == 2 || ggml_type_size(src->type) == 4); + if ((ggml_type_size(src->type) % 4) == 0) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_f32; + } else { + return ctx->device->pipeline_cpy_f32_f32; + } + } else { + if (contig) { + return ctx->device->pipeline_contig_cpy_f16_f16; + } else { + return ctx->device->pipeline_cpy_f16_f16; + } + } + } + + std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl; GGML_ABORT("fatal error"); } @@ -2966,23 +4831,54 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& const int tensor_type_size = ggml_type_size(tensor->type); const uint32_t ne = ggml_nelements(tensor); + std::array<uint32_t, 3> elements; + + if (ne > 262144) { + elements = { 512, 512, CEIL_DIV(ne, 262144) }; + } else if (ne > 512) { + elements = { 512, CEIL_DIV(ne, 512), 1 }; + } else { + elements = { ne, 1, 1 }; + } - const vk_op_unary_push_constants pc = { + vk_op_unary_push_constants pc = { (uint32_t)ne, (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size, (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]), 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }; + init_pushconst_fastdiv(pc); ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 }); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements); } -static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) { + switch(type) { + case GGML_TYPE_Q8_1: + return ctx->device->pipeline_quantize_q8_1; + default: + std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl; + GGML_ABORT("fatal error"); + } +} + +static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne) { + VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")"); + + vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1); + + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(uint32_t), &ne, { ne, 1, 1 }); +} + +static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << ")"); + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT const uint64_t ne00 = src0->ne[0]; @@ -3001,13 +4897,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const uint64_t r2 = ne12 / ne02; const uint64_t r3 = ne13 / ne03; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; - vk_buffer d_Qx; + vk_buffer d_Qx = nullptr; size_t qx_buf_offset = 0; - vk_buffer d_Qy; + vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; bool src0_uma = false; @@ -3020,43 +4916,119 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub src1_uma = d_Qy != nullptr; } - const bool x_non_contig = !ggml_vk_dim01_contiguous(src0); - const bool y_non_contig = !ggml_vk_dim01_contiguous(src1); + // Reformat and convert to fp16 if non-contiguous, or for coopmat2 for better perf + const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || + !ggml_vk_dim01_contiguous(src0); + const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || + !ggml_vk_dim01_contiguous(src1); + + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type); + bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0; - const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + // Check for mmq first + vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr; if (mmp == nullptr) { + // Fall back to f16 dequant mul mat + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); + quantize_y = false; + } + + const bool qx_needs_dequant = mmp == nullptr || x_non_contig; + const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig); + + if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat - mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16); + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const int x_ne = ne01 * ne00; - const int y_ne = ne11 * ne10; - const int d_ne = ne11 * ne01; + const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); + const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11)); - const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8; + vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); - const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10); + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11; + const int x_ne = ne01 * ne00; + const int y_ne = padded_n * ne10; + const int d_ne = ne11 * ne01; - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned); + const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline); const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; - const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; + const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne); const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; + vk_pipeline to_q8_1 = nullptr; + + if (x_non_contig) { + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); + } else { + to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); + } + if (y_non_contig) { + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); + } else { + to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); + } + GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT + GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + + if (quantize_y) { + to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1); + } + + if (dryrun) { + const uint64_t x_sz_upd = x_sz * ne02 * ne03; + const uint64_t y_sz_upd = y_sz * ne12 * ne13; + const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0; + if ( + (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || + (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size) || + (split_k > 1 && split_k_size > ctx->device->max_memory_allocation_size)) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) { + ctx->prealloc_size_x = x_sz_upd; + } + if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) { + ctx->prealloc_size_y = y_sz_upd; + } + if (split_k > 1 && ctx->prealloc_size_split_k < split_k_size) { + ctx->prealloc_size_split_k = split_k_size; + } + + // Request descriptor sets + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + if (qx_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1); + } + if (qy_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1); + } + if (quantize_y) { + ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1); + } + if (split_k > 1) { + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1); + } + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03); vk_buffer d_X; @@ -3064,13 +5036,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub vk_buffer d_Y; uint64_t y_buf_offset = 0; if (!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (qx_needs_dequant) { @@ -3083,41 +5055,16 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub } if (qy_needs_dequant) { d_Y = ctx->prealloc_y; - GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03); + GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13); + } else if (quantize_y) { + d_Y = ctx->prealloc_y; + GGML_ASSERT(d_Y->size >= y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)); } else { d_Y = d_Qy; y_buf_offset = qy_buf_offset; GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline to_fp16_vk_0 = nullptr; - vk_pipeline to_fp16_vk_1 = nullptr; - - if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16); - } else { - to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); - } - if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16); - } else { - to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); - } - GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT - GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT - - // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, 1); - if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_0, 1); - } - if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_1, 1); - } - if (split_k > 1) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1); - } - if (x_non_contig) { ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); } else if (qx_needs_dequant) { @@ -3128,6 +5075,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (y_non_contig) { ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }); } + if (quantize_y) { + ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13); + } uint32_t stride_batch_x = ne00*ne01; uint32_t stride_batch_y = ne10*ne11; @@ -3136,7 +5086,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub stride_batch_x = src0->nb[0] / ggml_type_size(src0->type); } - if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) { + if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant && !quantize_y) { stride_batch_y = src1->nb[0] / ggml_type_size(src1->type); } @@ -3147,15 +5097,16 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - split_k, ne12*ne13, ne02, ne12, r2, r3 + split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n ); // NOLINT } -static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << "),)"); + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT const uint64_t ne00 = src0->ne[0]; @@ -3168,8 +5119,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t ne12 = src1->ne[2]; const uint64_t ne13 = src1->ne[3]; - GGML_ASSERT(ne11 == 1); - const uint64_t ne20 = dst->ne[0]; const uint64_t ne21 = dst->ne[1]; const uint64_t ne22 = dst->ne[2]; @@ -3178,13 +5127,18 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t r2 = ne12 / ne02; const uint64_t r3 = ne13 / ne03; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + // batch_n indicates that we need to compute a few vector results, and this assumes + // ne12 and ne13 are 1. It overloads the batch_strides to hold the row strides. + GGML_ASSERT(ne11 == 1 || ne12 * ne13 == 1); + bool batch_n = ne11 > 1; + + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; - vk_buffer d_Qx; + vk_buffer d_Qx = nullptr; size_t qx_buf_offset = 0; - vk_buffer d_Qy; + vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; bool src0_uma = false; @@ -3218,21 +5172,62 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; + if (x_non_contig) { + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type); + } + if (y_non_contig) { + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type); + } else { + to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); + } + vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11); + GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT + GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + GGML_ASSERT(dmmv != nullptr); + + if (dryrun) { + const uint64_t x_sz_upd = x_sz * ne02 * ne03; + const uint64_t y_sz_upd = y_sz * ne12 * ne13; + if ( + (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || + (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) { + ctx->prealloc_size_x = x_sz_upd; + } + if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) { + ctx->prealloc_size_y = y_sz_upd; + } + + // Request descriptor sets + if (qx_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1); + } + if (qy_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1); + } + ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1); + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if(!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if(!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (qx_needs_dequant) { @@ -3250,30 +5245,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline to_fp16_vk_0 = nullptr; - vk_pipeline to_fp16_vk_1 = nullptr; - if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type); - } - if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type); - } else { - to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); - } - vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type); - GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT - GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT - GGML_ASSERT(dmmv != nullptr); - - // Allocate descriptor sets - if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_0, 1); - } - if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13); - } - ggml_pipeline_allocate_descriptor_sets(ctx->device, dmmv, ne12 * ne13); - if (x_non_contig) { GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment)); ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); @@ -3283,8 +5254,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }); } - uint32_t stride_batch_x = ne00*ne01; - uint32_t stride_batch_y = ne10*ne11; + // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride + uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01; + uint32_t stride_batch_y = batch_n ? ne10 : (ne10*ne11); + uint32_t stride_batch_d = batch_n ? ne20 : (ne20*ne21); if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) { stride_batch_x = src0->nb[0] / ggml_type_size(src0->type); @@ -3301,13 +5274,13 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute const vk_mat_vec_push_constants pc = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01, - stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21), + stride_batch_x, stride_batch_y, stride_batch_d, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3, }; ggml_vk_sync_buffers(subctx); @@ -3316,10 +5289,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z }); } -static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_p021_f16_f32(" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)"); + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << ")"); GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1)); GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // NOLINT GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // NOLINT @@ -3338,11 +5312,11 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c GGML_ASSERT(ne11 == 1); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; - vk_buffer d_Qy; + vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; bool src1_uma = false; @@ -3360,21 +5334,30 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + // With grouped query attention there are > 1 Q matrices per K, V matrix. + uint32_t gqa_ratio = (uint32_t)ne12 / (uint32_t)ne02; + if (gqa_ratio > 8 || gqa_ratio == 0 || ne12 != ne02 * gqa_ratio) { + gqa_ratio = 1; + } + + if (dryrun) { + // Request descriptor sets + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1); + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); - vk_buffer d_Qx = extra_src0->buffer_gpu.lock(); - const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs; + vk_buffer d_Qx = src0_buf_ctx->dev_buffer; + const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qx != nullptr); } - // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, 1); - const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset; @@ -3383,14 +5366,22 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c // compute const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; + + uint32_t workgroups_z = (uint32_t)ne12; + // When gqa_ratio > 1, each invocation does multiple rows and we can launch fewer workgroups + if (gqa_ratio > 1) { + workgroups_z /= gqa_ratio; + } + ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, workgroups_z }); } -static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)"); + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << ")"); GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); GGML_ASSERT(!ggml_is_permuted(src0)); @@ -3405,6 +5396,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t nb01 = src0->nb[1]; const uint64_t nb02 = src0->nb[2]; + const uint64_t nb12 = src1->nb[2]; + // const uint64_t ne10 = src1->ne[0]; const uint64_t ne11 = src1->ne[1]; const uint64_t ne12 = src1->ne[2]; @@ -3412,9 +5405,9 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con GGML_ASSERT(ne11 == 1); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; @@ -3430,26 +5423,30 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t); const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t); + const uint32_t channel_stride_y = nb12 / sizeof(float); const uint64_t qx_sz = ggml_nbytes(src0); const uint64_t qy_sz = ggml_nbytes(src1); const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + if (dryrun) { + // Request descriptor sets + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1); + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); - vk_buffer d_Qx = extra_src0->buffer_gpu.lock(); - const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs; + vk_buffer d_Qx = src0_buf_ctx->dev_buffer; + const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qx != nullptr); } - // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1); - const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset; @@ -3457,26 +5454,39 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset; // compute - const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; + const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); } -static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")"); - if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) { - ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst); - } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1) { - ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst); - } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) { - ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst); + if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1 && + // detect 0213 permutation, and batch size of 1 + src0->nb[0] <= src0->nb[2] && + src0->nb[2] <= src0->nb[1] && + src0->nb[1] <= src0->nb[3] && + src1->nb[0] <= src1->nb[2] && + src1->nb[2] <= src1->nb[1] && + src1->nb[1] <= src1->nb[3] && + src0->ne[3] == 1 && + src1->ne[3] == 1) { + ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst, dryrun); + } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 && + !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) { + ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun); + // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four) + // when ne12 and ne13 are one. + } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) && + (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) { + ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun); } else { - ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst); + ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun); } } -static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) { +static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3]; @@ -3496,7 +5506,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t nei0 = ids->ne[0]; const uint64_t nei1 = ids->ne[1]; - GGML_ASSERT(nei0 * nei1 <= 3072); + GGML_ASSERT(nei0 * nei1 <= 4096); const uint32_t nbi1 = ids->nb[1]; const uint32_t nbi2 = ids->nb[2]; @@ -3508,16 +5518,16 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t n_as = ne02; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; - ggml_tensor_extra_gpu * extra_ids = (ggml_tensor_extra_gpu *) ids->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; + ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context; - vk_buffer d_Qx; + vk_buffer d_Qx = nullptr; size_t qx_buf_offset = 0; - vk_buffer d_Qy; + vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; - vk_buffer d_ids; + vk_buffer d_ids = nullptr; size_t ids_buf_offset = 0; bool src0_uma = false; @@ -3533,31 +5543,41 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& ids_uma = d_ids != nullptr; } - const bool x_non_contig = !ggml_vk_dim01_contiguous(src0); - const bool y_non_contig = !ggml_vk_dim01_contiguous(src1); + // Reformat and convert to fp16 if non-contiguous, or for coopmat2 for better perf + const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || + !ggml_vk_dim01_contiguous(src0); + const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || + !ggml_vk_dim01_contiguous(src1); + + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type); + vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig; - if (mmp == nullptr) { - GGML_ABORT("fatal error"); + if (qx_needs_dequant) { + // Fall back to dequant + f16 mulmat + mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint64_t x_ne = ne01 * ne00; - const uint64_t y_ne = ne11 * ne10; - const uint64_t d_ne = ne21 * ne20; - - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1)); + const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned); + vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type); + + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; + const uint64_t x_ne = ne01 * ne00; + const uint64_t y_ne = padded_n * ne10; + const uint64_t d_ne = ne21 * ne20; const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); @@ -3566,26 +5586,68 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t ids_sz = nbi2; const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; + + if (x_non_contig) { + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); + } else { + to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); + } + if (y_non_contig) { + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); + } else { + to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); + } + GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT + GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + + if (dryrun) { + const uint64_t x_sz_upd = x_sz * ne02 * ne03; + const uint64_t y_sz_upd = y_sz * ne12 * ne13; + if ( + (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || + (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) { + ctx->prealloc_size_x = x_sz_upd; + } + if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) { + ctx->prealloc_size_y = y_sz_upd; + } + + // Request descriptor sets + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + if (qx_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1); + } + if (qy_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1); + } + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if (!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (!ids_uma) { - d_ids = extra_ids->buffer_gpu.lock(); - ids_buf_offset = extra_ids->offset + ids->view_offs; + d_ids = ids_buf_ctx->dev_buffer; + ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs; GGML_ASSERT(d_ids != nullptr); } if (qx_needs_dequant) { @@ -3598,38 +5660,13 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& } if (qy_needs_dequant) { d_Y = ctx->prealloc_y; - GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03); + GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13); } else { d_Y = d_Qy; y_buf_offset = qy_buf_offset; GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline to_fp16_vk_0 = nullptr; - vk_pipeline to_fp16_vk_1 = nullptr; - - if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16); - } else { - to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); - } - if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16); - } else { - to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); - } - GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT - GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT - - // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, 1); - if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_0, 1); - } - if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_1, 1); - } - if (x_non_contig) { ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); } else if (qx_needs_dequant) { @@ -3660,16 +5697,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz }, ne01, ne21, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11 + n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n ); // NOLINT } -static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) { +static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3]; - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << ")"); + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT GGML_ASSERT(ids->type == GGML_TYPE_I32); @@ -3695,16 +5733,16 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte const uint64_t ne22 = dst->ne[2]; const uint64_t ne23 = dst->ne[3]; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; - ggml_tensor_extra_gpu * extra_ids = (ggml_tensor_extra_gpu *) ids->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; + ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context; - vk_buffer d_Qx; + vk_buffer d_Qx = nullptr; size_t qx_buf_offset = 0; - vk_buffer d_Qy; + vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; - vk_buffer d_ids; + vk_buffer d_ids = nullptr; size_t ids_buf_offset = 0; bool src0_uma = false; @@ -3742,26 +5780,67 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte const uint64_t ids_sz = nbi2; const uint64_t d_sz = sizeof(float) * d_ne; - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; + if (x_non_contig) { + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type); + } + if (y_non_contig) { + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type); + } else { + to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); + } + vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type); + GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT + GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + GGML_ASSERT(dmmv != nullptr); + + if (dryrun) { + const uint64_t x_sz_upd = x_sz * ne02 * ne03; + const uint64_t y_sz_upd = y_sz * ne12 * ne13; + if ( + (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || + (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) { + ctx->prealloc_size_x = x_sz_upd; + } + if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) { + ctx->prealloc_size_y = y_sz_upd; + } + + // Request descriptor sets + if (qx_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1); + } + if (qy_needs_dequant) { + ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1); + } + ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1); + return; + } + + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if(!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if(!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if(!ids_uma) { - d_ids = extra_ids->buffer_gpu.lock(); - ids_buf_offset = extra_ids->offset + ids->view_offs; + d_ids = ids_buf_ctx->dev_buffer; + ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs; GGML_ASSERT(d_ids != nullptr); } if (qx_needs_dequant) { @@ -3779,30 +5858,6 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline to_fp16_vk_0 = nullptr; - vk_pipeline to_fp16_vk_1 = nullptr; - if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type); - } - if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type); - } else { - to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); - } - vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type); - GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT - GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT - GGML_ASSERT(dmmv != nullptr); - - // Allocate descriptor sets - if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_0, 1); - } - if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13); - } - ggml_pipeline_allocate_descriptor_sets(ctx->device, dmmv, ne12 * ne13); - if (x_non_contig) { GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment)); ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); @@ -3825,7 +5880,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute @@ -3841,84 +5896,375 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z }); } -static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { +static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_id(" << src0 << ", " << src1 << ", " << src2 << ", " << dst << ")"); if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) { - ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst); + ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun); } else { - ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst); + ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun); } } -static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - VK_LOG_DEBUG("ggml_vk_op_repeat(" << src0 << ", " << src1 << ", " << dst << ")"); - const uint64_t ne0 = dst->ne[0]; - const uint64_t ne1 = dst->ne[1]; - const uint64_t ne2 = dst->ne[2]; - const uint64_t ne3 = dst->ne[3]; +static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t D, bool f32acc) { + // Needs to be kept up to date on shader changes + const uint32_t wg_size = scalar_flash_attention_workgroup_size; + const uint32_t Br = scalar_flash_attention_num_large_rows; + const uint32_t Bc = scalar_flash_attention_Bc; - const uint64_t ne00 = src0->ne[0]; - const uint64_t ne01 = src0->ne[1]; - const uint64_t ne02 = src0->ne[2]; - const uint64_t ne03 = src0->ne[3]; + const uint32_t acctype = f32acc ? 4 : 2; + const uint32_t f16vec4 = 8; - const uint64_t nb0 = dst->nb[0]; - const uint64_t nb1 = dst->nb[1]; - const uint64_t nb2 = dst->nb[2]; - const uint64_t nb3 = dst->nb[3]; + const uint32_t tmpsh = wg_size * sizeof(float); + const uint32_t tmpshv4 = wg_size * 4 * acctype; - const uint64_t nb00 = src0->nb[0]; - const uint64_t nb01 = src0->nb[1]; - const uint64_t nb02 = src0->nb[2]; - const uint64_t nb03 = src0->nb[3]; + const uint32_t Qf = Br * (D / 4 + 2) * f16vec4; - // guaranteed to be an integer due to the check in ggml_can_repeat - const uint64_t nr0 = ne0/ne00; - const uint64_t nr1 = ne1/ne01; - const uint64_t nr2 = ne2/ne02; - const uint64_t nr3 = ne3/ne03; + const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br; + const uint32_t sfsh = Bc * sfshstride * acctype; - // TODO: support for transposed / permuted tensors - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); + const uint32_t kshstride = D / 4 + 2; + const uint32_t ksh = Bc * kshstride * f16vec4; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - - const vk_buffer src_buf = extra_src0->buffer_gpu.lock(); - const uint64_t src_offset = extra_src0->offset + src0->view_offs; - vk_buffer dst_buf = extra->buffer_gpu.lock(); - const uint64_t dst_offset = extra->offset + dst->view_offs; - - std::vector<vk::BufferCopy> copies; - - for (uint64_t i3 = 0; i3 < nr3; i3++) { - for (uint64_t k3 = 0; k3 < ne03; k3++) { - for (uint64_t i2 = 0; i2 < nr2; i2++) { - for (uint64_t k2 = 0; k2 < ne02; k2++) { - for (uint64_t i1 = 0; i1 < nr1; i1++) { - for (uint64_t k1 = 0; k1 < ne01; k1++) { - for (uint64_t i0 = 0; i0 < nr0; i0++) { - copies.push_back({ - src_offset + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01, - dst_offset + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0, - ne00*nb0, - }); - } - } - } - } - } + const uint32_t slope = Br * sizeof(float); + + const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope; + const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize; + + VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(D=" << D << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported); + + return supported; +} + +static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, ggml_tensor * dst, bool dryrun = false) { + VK_LOG_DEBUG("ggml_vk_flash_attn((" << q << ", name=" << q->name << ", type=" << q->type << ", ne0=" << q->ne[0] << ", ne1=" << q->ne[1] << ", ne2=" << q->ne[2] << ", ne3=" << q->ne[3] << ", nb0=" << q->nb[0] << ", nb1=" << q->nb[1] << ", nb2=" << q->nb[2] << ", nb3=" << q->nb[3]; + std::cerr << "), (" << k << ", name=" << k->name << ", type=" << k->type << ", ne0=" << k->ne[0] << ", ne1=" << k->ne[1] << ", ne2=" << k->ne[2] << ", ne3=" << k->ne[3] << ", nb0=" << k->nb[0] << ", nb1=" << k->nb[1] << ", nb2=" << k->nb[2] << ", nb3=" << k->nb[3]; + std::cerr << "), (" << v << ", name=" << v->name << ", type=" << v->type << ", ne0=" << v->ne[0] << ", ne1=" << v->ne[1] << ", ne2=" << v->ne[2] << ", ne3=" << v->ne[3] << ", nb0=" << v->nb[0] << ", nb1=" << v->nb[1] << ", nb2=" << v->nb[2] << ", nb3=" << v->nb[3]; + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << (dryrun ? "dryrun" : "") << ")"); + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const uint32_t nem1 = mask ? mask->ne[1] : 0; + const uint32_t nbm1 = mask ? mask->nb[1] : 0; + + const uint32_t D = neq0; + uint32_t N = neq1; + const uint32_t KV = nek1; + + GGML_ASSERT(ne0 == D); + GGML_ASSERT(ne2 == N); + + // input tensor rows must be contiguous + GGML_ASSERT(nbq0 == ggml_type_size(q->type)); + GGML_ASSERT(nbk0 == ggml_type_size(k->type)); + GGML_ASSERT(nbv0 == ggml_type_size(v->type)); + + GGML_ASSERT(neq0 == D); + GGML_ASSERT(nek0 == D); + GGML_ASSERT(nev0 == D); + + GGML_ASSERT(neq1 == N); + GGML_ASSERT(nev0 == D); + + GGML_ASSERT(nev1 == nek1); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + assert(dst->type == GGML_TYPE_F32); + assert(q->type == GGML_TYPE_F32); + assert(k->type == v->type); + + FaCodePath path = ctx->device->coopmat2 ? FA_COOPMAT2 : + ctx->device->coopmat1_fa_support ? FA_COOPMAT1 : FA_SCALAR; + + if (path == FA_COOPMAT1) { + const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) || + (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc); + + const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, D, dst->op_params[3] == GGML_PREC_F32); + + if (!coopmat_shape_supported || !coopmat_shmem_supported) { + path = FA_SCALAR; + } + } + + uint32_t gqa_ratio = 1; + uint32_t qk_ratio = neq2 / nek2; + uint32_t workgroups_x = (uint32_t)neq1; + uint32_t workgroups_y = (uint32_t)neq2; + uint32_t workgroups_z = (uint32_t)neq3; + + // For scalar/coopmat1 FA, we can use the "large" size to accommodate qga. + // For coopmat2 FA, we always use the small size (which is still pretty large for gqa). + uint32_t max_gqa; + switch (path) { + case FA_SCALAR: + case FA_COOPMAT1: + // We may switch from coopmat1 to scalar, so use the scalar limit for both + max_gqa = scalar_flash_attention_num_large_rows; + break; + case FA_COOPMAT2: + max_gqa = get_fa_num_small_rows(FA_COOPMAT2); + break; + default: + GGML_ASSERT(0); + } + + if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa && + qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) { + // grouped query attention - make the N dimension equal to gqa_ratio, reduce + // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1 + // and change addressing calculations to index Q's dimension 2. + gqa_ratio = qk_ratio; + N = gqa_ratio; + workgroups_y /= N; + } + + vk_pipeline *pipelines; + bool small_rows = N <= get_fa_num_small_rows(path); + + // coopmat1 does not actually support "small rows" (it needs 16 rows). + // So use scalar instead. + if (small_rows && path == FA_COOPMAT1) { + path = FA_SCALAR; + } + + // scalar is faster than coopmat2 when N==1 + if (N == 1 && path == FA_COOPMAT2) { + path = FA_SCALAR; + } + + bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32; + + switch (path) { + case FA_SCALAR: + switch (D) { + case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break; + case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break; + case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break; + case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break; + case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break; + case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break; + default: + GGML_ASSERT(!"unsupported D value"); + return; + } + break; + case FA_COOPMAT1: + switch (D) { + case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm1[k->type][f32acc][small_rows][0]; break; + case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm1[k->type][f32acc][small_rows][0]; break; + case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm1[k->type][f32acc][small_rows][0]; break; + case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm1[k->type][f32acc][small_rows][0]; break; + case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm1[k->type][f32acc][small_rows][0]; break; + case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm1[k->type][f32acc][small_rows][0]; break; + default: + GGML_ASSERT(!"unsupported D value"); + return; + } + break; + case FA_COOPMAT2: + switch (D) { + case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break; + case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break; + case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break; + case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break; + case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break; + case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break; + default: + GGML_ASSERT(!"unsupported D value"); + return; + } + break; + default: + GGML_ASSERT(0); + } + assert(pipelines); + + const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type)); + const uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type)); + const uint32_t v_stride = (uint32_t)(nbv1 / ggml_type_size(v->type)); + + bool aligned = (KV % pipelines[1]->align) == 0 && + // the "aligned" shader variant will forcibly align strides, for performance + (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0; + + // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads + GGML_ASSERT((nem1 % GGML_KQ_MASK_PAD) == 0); + + vk_pipeline pipeline = pipelines[aligned]; + assert(pipeline); + + uint32_t split_kv = KV; + uint32_t split_k = 1; + + // Use a placeholder core count if one isn't available. split_k is a big help for perf. + const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count : 16; + + // Try to use split_k when KV is large enough to be worth the overhead + if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) { + // Try to run two workgroups per SM. + split_k = ctx->device->shader_core_count * 2 / workgroups_y; + if (split_k > 1) { + // Try to evenly split KV into split_k chunks, but it needs to be a multiple + // of "align", so recompute split_k based on that. + split_kv = ROUNDUP_POW2(KV / split_k, pipelines[1]->align); + split_k = CEIL_DIV(KV, split_kv); + workgroups_x = split_k; } } + // Reserve space for split_k temporaries. For each split, we need to store the O matrix (D x ne1) + // and the per-row m and L values (ne1 rows). + const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k : 0; + if (split_k_size > ctx->device->max_memory_allocation_size) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (ctx->prealloc_size_split_k < split_k_size) { + ctx->prealloc_size_split_k = split_k_size; + } + + if (dryrun) { + // Request descriptor sets + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + if (split_k > 1) { + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1); + } + return; + } + + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; + + memcpy(&scale, (const float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (const float *) dst->op_params + 2, sizeof(float)); + + if (logit_softcap != 0) { + scale /= logit_softcap; + } + + const uint32_t n_head_kv = neq2; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr; + size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0; + + bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false; + + if (ctx->device->uma) { + ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset); + ggml_vk_host_get(ctx->device, k->data, d_K, k_buf_offset); + ggml_vk_host_get(ctx->device, v->data, d_V, v_buf_offset); + ggml_vk_host_get(ctx->device, dst->data, d_D, d_buf_offset); + Q_uma = d_Q != nullptr; + K_uma = d_K != nullptr; + V_uma = d_V != nullptr; + D_uma = d_D != nullptr; + if (mask) { + ggml_vk_host_get(ctx->device, mask->data, d_M, m_buf_offset); + M_uma = d_M != nullptr; + } + } + + + ggml_backend_vk_buffer_context * d_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * q_buf_ctx = (ggml_backend_vk_buffer_context *)q->buffer->context; + ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context; + ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context; + + if (!Q_uma) { + d_Q = q_buf_ctx->dev_buffer; + q_buf_offset = vk_tensor_offset(q) + q->view_offs; + } + if (!K_uma) { + d_K = k_buf_ctx->dev_buffer; + k_buf_offset = vk_tensor_offset(k) + k->view_offs; + } + if (!V_uma) { + d_V = v_buf_ctx->dev_buffer; + v_buf_offset = vk_tensor_offset(v) + v->view_offs; + } + if (!D_uma) { + d_D = d_buf_ctx->dev_buffer; + d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; + } + + if (!M_uma) { + d_M = d_Q; + m_buf_offset = q_buf_offset; + if (mask) { + ggml_backend_vk_buffer_context * m_buf_ctx = (ggml_backend_vk_buffer_context*)mask->buffer->context; + d_M = m_buf_ctx->dev_buffer; + m_buf_offset = vk_tensor_offset(mask) + mask->view_offs; + } + } + + const vk_flash_attn_push_constants pc = { N, KV, + (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, + (uint32_t)neq2, (uint32_t)neq3, + (uint32_t)nek2, (uint32_t)nek3, + (uint32_t)nev2, (uint32_t)nev3, + nem1, + q_stride, (uint32_t)nbq2, (uint32_t)nbq3, + k_stride, (uint32_t)nbk2, (uint32_t)nbk3, + v_stride, (uint32_t)nbv2, (uint32_t)nbv3, + nbm1, + scale, max_bias, logit_softcap, + mask != nullptr, n_head_log2, m0, m1, + gqa_ratio, split_kv, split_k }; + ggml_vk_sync_buffers(subctx); - subctx->s->buffer.copyBuffer(src_buf->buffer, dst_buf->buffer, copies); - GGML_UNUSED(ctx); - GGML_UNUSED(src1); -} + if (split_k > 1) { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, + { + vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE}, + }, + // We only use split_k when group query attention is enabled, which means + // there's no more than one tile of rows (i.e. workgroups_x would have been + // one). We reuse workgroups_x to mean the number of splits, so we need to + // cancel out the divide by wg_denoms[0]. + sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z }); + ggml_vk_sync_buffers(subctx); + const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k }; + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce, + { + vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE}, + vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE}, + }, + pc2.size() * uint32_t{sizeof(uint32_t)}, pc2.data(), { (uint32_t)ne1, 1, 1 }); + } else { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, + { + vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE}, + }, + sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x, workgroups_y, workgroups_z }); + } +} static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) { switch (op) { @@ -3932,21 +6278,37 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_ADD: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_add_f32; + case GGML_OP_SUB: + case GGML_OP_MUL: + case GGML_OP_DIV: + if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) || + (src1->type != GGML_TYPE_F32 && src1->type != GGML_TYPE_F16) || + (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16)) { + return nullptr; } - if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { - return ctx->device->pipeline_add_f16_f32_f16; + switch (op) { + case GGML_OP_ADD: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; } - return nullptr; - case GGML_OP_MUL: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_mul_f32; + case GGML_OP_SUB: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_sub_norepeat : ctx->device->pipeline_sub; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; } - return nullptr; - case GGML_OP_DIV: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_div_f32; + case GGML_OP_MUL: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_norepeat : ctx->device->pipeline_mul; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + case GGML_OP_DIV: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_norepeat : ctx->device->pipeline_div; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + default: + break; } return nullptr; case GGML_OP_CONCAT: @@ -3985,10 +6347,25 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_pad_f32; } return nullptr; + case GGML_OP_REPEAT: + if (ggml_type_size(src0->type) == sizeof(float) && ggml_type_size(dst->type) == sizeof(float)) { + return ctx->device->pipeline_repeat_f32; + } + return nullptr; + case GGML_OP_REPEAT_BACK: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_repeat_back_f32; + } + return nullptr; case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: - return ggml_vk_get_cpy_pipeline(ctx, src0->type, dst->type); + return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type); + case GGML_OP_SILU_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_silu_back_f32; + } + return nullptr; case GGML_OP_NORM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_norm_f32; @@ -4004,33 +6381,31 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_rms_norm_f32; } return nullptr; + case GGML_OP_RMS_NORM_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_rms_norm_back_f32; + } + return nullptr; case GGML_OP_UNARY: + if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) || + (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16) || + (src0->type != dst->type)) { + return nullptr; + } + switch (ggml_get_unary_op(dst)) { case GGML_UNARY_OP_SILU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_silu_f32; - } - break; + return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_GELU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_gelu_f32; - } - break; + return ctx->device->pipeline_gelu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_GELU_QUICK: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_gelu_quick_f32; - } - break; + return ctx->device->pipeline_gelu_quick[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_RELU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_relu_f32; - } - break; + return ctx->device->pipeline_relu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_TANH: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_tanh_f32; - } - break; + return ctx->device->pipeline_tanh[dst->type == GGML_TYPE_F16]; + case GGML_UNARY_OP_SIGMOID: + return ctx->device->pipeline_sigmoid[dst->type == GGML_TYPE_F16]; default: break; } @@ -4044,13 +6419,19 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_soft_max_f32; + return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32; } if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_soft_max_f32_f16; + return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16; + } + return nullptr; + case GGML_OP_SOFT_MAX_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_soft_max_back_f32; } return nullptr; case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: { const int mode = ((const int32_t *) dst->op_params)[2]; const bool is_neox = mode & 2; @@ -4077,11 +6458,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_argsort_f32; } return nullptr; + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_sum_rows_f32; } return nullptr; + case GGML_OP_ARGMAX: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) { + return ctx->device->pipeline_argmax_f32; + } + return nullptr; case GGML_OP_IM2COL: if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_im2col_f32; @@ -4095,6 +6482,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_timestep_embedding_f32; } return nullptr; + case GGML_OP_CONV_TRANSPOSE_1D: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_conv_transpose_1d_f32; + } + return nullptr; + case GGML_OP_POOL_2D: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_pool2d_f32; + } + return nullptr; case GGML_OP_LEAKY_RELU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_leaky_relu_f32; @@ -4107,36 +6504,83 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const GGML_UNUSED(src2); } -static ggml_vk_func_t ggml_vk_op_get_func(ggml_op op) { - switch(op) { - case GGML_OP_REPEAT: - return ggml_vk_op_repeat; - default: - return nullptr; - } -} - static bool ggml_vk_op_supports_incontiguous(ggml_op op) { switch (op) { case GGML_OP_CPY: case GGML_OP_GET_ROWS: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: case GGML_OP_UPSCALE: - case GGML_OP_SCALE: case GGML_OP_SQR: case GGML_OP_CLAMP: case GGML_OP_PAD: + case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + case GGML_OP_ROPE: + case GGML_OP_RMS_NORM: + case GGML_OP_IM2COL: return true; default: return false; } } +static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_tensor * t) +{ + return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));; +} + +template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { + GGML_UNUSED(p); + GGML_UNUSED(src0); + GGML_UNUSED(src1); + GGML_UNUSED(src2); + GGML_UNUSED(dst); + static_assert(!std::is_const<T>::value, "unexpected type"); + GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0); + GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0); + GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0); + GGML_ASSERT(!dst || get_misalign_bytes(ctx, dst) == 0); +} + +template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { + const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type); + const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type); + + p.misalign_offsets = (a_offset << 16) | d_offset; + + GGML_UNUSED(src1); + GGML_UNUSED(src2); +} + +template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { + const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type); + const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type); + const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type); + + GGML_ASSERT(dst->op != GGML_OP_GET_ROWS || (a_offset == 0 && b_offset == 0 && d_offset == 0)); + + p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset; + + GGML_UNUSED(src2); +} + +template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { + const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type); + const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type); + + p.a_offset = a_offset; + p.d_offset = d_offset; + + GGML_UNUSED(src1); + GGML_UNUSED(src2); +} + template<typename PC> -static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) { +static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; if (src1 != nullptr) { std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; @@ -4144,10 +6588,11 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co if (src2 != nullptr) { std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3]; } - std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")"); - GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type)))); // NOLINT + std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; + std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")"); + GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type)))); // NOLINT GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0)); // NOLINT - GGML_ASSERT(dst->extra != nullptr); + GGML_ASSERT(dst->buffer != nullptr); const uint64_t ne00 = src0->ne[0]; const uint64_t ne01 = src0->ne[1]; const uint64_t ne02 = src0->ne[2]; @@ -4175,30 +6620,30 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co const uint64_t ned3 = dst->ne[3]; const uint64_t ned = ned0 * ned1; + init_pushconst_fastdiv(pc); + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op); - ggml_vk_func_t op_func; if (pipeline == nullptr) { - op_func = ggml_vk_op_get_func(op); - if (op_func == nullptr) { - std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(op) << " for " << ggml_type_name(src0->type); - if (src1 != nullptr) { - std::cerr << " and " << ggml_type_name(src1->type); - } - std::cerr << " to " << ggml_type_name(dst->type) << std::endl; - GGML_ABORT("fatal error"); + std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(op) << " for " << ggml_type_name(src0->type); + if (src1 != nullptr) { + std::cerr << " and " << ggml_type_name(src1->type); } + std::cerr << " to " << ggml_type_name(dst->type) << std::endl; + GGML_ABORT("fatal error"); + } - op_func(ctx, subctx, src0, src1, dst); + if (dryrun) { + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); return; } const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; - ggml_tensor_extra_gpu * extra_src2 = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr; + ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr; vk_buffer d_X = nullptr; size_t x_buf_offset = 0; @@ -4229,7 +6674,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co uint64_t z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 : 0; uint64_t d_sz = ggml_type_size(dst->type) * ned; - vk_buffer d_D = extra->buffer_gpu.lock(); + vk_buffer d_D = dst_buf_ctx->dev_buffer; // Workaround for tiny tensor inputs on ROPE if (op == GGML_OP_ROPE && use_src1 && y_sz > d_D->size) { @@ -4237,23 +6682,28 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co } GGML_ASSERT(d_D != nullptr); - uint64_t d_buf_offset = ((extra->offset + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; - GGML_ASSERT(d_buf_offset == extra->offset || op == GGML_OP_CPY); // NOLINT + uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; if(!src0_uma) { - d_X = extra_src0->buffer_gpu.lock(); - x_buf_offset = extra_src0->offset + src0->view_offs; + d_X = src0_buf_ctx->dev_buffer; + x_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_X != nullptr); } if (use_src1 && !src1_uma) { - d_Y = extra_src1->buffer_gpu.lock(); - y_buf_offset = extra_src1->offset + src1->view_offs; + d_Y = src1_buf_ctx->dev_buffer; + y_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Y != nullptr); } if (use_src2 && !src2_uma) { - d_Z = extra_src2->buffer_gpu.lock(); - z_buf_offset = extra_src2->offset + src2->view_offs; + d_Z = src2_buf_ctx->dev_buffer; + z_buf_offset = vk_tensor_offset(src2) + src2->view_offs; GGML_ASSERT(d_Z != nullptr); } + // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets. + init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst); + x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1); + y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1); + z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1); + d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1); if (op_supports_incontiguous) { x_sz = ggml_nbytes(src0); @@ -4278,188 +6728,175 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co std::array<uint32_t, 3> elements; // Single call if dimension 2 is contiguous - if (op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, 1); + GGML_ASSERT(op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))); - switch (op) { - case GGML_OP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_SOFT_MAX: - case GGML_OP_SUM_ROWS: - { - const uint32_t nr = ggml_nrows(src0); - if (nr > 262144) { - elements = { 512, 512, CEIL_DIV(nr, 262144) }; - } else if (nr > 512) { - elements = { 512, CEIL_DIV(nr, 512), 1 }; - } else { - elements = { nr, 1, 1 }; - } - } break; - case GGML_OP_GROUP_NORM: - { - const uint32_t num_groups = dst->op_params[0]; - elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 }; - } break; - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_ROPE: - elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 }; - break; - case GGML_OP_GET_ROWS: - elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) }; - break; - case GGML_OP_ARGSORT: - elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 }; - break; - case GGML_OP_IM2COL: - { - const bool is_2D = dst->op_params[6] == 1; + switch (op) { + case GGML_OP_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: + case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + { + const uint32_t nr = ggml_nrows(src0); + if (nr > 262144) { + elements = { 512, 512, CEIL_DIV(nr, 262144) }; + } else if (nr > 512) { + elements = { 512, CEIL_DIV(nr, 512), 1 }; + } else { + elements = { nr, 1, 1 }; + } + } break; + case GGML_OP_RMS_NORM: + elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 }; + break; - const uint32_t IC = src1->ne[is_2D ? 2 : 1]; + case GGML_OP_SUM: + // We use GGML_OP_SUM_ROWS with 1 row. + elements = { 1, 1, 1 }; + break; + case GGML_OP_GROUP_NORM: + { + const uint32_t num_groups = dst->op_params[0]; + elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 }; + } break; + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: + elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 }; + break; + case GGML_OP_GET_ROWS: + elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) }; + break; + case GGML_OP_ARGSORT: + elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 }; + break; + case GGML_OP_IM2COL: + { + const bool is_2D = dst->op_params[6] == 1; - const uint32_t KH = is_2D ? src0->ne[1] : 1; - const uint32_t KW = src0->ne[0]; + const uint32_t IC = src1->ne[is_2D ? 2 : 1]; - const uint32_t OH = is_2D ? dst->ne[2] : 1; - const uint32_t OW = dst->ne[1]; + const uint32_t KH = is_2D ? src0->ne[1] : 1; + const uint32_t KW = src0->ne[0]; - const uint32_t batch = src1->ne[3]; + const uint32_t OH = is_2D ? dst->ne[2] : 1; + const uint32_t OW = dst->ne[1]; - elements = { OW * KW * KH, OH, batch * IC }; - } break; - case GGML_OP_TIMESTEP_EMBEDDING: - { - const uint32_t dim = dst->op_params[0]; - uint32_t half_ceil = (dim + 1) / 2; - elements = { half_ceil, (uint32_t)src0->ne[0], 1 }; - } break; - case GGML_OP_ADD: - case GGML_OP_DIV: - case GGML_OP_MUL: - case GGML_OP_SCALE: - case GGML_OP_SQR: - case GGML_OP_CLAMP: - case GGML_OP_PAD: - case GGML_OP_CPY: - case GGML_OP_CONCAT: - case GGML_OP_UPSCALE: - case GGML_OP_UNARY: - { - const uint32_t ne = ggml_nelements(dst); - if (ne > 262144) { - elements = { 512, 512, CEIL_DIV(ne, 262144) }; - } else if (ne > 512) { - elements = { 512, CEIL_DIV(ne, 512), 1 }; + const uint32_t batch = src1->ne[is_2D ? 3 : 2]; + + elements = { OW * KW * KH, OH, batch * IC }; + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + const uint32_t dim = dst->op_params[0]; + uint32_t half_ceil = (dim + 1) / 2; + elements = { half_ceil, (uint32_t)src0->ne[0], 1 }; + } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + elements = {uint32_t(src0->ne[1]), 1, 1}; // parallelize in {Cout, 1, 1} + } break; + case GGML_OP_POOL_2D: + { + const uint32_t N = dst->ne[3]; + const uint32_t OC = dst->ne[2]; + const uint32_t OH = dst->ne[1]; + const uint32_t OW = dst->ne[0]; + elements = { N * OC * OH * OW, 1, 1}; + } break; + case GGML_OP_ADD: + case GGML_OP_SUB: + case GGML_OP_DIV: + case GGML_OP_MUL: + case GGML_OP_SCALE: + case GGML_OP_SQR: + case GGML_OP_CLAMP: + case GGML_OP_PAD: + case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + case GGML_OP_CPY: + case GGML_OP_CONCAT: + case GGML_OP_UPSCALE: + case GGML_OP_UNARY: + { + uint32_t ne = ggml_nelements(dst); + if (op == GGML_OP_CPY && ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) { + // Convert from number of logical elements to 2- or 4-byte units. + ne /= ggml_blck_size(src0->type); + if ((ggml_type_size(src0->type) % 4) == 0) { + ne *= ggml_type_size(src0->type) / 4; } else { - elements = { ne, 1, 1 }; + ne *= ggml_type_size(src0->type) / 2; } - } break; - default: - elements = { (uint32_t)ggml_nelements(src0), 1, 1 }; - break; - } - - if (!op_supports_incontiguous) { - if (x_sz != VK_WHOLE_SIZE) { - x_sz *= ne02 * ne03; - } - if (use_src1 && y_sz != VK_WHOLE_SIZE) { - y_sz *= ne12 * ne13; - } - if (use_src2 && z_sz != VK_WHOLE_SIZE) { - z_sz *= ne22 * ne23; - } - if (d_sz != VK_WHOLE_SIZE) { - d_sz *= ned2 * ned3; } - } - - if (op == GGML_OP_SOFT_MAX) { - // Empty src1 is possible in soft_max, but the shader needs a buffer - vk_subbuffer subbuf_y; - if (use_src1) { - subbuf_y = { d_Y, y_buf_offset, y_sz }; + if (ne > 262144) { + elements = { 512, 512, CEIL_DIV(ne, 262144) }; + } else if (ne > 512) { + elements = { 512, CEIL_DIV(ne, 512), 1 }; } else { - subbuf_y = { d_X, 0, x_sz }; + elements = { ne, 1, 1 }; } + } break; + default: + elements = { (uint32_t)ggml_nelements(src0), 1, 1 }; + break; + } - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (op == GGML_OP_ROPE) { - // Empty src2 is possible in rope, but the shader needs a buffer - vk_subbuffer subbuf_z; - if (use_src2) { - subbuf_z = { d_Z, z_buf_offset, z_sz }; - } else { - subbuf_z = { d_X, 0, x_sz }; - } + if (!op_supports_incontiguous) { + if (x_sz != VK_WHOLE_SIZE) { + x_sz *= ne02 * ne03; + } + if (use_src1 && y_sz != VK_WHOLE_SIZE) { + y_sz *= ne12 * ne13; + } + if (use_src2 && z_sz != VK_WHOLE_SIZE) { + z_sz *= ne22 * ne23; + } + if (d_sz != VK_WHOLE_SIZE) { + d_sz *= ned2 * ned3; + } + } - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (op == GGML_OP_IM2COL) { - // im2col uses only src1 and dst buffers - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (use_src2) { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (use_src1) { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + if (op == GGML_OP_SOFT_MAX) { + // Empty src1 is possible in soft_max, but the shader needs a buffer + vk_subbuffer subbuf_y; + if (use_src1) { + subbuf_y = { d_Y, y_buf_offset, y_sz }; } else { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + subbuf_y = { d_X, 0, x_sz }; } - } else { - GGML_ASSERT(op != GGML_OP_SOFT_MAX); - GGML_ASSERT(op != GGML_OP_ARGSORT); - GGML_ASSERT(!use_src2); - - ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, ne02 * ne03); - switch (op) { - case GGML_OP_NORM: - case GGML_OP_GROUP_NORM: - case GGML_OP_RMS_NORM: - elements = { (uint32_t)ne01, 1, 1 }; - break; - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_ROPE: - elements = { (uint32_t)ne01, (uint32_t)ne00, 1 }; - break; - case GGML_OP_GET_ROWS: - elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) }; - break; - default: - elements = { (uint32_t)ne0, 1, 1 }; - break; + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) { + // Empty src2 is possible in rope, but the shader needs a buffer + vk_subbuffer subbuf_z; + if (use_src2) { + subbuf_z = { d_Z, z_buf_offset, z_sz }; + } else { + subbuf_z = { d_X, 0, x_sz }; } - for (uint64_t i03 = 0; i03 < ne03; i03++) { - for (uint64_t i02 = 0; i02 < ne02; i02++) { - const uint32_t it_idx0 = (i03 * ne02 + i02); - const uint32_t it_idx1 = use_src1 ? ((i03 % ne13) * ne12 + (i02 % ne12)) : 0; - const uint32_t x_offset = x_sz * it_idx0; - const uint32_t y_offset = y_sz * it_idx1; - const uint32_t d_offset = d_sz * it_idx0; - - if (use_src1) { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset + x_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset + y_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); - } else { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset + x_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); - } - } - } + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else if (op == GGML_OP_IM2COL) { + // im2col uses only src1 and dst buffers + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else if (use_src2) { + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else if (use_src1) { + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else { + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); } } -static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, {}); -} - -static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4471,10 +6908,30 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, 0, - }); + }, dryrun); +} + +static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t src1_type_size = ggml_type_size(src1->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 + int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 + // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused + int offset = dst->op_params[3] / 4; // offset in bytes + + ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ACC, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, offset, + }, dryrun); } -static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4486,10 +6943,25 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, 0, - }); + }, dryrun); +} + +static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t src1_type_size = ggml_type_size(src1->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SUB, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, 0, + }, dryrun); } -static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4501,10 +6973,10 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, 0, - }); + }, dryrun); } -static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4516,10 +6988,11 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, 0, - }); + }, dryrun); } -static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + +static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { int * op_params = (int *)dst->op_params; const uint32_t src0_type_size = ggml_type_size(src0->type); @@ -4533,10 +7006,10 @@ static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, co (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, op_params[0], - }); + }, dryrun); } -static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const float sf0 = (float)dst->ne[0] / src0->ne[0]; @@ -4545,14 +7018,14 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c const float sf3 = (float)dst->ne[3] / src0->ne[3]; ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, { - (uint32_t)ggml_nelements(dst), 0, + (uint32_t)ggml_nelements(dst), 0, 0, (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3], sf0, sf1, sf2, sf3, - }); + }, dryrun); } -static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4562,11 +7035,12 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, - op_params[0], 0.0f - }); + op_params[0], 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4576,10 +7050,12 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, - }); + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { + +static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4590,10 +7066,11 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, op_params[0], op_params[1], - }); + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4603,55 +7080,114 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, - }); + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); +} + +static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, { + (uint32_t)ggml_nelements(dst), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; +static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); - const uint32_t d_offset = ((extra->offset + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; + + ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, { + (uint32_t)ggml_nelements(dst), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); +} + +static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + uint32_t ne = (uint32_t)ggml_nelements(src0); + if (ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) { + // Convert from number of logical elements to 2- or 4-byte units. + ne /= ggml_blck_size(src0->type); + if ((ggml_type_size(src0->type) % 4) == 0) { + ne *= ggml_type_size(src0->type) / 4; + } else { + ne *= ggml_type_size(src0->type) / 2; + } + } ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, { - (uint32_t)ggml_nelements(src0), + ne, (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, - d_offset, + 0, 0.0f, 0.0f, - }); + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); +} + +static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }); + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun); } -static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - int * op_params = (int *)dst->op_params; +static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + const int * int_op_params = (const int *)dst->op_params; + const float * float_op_params = (const float *)dst->op_params; - uint32_t num_groups = op_params[0]; - uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - static const float eps = 1e-6f; + const uint32_t num_groups = int_op_params[0]; + const float eps = float_op_params[1]; + const uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }); + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun); +} + +static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + float * op_params = (float *)dst->op_params; + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + op_params[0], 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); } -static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }); + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun); } -static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }); +static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); } -static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { int32_t * op_params = (int32_t *)dst->op_params; - ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }); + ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun); } -static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; float scale = op_params[0]; @@ -4673,12 +7209,18 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, scale, max_bias, m0, m1, n_head_log2, - }); + nrows_x, + }, dryrun); +} + +static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + float * op_params = (float *)dst->op_params; + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], op_params[1] }, dryrun); } -static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { +static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) { const int n_dims = ((int32_t *) dst->op_params)[1]; - // const int mode = ((int32_t *) dst->op_params)[2]; + const int mode = ((int32_t *) dst->op_params)[2]; // const int n_ctx = ((int32_t *) dst->op_params)[3]; const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; const float freq_base = ((float *) dst->op_params)[5]; @@ -4687,20 +7229,28 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons const float attn_factor = ((float *) dst->op_params)[8]; const float beta_fast = ((float *) dst->op_params)[9]; const float beta_slow = ((float *) dst->op_params)[10]; + int sections[4] {}; + if (mode & 8) { + memcpy(sections, (int32_t *) dst->op_params + 11, sizeof(int)*4); + } float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); const float theta_scale = powf(freq_base, -2.0f/n_dims); + uint32_t s1 = src0->nb[1] / ggml_type_size(src0->type); + uint32_t s2 = src0->nb[2] / ggml_type_size(src0->type); + ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, - src2 != nullptr, - }); + src2 != nullptr, (uint32_t)src0->ne[2], s1, s2, + sections[0], sections[1], sections[2], sections[3], backprop + }, dryrun); } -static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { int32_t * op_params = (int32_t *)dst->op_params; uint32_t ncols = src0->ne[0]; @@ -4716,14 +7266,22 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c ncols, ncols_pad, op_params[0], - }); + }, dryrun); } -static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }); +static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); } -static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun); +} + +static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun); +} + +static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const int32_t s0 = dst->op_params[0]; const int32_t s1 = dst->op_params[1]; const int32_t p0 = dst->op_params[2]; @@ -4744,7 +7302,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co const uint32_t OW = dst->ne[1]; const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 - const uint32_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32 + const uint32_t batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32 const uint32_t pelements = OW * KW * KH; @@ -4754,22 +7312,82 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co pelements, IC * KH * KW, s0, s1, p0, p1, d0, d1, - }); + }, dryrun); } -static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t dim = dst->op_params[0]; const uint32_t max_period = dst->op_params[1]; const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type); ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, { nb1, dim, max_period, - }); + }, dryrun); } -static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + // src0: (K, Cout, Cin, 1) -- kernel + // src1: (L, Cin, 1, 1) -- input + // dst: (*, Cout, 1, 1) + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb10 == sizeof(float)); + + const int32_t s0 = dst->op_params[0]; + + vk_op_conv_transpose_1d_push_constants p{}; + p.Cout = static_cast<uint32_t>(ne01); + p.Cin = static_cast<uint32_t>(ne02); + p.K = static_cast<uint32_t>(ne00); + p.L = static_cast<uint32_t>(ne10); + p.KL = static_cast<uint32_t>(ne0); + p.nb01 = static_cast<uint32_t>(nb01 / nb00); + p.nb02 = static_cast<uint32_t>(nb02 / nb00); + p.nb11 = static_cast<uint32_t>(nb11 / nb10); + p.nb1 = static_cast<uint32_t>(nb1 / nb0); + p.s0 = static_cast<uint32_t>(s0); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun); +} + +static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + uint32_t op = static_cast<uint32_t>(dst->op_params[0]); + const int32_t k1 = dst->op_params[1]; + const int32_t k0 = dst->op_params[2]; + const int32_t s1 = dst->op_params[3]; + const int32_t s0 = dst->op_params[4]; + const int32_t p1 = dst->op_params[5]; + const int32_t p0 = dst->op_params[6]; + + const uint32_t IH = src0->ne[1]; + const uint32_t IW = src0->ne[0]; + + const uint32_t N = dst->ne[3]; + + const uint32_t OC = dst->ne[2]; + const uint32_t OH = dst->ne[1]; + const uint32_t OW = dst->ne[0]; + + const uint32_t parallel_elements = N * OC * OH * OW; + + ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_POOL_2D, { + IW, IH, OW, OH, OC, + parallel_elements, + op, + k0, k1, s0, s1, p0, p1, + }, dryrun); +} + + +static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const float * op_params = (const float *)dst->op_params; - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }); + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun); } #ifdef GGML_VULKAN_RUN_TESTS @@ -4823,10 +7441,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->a_s; shname = "F32_F16_ALIGNED_S"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->a_s; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_s; shname = "F16_F32_ALIGNED_S"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->a_s; + p = ctx->device->pipeline_matmul_f16.f32acc->a_s; shname = "F16_ALIGNED_S"; } else { GGML_ABORT("fatal error"); @@ -4839,10 +7457,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->a_m; shname = "F32_F16_ALIGNED_M"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->a_m; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_m; shname = "F16_F32_ALIGNED_M"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->a_m; + p = ctx->device->pipeline_matmul_f16.f32acc->a_m; shname = "F16_ALIGNED_M"; } else { GGML_ABORT("fatal error"); @@ -4855,10 +7473,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->a_l; shname = "F32_F16_ALIGNED_L"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->a_l; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_l; shname = "F16_F32_ALIGNED_L"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->a_l; + p = ctx->device->pipeline_matmul_f16.f32acc->a_l; shname = "F16_ALIGNED_L"; } else { GGML_ABORT("fatal error"); @@ -4878,10 +7496,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->s; shname = "F32_F16_S"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->s; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->s; shname = "F16_F32_S"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->s; + p = ctx->device->pipeline_matmul_f16.f32acc->s; shname = "F16_S"; } } else if (shader_size == 1) { @@ -4892,10 +7510,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->m; shname = "F32_F16_M"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->m; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->m; shname = "F16_F32_M"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->m; + p = ctx->device->pipeline_matmul_f16.f32acc->m; shname = "F16_M"; } } else if (shader_size == 2) { @@ -4906,18 +7524,18 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f32_f16->l; shname = "F32_F16_L"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16_f32->l; + p = ctx->device->pipeline_matmul_f16_f32.f32acc->l; shname = "F16_F32_L"; } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) { - p = ctx->device->pipeline_matmul_f16->l; + p = ctx->device->pipeline_matmul_f16.f32acc->l; shname = "F16_L"; } } } - ggml_pipeline_allocate_descriptor_sets(ctx->device, p, num_it); + ggml_pipeline_request_descriptor_sets(ctx, p, num_it); if (split_k > 1) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it); + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it); if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) { // Resize buffer @@ -4928,6 +7546,12 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t } } + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } + + ggml_pipeline_allocate_descriptor_sets(ctx); + vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal); @@ -4939,19 +7563,27 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t for (size_t i = 0; i < x_ne; i++) { if (std::is_same<float, X_TYPE>()) { x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // x[i] = 1.0f; + // x[i] = i + 1; + // x[i] = (i % k == i / k) ? 1.0f : 0.0f; } else if (std::is_same<ggml_fp16_t, X_TYPE>()) { x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + // x[i] = ggml_fp32_to_fp16(1.0f); + // x[i] = ggml_fp32_to_fp16(i + 1); + // x[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); } else { GGML_ABORT("fatal error"); } } for (size_t i = 0; i < y_ne; i++) { if (std::is_same<float, Y_TYPE>()) { - // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; - y[i] = (i % k == i / k) ? 1.0f : 0.0f; + y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // y[i] = (i % k == i / k) ? 1.0f : 0.0f; + // y[i] = i + 1; } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) { - // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); - y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); + y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + // y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); + // y[i] = ggml_fp32_to_fp16(i + 1); } else { GGML_ABORT("fatal error"); } @@ -4960,22 +7592,23 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t ggml_vk_buffer_write(d_X, 0, x, sizeof(X_TYPE) * k * m * batch); ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch); - vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); + vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); + ggml_vk_ctx_begin(ctx->device, subctx); for (size_t i = 0; i < num_it; i++) { - ggml_vk_ctx_begin(ctx->device, subctx); ggml_vk_matmul( ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 + split_k, batch, batch, batch, 1, 1, n ); - ggml_vk_ctx_end(subctx); } + ggml_vk_ctx_end(subctx); auto begin = std::chrono::high_resolution_clock::now(); ggml_vk_submit(subctx, ctx->fence); VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences"); ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_queue_command_pools_cleanup(ctx->device); auto end = std::chrono::high_resolution_clock::now(); double time = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0; @@ -5035,7 +7668,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t double err = std::fabs(d[i] - d_chk[i]); avg_err += err; - if (err > 0.05f && first_err_n == -1) { + if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) { first_err_b = i / (m * n); first_err_n = (i % (m * n)) / m; first_err_m = (i % (m * n)) % m; @@ -5044,14 +7677,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t avg_err /= m * n; - std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms avg_err=" << avg_err << std::endl; + double tflops = 2.0*m*n*k*batch*num_it / (time / 1000.0) / (1000.0*1000.0*1000.0*1000.0); - if (avg_err > 0.1) { + std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl; + + if (avg_err > 0.1 || std::isnan(avg_err)) { std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; std::cerr << "Actual result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); - std::cerr << std::endl; - ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b); std::cerr << "Expected result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); @@ -5077,16 +7710,13 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t free(d_chk); - ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue); - ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue); + ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool); + ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool); ggml_vk_destroy_buffer(d_X); ggml_vk_destroy_buffer(d_Y); ggml_vk_destroy_buffer(d_D); - ggml_pipeline_cleanup(p); - ggml_pipeline_cleanup(ctx->device->pipeline_matmul_split_k_reduce); - free(x); free(y); free(d); @@ -5164,14 +7794,20 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ ggml_vk_quantize_data(x, qx, ne, quant); ggml_vk_dequantize_data(qx, x_ref, ne, quant); - ggml_pipeline_allocate_descriptor_sets(ctx->device, p, 1); + ggml_pipeline_request_descriptor_sets(ctx, p, 1); + + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } + + ggml_pipeline_allocate_descriptor_sets(ctx); ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz); - vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); + vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); ggml_vk_ctx_begin(ctx->device, subctx); const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne }; - ggml_vk_dispatch_pipeline(ctx, subctx, p, { { qx_buf, 0, qx_sz }, { x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1}); ggml_vk_ctx_end(subctx); auto begin = std::chrono::high_resolution_clock::now(); @@ -5179,6 +7815,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ ggml_vk_submit(subctx, ctx->fence); VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences"); ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_queue_command_pools_cleanup(ctx->device); auto end = std::chrono::high_resolution_clock::now(); @@ -5223,71 +7860,204 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ free(x_chk); } -static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant) { +// This does not work without ggml q8_1 quantization support +// +// typedef uint16_t ggml_half; +// typedef uint32_t ggml_half2; +// +// #define QK8_1 32 +// typedef struct { +// union { +// struct { +// ggml_half d; // delta +// ggml_half s; // d * sum(qs[i]) +// } GGML_COMMON_AGGR_S; +// ggml_half2 ds; +// } GGML_COMMON_AGGR_U; +// int8_t qs[QK8_1]; // quants +// } block_q8_1; +// +// static void ggml_vk_test_quantize(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) { +// VK_LOG_DEBUG("ggml_vk_test_quantize(" << ne << ")"); +// GGML_ASSERT(quant == GGML_TYPE_Q8_1); +// +// const size_t x_sz = sizeof(float) * ne; +// const size_t qx_sz = ne * ggml_type_size(quant)/ggml_blck_size(quant); +// float * x = (float *) malloc(x_sz); +// block_q8_1 * qx = (block_q8_1 *)malloc(qx_sz); +// block_q8_1 * qx_res = (block_q8_1 *)malloc(qx_sz); +// vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); +// vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); +// +// for (size_t i = 0; i < ne; i++) { +// x[i] = rand() / (float)RAND_MAX; +// } +// +// vk_pipeline p = ggml_vk_get_quantize_pipeline(ctx, quant); +// +// ggml_pipeline_request_descriptor_sets(ctx, p, 1); +// +// if (ctx->device->need_compiles) { +// ggml_vk_load_shaders(ctx->device); +// } +// +// ggml_pipeline_allocate_descriptor_sets(ctx); +// +// ggml_vk_buffer_write(x_buf, 0, x, x_sz); +// +// vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); +// ggml_vk_ctx_begin(ctx->device, subctx); +// ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(x_buf), ggml_vk_subbuffer(qx_buf), ne); +// ggml_vk_ctx_end(subctx); +// +// auto begin = std::chrono::high_resolution_clock::now(); +// +// ggml_vk_submit(subctx, ctx->fence); +// VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_quantize waitForFences"); +// ctx->device->device.resetFences({ ctx->fence }); +// ggml_vk_queue_command_pools_cleanup(ctx->device); +// +// auto end = std::chrono::high_resolution_clock::now(); +// +// double ms_quant = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0; +// ggml_vk_buffer_read(qx_buf, 0, qx, qx_sz); +// +// ggml_vk_quantize_data(x, qx_res, ne, quant); +// +// int first_err = -1; +// +// for (size_t i = 0; i < ne / 32; i++) { +// double error = std::fabs(ggml_fp16_to_fp32(qx_res[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) - ggml_fp16_to_fp32(qx[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d)); +// +// if (first_err < 0 && error > 0.1) { +// first_err = i; +// } +// +// error = std::fabs(ggml_fp16_to_fp32(qx_res[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) - ggml_fp16_to_fp32(qx[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s)); +// +// if (first_err < 0 && error > 0.1) { +// first_err = i; +// } +// +// for (size_t j = 0; j < 32; j++) { +// uint64_t error = std::abs(qx_res[i].qs[j] - qx[i].qs[j]); +// +// if (first_err < 0 && error > 1) { +// first_err = i; +// } +// } +// } +// +// std::cerr << "TEST QUANTIZE " << ggml_type_name(quant) << " time=" << ms_quant << "ms " << (first_err == -1 ? "CORRECT" : "INCORRECT") << std::endl; +// +// if (first_err != -1) { +// std::cerr << "first_error = " << first_err << std::endl; +// std::cerr << "Actual result: " << std::endl << std::endl; +// std::cout << "d=" << ggml_fp16_to_fp32(qx[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) << " s=" << ggml_fp16_to_fp32(qx[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) << " "; +// for (size_t j = 0; j < 32; j++) { +// std::cout << " qs" << j << "=" << (uint32_t)qx[first_err].qs[j] << " "; +// } +// std::cerr << std::endl << std::endl << "Expected result: " << std::endl << std::endl; +// std::cout << "d=" << ggml_fp16_to_fp32(qx_res[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) << " s=" << ggml_fp16_to_fp32(qx_res[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) << " "; +// for (size_t j = 0; j < 32; j++) { +// std::cout << " qs" << j << "=" << (uint32_t)qx_res[first_err].qs[j] << " "; +// } +// std::cerr << std::endl; +// } +// +// ggml_vk_destroy_buffer(x_buf); +// ggml_vk_destroy_buffer(qx_buf); +// +// free(x); +// free(qx); +// free(qx_res); +// } + +static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant, bool mmq = false) { VK_LOG_DEBUG("ggml_vk_test_dequant_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << ggml_type_name(quant) << ")"); const size_t x_ne = m * k * batch; const size_t y_ne = k * n * batch; const size_t d_ne = m * n * batch; + vk_matmul_pipeline2 * pipelines; + + if (mmq) { + pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1; + } else { + pipelines = ctx->device->pipeline_dequant_mul_mat_mat; + } + + const bool fp16acc = ctx->device->fp16; + vk_pipeline p; std::string shname; if (shader_size == 0) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s; + p = fp16acc ? pipelines[quant].f16acc->a_s : pipelines[quant].f32acc->a_s; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S"; } else if (shader_size == 1) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m; + p = fp16acc ? pipelines[quant].f16acc->a_m : pipelines[quant].f32acc->a_m; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M"; } else if (shader_size == 2) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l; + p = fp16acc ? pipelines[quant].f16acc->a_l : pipelines[quant].f32acc->a_l; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L"; } else { GGML_ASSERT(0); } - const size_t kpad = ggml_vk_align_size(k, p->align); + const size_t kpad = mmq ? 0 : ggml_vk_align_size(k, p->align); - if (k != kpad) { + if (mmq || k != kpad) { if (shader_size == 0) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s; + p = fp16acc ? pipelines[quant].f16acc->s : pipelines[quant].f32acc->s; shname = std::string(ggml_type_name(quant)) + "_S"; } else if (shader_size == 1) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m; + p = fp16acc ? pipelines[quant].f16acc->m : pipelines[quant].f32acc->m; shname = std::string(ggml_type_name(quant)) + "_M"; } else if (shader_size == 2) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l; + p = fp16acc ? pipelines[quant].f16acc->l : pipelines[quant].f32acc->l; shname = std::string(ggml_type_name(quant)) + "_L"; } else { GGML_ASSERT(0); } } + if (p == nullptr) { + std::cerr << "error: no pipeline for ggml_vk_test_dequant_matmul " << ggml_type_name(quant) << std::endl; + return; + } + const size_t x_sz = sizeof(float) * x_ne; const size_t y_sz = sizeof(float) * y_ne; const size_t qx_sz = x_ne * ggml_type_size(quant)/ggml_blck_size(quant); + const size_t qy_sz = mmq ? y_ne * ggml_type_size(GGML_TYPE_Q8_1)/ggml_blck_size(GGML_TYPE_Q8_1) : y_sz; const size_t d_sz = sizeof(float) * d_ne; float * x = (float *) malloc(x_sz); float * y = (float *) malloc(y_sz); void * qx = malloc(qx_sz); vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer y_buf = ggml_vk_create_buffer_check(ctx->device, y_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); + vk_buffer qy_buf = ggml_vk_create_buffer_check(ctx->device, qy_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer d_buf = ggml_vk_create_buffer_check(ctx->device, d_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); float * d = (float *) malloc(d_sz); float * d_chk = (float *) malloc(d_sz); for (size_t i = 0; i < x_ne; i++) { x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // x[i] = (i % k == i / k) ? 1.0f : 0.0f; + // x[i] = i % k; } ggml_vk_quantize_data(x, qx, x_ne, quant); for (size_t i = 0; i < y_ne; i++) { - // y[i] = rand() / (float)RAND_MAX; - y[i] = (i % k == i / k) ? 1.0f : 0.0f; + y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // y[i] = (i % k == i / k) ? 1.0f : 0.0f; + // y[i] = i % k; } - ggml_pipeline_allocate_descriptor_sets(ctx->device, p, num_it); + ggml_pipeline_request_descriptor_sets(ctx, p, num_it); if (split_k > 1) { - ggml_pipeline_allocate_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it); + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it); if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) { // Resize buffer @@ -5297,27 +8067,49 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal); } } + if (mmq) { + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_quantize_q8_1, num_it); + } + + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } + + ggml_pipeline_allocate_descriptor_sets(ctx); ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz); ggml_vk_buffer_write(y_buf, 0, y, y_sz); - vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); - for (size_t i = 0; i < num_it; i++) { - ggml_vk_ctx_begin(ctx->device, subctx); - ggml_vk_matmul( - ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k), - m, n, k, - k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 - ); - ggml_vk_ctx_end(subctx); + vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); + ggml_vk_ctx_begin(ctx->device, subctx); + if (mmq) { + for (size_t i = 0; i < num_it; i++) { + ggml_vk_quantize_q8_1(ctx, subctx, { y_buf, 0, y_sz }, { qy_buf, 0, qy_sz }, y_ne); + ggml_vk_matmul( + ctx, subctx, p, { qx_buf, 0, qx_sz }, { qy_buf, 0, qy_sz }, { d_buf, 0, d_sz }, { ctx->prealloc_split_k, 0, ctx->prealloc_size_split_k }, + m, n, k, + k, k, m, k*m, k*n, m*n, + split_k, batch, batch, batch, 1, 1, n + ); + } + } else { + for (size_t i = 0; i < num_it; i++) { + ggml_vk_matmul( + ctx, subctx, p, { qx_buf, 0, qx_sz }, { y_buf, 0, y_sz }, { d_buf, 0, d_sz }, { ctx->prealloc_split_k, 0, ctx->prealloc_size_split_k }, + m, n, k, + k, k, m, k*m, k*n, m*n, + split_k, batch, batch, batch, 1, 1, n + ); + } } + ggml_vk_ctx_end(subctx); auto begin = std::chrono::high_resolution_clock::now(); ggml_vk_submit(subctx, ctx->fence); VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences"); ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_queue_command_pools_cleanup(ctx->device); auto end = std::chrono::high_resolution_clock::now(); @@ -5365,7 +8157,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, avg_err /= m * n; - std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms avg_err=" << avg_err << std::endl; + double tflops = 2.0*m*n*k*batch*num_it / (time_ms / 1000.0) / (1000.0*1000.0*1000.0*1000.0); + + std::cerr << "TEST dequant matmul " << shname; + if (mmq) { + std::cerr << " mmq"; + } + std::cerr << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl; if (avg_err > 0.01 || std::isnan(avg_err)) { std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; @@ -5375,6 +8173,12 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, std::cerr << "Expected result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + std::cerr << "src0: " << std::endl << std::endl; + ggml_vk_print_matrix_area(x, GGML_TYPE_F32, k, m, first_err_m, first_err_n, first_err_b); + std::cerr << std::endl; + std::cerr << "src1: " << std::endl << std::endl; + ggml_vk_print_matrix_area(y, GGML_TYPE_F32, k, n, first_err_m, first_err_n, first_err_b); + if (split_k > 1) { float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k); ggml_vk_buffer_read(ctx->prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k); @@ -5397,6 +8201,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, ggml_vk_destroy_buffer(qx_buf); ggml_vk_destroy_buffer(y_buf); + ggml_vk_destroy_buffer(qy_buf); ggml_vk_destroy_buffer(d_buf); free(x); @@ -5407,247 +8212,15 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, } #endif -static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor) { - VK_LOG_DEBUG("ggml_vk_create_extra(" << tensor << " (" << tensor->name << ", " << ggml_op_name(tensor->op) << "))"); - ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu; - extra->reset(); - tensor->extra = extra; - return extra; -} - -static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggml_tensor * node){ - VK_LOG_DEBUG("ggml_vk_preallocate_buffers_graph(" << node << ")"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra; - - if (extra == nullptr) { - return; - } - - ggml_tensor * src0 = node->src[0]; - ggml_tensor * src1 = node->src[1]; - - const bool use_src0 = src0 != nullptr; - const int64_t ne00 = use_src0 ? src0->ne[0] : 0; - const int64_t ne01 = use_src0 ? src0->ne[1] : 0; - const int64_t ne02 = use_src0 ? src0->ne[2] : 0; - const int64_t ne03 = use_src0 ? src0->ne[3] : 0; - const bool use_src1 = src1 != nullptr && node->op != GGML_OP_CPY && node->op != GGML_OP_CONT && node->op != GGML_OP_DUP; - const int64_t ne10 = use_src1 ? src1->ne[0] : 0; - const int64_t ne11 = use_src1 ? src1->ne[1] : 0; - const int64_t ne12 = use_src1 ? src1->ne[2] : 0; - const int64_t ne13 = use_src1 ? src1->ne[3] : 0; - const int64_t ne20 = node->ne[0]; - const int64_t ne21 = node->ne[1]; - const int64_t ne22 = node->ne[2]; - const int64_t ne23 = node->ne[3]; - - const ggml_type src0_type = (use_src0 && src0->type == GGML_TYPE_F32) ? src0->type : GGML_TYPE_F16; - const ggml_type src1_type = (use_src1 && src1->type == GGML_TYPE_F32) ? src1->type : GGML_TYPE_F16; - - const bool x_non_contig = use_src0 && !ggml_vk_dim01_contiguous(src0); - const bool y_non_contig = use_src1 && !ggml_vk_dim01_contiguous(src1); - - const bool y_f32_kernel = use_src1 && src1->type == GGML_TYPE_F32 && !y_non_contig; - - bool mmp = (use_src0 && use_src1 && (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID)) ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type) != nullptr : false; - - const bool qx_needs_dequant = use_src0 && (!mmp || x_non_contig); - const bool qy_needs_dequant = use_src1 && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig); - - int split_k; - if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) { - split_k = ggml_vk_guess_split_k(ne01, ne11, ne10); - } else { - split_k = 1; - } - const uint32_t x_ne = ne00 * ne01; - const uint32_t y_ne = ne10 * ne11; - const uint32_t d_ne = ne20 * ne21; - - const uint64_t x_sz = (use_src0 && qx_needs_dequant) ? ggml_vk_align_size(sizeof(src0_type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0; - const uint64_t y_sz = (use_src1 && qy_needs_dequant) ? ggml_vk_align_size(sizeof(src1_type) * y_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0; - uint64_t d_sz = ggml_vk_align_size(ggml_type_size(node->type) * d_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne22 * ne23; - const uint64_t split_k_size = split_k > 1 ? d_sz * 4 : 0; - - if (extra->buffer_gpu.expired()) { - // Workaround for CPU backend BLAS matmul calls - extra->buffer_gpu = ggml_vk_create_buffer_temp(ctx, d_sz); - } - - switch (node->op) { - case GGML_OP_REPEAT: - case GGML_OP_GET_ROWS: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - case GGML_OP_ADD: - case GGML_OP_SCALE: - case GGML_OP_SQR: - case GGML_OP_CLAMP: - case GGML_OP_PAD: - case GGML_OP_CPY: - case GGML_OP_CONT: - case GGML_OP_DUP: - case GGML_OP_MUL: - case GGML_OP_DIV: - case GGML_OP_CONCAT: - case GGML_OP_UPSCALE: - case GGML_OP_NORM: - case GGML_OP_GROUP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_SOFT_MAX: - case GGML_OP_ROPE: - case GGML_OP_ARGSORT: - case GGML_OP_SUM_ROWS: - case GGML_OP_IM2COL: - case GGML_OP_TIMESTEP_EMBEDDING: - case GGML_OP_LEAKY_RELU: - break; - case GGML_OP_UNARY: - switch (ggml_get_unary_op(node)) { - case GGML_UNARY_OP_SILU: - case GGML_UNARY_OP_GELU: - case GGML_UNARY_OP_GELU_QUICK: - case GGML_UNARY_OP_RELU: - case GGML_UNARY_OP_TANH: - break; - default: - return; - } - break; - case GGML_OP_MUL_MAT: - case GGML_OP_MUL_MAT_ID: - if ( - x_sz > ctx->device->max_memory_allocation_size || - y_sz > ctx->device->max_memory_allocation_size || - d_sz > ctx->device->max_memory_allocation_size || - split_k_size > ctx->device->max_memory_allocation_size) { - GGML_ABORT("Requested preallocation size is too large"); - } - if (ctx->prealloc_size_x < x_sz) { - ctx->prealloc_size_x = x_sz; - } - if (ctx->prealloc_size_y < y_sz) { - ctx->prealloc_size_y = y_sz; - } - if (ctx->prealloc_size_split_k < split_k_size) { - ctx->prealloc_size_split_k = split_k_size; - } - if (ctx->staging_size < x_sz + y_sz) { - ctx->staging_size = x_sz + y_sz; - } - break; - default: - return; - } -} - static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { #if defined(GGML_VULKAN_RUN_TESTS) - ctx->staging = ggml_vk_create_buffer_check(ctx->device, 100ul * 1024ul * 1024ul, - vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached, - vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_IQ4_NL); - - ggml_vk_test_matmul<ggml_fp16_t, ggml_fp16_t>(ctx, 512, 512, 100, 32, 100, 1, 2); - - ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0); - ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1); - ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2); - // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 0); - // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 1); - // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 2); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q2_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q2_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q2_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q3_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q3_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q3_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q6_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q6_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q6_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_IQ4_NL); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_IQ4_NL); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_IQ4_NL); - - std::cerr << std::endl; - const std::vector<size_t> vals { + 512, 512, 128, + 128, 512, 512, + 4096, 512, 4096, + 11008, 512, 4096, + 4096, 512, 11008, + 32000, 512, 4096, 8, 8, 8, 100, 46, 576, 623, 111, 128, @@ -5660,25 +8233,70 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { 49, 49, 128, 128, 49, 49, 4096, 49, 4096, - 11008, 49, 4096, - 4096, 49, 11008, - 32000, 49, 4096, - 512, 512, 128, - 128, 512, 512, - 4096, 512, 4096, - 11008, 512, 4096, - 4096, 512, 11008, - 32000, 512, 4096, }; - const size_t num_it = 1; + const size_t num_it = 100; + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q4_0); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q4_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q4_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q4_0, true); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q8_0); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q8_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q8_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q8_0, true); + + abort(); + for (size_t i = 0; i < vals.size(); i += 3) { ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0); ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1); ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2); - // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0); - // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1); - // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2); - std::cerr << std::endl; + std::cerr << '\n'; + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0); + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1); + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2); + std::cerr << '\n'; + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0); + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1); + ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2); + std::cerr << '\n' << std::endl; + + if (vals[i + 2] % 32 == 0) { + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_0); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_0); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_0); + std::cerr << '\n' << std::endl; + } + + if (vals[i + 2] % 256 == 0) { + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_K); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_K); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_K); + std::cerr << '\n' << std::endl; + } } GGML_ABORT("fatal error"); @@ -5708,32 +8326,24 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { } ctx->prealloc_split_k = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_split_k); } - if (ctx->staging == nullptr || (ctx->staging_size > 0 && ctx->staging->size < ctx->staging_size)) { - VK_LOG_MEMORY("ggml_vk_preallocate_buffers(staging_size: " << ctx->staging_size << ")"); - // Resize buffer - if (ctx->staging != nullptr) { - ggml_vk_destroy_buffer(ctx->staging); - } - ctx->staging = ggml_vk_create_buffer_check(ctx->device, ctx->staging_size, - vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached, - vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); - } } -static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, bool last_node){ - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra; +static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready); - if (ggml_is_empty(node) || extra == nullptr) { - return; +// Returns true if node has enqueued work into the queue, false otherwise +// If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution. +static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool almost_ready, bool submit){ + if (ggml_is_empty(node) || !node->buffer) { + return false; } VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")"); ctx->semaphore_idx = 0; - ctx->staging_offset = 0; const ggml_tensor * src0 = node->src[0]; const ggml_tensor * src1 = node->src[1]; const ggml_tensor * src2 = node->src[2]; + const ggml_tensor * src3 = node->src[3]; switch (node->op) { // Return on empty ops to avoid generating a compute_ctx and setting exit_tensor @@ -5742,7 +8352,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NONE: - return; + return false; case GGML_OP_UNARY: switch (ggml_get_unary_op(node)) { case GGML_UNARY_OP_SILU: @@ -5750,14 +8360,17 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_SIGMOID: break; default: - return; + return false; } break; case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: case GGML_OP_GET_ROWS: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -5769,97 +8382,177 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_CONV_TRANSPOSE_1D: + case GGML_OP_POOL_2D: case GGML_OP_LEAKY_RELU: + case GGML_OP_FLASH_ATTN_EXT: break; default: std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl; GGML_ABORT("fatal error"); - return; + return false; } vk_context compute_ctx; - if (ctx->compute_ctx.expired()) { - compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); - ctx->compute_ctx = compute_ctx; - ggml_vk_ctx_begin(ctx->device, compute_ctx); + if (!dryrun) { + if (ctx->compute_ctx.expired()) { + compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); + ctx->compute_ctx = compute_ctx; + ggml_vk_ctx_begin(ctx->device, compute_ctx); + } else { + compute_ctx = ctx->compute_ctx.lock(); + } } else { - compute_ctx = ctx->compute_ctx.lock(); + switch (node->op) { + case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + case GGML_OP_ACC: + case GGML_OP_GET_ROWS: + case GGML_OP_ADD: + case GGML_OP_SUB: + case GGML_OP_MUL: + case GGML_OP_DIV: + case GGML_OP_CONCAT: + case GGML_OP_UPSCALE: + case GGML_OP_SCALE: + case GGML_OP_SQR: + case GGML_OP_CLAMP: + case GGML_OP_PAD: + case GGML_OP_CPY: + case GGML_OP_CONT: + case GGML_OP_DUP: + case GGML_OP_SILU_BACK: + case GGML_OP_NORM: + case GGML_OP_GROUP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_UNARY: + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: + case GGML_OP_ARGSORT: + case GGML_OP_SUM: + case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + case GGML_OP_IM2COL: + case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_CONV_TRANSPOSE_1D: + case GGML_OP_POOL_2D: + case GGML_OP_LEAKY_RELU: + { + // These operations all go through ggml_vk_op_f32, so short-circuit and + // do the only thing needed for the dryrun. + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op); + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + return false; + } + default: + break; + } } switch (node->op) { case GGML_OP_REPEAT: - ggml_vk_repeat(ctx, compute_ctx, src0, node); + ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_REPEAT_BACK: + ggml_vk_repeat_back(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_ACC: + ggml_vk_acc(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_GET_ROWS: - ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node); + ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_ADD: - ggml_vk_add(ctx, compute_ctx, src0, src1, node); + ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun); + + break; + case GGML_OP_SUB: + ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_MUL: - ggml_vk_mul(ctx, compute_ctx, src0, src1, node); + ggml_vk_mul(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_DIV: - ggml_vk_div(ctx, compute_ctx, src0, src1, node); + ggml_vk_div(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_CONCAT: - ggml_vk_concat(ctx, compute_ctx, src0, src1, node); + ggml_vk_concat(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_UPSCALE: - ggml_vk_upscale(ctx, compute_ctx, src0, node); + ggml_vk_upscale(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_SCALE: - ggml_vk_scale(ctx, compute_ctx, src0, node); + ggml_vk_scale(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_SQR: - ggml_vk_sqr(ctx, compute_ctx, src0, node); + ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_CLAMP: - ggml_vk_clamp(ctx, compute_ctx, src0, node); + ggml_vk_clamp(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_PAD: - ggml_vk_pad(ctx, compute_ctx, src0, node); + ggml_vk_pad(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: - ggml_vk_cpy(ctx, compute_ctx, src0, node); + ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_SILU_BACK: + ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_NORM: - ggml_vk_norm(ctx, compute_ctx, src0, node); + ggml_vk_norm(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_GROUP_NORM: - ggml_vk_group_norm(ctx, compute_ctx, src0, node); + ggml_vk_group_norm(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_RMS_NORM: - ggml_vk_rms_norm(ctx, compute_ctx, src0, node); + ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_RMS_NORM_BACK: + ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_UNARY: @@ -5869,77 +8562,132 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: - ggml_vk_unary(ctx, compute_ctx, src0, node); + case GGML_UNARY_OP_SIGMOID: + ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun); break; default: - return; + return false; } break; case GGML_OP_DIAG_MASK_INF: - ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node); + ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_SOFT_MAX: - ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node); + ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun); + + break; + case GGML_OP_SOFT_MAX_BACK: + ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_ROPE: - ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node); + ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun); + + break; + case GGML_OP_ROPE_BACK: + ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, true, dryrun); break; case GGML_OP_ARGSORT: - ggml_vk_argsort(ctx, compute_ctx, src0, node); + ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_SUM: + ggml_vk_sum(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_SUM_ROWS: - ggml_vk_sum_rows(ctx, compute_ctx, src0, node); + ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_ARGMAX: + ggml_vk_argmax(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_IM2COL: - ggml_vk_im2col(ctx, compute_ctx, src0, src1, node); + ggml_vk_im2col(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_TIMESTEP_EMBEDDING: - ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node); + ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_CONV_TRANSPOSE_1D: + ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun); + + break; + case GGML_OP_POOL_2D: + ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_LEAKY_RELU: - ggml_vk_leaky_relu(ctx, compute_ctx, src0, node); + ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_MUL_MAT: - ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node); + ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node, dryrun); break; case GGML_OP_MUL_MAT_ID: - ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node); + ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node, dryrun); + + break; + + case GGML_OP_FLASH_ATTN_EXT: + ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node, dryrun); break; default: - return; + return false; + } + + if (dryrun) { + return false; } ctx->tensor_ctxs[node_idx] = compute_ctx; -#ifdef GGML_VULKAN_CHECK_RESULTS +#if defined(GGML_VULKAN_CHECK_RESULTS) // Force context reset on each node so that each tensor ends up in its own context // and can be run and compared to its CPU equivalent separately last_node = true; #endif - if (last_node) { + if (submit || last_node) { ggml_vk_ctx_end(compute_ctx); - compute_ctx->exit_tensor_idx = node_idx; + + // TODO probably it'd be better to pass a exit_node flag to ggml_vk_compute_forward + if (last_node) { + compute_ctx->exit_tensor_idx = node_idx_begin; + } + else { + compute_ctx->exit_tensor_idx = -1; + } + ctx->compute_ctx.reset(); + + bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false, almost_ready); + if (!ok) { + if (node->op == GGML_OP_UNARY) { + std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl; + } + else { + std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl; + } + } + } + return true; } -static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx){ - ggml_tensor_extra_gpu * extra = nullptr; +static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) { + ggml_backend_buffer * buf = nullptr; switch (tensor->op) { case GGML_OP_ADD: case GGML_OP_GET_ROWS: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -5951,24 +8699,33 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NONE: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_CONV_TRANSPOSE_1D: + case GGML_OP_POOL_2D: case GGML_OP_LEAKY_RELU: case GGML_OP_REPEAT: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + case GGML_OP_REPEAT_BACK: + buf = tensor->buffer; break; case GGML_OP_UNARY: @@ -5978,7 +8735,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + case GGML_UNARY_OP_SIGMOID: + buf = tensor->buffer; break; default: return false; @@ -5986,39 +8744,55 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * break; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + case GGML_OP_FLASH_ATTN_EXT: + buf = tensor->buffer; break; default: return false; } - if (extra == nullptr) { + if (buf == nullptr) { return false; } VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")"); -#ifdef GGML_VULKAN_CHECK_RESULTS - ggml_vk_check_results_0(tensor); -#endif - vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock(); + // always wait for the GPU work to be done for the last submit + if (tensor_idx == subctx->exit_tensor_idx) { + use_fence = true; + } + // Only run if ctx hasn't been submitted yet if (!subctx->seqs.empty()) { +#ifdef GGML_VULKAN_CHECK_RESULTS + ggml_vk_check_results_0(tensor); + use_fence = true; +#endif + // Do staging buffer copies for (auto& cpy : subctx->in_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - ggml_vk_submit(subctx, ctx->fence); + if (almost_ready && !ctx->almost_ready_fence_pending && !use_fence) { + ggml_vk_submit(subctx, ctx->almost_ready_fence); + ctx->almost_ready_fence_pending = true; + } else { + ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{}); + } + + if (use_fence) { + ggml_vk_wait_for_fence(ctx); + } +#ifdef GGML_VULKAN_CHECK_RESULTS + ggml_vk_check_results_1(tensor); +#endif } if (tensor_idx == subctx->exit_tensor_idx) { - VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); - ctx->device->device.resetFences({ ctx->fence }); - // Do staging buffer copies for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -6038,17 +8812,8 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) { } ctx->gc.temp_buffers.clear(); - for (auto& pipeline : ctx->device->pipelines) { - if (pipeline.expired()) { - continue; - } - - vk_pipeline pl = pipeline.lock(); - ggml_pipeline_cleanup(pl); - } - - ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue); - ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue); + ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool); + ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool); for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) { ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s }); @@ -6067,10 +8832,10 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) { ctx->device->device.resetEvent(event); } - ctx->staging_offset = 0; - ctx->tensor_ctxs.clear(); ctx->gc.contexts.clear(); + ctx->pipeline_descriptor_set_requirements = 0; + ctx->descriptor_set_idx = 0; } // Clean up on backend free @@ -6081,7 +8846,6 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ggml_vk_destroy_buffer(ctx->prealloc_x); ggml_vk_destroy_buffer(ctx->prealloc_y); ggml_vk_destroy_buffer(ctx->prealloc_split_k); - ggml_vk_destroy_buffer(ctx->staging); for (auto& buffer : ctx->buffer_pool) { ggml_vk_destroy_buffer(buffer); @@ -6090,7 +8854,6 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ctx->prealloc_size_x = 0; ctx->prealloc_size_y = 0; ctx->prealloc_size_split_k = 0; - ctx->staging_size = 0; for (auto& event : ctx->gc.events) { ctx->device->device.destroyEvent(event); @@ -6098,6 +8861,16 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ctx->gc.events.clear(); ctx->device->device.destroyFence(ctx->fence); + ctx->device->device.destroyFence(ctx->almost_ready_fence); + + for (auto& pool : ctx->descriptor_pools) { + ctx->device->device.destroyDescriptorPool(pool); + } + ctx->descriptor_pools.clear(); + ctx->descriptor_sets.clear(); + + ctx->compute_cmd_pool.destroy(ctx->device->device); + ctx->transfer_cmd_pool.destroy(ctx->device->device); } GGML_CALL static int ggml_vk_get_device_count() { @@ -6123,42 +8896,6 @@ GGML_CALL static void ggml_vk_get_device_description(int device, char * descript // device backend -static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT - -struct ggml_backend_vk_buffer_context { - vk_device_ref device; - vk_buffer dev_buffer; - ggml_tensor_extra_gpu * temp_tensor_extras = nullptr; - size_t temp_tensor_extra_index = 0; - std::string name; - - ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) : - device(device), - dev_buffer(dev_buffer), - name(name) { - } - - ~ggml_backend_vk_buffer_context() { - ggml_vk_destroy_buffer(dev_buffer); - if (temp_tensor_extras != nullptr) { - delete[] temp_tensor_extras; - } - } - - ggml_tensor_extra_gpu * ggml_vk_alloc_temp_tensor_extra() { - if (temp_tensor_extras == nullptr) { - temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_VK_MAX_NODES]; - } - - size_t alloc_index = temp_tensor_extra_index; - temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_VK_MAX_NODES; - ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index]; - extra->reset(); - - return extra; - } -}; - GGML_CALL static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) { ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; return ctx->name.c_str(); @@ -6183,51 +8920,46 @@ GGML_CALL static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t bu GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")"); - ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; - if (tensor->view_src != nullptr) { GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); - GGML_ASSERT(tensor->view_src->extra != nullptr); - tensor->extra = tensor->view_src->extra; - } else { - ggml_tensor_extra_gpu * extra = ctx->ggml_vk_alloc_temp_tensor_extra(); - extra->buffer_gpu = ctx->dev_buffer; - extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base; - tensor->extra = extra; } } -GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - VK_LOG_DEBUG("ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; +static void ggml_backend_vk_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + VK_LOG_DEBUG("ggml_backend_vk_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", " << offset << ", " << size << ")"); + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; + vk_buffer buf = buf_ctx->dev_buffer; - vk_buffer buf = extra->buffer_gpu.lock(); + uint32_t val32 = (uint32_t)value * 0x01010101; + ggml_vk_buffer_memset(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, val32, size); +} - ggml_vk_buffer_write(buf, extra->offset + tensor->view_offs + offset, data, size); +static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + VK_LOG_DEBUG("ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; + vk_buffer buf = buf_ctx->dev_buffer; - GGML_UNUSED(buffer); + ggml_vk_buffer_write(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { VK_LOG_DEBUG("ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - - vk_buffer buf = extra->buffer_gpu.lock(); + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; - ggml_vk_buffer_read(buf, extra->offset + tensor->view_offs + offset, data, size); + vk_buffer buf = buf_ctx->dev_buffer; - GGML_UNUSED(buffer); + ggml_vk_buffer_read(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { if (ggml_backend_buffer_is_vk(src->buffer)) { - ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; - vk_buffer src_buf = src_extra->buffer_gpu.lock(); - vk_buffer dst_buf = dst_extra->buffer_gpu.lock(); + vk_buffer src_buf = src_buf_ctx->dev_buffer; + vk_buffer dst_buf = dst_buf_ctx->dev_buffer; - ggml_vk_buffer_copy(dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src)); + ggml_vk_buffer_copy(dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src)); return true; } @@ -6247,6 +8979,7 @@ static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = { /* .free_buffer = */ ggml_backend_vk_buffer_free_buffer, /* .get_base = */ ggml_backend_vk_buffer_get_base, /* .init_tensor = */ ggml_backend_vk_buffer_init_tensor, + /* .memset_tensor = */ ggml_backend_vk_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_vk_buffer_set_tensor, /* .get_tensor = */ ggml_backend_vk_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_vk_buffer_cpy_tensor, @@ -6284,7 +9017,7 @@ GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_b GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; - return ctx->device->max_memory_allocation_size; + return ctx->device->suballocation_block_size; } GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { @@ -6330,8 +9063,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_bu try { ptr = ggml_vk_host_malloc(vk_instance.devices[0], size); } catch (vk::SystemError& e) { - std::cerr << "ggml_vulkan: Failed to allocate pinned memory." << std::endl; - std::cerr << "ggml_vulkan: " << e.what() << std::endl; + GGML_LOG_WARN("ggml_vulkan: Failed to allocate pinned memory (%s)\n", e.what()); // fallback to cpu buffer return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); } @@ -6352,6 +9084,12 @@ GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_back UNUSED(buft); } +static size_t ggml_backend_vk_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { + return vk_instance.devices[0]->suballocation_block_size; + + UNUSED(buft); +} + // Should be changed to return device-specific host buffer type // but that probably requires changes in llama.cpp GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { @@ -6360,7 +9098,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { /* .get_name = */ ggml_backend_vk_host_buffer_type_name, /* .alloc_buffer = */ ggml_backend_vk_host_buffer_type_alloc_buffer, /* .get_alignment = */ ggml_backend_vk_host_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_max_size = */ ggml_backend_vk_host_buffer_type_get_max_size, /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, @@ -6404,22 +9142,22 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; vk_context transfer_ctx; if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool); ctx->transfer_ctx = transfer_ctx; ggml_vk_ctx_begin(ctx->device, transfer_ctx); } else { transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer buf = extra->buffer_gpu.lock(); + vk_buffer buf = buf_ctx->dev_buffer; - ggml_vk_buffer_write_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); + ggml_vk_buffer_write_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { @@ -6427,46 +9165,46 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; vk_context transfer_ctx; if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool); ctx->transfer_ctx = transfer_ctx; ggml_vk_ctx_begin(ctx->device, transfer_ctx); } else { transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer buf = extra->buffer_gpu.lock(); + vk_buffer buf = buf_ctx->dev_buffer; - ggml_vk_buffer_read_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); + ggml_vk_buffer_read_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) { VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async()"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; if ((dst->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || dst->buffer->buft == ggml_backend_vk_host_buffer_type()) && ggml_backend_buffer_is_vk(src->buffer)) { - ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; vk_context transfer_ctx; if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool); ctx->transfer_ctx = transfer_ctx; ggml_vk_ctx_begin(ctx->device, transfer_ctx); } else { transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer src_buf = src_extra->buffer_gpu.lock(); - vk_buffer dst_buf = dst_extra->buffer_gpu.lock(); + vk_buffer src_buf = src_buf_ctx->dev_buffer; + vk_buffer dst_buf = dst_buf_ctx->dev_buffer; - ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src)); + ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src)); return true; } @@ -6489,8 +9227,7 @@ GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) { } ggml_vk_submit(transfer_ctx, ctx->fence); - VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences"); - ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_wait_for_fence(ctx); for (auto& cpy : transfer_ctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -6507,10 +9244,25 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; + if (vk_instance.debug_utils_support) { + vk::DebugUtilsLabelEXT dul = {}; + dul.pLabelName = "ggml_backend_vk_graph_compute"; + dul.color = std::array<float,4>{1.0f, 1.0f, 1.0f, 1.0f}; + vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul)); + } + + uint64_t total_mat_mul_bytes = 0; for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_vk_preallocate_buffers_graph(ctx, cgraph->nodes[i]); + ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false, false); + if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) { + total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]); + } + } + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); } ggml_vk_preallocate_buffers(ctx); + ggml_pipeline_allocate_descriptor_sets(ctx); int last_node = cgraph->n_nodes - 1; @@ -6522,31 +9274,111 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen // Reserve tensor context space for all nodes ctx->tensor_ctxs.resize(cgraph->n_nodes); - for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_vk_build_graph(ctx, cgraph->nodes[i], i, i == last_node); - } + bool first_node_in_batch = true; // true if next node will be first node in a batch + int submit_node_idx = 0; // index to first node in a batch + + vk_context compute_ctx; + if (vk_perf_logger_enabled) { + // allocate/resize the query pool + if (ctx->device->num_queries < cgraph->n_nodes + 1) { + if (ctx->device->query_pool) { + ctx->device->device.destroyQueryPool(ctx->device->query_pool); + } + vk::QueryPoolCreateInfo query_create_info; + query_create_info.queryType = vk::QueryType::eTimestamp; + query_create_info.queryCount = cgraph->n_nodes + 100; + ctx->device->query_pool = ctx->device->device.createQueryPool(query_create_info); + ctx->device->num_queries = query_create_info.queryCount; + } + + ctx->device->device.resetQueryPool(ctx->device->query_pool, 0, cgraph->n_nodes+1); + GGML_ASSERT(ctx->compute_ctx.expired()); + compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); + ctx->compute_ctx = compute_ctx; + ggml_vk_ctx_begin(ctx->device, compute_ctx); + compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, 0); + } + + // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution. + // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB + // (and scaled down based on model size, so smaller models submit earlier). + // Also submit at least every 100 nodes, in case there are workloads without as much matmul. + int nodes_per_submit = 100; + int submitted_nodes = 0; + int submit_count = 0; + uint64_t mul_mat_bytes = 0; + uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u); for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_tensor * node = cgraph->nodes[i]; + if (first_node_in_batch) { + submit_node_idx = i; + } - if (ggml_vk_is_empty(node)) { - continue; + if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) { + mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]); } - bool ok = ggml_vk_compute_forward(ctx, node, i); - if (!ok) { - if (node->op == GGML_OP_UNARY) { - std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl; + // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining) + bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5; + bool submit = (submitted_nodes >= nodes_per_submit) || + (mul_mat_bytes >= mul_mat_bytes_per_submit) || + (i == last_node) || + (almost_ready && !ctx->almost_ready_fence_pending); + + bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, almost_ready, submit); + + if (vk_perf_logger_enabled) { + if (ctx->compute_ctx.expired()) { + compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool); + ctx->compute_ctx = compute_ctx; + ggml_vk_ctx_begin(ctx->device, compute_ctx); } else { - std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl; + compute_ctx = ctx->compute_ctx.lock(); } + compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, i+1); } -#ifdef GGML_VULKAN_CHECK_RESULTS - else { - ggml_vk_check_results_1(node); - } + + if (enqueued) { + ++submitted_nodes; + +#ifndef GGML_VULKAN_CHECK_RESULTS + if (first_node_in_batch) { + first_node_in_batch = false; + } #endif - GGML_ASSERT(ok); + } + + if (submit && enqueued) { + first_node_in_batch = true; + submitted_nodes = 0; + mul_mat_bytes = 0; + if (submit_count < 3) { + mul_mat_bytes_per_submit *= 2; + } + submit_count++; + } + } + + if (vk_perf_logger_enabled) { + // End the command buffer and submit/wait + GGML_ASSERT(!ctx->compute_ctx.expired()); + compute_ctx = ctx->compute_ctx.lock(); + ggml_vk_ctx_end(compute_ctx); + + ggml_vk_submit(compute_ctx, ctx->device->fence); + VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences"); + ctx->device->device.resetFences({ ctx->device->fence }); + + // Get the results and pass them to the logger + std::vector<uint64_t> timestamps(cgraph->n_nodes + 1); + VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results"); + for (int i = 0; i < cgraph->n_nodes; i++) { + if (!ggml_vk_is_empty(cgraph->nodes[i])) { + ctx->device->perf_logger->log_timing(cgraph->nodes[i], uint64_t((timestamps[i+1] - timestamps[i]) * ctx->device->properties.limits.timestampPeriod)); + } + } + + ctx->device->perf_logger->print_timings(); } ggml_vk_graph_cleanup(ctx); @@ -6567,7 +9399,11 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: - return ggml_is_contiguous(op->src[0]); + case GGML_UNARY_OP_SIGMOID: + return ggml_is_contiguous(op->src[0]) && + (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && + (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && + (op->src[0]->type == op->type); default: return false; } @@ -6575,9 +9411,17 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: { - switch (op->src[0]->type) { + ggml_type src0_type = op->src[0]->type; + ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; + const vk_device& device = ctx->device; + if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) { + // If there's not enough shared memory for row_ids and the result tile, fallback to CPU + return false; + } + switch (src0_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -6588,6 +9432,14 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: break; default: @@ -6605,73 +9457,208 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const if (a->ne[3] != b->ne[3]) { return false; } + if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) || + !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) { + return false; + } + if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) { + // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader. + // So don't support this combination for now. + return false; + } + return true; } break; + case GGML_OP_FLASH_ATTN_EXT: + { + ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; + bool coopmat2 = ctx->device->coopmat2; + switch (op->src[0]->ne[0]) { + case 64: + case 80: + case 96: + case 112: + case 128: + case 256: + break; + default: + return false; + } + if (op->src[0]->type != GGML_TYPE_F32) { + return false; + } + if (op->type != GGML_TYPE_F32) { + return false; + } + if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) { + return false; + } + // It's straightforward to support different K/V dequant, but would + // significantly increase the number of pipelines + if (op->src[1]->type != op->src[2]->type) { + return false; + } + switch (op->src[1]->type) { + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q8_0: + // supported in scalar and coopmat2 paths + break; + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently + //case GGML_TYPE_Q2_K: + //case GGML_TYPE_Q3_K: + //case GGML_TYPE_Q4_K: + //case GGML_TYPE_Q5_K: + //case GGML_TYPE_Q6_K: + //case GGML_TYPE_IQ1_S: + //case GGML_TYPE_IQ1_M: + //case GGML_TYPE_IQ2_XXS: + //case GGML_TYPE_IQ2_XS: + //case GGML_TYPE_IQ2_S: + //case GGML_TYPE_IQ3_XXS: + //case GGML_TYPE_IQ3_S: + //case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ4_NL: + // currently supported only in coopmat2 path + if (!coopmat2) { + return false; + } + break; + default: + return false; + } + if (!coopmat2 && !ctx->device->subgroup_shuffle) { + // scalar FA uses subgroupShuffle + return false; + } + return true; + } case GGML_OP_GET_ROWS: { switch (op->src[0]->type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: return true; default: return false; } } break; + case GGML_OP_CONT: case GGML_OP_CPY: case GGML_OP_DUP: { ggml_type src0_type = op->src[0]->type; ggml_type src1_type = op->src[1] != nullptr ? op->src[1]->type : src0_type; - if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { - return true; + + if (src0_type == GGML_TYPE_F32) { + switch (src1_type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ4_NL: + return true; + default: + break; + } } - if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { - return true; + if (src1_type == GGML_TYPE_F32) { + switch (src0_type) { + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ4_NL: + return true; + default: + break; + } } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { return true; } + + // We can handle copying from a type to the same type if it's + // contiguous (memcpy). We use f16 or f32 shaders to do the copy, + // so the type/block size must be a multiple of 4. + if (src0_type == src1_type && + ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op) && + (ggml_type_size(src0_type) % 2) == 0) { + return true; + } return false; } break; case GGML_OP_REPEAT: - { - ggml_type src0_type = op->src[0]->type; - return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; - } break; + return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float); + case GGML_OP_REPEAT_BACK: + return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32; case GGML_OP_ROPE: - return ggml_is_contiguous(op->src[0]); + case GGML_OP_ROPE_BACK: case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: + case GGML_OP_RMS_NORM: + return true; case GGML_OP_NORM: case GGML_OP_GROUP_NORM: - case GGML_OP_RMS_NORM: + return ggml_is_contiguous(op->src[0]); case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: + return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && + (op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16) && + (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16); + case GGML_OP_SILU_BACK: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_SQR: + case GGML_OP_CLAMP: + return op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_ACC: case GGML_OP_CONCAT: case GGML_OP_UPSCALE: case GGML_OP_SCALE: - case GGML_OP_SQR: - case GGML_OP_CLAMP: case GGML_OP_PAD: - case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: return true; + case GGML_OP_CONV_TRANSPOSE_1D: + return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32; default: return false; } @@ -6828,6 +9815,39 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve UNUSED(instance_extensions); } +// Extension availability +static bool ggml_vk_instance_debug_utils_ext_available( + const std::vector<vk::ExtensionProperties> & instance_extensions) { + // Check for portability enumeration extension for MoltenVK support + for (const auto & properties : instance_extensions) { + if (strcmp("VK_EXT_debug_utils", properties.extensionName) == 0) { + return true; + } + } + + std::cerr << "ggml_vulkan: WARNING: Instance extension VK_EXT_debug_utils not found." << std::endl; + return false; + + UNUSED(instance_extensions); +} + +static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) { + switch (props.vendorID) { + case VK_VENDOR_ID_INTEL: + // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost, + // while some older hardware (ex. Arc A770) has performance regressions + return arch == vk_device_architecture::INTEL_XE2; + case VK_VENDOR_ID_AMD: + if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) { + // Workaround for AMD proprietary driver reporting support on all GPUs + return arch == vk_device_architecture::AMD_RDNA3; + } + return true; + default: + return true; + } +} + // checks #ifdef GGML_VULKAN_CHECK_RESULTS @@ -6894,10 +9914,10 @@ static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) const size_t tensor_size = ggml_nbytes(tensor); tensor_data = malloc(tensor_size); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - ggml_vk_buffer_read(buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size); + vk_buffer buffer_gpu = buf_ctx->dev_buffer; + ggml_vk_buffer_read(buffer_gpu, vk_tensor_offset(tensor) + tensor->view_offs, tensor_data, tensor_size); } std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl; @@ -6937,7 +9957,6 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { ggml_tensor * src0 = tensor->src[0]; ggml_tensor * src1 = tensor->src[1]; - ggml_tensor * src2 = tensor->src[2]; struct ggml_init_params iparams = { /*.mem_size =*/ 2ul*1024ul*1024ul*1024ul, @@ -6947,183 +9966,121 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { struct ggml_context * ggml_ctx = ggml_init(iparams); - struct ggml_tensor * src0_clone = nullptr; - struct ggml_tensor * src1_clone = nullptr; - struct ggml_tensor * src2_clone = nullptr; - struct ggml_tensor * tensor_clone = nullptr; - - size_t src0_size; - size_t src1_size; - size_t src2_size; - - void * src0_buffer = nullptr; - void * src1_buffer = nullptr; - void * src2_buffer = nullptr; - - if (src0 != nullptr) { - src0_clone = ggml_dup_tensor(ggml_ctx, src0); - - src0_size = ggml_nbytes(src0); - - src0_buffer = malloc(src0_size); - src0_clone->data = src0_buffer; - if (ggml_backend_buffer_is_host(src0->buffer)) { - memcpy(src0_clone->data, src0->data, src0_size); - memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src0->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src0->view_offs; - if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) { - for (int i3 = 0; i3 < src0->ne[3]; i3++) { - for (int i2 = 0; i2 < src0->ne[2]; i2++) { - const int idx = i3*src0->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]); - } - } - - src0_clone->nb[0] = src0->nb[0]; - src0_clone->nb[1] = src0->nb[1]; - for (int i = 2; i < GGML_MAX_DIMS; i++) { - src0_clone->nb[i] = src0_clone->nb[i - 1]*src0_clone->ne[i - 1]; - } - } else { - if (offset + src0_size >= buffer_gpu->size) { - src0_size = buffer_gpu->size - offset; - } - ggml_vk_buffer_read(buffer_gpu, offset, src0_clone->data, src0_size); - memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); - } - } else { - GGML_ABORT("fatal error"); - } + std::array<struct ggml_tensor *, 6> src_clone = {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr}; + std::array<size_t, 6> src_size = {0, 0, 0, 0, 0, 0}; + std::array<void *, 6> src_buffer = {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr}; + const char * srci_name[6] = {"src0", "src1", "src2", "src3", "src4", "src5"}; - if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src0, "src0"); - } - } - if (src1 != nullptr) { - src1_clone = ggml_dup_tensor(ggml_ctx, src1); - - src1_size = ggml_nbytes(src1); - - src1_buffer = malloc(src1_size); - src1_clone->data = src1_buffer; - if (ggml_backend_buffer_is_host(src1->buffer)) { - memcpy(src1_clone->data, src1->data, src1_size); - memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src1->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src1->view_offs; - if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) { - for (int i3 = 0; i3 < src1->ne[3]; i3++) { - for (int i2 = 0; i2 < src1->ne[2]; i2++) { - const int idx = i3*src1->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]); - } - } - - src1_clone->nb[0] = src1->nb[0]; - src1_clone->nb[1] = src1->nb[1]; - for (int i = 2; i < GGML_MAX_DIMS; i++) { - src1_clone->nb[i] = src1_clone->nb[i - 1]*src1_clone->ne[i - 1]; - } - } else { - if (offset + src1_size >= buffer_gpu->size) { - src1_size = buffer_gpu->size - offset; - } - ggml_vk_buffer_read(buffer_gpu, offset, src1_clone->data, src1_size); - memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); - } - } else { - GGML_ABORT("fatal error"); - } + struct ggml_tensor * tensor_clone = nullptr; - if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src1, "src1"); + for (int i = 0; i < 6; i++) { + ggml_tensor * srci = tensor->src[i]; + if (srci == nullptr) { + continue; } - } - if (src2 != nullptr) { - src2_clone = ggml_dup_tensor(ggml_ctx, src2); - - src2_size = ggml_nbytes(src2); - - src2_buffer = malloc(src2_size); - src2_clone->data = src2_buffer; - if (ggml_backend_buffer_is_host(src2->buffer)) { - memcpy(src2_clone->data, src2->data, src2_size); - memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src2->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src2->view_offs; - if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) { - for (int i3 = 0; i3 < src2->ne[3]; i3++) { - for (int i2 = 0; i2 < src2->ne[2]; i2++) { - const int idx = i3*src2->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]); + ggml_tensor * srci_clone = ggml_dup_tensor(ggml_ctx, srci); + size_t srci_size = ggml_nbytes(srci); + + src_clone[i] = srci_clone; + src_size[i] = ggml_nbytes(srci); + src_buffer[i] = malloc(srci_size); + + srci_clone->data = src_buffer[i]; + if (ggml_backend_buffer_is_host(srci->buffer)) { + memcpy(srci_clone->data, srci->data, srci_size); + memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS); + } else if (ggml_backend_buffer_is_vk(srci->buffer)) { + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)srci->buffer->context; + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(srci) + srci->view_offs; + if (!ggml_is_contiguous(srci) && ggml_vk_dim01_contiguous(srci)) { + for (int i3 = 0; i3 < srci->ne[3]; i3++) { + for (int i2 = 0; i2 < srci->ne[2]; i2++) { + const int idx = i3*srci->ne[2] + i2; + ggml_vk_buffer_read(buffer_gpu, offset + idx * srci->nb[2], ((char *)srci_clone->data + idx * srci_clone->nb[2]), srci->ne[1] * srci->nb[1]); } } - src2_clone->nb[0] = src2->nb[0]; - src2_clone->nb[1] = src2->nb[1]; + srci_clone->nb[0] = srci->nb[0]; + srci_clone->nb[1] = srci->nb[1]; for (int i = 2; i < GGML_MAX_DIMS; i++) { - src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1]; + srci_clone->nb[i] = srci_clone->nb[i - 1]*srci_clone->ne[i - 1]; } } else { - if (offset + src2_size >= buffer_gpu->size) { - src2_size = buffer_gpu->size - offset; + if (offset + srci_size >= buffer_gpu->size) { + srci_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(buffer_gpu, offset, src2_clone->data, src2_size); - memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); + ggml_vk_buffer_read(buffer_gpu, offset, srci_clone->data, srci_size); + memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { GGML_ABORT("fatal error"); } if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src2, "src2"); + ggml_vk_print_tensor(srci, srci_name[i]); } } - if (tensor->op == GGML_OP_MUL_MAT) { - tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone); + if (tensor->op == GGML_OP_FLASH_ATTN_EXT) { + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_flash_attn_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], params[0], params[1], params[2]); + } else if (tensor->op == GGML_OP_MUL_MAT) { + tensor_clone = ggml_mul_mat(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_MUL_MAT_ID) { - tensor_clone = ggml_mul_mat_id(ggml_ctx, src0_clone, src1_clone, src2_clone); + tensor_clone = ggml_mul_mat_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]); + } else if (tensor->op == GGML_OP_SUB) { + tensor_clone = ggml_sub(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_MUL) { - tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_DIV) { - tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_div(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_CONCAT) { - tensor_clone = ggml_concat(ggml_ctx, src0_clone, src1_clone, *(int *)tensor->op_params); + tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params); } else if (tensor->op == GGML_OP_UPSCALE) { - tensor_clone = ggml_upscale_ext(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); } else if (tensor->op == GGML_OP_SCALE) { - tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]); } else if (tensor->op == GGML_OP_SQR) { - tensor_clone = ggml_sqr(ggml_ctx, src0_clone); - } else if (tensor->op == GGML_OP_CLAMP) { - tensor_clone = ggml_clamp(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); + tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_CLAMP) { + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_clamp(ggml_ctx, src_clone[0], params[0], params[1]); } else if (tensor->op == GGML_OP_PAD) { - tensor_clone = ggml_pad(ggml_ctx, src0_clone, tensor->ne[0] - src0_clone->ne[0], tensor->ne[1] - src0_clone->ne[1], tensor->ne[2] - src0_clone->ne[2], tensor->ne[3] - src0_clone->ne[3]); + tensor_clone = ggml_pad(ggml_ctx, src_clone[0], tensor->ne[0] - src_clone[0]->ne[0], tensor->ne[1] - src_clone[0]->ne[1], tensor->ne[2] - src_clone[0]->ne[2], tensor->ne[3] - src_clone[0]->ne[3]); + } else if (tensor->op == GGML_OP_REPEAT) { + tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor); + } else if (tensor->op == GGML_OP_REPEAT_BACK) { + tensor_clone = ggml_repeat_back(ggml_ctx, src_clone[0], tensor); } else if (tensor->op == GGML_OP_ADD) { - tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]); + } else if (tensor->op == GGML_OP_ACC) { + tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]); } else if (tensor->op == GGML_OP_NORM) { - tensor_clone = ggml_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); + tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params); } else if (tensor->op == GGML_OP_GROUP_NORM) { - tensor_clone = ggml_group_norm(ggml_ctx, src0_clone, *(int *)tensor->op_params); + const float * float_params = (const float *)tensor->op_params; + tensor_clone = ggml_group_norm(ggml_ctx, src_clone[0], tensor->op_params[0], float_params[1]); } else if (tensor->op == GGML_OP_RMS_NORM) { - tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); + tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params); + } else if (tensor->op == GGML_OP_RMS_NORM_BACK) { + const float eps = ((float *) tensor->op_params)[0]; + tensor_clone = ggml_rms_norm_back(ggml_ctx, src_clone[0], src_clone[1], eps); + } else if (tensor->op == GGML_OP_SILU_BACK) { + tensor_clone = ggml_silu_back(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_SOFT_MAX) { if (src1 != nullptr) { - tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_soft_max_ext(ggml_ctx, src_clone[0], src_clone[1], params[0], params[1]); } else { - tensor_clone = ggml_soft_max(ggml_ctx, src0_clone); + tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]); } + } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) { + tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); } else if (tensor->op == GGML_OP_DIAG_MASK_INF) { - tensor_clone = ggml_diag_mask_inf(ggml_ctx, src0_clone, *(int *)tensor->op_params); - } else if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], tensor->op_params[0]); + } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) { const int n_dims = ((int32_t *) tensor->op_params)[1]; const int mode = ((int32_t *) tensor->op_params)[2]; //const int n_ctx_ggml = ((int32_t *) tensor->op_params)[3]; @@ -7134,23 +10091,39 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const float attn_factor = ((float *) tensor->op_params)[8]; const float beta_fast = ((float *) tensor->op_params)[9]; const float beta_slow = ((float *) tensor->op_params)[10]; - tensor_clone = ggml_rope_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + if (mode & GGML_ROPE_TYPE_MROPE) { + int32_t *sections = ((int32_t *) tensor->op_params) + 11; + if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } else { + tensor_clone = ggml_rope_multi_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } + } else { + if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } else { + tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } + } } else if (tensor->op == GGML_OP_UNARY) { switch (ggml_get_unary_op(tensor)) { case GGML_UNARY_OP_SILU: - tensor_clone = ggml_silu(ggml_ctx, src0_clone); + tensor_clone = ggml_silu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_GELU: - tensor_clone = ggml_gelu(ggml_ctx, src0_clone); + tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_GELU_QUICK: - tensor_clone = ggml_gelu_quick(ggml_ctx, src0_clone); + tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_RELU: - tensor_clone = ggml_relu(ggml_ctx, src0_clone); + tensor_clone = ggml_relu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_TANH: - tensor_clone = ggml_tanh(ggml_ctx, src0_clone); + tensor_clone = ggml_tanh(ggml_ctx, src_clone[0]); + break; + case GGML_UNARY_OP_SIGMOID: + tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]); break; default: std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; @@ -7158,29 +10131,33 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { } } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) { if (src1 == nullptr) { - tensor_clone = ggml_dup(ggml_ctx, src0_clone); + tensor_clone = ggml_dup(ggml_ctx, src_clone[0]); tensor_clone->type = tensor->type; } else { - tensor_clone = ggml_cpy(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]); } } else if (tensor->op == GGML_OP_CONT) { - tensor_clone = ggml_cont_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); } else if (tensor->op == GGML_OP_RESHAPE) { - tensor_clone = ggml_reshape_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_reshape_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); } else if (tensor->op == GGML_OP_VIEW) { - tensor_clone = ggml_view_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]); + tensor_clone = ggml_view_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]); } else if (tensor->op == GGML_OP_PERMUTE) { int32_t * params = (int32_t *)tensor->op_params; - tensor_clone = ggml_permute(ggml_ctx, src0_clone, params[0], params[1], params[2], params[3]); + tensor_clone = ggml_permute(ggml_ctx, src_clone[0], params[0], params[1], params[2], params[3]); } else if (tensor->op == GGML_OP_TRANSPOSE) { - tensor_clone = ggml_transpose(ggml_ctx, src0_clone); + tensor_clone = ggml_transpose(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_GET_ROWS) { - tensor_clone = ggml_get_rows(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_get_rows(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_ARGSORT) { - tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params); + tensor_clone = ggml_argsort(ggml_ctx, src_clone[0], (ggml_sort_order) *(int *)tensor->op_params); + } else if (tensor->op == GGML_OP_SUM) { + tensor_clone = ggml_sum(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_SUM_ROWS) { - tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone); - } else if (tensor->op == GGML_OP_IM2COL) { + tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_ARGMAX) { + tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_IM2COL) { const int32_t s0 = tensor->op_params[0]; const int32_t s1 = tensor->op_params[1]; const int32_t p0 = tensor->op_params[2]; @@ -7189,15 +10166,31 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const int32_t d1 = tensor->op_params[5]; const bool is_2D = tensor->op_params[6] == 1; - tensor_clone = ggml_im2col(ggml_ctx, src0_clone, src1_clone, s0, s1, p0, p1, d0, d1, is_2D, tensor->type); + tensor_clone = ggml_im2col(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1, is_2D, tensor->type); } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) { const int32_t dim = tensor->op_params[0]; const int32_t max_period = tensor->op_params[1]; - tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period); + tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period); + } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){ + const int32_t s0 = tensor->op_params[0]; + const int32_t p0 = tensor->op_params[1]; + const int32_t d0 = tensor->op_params[2]; + tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0); + } else if (tensor->op == GGML_OP_POOL_2D) { + enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]); + const int32_t k0 = tensor->op_params[1]; + const int32_t k1 = tensor->op_params[2]; + const int32_t s0 = tensor->op_params[3]; + const int32_t s1 = tensor->op_params[4]; + const int32_t p0 = tensor->op_params[5]; + const int32_t p1 = tensor->op_params[6]; + + tensor_clone = ggml_pool_2d(ggml_ctx, src_clone[0], op, k0, k1, s0, s1, p0, p1); } else if (tensor->op == GGML_OP_LEAKY_RELU) { const float * op_params = (const float *)tensor->op_params; - tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false); - } else { + tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false); + } + else { std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; GGML_ABORT("fatal error"); } @@ -7217,11 +10210,10 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { memcpy(comp_result, tensor_clone->data, comp_size); memcpy(comp_nb, tensor_clone->nb, sizeof(size_t) * GGML_MAX_DIMS); - if (src0 != nullptr) { - free(src0_buffer); - } - if (src1 != nullptr) { - free(src1_buffer); + for (int i = 0; i < 6; i++) { + if (src_buffer[i] != nullptr) { + free(src_buffer[i]); + } } ggml_free(ggml_ctx); @@ -7242,6 +10234,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { ggml_tensor * src0 = tensor->src[0]; ggml_tensor * src1 = tensor->src[1]; ggml_tensor * src2 = tensor->src[2]; + ggml_tensor * src3 = tensor->src[3]; void * tensor_data = tensor->data; @@ -7249,14 +10242,15 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { size_t tensor_size = ggml_nbytes(tensor); tensor_data = malloc(tensor_size); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - if (extra->offset + tensor->view_offs + tensor_size >= buffer_gpu->size) { - tensor_size = buffer_gpu->size - (extra->offset + tensor->view_offs); + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(tensor) + tensor->view_offs; + if (offset + tensor_size >= buffer_gpu->size) { + tensor_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size); + ggml_vk_buffer_read(buffer_gpu, offset, tensor_data, tensor_size); } float first_error_result = -1.0f; @@ -7283,6 +10277,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { } else if (tensor->type == GGML_TYPE_I32) { correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]); result = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]); + } else if (tensor->type == GGML_TYPE_I64) { + correct = *(int64_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]); + result = *(int64_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]); } else { std::cerr << "Results check not implemented for type " << ggml_type_name(tensor->type) << std::endl; } @@ -7303,6 +10300,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { if (src2 != nullptr) { std::cerr << "src2=" << src2 << " src2->name=" << src2->name << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl; } + if (src3 != nullptr) { + std::cerr << "src3=" << src3 << " src3->name=" << src3->name << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl; + } std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl; std::cerr << std::endl << "Result:" << std::endl; ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3); @@ -7313,7 +10313,8 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { ggml_vk_print_graph_origin(tensor, done); GGML_ABORT("fatal error"); } - if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) { + const double denom = std::fabs(correct) > 1.0f ? (std::fabs(correct) > 1e-8 ? std::fabs(correct) : 1e-8) : 1.0f; + if (first_error[0] == -1 && std::fabs(correct - result) / denom > 0.5) { first_error[0] = i0; first_error[1] = i1; first_error[2] = i2; @@ -7325,7 +10326,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { // Special case, value is infinite, avoid NaN result in avg_err // NaN also appears in results, if both are nan error is 0 if (!std::isinf(correct) && !std::isinf(result) && !std::isnan(correct) && !std::isnan(result)) { - avg_err += std::fabs(correct - result); + avg_err += std::fabs(correct - result) / denom; } counter++; } @@ -7347,6 +10348,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { if (src2 != nullptr) { std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl; } + if (src3 != nullptr) { + std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl; + } std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl; std::cerr << std::endl << "Result:" << std::endl; ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0); @@ -7357,7 +10361,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { ggml_vk_print_graph_origin(tensor, done); } - if (avg_err > 0.05 || std::isnan(avg_err)) { + if (avg_err > 0.5 || std::isnan(avg_err)) { std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl; std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl; if (src0 != nullptr) { @@ -7369,6 +10373,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { if (src2 != nullptr) { std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl; } + if (src3 != nullptr) { + std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl; + } std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl; std::cerr << std::endl << "Result:" << std::endl; ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index d06f6553..97b5bff7 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -6310,6 +6310,7 @@ struct ggml_tensor * ggml_argmax( return result; } + // ggml_repeat struct ggml_tensor * ggml_repeat( diff --git a/ggml/src/vulkan-shaders/CMakeLists.txt b/ggml/src/vulkan-shaders/CMakeLists.txt index 10075db3..a22ea817 100644 --- a/ggml/src/vulkan-shaders/CMakeLists.txt +++ b/ggml/src/vulkan-shaders/CMakeLists.txt @@ -1,5 +1,29 @@ +cmake_minimum_required(VERSION 3.19) +project("vulkan-shaders-gen" C CXX) + find_package (Threads REQUIRED) +if (GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + message(STATUS "Enabling coopmat glslc support") +endif() +if (GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + message(STATUS "Enabling coopmat2 glslc support") +endif() +if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + message(STATUS "Enabling dot glslc support") +endif() +if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + message(STATUS "Enabling bfloat16 glslc support") +endif() +if (GGML_VULKAN_SHADER_DEBUG_INFO) + add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO) + message(STATUS "Enabling shader debug info") +endif() + set(TARGET vulkan-shaders-gen) add_executable(${TARGET} vulkan-shaders-gen.cpp) install(TARGETS ${TARGET} RUNTIME) diff --git a/ggml/src/vulkan-shaders/acc.comp b/ggml/src/vulkan-shaders/acc.comp new file mode 100644 index 00000000..d896f1ef --- /dev/null +++ b/ggml/src/vulkan-shaders/acc.comp @@ -0,0 +1,29 @@ +#version 450 + +#include "types.comp" +#include "generic_binary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +void main() { + const uint idx = gl_GlobalInvocationID.x; + if (idx >= p.ne) { + return; + } + + const uint offset = p.param3; + const uint src1_i = idx - offset; + const uint oz = src1_i / p.nb02; + const uint oy = (src1_i - (oz * p.nb02)) / p.nb01; + const uint ox = src1_i % p.nb01; + + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) { + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + ox + oy * p.ne10 + oz * p.ne10 * p.ne11])); + } else { + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)])); + } +} + diff --git a/ggml/src/vulkan-shaders/add.comp b/ggml/src/vulkan-shaders/add.comp index 3974845d..2b4085c4 100644 --- a/ggml/src/vulkan-shaders/add.comp +++ b/ggml/src/vulkan-shaders/add.comp @@ -1,14 +1,29 @@ #version 450 +#extension GL_EXT_shader_16bit_storage : require + #include "types.comp" #include "generic_binary_head.comp" +const uint num_threads = 256; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + void main() { - const uint idx = get_idx(); + uint idx = get_idx(); - if (idx >= p.ne) { - return; - } + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 2; - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)])); + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)])); + + idx += num_threads; + } } diff --git a/ggml/src/vulkan-shaders/argmax.comp b/ggml/src/vulkan-shaders/argmax.comp new file mode 100644 index 00000000..eaf4da34 --- /dev/null +++ b/ggml/src/vulkan-shaders/argmax.comp @@ -0,0 +1,51 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +layout (constant_id = 0) const uint BLOCK_SIZE = 32; + +shared FLOAT_TYPE tmpmax[BLOCK_SIZE]; +shared uint tmp[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint col = gl_LocalInvocationID.x; + + if (col >= p.KX) { + return; + } + A_TYPE amax = data_a[row*p.KX + col]; + tmp[col] = col; + + for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) { + A_TYPE val = data_a[row*p.KX + i]; + if (val > amax) { + amax = val; + tmp[col] = i; + } + } + tmpmax[col] = amax; + + barrier(); + [[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) { + if (col < s && col + s < p.KX) { + if (tmpmax[col] < tmpmax[col + s]) { + tmpmax[col] = tmpmax[col + s]; + tmp[col] = tmp[col + s]; + } + } + barrier(); + } + + if (col == 0) { + data_d[row] = D_TYPE(tmp[0]); + } +} diff --git a/ggml/src/vulkan-shaders/argsort.comp b/ggml/src/vulkan-shaders/argsort.comp index e55414b0..d4fa45b1 100644 --- a/ggml/src/vulkan-shaders/argsort.comp +++ b/ggml/src/vulkan-shaders/argsort.comp @@ -29,20 +29,18 @@ void main() { const int col = int(gl_LocalInvocationID.x); const uint row = gl_WorkGroupID.y; - if (col >= p.ncols_pad) { - return; - } - const uint row_offset = row * p.ncols; // initialize indices - dst_row[col] = col; + if (col < p.ncols_pad) { + dst_row[col] = col; + } barrier(); for (uint k = 2; k <= p.ncols_pad; k *= 2) { for (uint j = k / 2; j > 0; j /= 2) { const uint ixj = col ^ j; - if (ixj > col) { + if (col < p.ncols_pad && ixj > col) { if ((col & k) == 0) { if (dst_row[col] >= p.ncols || (dst_row[ixj] < p.ncols && (p.order == ASC ? diff --git a/ggml/src/vulkan-shaders/clamp.comp b/ggml/src/vulkan-shaders/clamp.comp index 7071302a..1e5cb8da 100644 --- a/ggml/src/vulkan-shaders/clamp.comp +++ b/ggml/src/vulkan-shaders/clamp.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_unary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint idx = get_idx(); @@ -10,6 +12,6 @@ void main() { return; } - const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]); - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val)); + const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val)); } diff --git a/ggml/src/vulkan-shaders/concat.comp b/ggml/src/vulkan-shaders/concat.comp index 08ab5514..9ee2f1fa 100644 --- a/ggml/src/vulkan-shaders/concat.comp +++ b/ggml/src/vulkan-shaders/concat.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_binary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; const int dim = p.param3; @@ -28,8 +30,12 @@ void main() { const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03; #ifndef OPTIMIZATION_ERROR_WORKAROUND - data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]); + data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : data_b[get_boffset() + src1_idx]); #else - data_d[p.d_offset + dst_idx] = is_src0 ? data_a[src0_idx] : data_b[src1_idx]; + if (is_src0) { + data_d[get_doffset() + dst_idx] = data_a[get_aoffset() + src0_idx]; + } else { + data_d[get_doffset() + dst_idx] = data_b[get_boffset() + src1_idx]; + } #endif } diff --git a/ggml/src/vulkan-shaders/contig_copy.comp b/ggml/src/vulkan-shaders/contig_copy.comp new file mode 100644 index 00000000..6567a8c5 --- /dev/null +++ b/ggml/src/vulkan-shaders/contig_copy.comp @@ -0,0 +1,49 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +#extension GL_EXT_control_flow_attributes : require + +const uint num_threads = 128; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + +void main() { + uint idx = get_idx(); + + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 4; + + // fast path for when all four iterations are in-bounds + if (idx + (num_iter-1)*num_threads < p.ne) { + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) + data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); +#else + data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; +#endif + idx += num_threads; + } + } else { + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) + data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); +#else + data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; +#endif + idx += num_threads; + } + } +} diff --git a/ggml/src/vulkan-shaders/conv_transpose_1d.comp b/ggml/src/vulkan-shaders/conv_transpose_1d.comp new file mode 100644 index 00000000..b17b4e83 --- /dev/null +++ b/ggml/src/vulkan-shaders/conv_transpose_1d.comp @@ -0,0 +1,98 @@ +#version 450 + +#include "types.comp" + +layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; // src0 - kernel: [K, Cout, Cin] +layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; // src1 - input: [L, Cin] +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; // dst - result [KL, Cout] + +layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in; + +layout (push_constant) uniform parameter { + uint32_t Cout; + uint32_t Cin; + uint32_t K; + uint32_t L; + uint32_t KL; + + uint32_t nb01; + uint32_t nb02; + uint32_t nb11; + uint32_t nb1; + + int32_t s0; +} p; + + +uint32_t Cout_idx = gl_WorkGroupID.x; +const uint32_t bs = gl_WorkGroupSize.x; +uint32_t tid = gl_LocalInvocationID.x; +// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K. +uint32_t tmp_len = bs*p.s0+p.K; +shared D_TYPE tmp[4096]; + +uint splitWork(uint workSize){ + return (bs + workSize -1) / bs; +} + +void main(){ + for(uint32_t i = 0; i < splitWork(tmp_len); i++){ + uint32_t idx = i*bs+tid; + if(idx < tmp_len){ + tmp[idx] = 0.0; + } + } + + uint32_t L_blocks = splitWork(p.L); + for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){ + if(L_block_id > 0){ + barrier(); + // Shift values in tmp to the current processing window + for(int i = 0; i < splitWork(tmp_len); i++){ + uint32_t idx = i*bs+tid; + if(idx >= bs*p.s0 && idx < tmp_len){ + tmp[idx-bs*p.s0] = tmp[idx]; + tmp[idx] = 0.0; + }else if(idx >= p.K && idx < bs*p.s0){ + tmp[idx] = 0.0; + } + } + } + barrier(); + + // Save contributions of the block to tmp + uint32_t L_idx = L_block_id*bs + tid; + for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){ + D_TYPE dp = 0.0; + for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){ + A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02]; + if(L_idx < p.L){ + B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11]; + dp = fma(elemKrn, elemInp, dp); + } + } + tmp[tid*p.s0 + K_idx] += dp; + barrier(); + } + + // Save the computed values except the last block that can have different size + uint32_t KLb_idx = L_block_id*bs*p.s0; + if(L_block_id < L_blocks-1){ + for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){ + uint32_t sh_idx = p.s0*tid+s0_idx; + uint32_t KL_idx = KLb_idx+sh_idx; + if(KL_idx < p.KL){ + data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx]; + } + } + } + } + + for(uint32_t i = 0; i < splitWork(tmp_len); i++){ + uint32_t idx = i*bs+tid; + uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx; + if(KL_idx < p.KL){ + data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx]; + } + } +} diff --git a/ggml/src/vulkan-shaders/copy.comp b/ggml/src/vulkan-shaders/copy.comp index c26917c0..f476a2e3 100644 --- a/ggml/src/vulkan-shaders/copy.comp +++ b/ggml/src/vulkan-shaders/copy.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_unary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint idx = get_idx(); @@ -10,9 +12,12 @@ void main() { return; } -#ifndef OPTIMIZATION_ERROR_WORKAROUND - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]); +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]); #else - data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)]; + data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)]; #endif } diff --git a/ggml/src/vulkan-shaders/copy_from_quant.comp b/ggml/src/vulkan-shaders/copy_from_quant.comp new file mode 100644 index 00000000..dbc7daa3 --- /dev/null +++ b/ggml/src/vulkan-shaders/copy_from_quant.comp @@ -0,0 +1,51 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" +#include "dequant_funcs.comp" + +#if defined(DATA_A_IQ4_NL) +// 16 invocations needed for init_iq4nl_shmem +layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in; +#else +layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in; +#endif + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); + if (gl_LocalInvocationIndex.x != 0) { + return; + } +#endif + + const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K; + + if (idx >= p.ne) { + return; + } + + uint dst_idx = get_doffset() + dst_idx(idx); + uint src_idx = src0_idx_quant(idx, QUANT_K); + + const uint a_offset = 0; + const uint ib = src_idx; + const vec2 dm = get_dm(ib, a_offset); + + [[unroll]] for (int j = 0; j < QUANT_K; j += 4) { + vec4 v = dequantize4(ib, j / QUANT_R, a_offset); + v = v * dm.x + vec4(dm.y); + +#if QUANT_R == 2 + data_d[dst_idx + j/2 + 0] = v[0]; + data_d[dst_idx + j/2 + QUANT_K/2 + 0] = v[1]; + data_d[dst_idx + j/2 + 1] = v[2]; + data_d[dst_idx + j/2 + QUANT_K/2 + 1] = v[3]; +#else + data_d[dst_idx + j + 0] = v[0]; + data_d[dst_idx + j + 1] = v[1]; + data_d[dst_idx + j + 2] = v[2]; + data_d[dst_idx + j + 3] = v[3]; +#endif + } +} diff --git a/ggml/src/vulkan-shaders/copy_to_quant.comp b/ggml/src/vulkan-shaders/copy_to_quant.comp new file mode 100644 index 00000000..9c76437d --- /dev/null +++ b/ggml/src/vulkan-shaders/copy_to_quant.comp @@ -0,0 +1,242 @@ +#version 450 + +#if RTE16 +#extension GL_EXT_spirv_intrinsics : enable +spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits +#endif // RTE16 + +#include "types.comp" +#include "generic_unary_head.comp" + +#if defined(DATA_A_IQ4_NL) +// 16 invocations needed for init_iq4nl_shmem +layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in; +#else +layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in; +#endif + +layout (binding = 0) readonly buffer S {float data_s[];}; +layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];}; + +#if defined(DATA_A_Q4_0) +void quantize(uint dst_idx, uint src_idx) +{ + float amax = 0.0; + float vmax = 0.0; + + [[unroll]] for (int j = 0; j < QUANT_K_Q4_0; ++j) { + const float v = data_s[src_idx + j]; + if (amax < abs(v)) { + amax = abs(v); + vmax = v; + } + } + + const float d = vmax / -8; + const float id = (d != 0.0) ? 1.0/d : 0.0; + + data_q[dst_idx].d = float16_t(d); + + [[unroll]] for (int j = 0; j < QUANT_K_Q4_0/2; ++j) { + const float x0 = data_s[src_idx + 0 + j]*id; + const float x1 = data_s[src_idx + QUANT_K_Q4_0/2 + j]*id; + + const uint xi0 = min(15, int(x0 + 8.5)); + const uint xi1 = min(15, int(x1 + 8.5)); + + data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4)); + } +} +#endif + +#if defined(DATA_A_Q4_1) +void quantize(uint dst_idx, uint src_idx) +{ + float vmin = 1.0/0.0; + float vmax = -vmin; + + [[unroll]] for (int j = 0; j < QUANT_K_Q4_1; ++j) { + const float v = data_s[src_idx + j]; + + if (v < vmin) vmin = v; + if (v > vmax) vmax = v; + } + + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = (d != 0.0) ? 1.0/d : 0.0; + + data_q[dst_idx].d = float16_t(d); + data_q[dst_idx].m = float16_t(vmin); + + [[unroll]] for (int j = 0; j < QUANT_K_Q4_1/2; ++j) { + const float x0 = (data_s[src_idx + 0 + j] - vmin)*id; + const float x1 = (data_s[src_idx + QUANT_K_Q4_1/2 + j] - vmin)*id; + + const uint xi0 = min(15, int(x0 + 0.5)); + const uint xi1 = min(15, int(x1 + 0.5)); + + data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4)); + } +} +#endif + +#if defined(DATA_A_Q5_0) +void quantize(uint dst_idx, uint src_idx) +{ + float amax = 0.0; + float vmax = 0.0; + + [[unroll]] for (int j = 0; j < QUANT_K_Q5_0; ++j) { + const float v = data_s[src_idx + j]; + if (amax < abs(v)) { + amax = abs(v); + vmax = v; + } + } + + const float d = vmax / -16; + const float id = (d != 0.0) ? 1.0/d : 0.0; + + data_q[dst_idx].d = float16_t(d); + + uint32_t qh = 0; + [[unroll]] for (int j = 0; j < QUANT_K_Q5_0/2; ++j) { + const float x0 = data_s[src_idx + 0 + j]*id; + const float x1 = data_s[src_idx + QUANT_K_Q5_0/2 + j]*id; + + const uint xi0 = min(31, int(x0 + 16.5)); + const uint xi1 = min(31, int(x1 + 16.5)); + + data_q[dst_idx].qs[j] = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4)); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_0/2); + } + data_q[dst_idx].qh[0] = uint16_t(qh & 0xFFFF); + data_q[dst_idx].qh[1] = uint16_t(qh >> 16); +} +#endif + +#if defined(DATA_A_Q5_1) +void quantize(uint dst_idx, uint src_idx) +{ + float min = data_s[src_idx + 0]; + float max = min; + + [[unroll]] for (int j = 1; j < QUANT_K_Q5_1; ++j) { + const float v = data_s[src_idx + j]; + min = v < min ? v : min; + max = v > max ? v : max; + } + + const float d = (max - min) / 31; + const float id = (d != 0) ? 1.0/d : 0.0; + + data_q[dst_idx].d = float16_t(d); + data_q[dst_idx].m = float16_t(min); + + uint32_t qh = 0; + [[unroll]] for (int j = 0; j < QUANT_K_Q5_1/2; ++j) { + const float x0 = (data_s[src_idx + 0 + j] - min)*id; + const float x1 = (data_s[src_idx + QUANT_K_Q5_1/2 + j] - min)*id; + + const uint xi0 = uint(x0 + 0.5); + const uint xi1 = uint(x1 + 0.5); + + data_q[dst_idx].qs[j] = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4)); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_1/2); + } + data_q[dst_idx].qh = qh; +} +#endif + +#if defined(DATA_A_Q8_0) +void quantize(uint dst_idx, uint src_idx) +{ + float amax = 0.0; // absolute max + + [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; j++) { + const float v = data_s[src_idx + j]; + amax = max(amax, abs(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = (d != 0.0) ? 1.0/d : 0.0; + + data_q[dst_idx].d = float16_t(d); + + [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; ++j) { + const float x0 = data_s[src_idx + j]*id; + + data_q[dst_idx].qs[j] = int8_t(round(x0)); + } +} +#endif + +#if defined(DATA_A_IQ4_NL) +uint best_index(float x) { + if (x <= kvalues_iq4nl[0]) return 0; + if (x >= kvalues_iq4nl[15]) return 15; + int ml = 0, mu = 15; + while (mu-ml > 1) { + int mav = (ml+mu)/2; + if (x < kvalues_iq4nl[mav]) mu = mav; else ml = mav; + } + return x - kvalues_iq4nl[mu-1] < kvalues_iq4nl[mu] - x ? mu-1 : mu; +} + +void quantize(uint dst_idx, uint src_idx) +{ + float amax = 0.0; + float vmax = 0.0; + + [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL; ++j) { + const float v = data_s[src_idx + j]; + if (amax < abs(v)) { + amax = abs(v); + vmax = v; + } + } + + float d = vmax / kvalues_iq4nl[0]; + const float id = (d != 0.0) ? 1.0/d : 0.0; + + float sumqx = 0, sumq2 = 0; + [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL/2; ++j) { + const float x0 = data_s[src_idx + 0 + j]*id; + const float x1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*id; + const uint xi0 = best_index(x0); + const uint xi1 = best_index(x1); + data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4)); + const float v0 = kvalues_iq4nl[xi0]; + const float v1 = kvalues_iq4nl[xi1]; + const float w0 = data_s[src_idx + 0 + j]*data_s[src_idx + 0 + j]; + const float w1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*data_s[src_idx + QUANT_K_IQ4_NL/2 + j]; + sumqx += w0*v0*data_s[src_idx + j] + w1*v1*data_s[src_idx + QUANT_K_IQ4_NL/2 + j]; + sumq2 += w0*v0*v0 + w1*v1*v1; + } + + data_q[dst_idx].d = float16_t(sumq2 > 0 ? sumqx/sumq2 : d); + +} +#endif + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); + if (gl_LocalInvocationIndex.x != 0) { + return; + } +#endif + + const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K; + + if (idx >= p.ne) { + return; + } + + uint dst_idx = dst_idx_quant(idx, QUANT_K); + uint src_idx = get_aoffset() + src0_idx(idx); + + quantize(dst_idx, src_idx); +} diff --git a/ggml/src/vulkan-shaders/cos.comp b/ggml/src/vulkan-shaders/cos.comp new file mode 100644 index 00000000..0b8d02f5 --- /dev/null +++ b/ggml/src/vulkan-shaders/cos.comp @@ -0,0 +1,17 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +void main() { + const uint idx = get_idx(); + + if (idx >= p.ne) { + return; + } + + const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val)); +} diff --git a/ggml/src/vulkan-shaders/count_equal.comp b/ggml/src/vulkan-shaders/count_equal.comp new file mode 100644 index 00000000..d9345497 --- /dev/null +++ b/ggml/src/vulkan-shaders/count_equal.comp @@ -0,0 +1,31 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#include "types.comp" +#include "generic_head.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer Y {B_TYPE data_b[];}; +layout (binding = 2) buffer D {D_TYPE data_d[];}; + +const uint CHUNK_SIZE = 512; + +void main() { + const uint base = gl_WorkGroupID.x * CHUNK_SIZE; + const uint col = gl_LocalInvocationID.x; + + uint count = 0; + [[unroll]] + for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) { + const uint idx = base + i + col; + if (idx >= p.KX) { + break; + } + count += uint(data_a[idx] == data_b[idx]); + } + + atomicAdd(data_d[0], D_TYPE(count)); +} diff --git a/ggml/src/vulkan-shaders/dequant_funcs.comp b/ggml/src/vulkan-shaders/dequant_funcs.comp index d5b98973..0d9739d4 100644 --- a/ggml/src/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/vulkan-shaders/dequant_funcs.comp @@ -2,6 +2,15 @@ #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require #endif +#include "types.comp" + +#if defined(A_TYPE_PACKED16) +layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];}; +#endif +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif + #if defined(DATA_A_F32) vec2 dequantize(uint ib, uint iqs, uint a_offset) { return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]); @@ -14,55 +23,440 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { } #endif +#if defined(DATA_A_BF16) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1])); +} +#endif + #if defined(DATA_A_Q4_0) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); const uint vui = uint(data_a[a_offset + ib].qs[iqs]); - return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; + return (vec2(vui & 0xF, vui >> 4) - 8.0f); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) - 8.0f); } #endif #if defined(DATA_A_Q4_1) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); - const float m = float(data_a[a_offset + ib].m); const uint vui = uint(data_a[a_offset + ib].qs[iqs]); - return vec2(vui & 0xF, vui >> 4) * d + m; + return vec2(vui & 0xF, vui >> 4); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12); } #endif #if defined(DATA_A_Q5_0) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0]; const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); const uint vui = uint(data_a[a_offset + ib].qs[iqs]); - return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d; + return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0]; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f); } #endif #if defined(DATA_A_Q5_1) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); - const float m = float(data_a[a_offset + ib].m); const uint uint_qh = data_a[a_offset + ib].qh; const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); const uint vui = uint(data_a[a_offset + ib].qs[iqs]); - return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m; + return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint uint_qh = data_a_packed16[a_offset + ib].qh; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y); } #endif #if defined(DATA_A_Q8_0) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); - return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d; + return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy; + return vec4(v0.x, v0.y, v1.x, v1.y); +} +#endif + +#if defined(DATA_A_IQ1_S) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + const int i8 = int(iqs % 8); + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const float dl = float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint idxhi = bitfieldExtract(qh, 3 * int(ib8 & 3), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]); + // Signed bitfield extract. + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + return dl * (vec2(gvec) + delta); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + const int i8 = int(iqs % 8); + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const float dl = 2 * bitfieldExtract(qh, 12, 3) + 1; + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]); + // Signed bitfield extract. + const ivec4 gvec = ivec4( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2), + bitfieldExtract(grid, 2 * (i8 + 2), 2), + bitfieldExtract(grid, 2 * (i8 + 3), 2) + ); + return dl * (vec4(gvec) + delta); +} +#endif + +#if defined(DATA_A_IQ1_M) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib8 = iqs / 8; + const uint ib16 = iqs / 16; + const int i8 = int(iqs % 8); + const uint sc = data_a[a_offset + ib].scales[iqs / 64]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + // Signed bitfield extract. + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + return dl * (vec2(gvec) + delta); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib8 = iqs / 8; + const uint ib16 = iqs / 16; + const int i8 = int(iqs % 8); + const uint sc = data_a[a_offset + ib].scales[iqs / 64]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + // Signed bitfield extract. + const ivec4 gvec = ivec4( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2), + bitfieldExtract(grid, 2 * (i8 + 2), 2), + bitfieldExtract(grid, 2 * (i8 + 3), 2) + ); + return dl * (vec4(gvec) + delta); +} +#endif + +#if defined(DATA_A_IQ2_XXS) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = (iqs / 8) % 4; + const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8]; + // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale) + const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2], + data_a_packed16[a_offset + ib].qs[4 * ib32 + 3])); + const float db = 0.25 * (0.5 + (signs >> 28)); + const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7); + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4))); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + return db * vec2( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0) + ); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = (iqs / 8) % 4; + const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8]; + // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale) + const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2], + data_a_packed16[a_offset + ib].qs[4 * ib32 + 3])); + const float db = 0.25 * (0.5 + (signs >> 28)); + const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7); + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4))); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + bool sign2 = (sign & 4) != 0; + bool sign3 = (sign & 8) != 0; + return db * vec4( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0), + grid.z * (sign2 ? -1.0 : 1.0), + grid.w * (sign3 ? -1.0 : 1.0) + ); +} +#endif + +#if defined(DATA_A_IQ2_XS) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf; + const uint qs = data_a[a_offset + ib].qs[iqs / 8]; + const float db = 0.25 * (0.5 + scale); + const uint sign7 = qs >> 9; + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4))); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + return db * vec2( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0) + ); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf; + const uint qs = data_a[a_offset + ib].qs[iqs / 8]; + const float db = 0.25 * (0.5 + scale); + const uint sign7 = qs >> 9; + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4))); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + bool sign2 = (sign & 4) != 0; + bool sign3 = (sign & 8) != 0; + return db * vec4( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0), + grid.z * (sign2 ? -1.0 : 1.0), + grid.w * (sign3 ? -1.0 : 1.0) + ); +} +#endif + +#if defined(DATA_A_IQ2_S) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + + const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qhshift = 2 * (ib8 % 4); + const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8); + + const float db = 0.25 * (0.5 + scale); + const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + return db * vec2( + grid[iqs % 4] * (sign0 ? -1.0 : 1.0), + grid[(iqs % 4) + 1] * (sign1 ? -1.0 : 1.0) + ); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + + const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qhshift = 2 * (ib8 % 4); + const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8); + + const float db = 0.25 * (0.5 + scale); + const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + bool sign2 = (sign & 4) != 0; + bool sign3 = (sign & 8) != 0; + return db * vec4( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0), + grid.z * (sign2 ? -1.0 : 1.0), + grid.w * (sign3 ? -1.0 : 1.0) + ); +} +#endif + +#if defined(DATA_A_IQ3_XXS) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib4 = iqs / 4; + const uint ib32 = iqs / 32; + const uint is = QUANT_K / 4 + 4 * ib32; + const uint qs = data_a[a_offset + ib].qs[ib4]; + // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale) + const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2], + data_a_packed16[a_offset + ib].qs[is / 2 + 1])); + const float db = 0.5 * (0.5 + (signs >> 28)); + const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7); + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq3xxs_grid[qs] >> (8 * (iqs % 4))); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + return db * vec2( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0) + ); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib4 = iqs / 4; + const uint ib32 = iqs / 32; + const uint is = QUANT_K / 4 + 4 * ib32; + const uint qs = data_a[a_offset + ib].qs[ib4]; + const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2], + data_a_packed16[a_offset + ib].qs[is / 2 + 1])); + const float db = 0.5 * (0.5 + (signs >> 28)); + const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7); + // Add parity bit + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const uint sign = sign8 >> (iqs % 8); + const u8vec4 grid = unpack8(iq3xxs_grid[qs]); + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + bool sign2 = (sign & 4) != 0; + bool sign3 = (sign & 8) != 0; + return db * vec4( + grid.x * (sign0 ? -1.0 : 1.0), + grid.y * (sign1 ? -1.0 : 1.0), + grid.z * (sign2 ? -1.0 : 1.0), + grid.w * (sign3 ? -1.0 : 1.0) + ); +} +#endif + +#if defined(DATA_A_IQ3_S) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint qs = data_a[a_offset + ib].qs[iqs / 4]; + const uint qh = data_a[a_offset + ib].qh[iqs / 32]; + const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8); + const uint scale = data_a[a_offset + ib].scales[iqs / 64]; + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + const float db = 1 + 2 * ((scale >> (4 * ((iqs / 32) & 1))) & 0xf); + const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ((iqs / 4) % 8))) & 256)] >> (8 * (iqs % 4)); + return db * vec2( + int(grid & 0xFF) * (sign0 ? -1.0 : 1.0), + int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0) + ); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib4 = iqs / 4; + const uint ib32 = iqs / 32; + const uint qs = data_a[a_offset + ib].qs[ib4]; + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8); + const uint scale = data_a[a_offset + ib].scales[ib32 / 2]; + bool sign0 = (sign & 1) != 0; + bool sign1 = (sign & 2) != 0; + bool sign2 = (sign & 4) != 0; + bool sign3 = (sign & 8) != 0; + const float db = 1 + 2 * ((scale >> (4 * (ib32 & 1))) & 0xf); + const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ib4 % 8)) & 256)] >> (8 * (iqs % 4)); + return db * vec4( + int(grid & 0xFF) * (sign0 ? -1.0 : 1.0), + int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0), + int((grid >> 16) & 0xFF) * (sign2 ? -1.0 : 1.0), + int((grid >> 24) & 0xFF) * (sign3 ? -1.0 : 1.0) + ); +} +#endif + +#if defined(DATA_A_IQ4_XS) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint iq = 16 * ib32 + (iqs % 16); + + const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3; + const uint qshift = (iqs & 16) >> 2; + u8vec2 qs = u8vec2(data_a[a_offset + ib].qs[iq], data_a[a_offset + ib].qs[iq + 1]); + qs = (qs >> qshift) & uint8_t(0xF); + + const float dl = float(int(sl | (sh << 4)) - 32); + return dl * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint iq = 16 * ib32 + (iqs % 16); + + const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3; + const uint qshift = (iqs & 16) >> 2; + u8vec4 qs = u8vec4( + data_a[a_offset + ib].qs[iq + 0], + data_a[a_offset + ib].qs[iq + 1], + data_a[a_offset + ib].qs[iq + 2], + data_a[a_offset + ib].qs[iq + 3] + ); + qs = (qs >> qshift) & uint8_t(0xF); + + const float dl = float(int(sl | (sh << 4)) - 32); + return dl * vec4( + kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y], + kvalues_iq4nl[qs.z], kvalues_iq4nl[qs.w]); } #endif #if defined(DATA_A_IQ4_NL) vec2 dequantize(uint ib, uint iqs, uint a_offset) { - const float d = float(data_a[a_offset + ib].d); const uint vui = uint(data_a[a_offset + ib].qs[iqs]); - return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d; + return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[vui >> 12]); +} +#endif + +#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16) +vec2 get_dm(uint ib, uint a_offset) { + return vec2(0, 0); +} +#endif + +#if defined(DATA_A_IQ1_M) +vec2 get_dm(uint ib, uint a_offset) { + const uint16_t[4] scales = data_a[a_offset + ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + return vec2(d, 0); +} +#endif + +#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +vec2 get_dm(uint ib, uint a_offset) { + return vec2(float(data_a[a_offset + ib].d), 0); +} +#endif + +#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1) +vec2 get_dm(uint ib, uint a_offset) { + return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m)); } #endif diff --git a/ggml/src/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/vulkan-shaders/dequant_funcs_cm2.comp new file mode 100644 index 00000000..9cb7da2d --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_funcs_cm2.comp @@ -0,0 +1,699 @@ + +#include "types.comp" + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 { + block_q4_0_packed16 block; +}; + +float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + const uint shift = (idx & 0x10) >> 2; + uint32_t qs = uint32_t(bl.block.qs[(idx & 0xE) >> 1]); + qs >>= shift; + qs &= 0x0F0F; + qs = unpack8(qs)[idx & 1]; + float16_t ret = (float16_t(qs) - float16_t(8)) * d; + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ4_1 { + block_q4_1 block; +}; + +float16_t dequantFuncQ4_1(const in decodeBufQ4_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const float16_t m = bl.block.m; + const uint idx = coordInBlock[1]; + const uint iqs = idx & 0xF; + const uint shift = (idx & 0x10) >> 2; + uint32_t qs = bl.block.qs[iqs]; + qs >>= shift; + qs &= 0xF; + float16_t ret = float16_t(qs) * d + m; + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ5_0 { + block_q5_0 block; +}; + +float16_t dequantFuncQ5_0(const in decodeBufQ5_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + const uint iqs = idx & 0xF; + + const uint uint_qh = uint(bl.block.qh[1]) << 16 | bl.block.qh[0]; + const uint qh = ((uint_qh >> idx) << 4) & 0x10; + + const uint shift = (idx & 0x10) >> 2; + uint32_t qs = bl.block.qs[iqs]; + qs >>= shift; + qs &= 0xF; + + float16_t ret = (float16_t(qs | qh) - float16_t(16)) * d; + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufQ5_1 { + block_q5_1 block; +}; + +float16_t dequantFuncQ5_1(const in decodeBufQ5_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const float16_t m = bl.block.m; + const uint idx = coordInBlock[1]; + const uint iqs = idx & 0xF; + + const uint uint_qh = bl.block.qh; + const uint qh = ((uint_qh >> idx) << 4) & 0x10; + + const uint shift = (idx & 0x10) >> 2; + uint32_t qs = bl.block.qs[iqs]; + qs >>= shift; + qs &= 0xF; + + float16_t ret = float16_t(qs | qh) * d + m; + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ8_0 { + block_q8_0_packed16 block; +}; + +float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + const uint iqs = idx; + + // Load 16b and select the byte for this element + int32_t qs = unpack8(bl.block.qs[(iqs & 0x1E) >> 1])[iqs & 1]; + float16_t ret = float16_t(qs) * d; + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_K { + block_q2_K block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2_K_packed16 { + block_q2_K_packed16 block; +}; + +float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl); + const f16vec2 d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint scalesi = (idx & 0xF0) >> 4; // 0..15 + const uint qsshift = (idx & 0x60) >> 4; // 0,2,4,6 + + uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]); + qs = (qs >> qsshift) & 0x0303; + qs = unpack8(qs)[idx & 1]; + + const uint scales = bl.block.scales[scalesi]; + float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4); + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ3_K { + block_q3_K block; +}; + +float16_t dequantFuncQ3_K(const in decodeBufQ3_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const uint idx = coordInBlock[1]; + const uint iqs = idx; + + const uint n = iqs / 128; // 0,1 + const uint qsi = n * 32 + (iqs % 32); // 0..63 + const uint hmi = (iqs % 32); // 0..31 + const uint j = (iqs % 128) / 8; // 0..15 + const uint is = iqs / 16; // 0..15 + const uint halfsplit = ((iqs % 128) / 32); // 0,1,2,3 + const uint qsshift = halfsplit * 2; // 0,2,4,6 + const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128 + + uint32_t scaleidx0 = (is < 8) ? is : (is-8); + uint32_t scaleidx0shift = (is < 8) ? 0 : 4; + uint32_t scaleidx1 = is + 8 - (is/4)*4; + uint32_t scaleidx1shift = (is/4)*2; + + const int8_t us = int8_t(((bl.block.scales[scaleidx0] >> scaleidx0shift) & 0xF) | (((bl.block.scales[scaleidx1] >> scaleidx1shift) & 3) << 4)); + + const float16_t dl = bl.block.d * float16_t(us - 32); + + float16_t ret = dl * float16_t(int8_t((bl.block.qs[qsi ] >> qsshift) & 3) - (((bl.block.hmask[hmi ] & m) != 0) ? 0 : 4)); + + return ret; +} + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K { + block_q4_K block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed16 { + block_q4_K_packed16 block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed128 { + block_q4_K_packed128 block; +}; + +#if defined(IS_MUL_MM2) + +// For Q4_K and Q5_K in the mat-mul shader, we decode a tile's worth of scales +// into shared memory and then process the whole tile using those scales. +// There is a fetch function that loads into private variables and then a store +// function that stores into shared memory. +// Q4_K and Q5_K have the same encoding of scales, so everything is shared except +// the part that fetches from the structure (which has a different block layout). +#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K) +const uint shAscales_stride = (BM + 2); +// 1 scale per 32 elements -> 8 scales per block, per row +shared vec2 shAscales[8 * shAscales_stride]; +uvec4 row_v; +#endif + +#if defined(DATA_A_Q4_K) +layout (binding = 0) readonly buffer A_Q4_K_128 {block_q4_K_packed128 data_a_q4_k_packed128[];}; + +void fetch_scalesQ4_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds) +{ + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + uint row = ir_BM + tid_row; + uint block_index = pos_a + row * stride_a + (block_k / QUANT_K); + if (in_bounds || row < p.M) { + row_v = data_a_q4_k_packed128[block_index].q4k[0]; + } +} +#endif +#if defined(DATA_A_Q5_K) +layout (binding = 0) readonly buffer A_Q5_K_128 {block_q5_K_packed128 data_a_q5_k_packed128[];}; + +void fetch_scalesQ5_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds) +{ + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + uint row = ir_BM + tid_row; + uint block_index = pos_a + row * stride_a + (block_k / QUANT_K); + if (in_bounds || row < p.M) { + row_v = data_a_q5_k_packed128[block_index].q5k[0]; + } +} +#endif + +#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K) +void store_scalesQ4_K(uint tid) +{ + barrier(); + + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + [[unroll]] for (uint idx = 0; idx < is_per_tid; ++idx) { + uint is = idx + is_start; + uvec4 v = row_v; + const vec2 loadd = vec2(unpackFloat2x16(v.x)); + + uint32_t sc; + uint32_t mbyte; + + uint32_t scale0 = v.y; + uint32_t scale4 = v.z; + uint32_t scale8 = v.w; + + uint32_t sc_lo = scale0; + uint32_t mb_lo = scale4; + uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2); + uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2); + + sc = is < 4 ? sc_lo : sc_hi; + mbyte = is < 4 ? mb_lo : mb_hi; + sc = sc >> (8 * (is & 3)); + mbyte = mbyte >> (8 * (is & 3)); + sc &= 0x3F; + mbyte &= 0x3F; + + const float d = loadd.x * float(sc); + const float m = loadd.y * float(mbyte); + shAscales[is * shAscales_stride + tid_row] = vec2(d,m); + } + + barrier(); +} +#endif + +#endif + +float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl); + decodeBufQ4_K_packed128 bl128 = decodeBufQ4_K_packed128(bl); + const uint idx = coordInBlock[1]; + + const uint b = (idx & 0x20) >> 5; // 0,1 + const uint is = (idx & 0xE0) >> 5; // 0..7 + +#if defined(IS_MUL_MM2) && defined(DATA_A_Q4_K) + vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)]; + float d = v.x; + float m = v.y; +#else + uvec4 v = bl128.block.q4k[0]; + const vec2 loadd = vec2(unpackFloat2x16(v.x)); + + uint32_t sc; + uint32_t mbyte; + + uint32_t scale0 = v.y; + uint32_t scale4 = v.z; + uint32_t scale8 = v.w; + + uint32_t sc_lo = scale0; + uint32_t mb_lo = scale4; + uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2); + uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2); + + sc = is < 4 ? sc_lo : sc_hi; + mbyte = is < 4 ? mb_lo : mb_hi; + sc = sc >> (8 * (is & 3)); + mbyte = mbyte >> (8 * (is & 3)); + sc &= 0x3F; + mbyte &= 0x3F; + + const float d = loadd.x * float(sc); + const float m = loadd.y * float(mbyte); +#endif + + uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]); + qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF; + + float ret = d * float(qs) - m; + + return float16_t(ret); +} + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K { + block_q5_K block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed16 { + block_q5_K_packed16 block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed128 { + block_q5_K_packed128 block; +}; + +float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl); + decodeBufQ5_K_packed128 bl128 = decodeBufQ5_K_packed128(bl); + const uint idx = coordInBlock[1]; + + const uint b = (idx & 0x20) >> 5; // 0,1 + const uint is = (idx & 0xE0) >> 5; // 0..7 + +#if defined(IS_MUL_MM2) && defined(DATA_A_Q5_K) + vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)]; + float d = v.x; + float m = v.y; +#else + uvec4 v = bl128.block.q5k[0]; + + const f16vec2 loadd = unpackFloat2x16(v.x); + + uint32_t sc; + uint32_t mbyte; + + uint32_t scale0 = v.y; + uint32_t scale4 = v.z; + uint32_t scale8 = v.w; + + uint32_t sc_lo = scale0; + uint32_t mb_lo = scale4; + uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2); + uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2); + + sc = is < 4 ? sc_lo : sc_hi; + mbyte = is < 4 ? mb_lo : mb_hi; + sc = sc >> (8 * (is & 3)); + mbyte = mbyte >> (8 * (is & 3)); + sc &= 0x3F; + mbyte &= 0x3F; + + const float16_t d = loadd.x * float16_t(sc); + const float16_t m = loadd.y * float16_t(mbyte); +#endif + + uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]); + qh = ((qh >> is) & 0x101) << 4; + + uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]); + qs = (qs >> (b * 4)) & 0x0F0F; + qs = unpack8(qs | qh)[idx & 1]; + + float ret = d * float(qs) - m; + + return float16_t(ret); +} + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_K { + block_q6_K block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ6_K_packed16 { + block_q6_K_packed16 block; +}; + +float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufQ6_K_packed16 bl16 = decodeBufQ6_K_packed16(bl); + const uint idx = coordInBlock[1]; + + const uint b = (idx & 0x40) >> 6; // 0,1 + const uint qhshift = (idx & 0x60) >> 4; // 0,2,4,6 + const uint is = (idx & 0xF0) >> 4; // 0..15 + + const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]); + + uint ql = uint32_t(bl16.block.ql[((idx & 0x80) >> 2) + ((idx & 0x3E) >> 1)]); + ql = (ql >> (b * 4)) & 0x0F0F; + + uint qh = uint32_t(bl16.block.qh[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]); + qh = ((qh >> qhshift) & 0x0303) << 4; + + int q = unpack8(ql | qh)[idx & 1]; + + float16_t ret = dscale * float16_t(q - 32); + + return ret; +} + +#if defined(DATA_A_IQ1_S) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S { + block_iq1_s block; +}; + +float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint ib32 = (idx & 0xE0) >> 5; + const uint ib8 = (idx & 0xF8) >> 3; + + const uint qh = bl.block.qh[ib32]; + const uint qs = bl.block.qs[ib8]; + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]; + + float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta)); + return ret; +} +#endif + +#if defined(DATA_A_IQ1_M) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M { + block_iq1_m block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufIQ1_M_packed64 { + block_iq1_m_packed64 block; +}; + +float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufIQ1_M_packed64 bl64 = decodeBufIQ1_M_packed64(bl); + const uint idx = coordInBlock[1]; + + uvec2 scales = unpack32(bl64.block.scales); + const float16_t d = uint16BitsToHalf(uint16_t(((scales.x & 0xF000) >> 12) | ((scales.x & 0xF0000000) >> 24) | ((scales.y & 0xF000) >> 4) | ((scales.y & 0xF0000000) >> 16))); + + const uint ib8 = (idx & 0xF8) >> 3; + const uint ib16 = (idx & 0xF0) >> 4; + const int i8 = int(idx % 8); + const uint sc = bl.block.scales[ib8 / 8]; + const uint qs = bl.block.qs[ib8]; + const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint grid = iq1s_grid[qs | ((qh & 7) << 8)]; + + float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta)); + return ret; +} +#endif + +#if defined(DATA_A_IQ2_XXS) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS { + block_iq2_xxs block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS_packed16 { + block_iq2_xxs_packed16 block; +}; + +float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufIQ2_XXS_packed16 bl16 = decodeBufIQ2_XXS_packed16(bl); + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint ib32 = (idx & 0xE0) >> 5; // 0..7 + const uint ib8 = (idx & 0x18) >> 3; // 0..3 + const uint iqs = 8 * ib32 + ib8; + + const uint qs = bl.block.qs[iqs]; + const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3])); + + const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28)); + uint sign = bitfieldExtract(signscale, 7 * int(ib8), 7); + sign |= bitCount(sign) << 7; + + uint g2 = iq2xxs_grid[qs][(idx & 4) >> 2]; + g2 >>= (idx & 2) * 8; + const vec2 g = vec2(unpack8(g2)); + + vec2 ret = dscale * g * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf); + return float16_t(ret[idx & 1]); +} +#endif + +#if defined(DATA_A_IQ2_XS) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XS { + block_iq2_xs block; +}; + +float16_t dequantFuncIQ2_XS(const in decodeBufIQ2_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint is = (idx & 0xE0) >> 5; // 0..8 + const uint sshift = (idx & 0x10) >> 2; // 0,4 + const uint iqs = (idx & 0xF8) >> 3; // 0..63 + + const uint16_t qs = bl.block.qs[iqs]; + const float dscale = float(bl.block.d) * 0.25 * (0.5 + float((bl.block.scales[is] >> sshift) & 0xF)); + + uint sign = uint(qs >> 9); + sign |= bitCount(sign) << 7; + uint g2 = iq2xs_grid[qs & 0x1FF][(idx & 4) >> 2]; + g2 >>= (idx & 2) * 8; + const vec2 g = vec2(unpack8(g2)); + + vec2 ret = dscale * g * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf); + return float16_t(ret[idx & 1]); +} +#endif + +#if defined(DATA_A_IQ2_S) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_S { + block_iq2_s block; +}; + +float16_t dequantFuncIQ2_S(const in decodeBufIQ2_S bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + uint idx = coordInBlock[1]; + + const uint ib32 = (idx & 0xE0) >> 5; // 0..7 + const uint ib8 = (idx & 0xF8) >> 3; // 0..31 + const uint qhshift = 2 * (ib8 % 4); + + const uint scale = (bl.block.scales[ib32] >> ((idx & 0x10) >> 2)) & 0xf; + const uint qs = bl.block.qs[ib8]; + const uint qh = bl.block.qh[ib32]; + const uint sign = bl.block.qs[QUANT_K / 8 + ib8] >> (idx & 0x6); + + const float d = float(bl.block.d); + const float db = d * 0.25 * (0.5 + scale); + const ivec2 sign01 = 1 - (2 & ivec2(sign << 1, sign)); + uint g2 = iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 4) >> 2]; + g2 >>= (idx & 2) * 8; + const vec2 v = db * vec2(sign01) * vec2(unpack8(g2)); + return float16_t(v[idx & 1]); +} +#endif + +#if defined(DATA_A_IQ3_XXS) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS { + block_iq3_xxs block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS_packed16 { + block_iq3_xxs_packed16 block; +}; + +float16_t dequantFuncIQ3_XXS(const in decodeBufIQ3_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufIQ3_XXS_packed16 bl16 = decodeBufIQ3_XXS_packed16(bl); + uint idx = coordInBlock[1]; + + const uint iqs = (idx & 0xFC) >> 2; // 0..63 + const uint is = QUANT_K / 4 + ((idx & 0xE0) >> 3);// 8 values + + const float d = float(bl.block.d); + const uint qs = bl.block.qs[iqs]; + const uint signs = pack32(u16vec2( + bl16.block.qs[is/2+0], + bl16.block.qs[is/2+1] + )); + const float db = d * 0.5 * (0.5 + (signs >> 28)); + const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7); + const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (idx & 0x6); + const ivec2 sign01 = ivec2(1 - (2 & ivec2(sign << 1, sign))); + const uint grid = iq3xxs_grid[qs] >> (16 * ((idx & 2) >> 1)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + return float16_t(v[idx & 1]); +} +#endif + +#if defined(DATA_A_IQ3_S) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_S { + block_iq3_s block; +}; + +float16_t dequantFuncIQ3_S(const in decodeBufIQ3_S bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + uint idx = coordInBlock[1]; + + const uint iqs = (idx & 0xFC) >> 2; // 0..63 + const uint iqh = (idx & 0xE0) >> 5; + + const float d = float(bl.block.d); + const uint qs = bl.block.qs[iqs]; + const uint qh = bl.block.qh[iqh]; + const int8_t sign = int8_t(bl.block.signs[iqs / 2] >> (idx & 0x6)); + const uint scale = bl.block.scales[iqs / 16]; + const ivec2 sign01 = ivec2(1 - (2 & ivec2(sign << 1, sign))); + const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf)); + const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> ((idx & 2) << 3); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + + return float16_t(v[idx & 1]); +} +#endif + +#if defined(DATA_A_IQ4_XS) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_XS { + block_iq4_xs block; +}; + +float16_t dequantFuncIQ4_XS(const in decodeBufIQ4_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint ib32 = (idx & 0xE0) >> 5; // 0..7 + + const uint sl = (bl.block.scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const uint sh = ((bl.block.scales_h) >> (2 * ib32)) & 3; + const uint qshift = (idx & 16) >> 2; + const uint q = (bl.block.qs[16 * ib32 + (idx % 16)] >> qshift) & 0xF; + + float16_t ret = d * float16_t(int(sl | (sh << 4)) - 32) * float16_t(kvalues_iq4nl[q]); + return ret; +} +#endif + +#if defined(DATA_A_IQ4_NL) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_NL { + block_iq4_nl block; +}; + +float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + const uint iqs = idx & 0xF; + const uint shift = (idx & 0x10) >> 2; + uint32_t qs = bl.block.qs[iqs]; + qs >>= shift; + qs &= 0xF; + float16_t ret = float16_t(kvalues_iq4nl[qs]) * d; + return ret; +} +#endif + +#if defined(DATA_A_Q4_0) +#define dequantFuncA dequantFuncQ4_0 +#elif defined(DATA_A_Q4_1) +#define dequantFuncA dequantFuncQ4_1 +#elif defined(DATA_A_Q5_0) +#define dequantFuncA dequantFuncQ5_0 +#elif defined(DATA_A_Q5_1) +#define dequantFuncA dequantFuncQ5_1 +#elif defined(DATA_A_Q8_0) +#define dequantFuncA dequantFuncQ8_0 +#elif defined(DATA_A_Q2_K) +#define dequantFuncA dequantFuncQ2_K +#elif defined(DATA_A_Q3_K) +#define dequantFuncA dequantFuncQ3_K +#elif defined(DATA_A_Q4_K) +#define dequantFuncA dequantFuncQ4_K +#define fetch_scales fetch_scalesQ4_K +#define store_scales store_scalesQ4_K +#elif defined(DATA_A_Q5_K) +#define dequantFuncA dequantFuncQ5_K +#define fetch_scales fetch_scalesQ5_K +#define store_scales store_scalesQ4_K +#elif defined(DATA_A_Q6_K) +#define dequantFuncA dequantFuncQ6_K +#elif defined(DATA_A_IQ1_S) +#define dequantFuncA dequantFuncIQ1_S +#elif defined(DATA_A_IQ1_M) +#define dequantFuncA dequantFuncIQ1_M +#elif defined(DATA_A_IQ2_XXS) +#define dequantFuncA dequantFuncIQ2_XXS +#elif defined(DATA_A_IQ2_XS) +#define dequantFuncA dequantFuncIQ2_XS +#elif defined(DATA_A_IQ2_S) +#define dequantFuncA dequantFuncIQ2_S +#elif defined(DATA_A_IQ3_XXS) +#define dequantFuncA dequantFuncIQ3_XXS +#elif defined(DATA_A_IQ3_S) +#define dequantFuncA dequantFuncIQ3_S +#elif defined(DATA_A_IQ4_XS) +#define dequantFuncA dequantFuncIQ4_XS +#elif defined(DATA_A_IQ4_NL) +#define dequantFuncA dequantFuncIQ4_NL +#endif diff --git a/ggml/src/vulkan-shaders/dequant_iq1_m.comp b/ggml/src/vulkan-shaders/dequant_iq1_m.comp new file mode 100644 index 00000000..b604c188 --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq1_m.comp @@ -0,0 +1,42 @@ +#version 450 + +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq1_m data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint ib64 = ib32 / 2; + const uint b_idx = 256 * ib + 32 * ib32; + + const uint16_t[4] scales = data_a[ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + + const uint sc = data_a[ib].scales[ib64]; + [[unroll]] for (int l = 0; l < 4; ++l) { + const uint ib16 = 2 * ib32 + l / 2; + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1); + const uint qh = data_a[ib].qh[ib16] >> (4 * (l & 1)); + const uint qs = data_a[ib].qs[4 * ib32 + l]; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + [[unroll]] for (int j = 0; j < 8; ++j) { + data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta)); + } + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq1_s.comp b/ggml/src/vulkan-shaders/dequant_iq1_s.comp new file mode 100644 index 00000000..fd1e4e30 --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq1_s.comp @@ -0,0 +1,35 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq1_s data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * ib32; + + uint qh = data_a[ib].qh[ib32]; + const float d = float(data_a[ib].d); + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qs = data_a[ib].qs[4 * ib32 + l]; + const uint hi = bitfieldExtract(qh, 3 * int(l), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]); + [[unroll]] for (int j = 0; j < 8; ++j) { + data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta)); + } + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq2_s.comp b/ggml/src/vulkan-shaders/dequant_iq2_s.comp new file mode 100644 index 00000000..48f6b65b --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq2_s.comp @@ -0,0 +1,44 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq2_s data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * ib32; + + const float d = float(data_a[ib].d); + const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4); + const vec2 db = d * (0.5 + scale) * 0.25; + + uint qh = data_a[ib].qh[ib32]; + [[unroll]] for (uint l = 0; l < 4; ++l) { + uint qs = data_a[ib].qs[4 * ib32 + l]; + const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l]; + qs |= (qh << (8 - 2 * l)) & 0x300; + const uvec2 grid = iq2s_grid[qs & 511]; + const u8vec4 grid0 = unpack8(grid.x); + const u8vec4 grid1 = unpack8(grid.y); + data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign & 2) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign & 4) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign & 8) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign & 16) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign & 32) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign & 64) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign & 128) != 0 ? -1.0 : 1.0)); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq2_xs.comp b/ggml/src/vulkan-shaders/dequant_iq2_xs.comp new file mode 100644 index 00000000..a08331c4 --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq2_xs.comp @@ -0,0 +1,43 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq2_xs data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * ib32; + + const float d = float(data_a[ib].d); + const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4); + const vec2 db = d * (0.5 + scale) * 0.25; + + [[unroll]] for (uint l = 0; l < 4; ++l) { + uint16_t qs = data_a[ib].qs[4 * ib32 + l]; + const uint sign7 = qs >> 9; + const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit + const uvec2 grid = iq2xs_grid[qs & 511]; + const u8vec4 grid0 = unpack8(grid.x); + const u8vec4 grid1 = unpack8(grid.y); + data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0)); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq2_xxs.comp b/ggml/src/vulkan-shaders/dequant_iq2_xxs.comp new file mode 100644 index 00000000..e370690b --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq2_xxs.comp @@ -0,0 +1,48 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq2_xxs data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 scale block (32 values) + // Each block is described by 4 lattice indices, 4x7 sign bits and 4 scale bits + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint is = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * is; + + const float d = float(data_a[ib].d); + uint signscale = pack32(u8vec4( + data_a[ib].qs[8*is + 4], + data_a[ib].qs[8*is + 5], + data_a[ib].qs[8*is + 6], + data_a[ib].qs[8*is + 7] + )); + const float db = d * (0.5 + (signscale >> 28)) * 0.25; + + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7); + const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit + const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]]; + const u8vec4 grid0 = unpack8(grid.x); + const u8vec4 grid1 = unpack8(grid.y); + data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 3] = D_TYPE(db * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 4] = D_TYPE(db * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 5] = D_TYPE(db * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 6] = D_TYPE(db * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 7] = D_TYPE(db * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0)); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq3_s.comp b/ggml/src/vulkan-shaders/dequant_iq3_s.comp new file mode 100644 index 00000000..c3f4bca5 --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq3_s.comp @@ -0,0 +1,39 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq3_s data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 scale nibble. + // Each block contains 4 scale bytes (8 scales) for 256 output values. + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint is = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * is; + + const float d = float(data_a[ib].d); + const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf)); + + // We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes. + uint qh = data_a[ib].qh[is]; + [[unroll]] for (uint l = 0; l < 8; ++l) { + uint qs = data_a[ib].qs[8 * is + l]; + uint gidx = qs | ((qh << (8 - l)) & 256); + uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1)); + u8vec4 grid = unpack8(iq3s_grid[gidx]); + data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 4 * l + 3] = D_TYPE(db * grid.w * ((signs & 8) != 0 ? -1.0 : 1.0)); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq3_xxs.comp b/ggml/src/vulkan-shaders/dequant_iq3_xxs.comp new file mode 100644 index 00000000..a92b8296 --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq3_xxs.comp @@ -0,0 +1,49 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq3_xxs data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 scale block (32 values) + // 8 threads handle 1 superblock + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint is = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * is; + const uint s_idx = QUANT_K / 4 + 4 * is; + + const float d = float(data_a[ib].d); + uint signscale = pack32(u8vec4( + data_a[ib].qs[s_idx + 0], + data_a[ib].qs[s_idx + 1], + data_a[ib].qs[s_idx + 2], + data_a[ib].qs[s_idx + 3] + )); + const float db = d * (0.5 + (signscale >> 28)) * 0.5; + + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7); + // Restore parity bit. + const uint sign8 = sign7 | (bitCount(sign7) << 7); + const u8vec4 grid0 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l]]); + const u8vec4 grid1 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l + 1]]); + data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 3] = D_TYPE(db * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 4] = D_TYPE(db * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 5] = D_TYPE(db * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 6] = D_TYPE(db * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0)); + data_b[b_idx + 8 * l + 7] = D_TYPE(db * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0)); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/vulkan-shaders/dequant_iq4_nl.comp index 34ef3da3..46d9ad15 100644 --- a/ggml/src/vulkan-shaders/dequant_iq4_nl.comp +++ b/ggml/src/vulkan-shaders/dequant_iq4_nl.comp @@ -10,6 +10,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; void main() { const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64; + init_iq_shmem(gl_WorkGroupSize); + const uint tid = gl_LocalInvocationID.x % 64; const uint il = tid/32; const uint ir = tid%32; diff --git a/ggml/src/vulkan-shaders/dequant_iq4_xs.comp b/ggml/src/vulkan-shaders/dequant_iq4_xs.comp new file mode 100644 index 00000000..f930852a --- /dev/null +++ b/ggml/src/vulkan-shaders/dequant_iq4_xs.comp @@ -0,0 +1,34 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq4_xs data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (1 scale and 32 quantized values) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + + const float d = float(data_a[ib].d); + // Scales are 6 bits + const uint scale = ((data_a[ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF) + | (((data_a[ib].scales_h >> (2 * ib32)) & 3) << 4); + const float dl = d * (int(scale) - 32); + + const uint b_idx = 256 * ib + 32 * ib32; + const uint q_idx = 16 * ib32; + [[unroll]] for (uint l = 0; l < 16; ++l) { + data_b[b_idx + l + 0] = D_TYPE(dl * kvalues_iq4nl[data_a[ib].qs[q_idx + l] & 0xF]); + data_b[b_idx + l + 16] = D_TYPE(dl * kvalues_iq4nl[data_a[ib].qs[q_idx + l] >> 4]); + } +} diff --git a/ggml/src/vulkan-shaders/dequant_q4_k.comp b/ggml/src/vulkan-shaders/dequant_q4_k.comp index 92acb754..987f113a 100644 --- a/ggml/src/vulkan-shaders/dequant_q4_k.comp +++ b/ggml/src/vulkan-shaders/dequant_q4_k.comp @@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; void main() { [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) { - const uint i = gl_WorkGroupID.x * 256 + wgy; - if (i >= p.M * p.K / QUANT_K) { + const uint ib = gl_WorkGroupID.x * 256 + wgy; + if (ib >= p.M * p.K / QUANT_K) { return; } @@ -20,37 +20,49 @@ void main() { const uint is = 2 * il; const uint n = 4; - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y); + const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y); - const uint y_idx = i * QUANT_K + 64 * il + n * ir; + const uint y_idx = ib * QUANT_K + 64 * il + n * ir; const uint qs_idx = 32*il + n * ir; - uint8_t sc; - uint8_t m; - if (is < 4) { - sc = uint8_t(data_a[i].scales[is] & 63); - m = uint8_t(data_a[i].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4)); - m = uint8_t((data_a[i].scales[is + 4] >> 4) | ((data_a[i].scales[is ] >> 6) << 4)); - } + uint scidx0 = (is < 4) ? is : (is + 4); + uint scidx1 = (is < 4) ? is : (is - 4); + uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + uint scidxshift1 = (is < 4) ? 0 : 2; + uint mbidx0 = is + 4; + uint mbidx1 = (is < 4) ? is + 4 : is; + uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + uint mbidxshift0 = (is < 4) ? 0 : 4; + uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + uint mbidxshift1 = (is < 4) ? 0 : 2; + + uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const FLOAT_TYPE d1 = dall * sc; - const FLOAT_TYPE m1 = dmin * m; - - if (is < 4) { - sc = uint8_t(data_a[i].scales[is + 1] & 63); - m = uint8_t(data_a[i].scales[is + 5] & 63); - } else { - sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4)); - m = uint8_t((data_a[i].scales[is + 5] >> 4) | ((data_a[i].scales[is + 1] >> 6) << 4)); - } + const FLOAT_TYPE m1 = dmin * mbyte; + + scidx0 = (is < 4) ? is + 1 : (is + 5); + scidx1 = (is < 4) ? is + 1 : (is - 3); + scidxmask1 = (is < 4) ? 0x30 : 0xC0; + scidxshift1 = (is < 4) ? 0 : 2; + mbidx0 = is + 5; + mbidx1 = (is < 4) ? is + 5 : is + 1; + mbidxmask0 = (is < 4) ? 0xF : 0xF0; + mbidxshift0 = (is < 4) ? 0 : 4; + mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + mbidxshift1 = (is < 4) ? 0 : 2; + + sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const FLOAT_TYPE d2 = dall * sc; - const FLOAT_TYPE m2 = dmin * m; + const FLOAT_TYPE m2 = dmin * mbyte; [[unroll]] for (uint l = 0; l < n; ++l) { - data_b[y_idx + l ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1); - data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >> 4) - m2); + data_b[y_idx + l ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1); + data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >> 4) - m2); } } } diff --git a/ggml/src/vulkan-shaders/dequant_q5_k.comp b/ggml/src/vulkan-shaders/dequant_q5_k.comp index f314a76d..6db5403b 100644 --- a/ggml/src/vulkan-shaders/dequant_q5_k.comp +++ b/ggml/src/vulkan-shaders/dequant_q5_k.comp @@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; void main() { [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) { - const uint i = gl_WorkGroupID.x * 256 + wgy; - if (i >= p.M * p.K / QUANT_K) { + const uint ib = gl_WorkGroupID.x * 256 + wgy; + if (ib >= p.M * p.K / QUANT_K) { return; } @@ -19,40 +19,52 @@ void main() { const uint ir = tid % 16; const uint is = 2 * il; - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y); + const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y); - const uint y_idx = i * QUANT_K + 64 * il + 2 * ir; + const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir; const uint qs_idx = 32*il + 2 * ir; const uint qh_idx = 2 * ir; - uint8_t sc; - uint8_t m; - if (is < 4) { - sc = uint8_t(data_a[i].scales[is] & 63); - m = uint8_t(data_a[i].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4)); - m = uint8_t((data_a[i].scales[is + 4] >> 4) | ((data_a[i].scales[is ] >> 6) << 4)); - } + uint scidx0 = (is < 4) ? is : (is + 4); + uint scidx1 = (is < 4) ? is : (is - 4); + uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + uint scidxshift1 = (is < 4) ? 0 : 2; + uint mbidx0 = is + 4; + uint mbidx1 = (is < 4) ? is + 4 : is; + uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + uint mbidxshift0 = (is < 4) ? 0 : 4; + uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + uint mbidxshift1 = (is < 4) ? 0 : 2; + + uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const FLOAT_TYPE d1 = dall * sc; - const FLOAT_TYPE m1 = dmin * m; - - if (is < 4) { - sc = uint8_t(data_a[i].scales[is + 1] & 63); - m = uint8_t(data_a[i].scales[is + 5] & 63); - } else { - sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4)); - m = uint8_t((data_a[i].scales[is + 5] >> 4) | ((data_a[i].scales[is + 1] >> 6) << 4)); - } + const FLOAT_TYPE m1 = dmin * mbyte; + + scidx0 = (is < 4) ? is + 1 : (is + 5); + scidx1 = (is < 4) ? is + 1 : (is - 3); + scidxmask1 = (is < 4) ? 0x30 : 0xC0; + scidxshift1 = (is < 4) ? 0 : 2; + mbidx0 = is + 5; + mbidx1 = (is < 4) ? is + 5 : is + 1; + mbidxmask0 = (is < 4) ? 0xF : 0xF0; + mbidxshift0 = (is < 4) ? 0 : 4; + mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + mbidxshift1 = (is < 4) ? 0 : 2; + + sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const FLOAT_TYPE d2 = dall * sc; - const FLOAT_TYPE m2 = dmin * m; + const FLOAT_TYPE m2 = dmin * mbyte; const uint8_t hm1 = uint8_t(1 << (2 * il )); const uint8_t hm2 = uint8_t(1 << (2 * il + 1)); - data_b[y_idx ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx ] & 0xF) + (((data_a[i].qh[qh_idx ] & hm1) != 0) ? 16 : 0)) - m1); - data_b[y_idx + 1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1); - data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx ] >> 4) + (((data_a[i].qh[qh_idx ] & hm2) != 0) ? 16 : 0)) - m2); - data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2); + data_b[y_idx ] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx ] & 0xF) + (((data_a[ib].qh[qh_idx ] & hm1) != 0) ? 16 : 0)) - m1); + data_b[y_idx + 1] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1] & 0xF) + (((data_a[ib].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1); + data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx ] >> 4) + (((data_a[ib].qh[qh_idx ] & hm2) != 0) ? 16 : 0)) - m2); + data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1] >> 4) + (((data_a[ib].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2); } } diff --git a/ggml/src/vulkan-shaders/diag_mask_inf.comp b/ggml/src/vulkan-shaders/diag_mask_inf.comp index 4e68742b..26d8bc22 100644 --- a/ggml/src/vulkan-shaders/diag_mask_inf.comp +++ b/ggml/src/vulkan-shaders/diag_mask_inf.comp @@ -12,7 +12,7 @@ layout (push_constant) uniform parameter #include "types.comp" -layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x = 1, local_size_y = 512, local_size_z = 1) in; layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; diff --git a/ggml/src/vulkan-shaders/div.comp b/ggml/src/vulkan-shaders/div.comp index 8cfce58b..9fb69c6c 100644 --- a/ggml/src/vulkan-shaders/div.comp +++ b/ggml/src/vulkan-shaders/div.comp @@ -3,12 +3,25 @@ #include "types.comp" #include "generic_binary_head.comp" +const uint num_threads = 256; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + void main() { - const uint idx = get_idx(); + uint idx = get_idx(); - if (idx >= p.ne) { - return; - } + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 2; - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)])); + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)])); + + idx += num_threads; + } } diff --git a/ggml/src/vulkan-shaders/flash_attn.comp b/ggml/src/vulkan-shaders/flash_attn.comp new file mode 100644 index 00000000..ce230a8f --- /dev/null +++ b/ggml/src/vulkan-shaders/flash_attn.comp @@ -0,0 +1,337 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#extension GL_KHR_shader_subgroup_shuffle : enable + +#include "types.comp" +#include "flash_attn_base.comp" + +const uint32_t D_per_thread = D / D_split; + +const uint32_t cols_per_iter = WorkGroupSize / D_split; +const uint32_t cols_per_thread = Bc / cols_per_iter; + + +layout (binding = 0) readonly buffer Q {float data_q[];}; +layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];}; +layout (binding = 1) readonly buffer K {float16_t data_k[];}; +layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];}; +layout (binding = 2) readonly buffer V {float16_t data_v[];}; +layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];}; +layout (binding = 3) readonly buffer M {float16_t data_m[];}; + +// Store the output when doing grouped query attention. +// Rows index by Q's dimension 2, and the first N rows are valid. +D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + uint32_t offset = (iq2 + r) * D + c; + data_o[o_offset + offset] = D_TYPE(elem); + return elem; +} + +shared FLOAT_TYPE tmpsh[WorkGroupSize]; +shared vec4 tmpshv4[WorkGroupSize]; + +shared float masksh[Bc][Br]; +shared vec4 Qf[Br][D / 4]; + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + init_indices(); + + const uint32_t tid = gl_LocalInvocationIndex; + const uint32_t d_tid = gl_LocalInvocationIndex % D_split; + const uint32_t col_tid = gl_LocalInvocationIndex / D_split; + + uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4; + + [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) { + uint32_t d = (idx + tid) % (D / 4); + uint32_t r = (idx + tid) / (D / 4); + if (r < Br && d < D / 4 && + i * Br + r < N) { + Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale; + } + } + barrier(); + + vec4 Of[Br][D_per_thread / 4]; + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] = vec4(0.0); + } + } + + float Lf[Br], Mf[Br]; + + // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M. + const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF); + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Lf[r] = 0; + Mf[r] = NEG_FLT_MAX_OVER_2; + } + + float slope[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + slope[r] = 1.0; + } + + // ALiBi + if (p.max_bias > 0.0f) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2); + } + } + +#if BLOCK_SIZE > 1 + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE; +#else + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2; +#endif + + [[dont_unroll]] + for (uint32_t j = start_j; j < end_j; ++j) { + + float Sf[Br][cols_per_thread]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Sf[r][c] = 0.0; + } + } + + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid); + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K); +#else + vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]); +#endif + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf); + } + } + } + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + // Compute sum across the D_split + [[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Sf[r][c] += subgroupShuffleXor(Sf[r][c], s); + } + } + } + + if (p.logit_softcap != 0.0f) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]); + } + } + } + + if (p.mask != 0) { + + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) % Bc; + uint32_t r = (idx + tid) / Bc; + if (idx + tid < Bc * Br) { + masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]); + } + } + barrier(); + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + float mvf = masksh[c * cols_per_iter + col_tid][r]; + + Sf[r][c] += slope[r]*mvf; + } + } + barrier(); + } + + float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + rowmaxf[r] = Sf[r][0]; + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + rowmaxf[r] = max(rowmaxf[r], Sf[r][c]); + } + Moldf[r] = Mf[r]; + + // M = max(rowmax, Mold) + // P = e^(S - M) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf[r], Moldf[r]); + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Pf[r][c] = exp(Sf[r][c] - Mf[r]); + } + eMf[r] = exp(Moldf[r] - Mf[r]); + + // Compute sum across row of P + rowsumf[r] = 0.0; + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + rowsumf[r] += Pf[r][c]; + } + + Lf[r] = eMf[r]*Lf[r] + rowsumf[r]; + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] = eMf[r] * Of[r][d]; + } + } + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid); + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V); +#else + vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]); +#endif + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] += Pf[r][c] * Vf; + } + } + } + + barrier(); + } + + // reduce across threads + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + float rowmaxf, eMf; + + tmpsh[tid] = Mf[r]; + // Compute max across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]); + } + barrier(); + } + rowmaxf = tmpsh[d_tid]; + barrier(); + + float Moldf = Mf[r]; + + // M = max(rowmax, Mold) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf, Moldf); + eMf = exp(Moldf - Mf[r]); + + Lf[r] = eMf*Lf[r]; + + tmpsh[tid] = Lf[r]; + + // Compute sum across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s]; + } + barrier(); + } + Lf[r] = tmpsh[d_tid]; + barrier(); + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + + Of[r][d] = eMf * Of[r][d]; + tmpshv4[tid] = Of[r][d]; + + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + Of[r][d] += tmpshv4[tid + s]; + tmpshv4[tid] = Of[r][d]; + } + barrier(); + } + Of[r][d] = tmpshv4[d_tid]; + barrier(); + } + } + + + // If there is split_k, then the split_k resolve shader does the final + // division by L. Store the intermediate O value and per-row m and L values. + if (p.k_num > 1) { + uint32_t o_offset = D * p.ne1 * split_k_index; + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); + } + } + } + } + + o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N); + perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N); + } + } + + return; + } + + float Lfrcp[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Lfrcp[r] = 1.0 / Lf[r]; + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] *= Lfrcp[r]; + } + } + + uint32_t o_offset = iq3*p.ne2*p.ne1; + + if (p.gqa_ratio > 1) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); + } + } + } + } + } else { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (i * Br + r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]); + } + } + } + } + } +} diff --git a/ggml/src/vulkan-shaders/flash_attn_base.comp b/ggml/src/vulkan-shaders/flash_attn_base.comp new file mode 100644 index 00000000..61d90e2d --- /dev/null +++ b/ggml/src/vulkan-shaders/flash_attn_base.comp @@ -0,0 +1,162 @@ + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (constant_id = 0) const uint32_t WorkGroupSize = 128; +layout (constant_id = 1) const uint32_t Br = 1; +layout (constant_id = 2) const uint32_t Bc = 32; +layout (constant_id = 3) const uint32_t D = 32; +layout (constant_id = 4) const uint32_t Clamp = 0; +layout (constant_id = 5) const uint32_t D_split = 16; + + +layout (push_constant) uniform parameter { + uint32_t N; + uint32_t KV; + + uint32_t ne1; + uint32_t ne2; + uint32_t ne3; + + uint32_t neq2; + uint32_t neq3; + uint32_t nek2; + uint32_t nek3; + uint32_t nev2; + uint32_t nev3; + uint32_t nem1; + + uint32_t nb01; + uint32_t nb02; + uint32_t nb03; + uint32_t nb11; + uint32_t nb12; + uint32_t nb13; + uint32_t nb21; + uint32_t nb22; + uint32_t nb23; + uint32_t nb31; + + float scale; + float max_bias; + float logit_softcap; + + uint32_t mask; + uint32_t n_head_log2; + float m0; + float m1; + + uint32_t gqa_ratio; + uint32_t split_kv; + uint32_t k_num; +} p; + +layout (binding = 4) writeonly buffer O {D_TYPE data_o[];}; + +#if defined(A_TYPE_PACKED16) +#define BINDING_IDX_K 0 +#define BINDING_IDX_V 1 +layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2]; +#endif + +#if defined(DATA_A_Q4_0) +#define BLOCK_BYTE_SIZE 18 + +vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) { + uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]); + uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]); + uint shift = (iqs & 0x10) >> 2; + vui_lo >>= shift; + vui_hi >>= shift; + + return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f); +} +#endif + +#if defined(DATA_A_Q8_0) +#define BLOCK_BYTE_SIZE 34 +vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) { + const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy; + + return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y); +} +#endif + +#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b)) + + +// Store column zero. This is used to save per-row m and L values for split_k. +ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + if (r < N && c == 0) { + uint32_t offset = iq2 + r; + data_o[o_offset + offset] = D_TYPE(elem); + } + return elem; +} + +// Load the slope matrix, indexed by Q's dimension 2. +ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2) +{ + const uint32_t h = iq2 + (r % p.gqa_ratio); + + const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1); + const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1); + + return ACC_TYPE(pow(base, ACC_TYPE(exph))); +} + +uint32_t i, N, KV, split_k_index, Tr, start_j, end_j, + iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3, + q_stride, k_stride, v_stride, m_stride; + +void init_indices() +{ + N = p.N; + KV = p.KV; + + i = gl_WorkGroupID.x; + split_k_index = 0; + + if (p.k_num > 1) { + i = 0; + split_k_index = gl_WorkGroupID.x; + } + + Tr = CEIL_DIV(N, Br); + + start_j = split_k_index * p.split_kv / Bc; + end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc); + + // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y. + // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2. + iq2 = gl_WorkGroupID.y * p.gqa_ratio; + iq3 = gl_WorkGroupID.z; + + // broadcast factors + rk2 = p.neq2/p.nek2; + rk3 = p.neq3/p.nek3; + + rv2 = p.neq2/p.nev2; + rv3 = p.neq3/p.nev3; + + // k indices + ik3 = iq3 / rk3; + ik2 = iq2 / rk2; + + // v indices + iv3 = iq3 / rv3; + iv2 = iq2 / rv2; + + // nb?1 are already divided by the type size and are in units of elements. + // When using grouped query attention, Q is indexed by iq2, so the stride + // should be nb02 (which is in bytes). + q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01; + k_stride = p.nb11; + v_stride = p.nb21; + // When using grouped query attention, all rows use the same mask (stride 0). + // "p.gqa_ratio >> 16" is just a roundabout way of writing zero + // that prevents the compiler from folding the "&" through the select + // and breaking the alignment detection. + m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV; +} diff --git a/ggml/src/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/vulkan-shaders/flash_attn_cm1.comp new file mode 100644 index 00000000..da478be2 --- /dev/null +++ b/ggml/src/vulkan-shaders/flash_attn_cm1.comp @@ -0,0 +1,360 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#extension GL_KHR_shader_subgroup_basic : enable +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_cooperative_matrix : enable + +#include "types.comp" +#include "flash_attn_base.comp" + +const uint32_t D_per_thread = D / D_split; +const uint32_t row_split = 4; +const uint32_t rows_per_thread = Br / row_split; +const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split; +const uint32_t cols_per_thread = Bc / cols_per_iter; + + +layout (binding = 0) readonly buffer Q {float data_q[];}; +layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];}; +layout (binding = 1) readonly buffer K {float16_t data_k[];}; +layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];}; +layout (binding = 2) readonly buffer V {float16_t data_v[];}; +layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];}; +layout (binding = 3) readonly buffer M {float16_t data_m[];}; + +// Store the output when doing grouped query attention. +// Rows index by Q's dimension 2, and the first N rows are valid. +D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + uint32_t offset = (iq2 + r) * D + c; + data_o[o_offset + offset] = D_TYPE(elem); + return elem; +} + +// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd +const uint32_t MatBr = 16; +const uint32_t MatBc = 16; + +shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x]; +shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x]; + +const uint32_t qstride = D / 4 + 2; // in units of f16vec4 +shared f16vec4 Qf[Br * qstride]; + +// Avoid padding for D==256 to make it fit in 48KB shmem. +const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br; +shared ACC_TYPE sfsh[Bc * sfshstride]; + +const uint32_t kshstride = D / 4 + 2; // in units of f16vec4 +shared f16vec4 ksh[Bc * kshstride]; + +shared float slope[Br]; + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + init_indices(); + + const uint32_t tid = gl_LocalInvocationIndex; + + const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split; + const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup; + const uint32_t d_tid = gl_LocalInvocationIndex % D_split; + const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split; + +#define tile_row(r) (row_tid * rows_per_thread + (r)) + + uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4; + + [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) { + uint32_t d = (idx + tid) % (D / 4); + uint32_t r = (idx + tid) / (D / 4); + if (r < Br && d < D / 4 && + i * Br + r < N) { + Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale); + } + } + barrier(); + + ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4]; + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Of[r][d] = ACC_TYPEV4(0.0); + } + } + + float Lf[rows_per_thread], Mf[rows_per_thread]; + + // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M. + const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF); + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Lf[r] = 0; + Mf[r] = NEG_FLT_MAX_OVER_2; + } + + // ALiBi + if (p.max_bias > 0.0f) { + if (tid < Br) { + uint r = tid; + slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2); + } + barrier(); + } else { + if (tid < Br) { + uint r = tid; + slope[r] = 1.0; + } + barrier(); + } + +#if BLOCK_SIZE > 1 + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE; +#else + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2; +#endif + + [[dont_unroll]] + for (uint32_t j = start_j; j < end_j; ++j) { + + [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) { + uint32_t d = (idx + tid) % (D / 4); + uint32_t c = (idx + tid) / (D / 4); + if (c < Bc && d < D / 4) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d; + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + f16vec4 K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K)); +#else + f16vec4 K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]); +#endif + + ksh[c * kshstride + d] = K_Tf; + } + } + barrier(); + + // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br + // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16 + // This is written transposed in order to allow for N being 8 if implementations need it + coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0); + coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat; + coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat; + + for (uint32_t d = 0; d < D / 16; ++d) { + coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor); + + uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4; + coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor); + + SfMat = coopMatMulAdd(KMat, QMat, SfMat); + } + + uint coord = gl_SubgroupID * MatBc * sfshstride; + coopMatStore(SfMat, sfsh, coord, sfshstride, gl_CooperativeMatrixLayoutRowMajor); + barrier(); + + if (p.logit_softcap != 0.0f) { + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) / Br; + uint32_t r = (idx + tid) % Br; + if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) { + sfsh[c * sfshstride + r] = ACC_TYPE(p.logit_softcap * tanh(sfsh[c * sfshstride + r])); + } + } + barrier(); + } + + if (p.mask != 0) { + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) % Bc; + uint32_t r = (idx + tid) / Bc; + if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) { + sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)])); + } + } + barrier(); + } + + float eMf[rows_per_thread]; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + float rowmaxf = sfsh[tile_row(r) + (0 * cols_per_iter + col_tid) * sfshstride]; + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride])); + } + float Moldf = Mf[r]; + + // M = max(rowmax, Mold) + // P = e^(S - M) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf, Moldf); + eMf[r] = exp(Moldf - Mf[r]); + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Of[r][d] = float16_t(eMf[r]) * Of[r][d]; + } + } + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Lf[r] = eMf[r]*Lf[r]; + } + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + float Pf[rows_per_thread]; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]); + Lf[r] += Pf[r]; + } + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid); + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V); +#else + vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]); +#endif + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf); + } + } + } + + barrier(); + } + + // reduce across threads + + float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread]; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + FLOAT_TYPE M = Mf[r]; + tmpsh[tid] = M; + // Compute max across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { + M = max(M, tmpsh[tid ^ s]); + barrier(); + tmpsh[tid] = M; + barrier(); + } + rowmaxf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup]; + barrier(); + } + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Moldf[r] = Mf[r]; + + // M = max(rowmax, Mold) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf[r], Moldf[r]); + eMf[r] = exp(Moldf[r] - Mf[r]); + + Lf[r] = eMf[r]*Lf[r]; + } + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + FLOAT_TYPE L = Lf[r]; + tmpsh[tid] = L; + // Compute sum across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { + L += tmpsh[tid ^ s]; + barrier(); + tmpsh[tid] = L; + barrier(); + } + Lf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup]; + barrier(); + } + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + + Of[r][d] = float16_t(eMf[r]) * Of[r][d]; + tmpshv4[tid] = Of[r][d]; + + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { + Of[r][d] += tmpshv4[tid ^ s]; + barrier(); + tmpshv4[tid] = Of[r][d]; + barrier(); + } + Of[r][d] = tmpshv4[d_tid + row_tid * threads_per_rowgroup]; + barrier(); + } + } + + // If there is split_k, then the split_k resolve shader does the final + // division by L. Store the intermediate O value and per-row m and L values. + if (p.k_num > 1) { + uint32_t o_offset = D * p.ne1 * split_k_index; + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + if (tile_row(r) < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N); + } + } + } + } + + o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + if (tile_row(r) < N) { + perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N); + perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N); + } + } + + return; + } + + float Lfrcp[rows_per_thread]; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Lfrcp[r] = 1.0 / Lf[r]; + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + Of[r][d] *= float16_t(Lfrcp[r]); + } + } + + uint32_t o_offset = iq3*p.ne2*p.ne1; + + if (p.gqa_ratio > 1) { + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + if (tile_row(r) < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N); + } + } + } + } + } else { + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + if (i * Br + tile_row(r) < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]); + } + } + } + } + } +} diff --git a/ggml/src/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/vulkan-shaders/flash_attn_cm2.comp new file mode 100644 index 00000000..6acf67a0 --- /dev/null +++ b/ggml/src/vulkan-shaders/flash_attn_cm2.comp @@ -0,0 +1,267 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require + +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_cooperative_matrix : enable +#extension GL_NV_cooperative_matrix2 : enable +#extension GL_EXT_buffer_reference : enable +#extension GL_KHR_shader_subgroup_ballot : enable +#extension GL_KHR_shader_subgroup_vote : enable +#extension GL_EXT_null_initializer : enable + +#include "types.comp" +#include "dequant_funcs_cm2.comp" +#include "flash_attn_base.comp" + +layout (binding = 0) readonly buffer Q {uint8_t data_q[];}; +layout (binding = 1) readonly buffer K {uint8_t data_k[];}; +layout (binding = 2) readonly buffer V {uint8_t data_v[];}; +layout (binding = 3) readonly buffer M {uint8_t data_m[];}; + +ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) { + return max(x, y); +} + +ACC_TYPE smearReduce(const in ACC_TYPE x, const in ACC_TYPE y) { + return x; +} + +// Replace matrix elements >= numRows or numCols with 'replace' +ACC_TYPE replacePadding(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem, const in ACC_TYPE replace, const in uint32_t numRows, const in uint32_t numCols) { + if (row >= numRows || col >= numCols) { + return replace; + } + return elem; +} + +ACC_TYPE Exp(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem) +{ + return exp(elem); +} + +ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem0, const in ACC_TYPE elem1) +{ + return max(elem0, elem1); +} + +#if defined(BLOCK_SIZE) +#define DECODEFUNC , DEQUANTFUNC +#else +#define DECODEFUNC +#endif + +// Store the output when doing grouped query attention. +// Rows index by Q's dimension 2, and the first N rows are valid. +D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + if (r < N && c < D) { + uint32_t offset = (iq2 + r) * D + c; + data_o[o_offset + offset] = D_TYPE(elem); + } + return elem; +} + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + init_indices(); + + tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutNV<2, Clamp> tensorLayoutK = createTensorLayoutNV(2, Clamp); + tensorLayoutNV<2, Clamp> tensorLayoutV = createTensorLayoutNV(2, Clamp); + + tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0); + +#if defined(BLOCK_SIZE) + tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE); + tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE); +#endif + + tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D); + tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D); + tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D); + + // hint to the compiler that strides are aligned for the aligned variant of the shader + if (Clamp != gl_CooperativeMatrixClampModeConstantNV) + { + q_stride &= ~7; +#if !defined(BLOCK_SIZE) + k_stride &= ~7; + v_stride &= ~7; +#endif + m_stride &= ~7; + } + tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1); + tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1); + tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1); + + coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q; + coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16; + + uint32_t q_offset = iq2*p.nb02+iq3*p.nb03; + coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D)); + + Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q); + Qf16 *= float16_t(p.scale); + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0); + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M; + + // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M. + const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF); + + L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0); + M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2); + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0); + + // ALiBi + if (p.max_bias > 0.0f) { + coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2); + } + + [[dont_unroll]] + for (uint32_t j = start_j; j < end_j; ++j) { + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0); + + coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T; + + uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13; + coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC); + S = coopMatMulAdd(Qf16, K_T, S); + + if (p.logit_softcap != 0.0f) { + [[unroll]] + for (int k = 0; k < S.length(); ++k) { + S[k] = ACC_TYPE(p.logit_softcap)*tanh(S[k]); + } + } + + if (p.mask != 0) { + tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp); + tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV); + tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1); + + coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv; + + coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc)); + + S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv); + } + + // Clear padding elements to -inf, so they don't contribute to rowmax + if (Clamp != 0 && + ((j + 1) * Bc > KV || + (i + 1) * Br > N)) { + + uint R = ((i + 1) * Br > N) ? (N % Br) : Br; + uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc; + + coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(NEG_FLT_MAX_OVER_2), R, C); + } + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> rowmax, P, rowsum, eM; + + coopMatReduceNV(rowmax, S, gl_CooperativeMatrixReduceRowNV, maxReduce); + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> Mold = M; + + // M = max(rowmax, Mold) + // P = e^(S - M) + // eM = e^(Mold - M) + coopMatPerElementNV(M, rowmax, Max, Mold); + coopMatPerElementNV(P, S - M, Exp); + coopMatPerElementNV(eM, Mold - M, Exp); + + // Clear padding elements to 0, so they don't contribute to rowsum + if (Clamp != 0 && + ((j + 1) * Bc > KV || + (i + 1) * Br > N)) { + + uint R = ((i + 1) * Br > N) ? (N % Br) : Br; + uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc; + + coopMatPerElementNV(P, P, replacePadding, ACC_TYPE(0.0), R, C); + } + + coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA> P_A = coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA>(P); + + // compute rowsum by multiplying by matrix of all ones. + coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB> One = coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB>(1.0); + + rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0); + rowsum = coopMatMulAdd(P_A, One, rowsum); + + coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V; + uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23; + coopMatLoadTensorNV(V, data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC); + + L = eM*L + rowsum; + + // This is the "diagonal" matrix in the paper, but since we do componentwise + // multiply rather than matrix multiply it has the diagonal element smeared + // across the row + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag; + + // resize eM by using smear/reduce + coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce); + + // multiply with fp16 accumulation, then add to O. + coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0); + PV = coopMatMulAdd(P_A, V, PV); + + O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV); + } + + // If there is split_k, then the split_k resolve shader does the final + // division by L. Store the intermediate O value and per-row m and L values. + if (p.k_num > 1) { + coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O); + + uint32_t o_offset = D * p.ne1 * split_k_index; + coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N); + + o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2; + coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N); + coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N); + return; + } + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag; + + // resize L by using smear/reduce + coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce); + + [[unroll]] + for (int k = 0; k < Ldiag.length(); ++k) { + Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k]; + } + + O = Ldiag*O; + + uint32_t o_offset = iq3*p.ne2*p.ne1; + + coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O); + if (p.gqa_ratio > 1) { + coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N); + } else { + tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D); + + // permute dimensions + tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2); + + coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute); + } +} diff --git a/ggml/src/vulkan-shaders/flash_attn_split_k_reduce.comp b/ggml/src/vulkan-shaders/flash_attn_split_k_reduce.comp new file mode 100644 index 00000000..a7e39568 --- /dev/null +++ b/ggml/src/vulkan-shaders/flash_attn_split_k_reduce.comp @@ -0,0 +1,59 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#define BLOCK_SIZE 32 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {float data_a[];}; +layout (binding = 1) writeonly buffer D {float data_d[];}; + +layout (push_constant) uniform parameter { + uint D; + uint N; + uint k_num; +} p; + +void main() { + // Each workgroup handles a row + const uint n = gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + uint D = p.D; + uint N = p.N; + uint k_num = p.k_num; + + uint l_offset = D * N * k_num + n; + uint m_offset = D * N * k_num + N + n; + uint lm_stride = N * 2; + + // Compute the max m value for the row + float m_max = -1.0/0.0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + float m = data_a[m_offset + k * lm_stride]; + m_max = max(m_max, m); + } + + // Compute L based on m_max + float L = 0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + float l = data_a[l_offset + k * lm_stride]; + float m = data_a[m_offset + k * lm_stride]; + L += exp(m - m_max) * l; + } + + L = 1.0 / L; + + // Scale and sum the O contributions based on m_max and store the result to memory + for (uint d = tid; d < D; d += BLOCK_SIZE) { + float O = 0.0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + uint o_offset = D * N * k + D * n + d; + float m = data_a[m_offset + k * lm_stride]; + O += exp(m - m_max) * data_a[o_offset]; + } + O *= L; + data_d[D * n + d] = O; + } +} diff --git a/ggml/src/vulkan-shaders/generic_binary_head.comp b/ggml/src/vulkan-shaders/generic_binary_head.comp index b6beaff1..062e2a4c 100644 --- a/ggml/src/vulkan-shaders/generic_binary_head.comp +++ b/ggml/src/vulkan-shaders/generic_binary_head.comp @@ -1,4 +1,5 @@ #extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_control_flow_attributes : require layout (push_constant) uniform parameter { @@ -6,47 +7,58 @@ layout (push_constant) uniform parameter uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03; uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13; uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23; - uint d_offset; + uint misalign_offsets; float param1; float param2; int param3; } p; -layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; - layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; +// true if src0/src1 are the same shape and the indices can be reused without additional modulus +layout(constant_id = 0) const bool norepeat = false; + uint get_idx() { return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; } -uint src0_idx(uint idx) { - const uint i03 = idx / (p.ne02*p.ne01*p.ne00); - const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00; - const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00); - const uint i02_offset = i02*p.ne01*p.ne00; - const uint i01 = (idx - i03_offset - i02_offset) / p.ne00; - const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00; - return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00; +uint get_aoffset() { return p.misalign_offsets >> 16; } +uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; } +uint get_doffset() { return p.misalign_offsets & 0xFF; } + +// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1 +uint fastmod(uint a, uint b) { + if ((b & (b-1)) == 0) { + return a & (b-1); + } + return a % b; } -uint src1_idx(uint idx) { - const uint i03 = idx / (p.ne02*p.ne01*p.ne00); +uint fastdiv(uint a, uint b) { + return (a < b) ? 0 : (a / b); +} + +void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) { + i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00)); const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00; - const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00); + i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00)); const uint i02_offset = i02*p.ne01*p.ne00; - const uint i01 = (idx - i03_offset - i02_offset) / p.ne00; - const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00; + i01 = (idx - i03_offset - i02_offset) / p.ne00; + i00 = idx - i03_offset - i02_offset - i01*p.ne00; +} + +uint src0_idx(uint i00, uint i01, uint i02, uint i03) { + return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00; +} - return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10; +uint src1_idx(uint i00, uint i01, uint i02, uint i03) { + if (norepeat) { + return i03*p.nb13 + i02*p.nb12 + i01*p.nb11 + i00*p.nb10; + } else { + return fastmod(i03, p.ne13)*p.nb13 + fastmod(i02, p.ne12)*p.nb12 + fastmod(i01, p.ne11)*p.nb11 + fastmod(i00, p.ne10)*p.nb10; + } } -uint dst_idx(uint idx) { - const uint i23 = idx / (p.ne22*p.ne21*p.ne20); - const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20; - const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20); - const uint i22_offset = i22*p.ne21*p.ne20; - const uint i21 = (idx - i23_offset - i22_offset) / p.ne20; - const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20; - return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20; +uint dst_idx(uint i00, uint i01, uint i02, uint i03) { + return i03*p.nb23 + i02*p.nb22 + i01*p.nb21 + i00*p.nb20; } diff --git a/ggml/src/vulkan-shaders/generic_unary_head.comp b/ggml/src/vulkan-shaders/generic_unary_head.comp index eacdefc7..8dc9d360 100644 --- a/ggml/src/vulkan-shaders/generic_unary_head.comp +++ b/ggml/src/vulkan-shaders/generic_unary_head.comp @@ -1,15 +1,21 @@ #extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_control_flow_attributes : require layout (push_constant) uniform parameter { uint ne; uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03; uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13; - uint d_offset; + uint misalign_offsets; float param1; float param2; -} p; -layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + uint ne0_012mp; uint ne0_012L; + uint ne0_01mp; uint ne0_01L; + uint ne0_0mp; uint ne0_0L; + uint ne1_012mp; uint ne1_012L; + uint ne1_01mp; uint ne1_01L; + uint ne1_0mp; uint ne1_0L; +} p; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; @@ -18,22 +24,53 @@ uint get_idx() { return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; } +uint get_aoffset() { return p.misalign_offsets >> 16; } +uint get_doffset() { return p.misalign_offsets & 0xFFFF; } + +// see init_fastdiv_values in ggml-vulkan.cpp +uint fastdiv(uint n, uint mp, uint L) { + uint msbs, lsbs; + // msbs = mulhi(n, mp) + umulExtended(n, mp, msbs, lsbs); + return (msbs + n) >> L; +} + uint src0_idx(uint idx) { - const uint i03 = idx / (p.ne02*p.ne01*p.ne00); + const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L); const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00; - const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00); + const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L); const uint i02_offset = i02*p.ne01*p.ne00; - const uint i01 = (idx - i03_offset - i02_offset) / p.ne00; + const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L); const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00; return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00; } uint dst_idx(uint idx) { - const uint i13 = idx / (p.ne12*p.ne11*p.ne10); + const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L); const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; - const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10); + const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L); const uint i12_offset = i12*p.ne11*p.ne10; - const uint i11 = (idx - i13_offset - i12_offset) / p.ne10; + const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L); const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10; } + +uint src0_idx_quant(uint idx, uint qk) { + const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L); + const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00; + const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L); + const uint i02_offset = i02*p.ne01*p.ne00; + const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L); + const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00; + return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + (i00/qk)*p.nb00; +} + +uint dst_idx_quant(uint idx, uint qk) { + const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L); + const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; + const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L); + const uint i12_offset = i12*p.ne11*p.ne10; + const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L); + const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; + return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + (i10/qk)*p.nb10; +} diff --git a/ggml/src/vulkan-shaders/get_rows.comp b/ggml/src/vulkan-shaders/get_rows.comp index e9ff22ef..ee6b86a1 100644 --- a/ggml/src/vulkan-shaders/get_rows.comp +++ b/ggml/src/vulkan-shaders/get_rows.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_binary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint i00 = gl_GlobalInvocationID.x; const uint i10 = gl_GlobalInvocationID.y; @@ -13,14 +15,19 @@ void main() { return; } - const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12]; + const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12]; - const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03; - const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23; + const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03; + const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23; +#if defined(DATA_A_BF16) + FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00])); +#else + FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]); +#endif #ifndef OPTIMIZATION_ERROR_WORKAROUND - data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]); + data_d[d_offset + i00] = D_TYPE(v); #else - data_d[d_offset + i00] = data_a[a_offset + i00]; + data_d[d_offset + i00] = D_TYPE(v); #endif } diff --git a/ggml/src/vulkan-shaders/get_rows_quant.comp b/ggml/src/vulkan-shaders/get_rows_quant.comp index 53a9a96f..cfd645a3 100644 --- a/ggml/src/vulkan-shaders/get_rows_quant.comp +++ b/ggml/src/vulkan-shaders/get_rows_quant.comp @@ -1,15 +1,23 @@ #version 450 +#extension GL_EXT_control_flow_attributes : enable + #include "types.comp" #include "generic_binary_head.comp" #include "dequant_funcs.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint i00 = (gl_GlobalInvocationID.x)*2; const uint i10 = gl_GlobalInvocationID.y; const uint i11 = (gl_GlobalInvocationID.z)/p.ne12; const uint i12 = (gl_GlobalInvocationID.z)%p.ne12; +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + if (i00 >= p.ne00) { return; } @@ -25,6 +33,8 @@ void main() { const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2; vec2 v = dequantize(ib, iqs, 0); + const vec2 dm = get_dm(ib, 0); + v = v * dm.x + dm.y; data_d[d_offset + iybs + iqs ] = D_TYPE(v.x); data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y); diff --git a/ggml/src/vulkan-shaders/group_norm.comp b/ggml/src/vulkan-shaders/group_norm.comp index 5ad9b28d..b6a0d564 100644 --- a/ggml/src/vulkan-shaders/group_norm.comp +++ b/ggml/src/vulkan-shaders/group_norm.comp @@ -19,7 +19,7 @@ void main() { const uint tid = gl_LocalInvocationID.x; const uint start = gl_WorkGroupID.x * group_size + tid; - const uint end = start + group_size; + const uint end = (gl_WorkGroupID.x + 1) * group_size; tmp[tid] = 0.0f; diff --git a/ggml/src/vulkan-shaders/im2col.comp b/ggml/src/vulkan-shaders/im2col.comp index 4d48610a..09aa849e 100644 --- a/ggml/src/vulkan-shaders/im2col.comp +++ b/ggml/src/vulkan-shaders/im2col.comp @@ -1,6 +1,12 @@ #version 450 #extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_spirv_intrinsics: enable +#extension GL_EXT_control_flow_attributes : require + +#if RTE16 +spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits +#endif layout (push_constant) uniform parameter { @@ -18,40 +24,77 @@ layout (push_constant) uniform parameter #include "types.comp" -#define BLOCK_SIZE 256 +layout(constant_id = 0) const uint BLOCK_SIZE = 32; + +const uint NUM_ITER = 512 / BLOCK_SIZE; -layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; void main() { - const uint i = gl_GlobalInvocationID.x; - if (i >= p.pelements) { - return; - } - - const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1); - const uint kx = i / ksize; - const uint kd = kx * ksize; - const uint ky = (i - kd) / p.OW; - const uint ix = i % p.OW; + const uint gidx = gl_GlobalInvocationID.x; const uint oh = gl_GlobalInvocationID.y; const uint batch = gl_GlobalInvocationID.z / p.IC; const uint ic = gl_GlobalInvocationID.z % p.IC; - const uint iiw = ix * p.s0 + kx * p.d0 - p.p0; - const uint iih = oh * p.s1 + ky * p.d1 - p.p1; + const uint src_base = ic * p.offset_delta + batch * p.batch_offset; + const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH); + const int oh_s1 = int(oh) * p.s1; + const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1); + + const uint base_linear_idx = gidx * NUM_ITER; + + const uint max_ky = ksize / p.OW; + + uint current_kx = base_linear_idx / ksize; + const uint rem = base_linear_idx - (current_kx * ksize); + uint current_ky = rem / p.OW; + uint current_ix = rem % p.OW; + + A_TYPE values[NUM_ITER]; + uint offset_dst[NUM_ITER]; + [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { + values[idx] = A_TYPE(0); + } + + [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { + + const uint linear_idx = base_linear_idx + idx; + + if (linear_idx >= p.pelements) { + continue; + } - const uint offset_dst = - ((batch * p.OH + oh) * p.OW + ix) * p.CHW + - (ic * (p.KW * p.KH) + ky * p.KW + kx); + const uint iiw = current_ix * p.s0 + current_kx * p.d0 - p.p0; + const uint iih = oh_s1 + current_ky * p.d1 - p.p1; - if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) { - data_d[offset_dst] = D_TYPE(0.0f); - } else { - const uint offset_src = ic * p.offset_delta + batch * p.batch_offset; - data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]); + offset_dst[idx] = dst_base + current_ix * p.CHW + current_ky * p.KW + current_kx; + + if ((iih < p.IH) && (iiw < p.IW)) { + values[idx] = data_a[src_base + iih * p.IW + iiw]; + } + + if (++current_ix == p.OW) { + current_ix = 0; + if (++current_ky == max_ky) { + current_ky = 0; + current_kx++; + } + } } + + [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { + + const uint linear_idx = base_linear_idx + idx; + + if (linear_idx >= p.pelements) { + continue; + } + + data_d[offset_dst[idx]] = D_TYPE(values[idx]); + } + } diff --git a/ggml/src/vulkan-shaders/mul.comp b/ggml/src/vulkan-shaders/mul.comp index bfb61c92..43de19df 100644 --- a/ggml/src/vulkan-shaders/mul.comp +++ b/ggml/src/vulkan-shaders/mul.comp @@ -3,12 +3,25 @@ #include "types.comp" #include "generic_binary_head.comp" +const uint num_threads = 256; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + void main() { - const uint idx = get_idx(); + uint idx = get_idx(); - if (idx >= p.ne) { - return; - } + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 2; - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)])); + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)])); + + idx += num_threads; + } } diff --git a/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp b/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp index 825b9103..4c64fd47 100644 --- a/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp +++ b/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp @@ -5,7 +5,9 @@ layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {float data_a[];}; +layout (binding = 0) readonly buffer A4 {vec4 data_a4[];}; layout (binding = 1) writeonly buffer D {float data_d[];}; +layout (binding = 1) writeonly buffer D4 {vec4 data_d4[];}; layout (push_constant) uniform parameter { uint ne; @@ -13,17 +15,34 @@ layout (push_constant) uniform parameter { } p; void main() { - const uint idx = gl_GlobalInvocationID.x; + // Each invocation handles four consecutive components + const uint idx = gl_GlobalInvocationID.x * 4; if (idx >= p.ne) { return; } - float result = 0.0f; + // Check if all four components are in bounds and aligned, + // then use vector loads + if (idx + 3 < p.ne && (p.ne % 4) == 0) { + vec4 result = vec4(0.0f); - [[unroll]] for (uint i = 0; i < p.k_num; i++) { - result += data_a[i * p.ne + idx]; - } + [[unroll]] for (uint i = 0; i < p.k_num; i++) { + result += data_a4[(i * p.ne + idx) / 4]; + } + + data_d4[idx / 4] = result; + } else { + [[unroll]] for (uint j = 0; j < 4; ++j) { + if (idx + j < p.ne) { + float result = 0.0f; - data_d[idx] = result; + [[unroll]] for (uint i = 0; i < p.k_num; i++) { + result += data_a[i * p.ne + idx + j]; + } + + data_d[idx + j] = result; + } + } + } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec.comp b/ggml/src/vulkan-shaders/mul_mat_vec.comp index 46a6369b..bb429dd5 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec.comp @@ -1,57 +1,169 @@ #version 450 -#ifdef FLOAT16 -#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require -#endif +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require #include "mul_mat_vec_base.comp" layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -layout (constant_id = 0) const uint BLOCK_SIZE = 32; +#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16) +#define K_PER_ITER 8 +#else +#define K_PER_ITER 2 +#endif -shared FLOAT_TYPE tmp[BLOCK_SIZE]; -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; - const uint tid = gl_LocalInvocationID.x; +uint a_offset, b_offset, d_offset, y_offset; + +void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter) +{ + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const uint col = i*BLOCK_SIZE + K_PER_ITER*tid; + const uint iqs = (col%QUANT_K)/QUANT_R; // quant index + const uint iybs = col - col%QUANT_K; // y block start index + +#if K_PER_ITER == 8 +#if QUANT_R == 2 + const vec4 bv02 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]); + const vec4 bv13 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4]); + const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y); + const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w); +#else + const vec4 bv0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]); + const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]); +#endif +#else + // Check if the second of the pair of elements is OOB, and don't fetch B or + // accumulate it. We still fetch a pair of elements for A, which is fine for + // quantized formats since they'll be within the same block. We should + // probably skip fetching the second element for F16/F32, but as of now we + // still do. + const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols); + + FLOAT_TYPE b0 = 0, b1 = 0; + b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]); + if (!OOB) { + b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]); + } +#endif + uint ibi = first_row*p.ncols; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib = (ibi + col)/QUANT_K; // block index + ibi += p.ncols; + +#if K_PER_ITER == 8 + vec4 v = dequantize4(ib, iqs, a_offset); + vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset); + + const vec2 dm = get_dm(ib, a_offset); + if (dm.y != 0) { // quant has min component + v = v * dm.x + dm.y; + v2 = v2 * dm.x + dm.y; + } + + // matrix multiplication + FLOAT_TYPE rowtmp = dot(bv0, v); + rowtmp += dot(bv1, v2); + + if (dm.y == 0) + rowtmp *= dm.x; + + temp[j][n] += rowtmp; +#else + const vec2 v = dequantize(ib, iqs, a_offset); - // There are not enough cols to use all threads - if (tid >= p.ncols) { - return; + // matrix multiplication + temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]); + if (!OOB) { + temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]); + } +#endif + } } +} - const uint block_size = min(p.ncols, BLOCK_SIZE); +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + const uint tid = gl_LocalInvocationID.x; - uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); + a_offset /= QUANT_K; - const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2; + y_offset = QUANT_R == 1 ? 1 : QUANT_K/2; - tmp[tid] = FLOAT_TYPE(0.0f); + FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; - [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) { - const uint col = i*block_size + 2*tid; - const uint ib = (row*p.ncols + col)/QUANT_K; // block index - const uint iqs = (col%QUANT_K)/QUANT_R; // quant index - const uint iybs = col - col%QUANT_K; // y block start index + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } - vec2 v = dequantize(ib, iqs, a_offset / QUANT_K); + uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE); + if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) { + num_iters++; + } + int unroll_count = 4; + uint unrolled_iters = num_iters & ~(unroll_count - 1); - // matrix multiplication - tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) + - FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]); +#if K_PER_ITER == 2 + // If the K dimension is odd, we need lastiter==true on the last iteration + // so OOB is computed correctly. Skip some unrolling to make that happen. + if ((p.ncols & 1) != 0 && + unrolled_iters == num_iters && + unrolled_iters > 0) { + unrolled_iters -= unroll_count; } +#endif - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; + uint i = 0; + while (i < unrolled_iters) { + // Manually partially unroll the loop + [[unroll]] for (uint k = 0; k < unroll_count; ++k) { + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; } - barrier(); } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + + unroll_count = 2; + unrolled_iters = num_iters & ~(unroll_count - 1); + +#if K_PER_ITER == 2 + if ((p.ncols & 1) != 0 && + unrolled_iters == num_iters && + unrolled_iters > 0) { + unrolled_iters -= unroll_count; + } +#endif + + while (i < unrolled_iters) { + // Manually partially unroll the loop + [[unroll]] for (uint k = 0; k < unroll_count; ++k) { + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; + } + } + while (i < num_iters) { + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true); + i++; + } + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/vulkan-shaders/mul_mat_vec_base.comp index 5920bc93..903753c7 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_base.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_base.comp @@ -2,8 +2,6 @@ #extension GL_EXT_shader_16bit_storage : require #extension GL_EXT_shader_8bit_storage : require -#define K_QUANTS_PER_ITERATION 2 - #ifdef MUL_MAT_ID #define EXPERT_COUNT 8 #endif @@ -12,6 +10,9 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; +layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; @@ -49,13 +50,16 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) { #endif #ifndef MUL_MAT_ID - const uint i13 = batch_idx / p.ne12; - const uint i12 = batch_idx % p.ne12; + uint batch_idx_a = 0; + if (batch_idx != 0) { + const uint i13 = batch_idx / p.ne12; + const uint i12 = batch_idx % p.ne12; - const uint i03 = i13 / p.broadcast3; - const uint i02 = i12 / p.broadcast2; + const uint i03 = i13 / p.broadcast3; + const uint i02 = i12 / p.broadcast2; - const uint batch_idx_a = i03 * p.ne02 + i02; + batch_idx_a = i03 * p.ne02 + i02; + } #else const uint expert_id = data_ids[expert_idx]; #endif @@ -79,3 +83,36 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) { batch_idx * p.batch_stride_d; #endif } + +layout (constant_id = 0) const uint BLOCK_SIZE = 32; +layout (constant_id = 1) const uint NUM_ROWS = 1; +layout (constant_id = 2) const uint NUM_COLS = 1; + +shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE]; + +void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) { + // sum up partial sums and write back result + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + tmpsh[j][n][tid] = temp[j][n]; + } + } + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) { + if (tid < s) { + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + tmpsh[j][n][tid] += tmpsh[j][n][tid + s]; + } + } + } + barrier(); + } + if (tid == 0) { + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]); + } + } + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq1_m.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq1_m.comp new file mode 100644 index 00000000..e4acbd4f --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq1_m.comp @@ -0,0 +1,82 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint16_t[4] scales = data_a[ibi].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + + const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1)); + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1)); + const uint qs = data_a[ibi].qs[4 * ib32 + l]; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1); + + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (int k = 0; k < 4; ++k) { + sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta, + fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum)); + } + temp[j][n] = fma(dl, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq1_s.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq1_s.comp new file mode 100644 index 00000000..309da099 --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq1_s.comp @@ -0,0 +1,79 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint qh = data_a[ibi].qh[ib32]; + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qs = data_a[ibi].qs[4 * ib32 + l]; + const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (int k = 0; k < 4; ++k) { + sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta, + fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum)); + } + temp[j][n] = fma(dl, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq2_s.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_s.comp new file mode 100644 index 00000000..8d01536f --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_s.comp @@ -0,0 +1,90 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint nibble_shift = 4 * (itid & 1); + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF; + const float db = d * (0.5 + scale) * 0.25; + + const uint qh = data_a[ibi].qh[ib32]; + const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147 + const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy; + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint8_t sign = sign16[l]; + const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300); + const uvec2 grid = iq2s_grid[qs]; + const vec4 grid0 = vec4(unpack8(grid.x)); + const vec4 grid1 = vec4(unpack8(grid.y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xs.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xs.comp new file mode 100644 index 00000000..c4960432 --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xs.comp @@ -0,0 +1,87 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint nibble_shift = 4 * (itid & 1); + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF; + const float db = d * (0.5 + scale) * 0.25; + + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs = data_a[ibi].qs[2 * itid + l]; + const uint sign = qs >> 9; + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x)); + const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xxs.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xxs.comp new file mode 100644 index 00000000..94d4b92e --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq2_xxs.comp @@ -0,0 +1,87 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint signscale = pack32(u16vec2( + data_a_packed16[ibi].qs[4 * ib32 + 2], + data_a_packed16[ibi].qs[4 * ib32 + 3])); + const float db = d * 0.25 * (0.5 + (signscale >> 28)); + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs = data_a[ibi].qs[8 * ib32 + 2 * (itid & 1) + l]; + const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7); + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq2xxs_grid[qs].x)); + const vec4 grid1 = vec4(unpack8(iq2xxs_grid[qs].y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq3_s.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq3_s.comp new file mode 100644 index 00000000..f021e404 --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq3_s.comp @@ -0,0 +1,90 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const float dscale = d * (1 + 2 * scale); + const uint qh = data_a[ibi].qh[ib32]; + FLOAT_TYPE sum[NUM_COLS]; + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + sum[j] = 0.0; + } + [[unroll]] for (uint l = 0; l < 4; ++l) { + const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147 + const uint sign = data_a[ibi].signs[4 * ib32 + l]; + const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)])); + const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)])); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + sum[j] = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w), + sum[j])))))))); + } + } + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + temp[j][n] = fma(dscale, sum[j], temp[j][n]); + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_iq3_xxs.comp b/ggml/src/vulkan-shaders/mul_mat_vec_iq3_xxs.comp new file mode 100644 index 00000000..3fe9dc3a --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mat_vec_iq3_xxs.comp @@ -0,0 +1,88 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint signscale = pack32(u16vec2( + data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32], + data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32 + 1])); + const float db = d * 0.5 * (0.5 + (signscale >> 28)); + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs0 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l]; + const uint qs1 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l + 1]; + const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7); + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq3xxs_grid[qs0])); + const vec4 grid1 = vec4(unpack8(iq3xxs_grid[qs1])); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp index cb3f3c0d..bc633369 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp @@ -12,13 +12,18 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE dst[];}; +layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + layout (push_constant) uniform parameter { uint ncols_x; uint nrows_x; uint row_stride_x; uint channel_stride_x; + uint channel_stride_y; uint channel_x_divisor; + uint ne12; uint b_offset; uint d_offset; } p; @@ -30,6 +35,7 @@ void main() { const uint row_x = gl_GlobalInvocationID.y; const uint channel = gl_GlobalInvocationID.z; const uint channel_x = channel / p.channel_x_divisor; + const uint channel_y = channel % p.ne12; const uint nrows_y = p.ncols_x; const uint nrows_dst = p.nrows_x; @@ -37,25 +43,66 @@ void main() { const uint idst = channel*nrows_dst + row_dst; - tmp[tid] = 0.0f; + FLOAT_TYPE temp = 0.0f; - for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) { - const uint col_x = col_x0 + tid; + // Detect alignment for vector loads + bool is_aligned = (p.ncols_x % 4) == 0 && (p.row_stride_x % 4) == 0 && (p.channel_stride_x % 4) == 0; - if (col_x >= p.ncols_x) { - break; - } + for (uint col_x0 = 0; col_x0 < p.ncols_x;) { + + // Unroll 2x and do vec4 loads if aligned + const uint unroll_count = 2; + if (col_x0 + unroll_count * 4 * BLOCK_SIZE <= p.ncols_x && is_aligned) { + [[unroll]] for (uint i = 0; i < unroll_count; ++i) { + const uint col_x = col_x0 + 4*tid; + + const uint row_y = col_x; + + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel_y*p.channel_stride_y + row_y; + + const vec4 av4 = vec4(data_a_v4[ix / 4]); + const vec4 bv4 = vec4(data_b_v4[iy / 4]); + + temp += dot(av4, bv4); + + col_x0 += 4*BLOCK_SIZE; + } + // do vec4 loads if aligned + } else if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) { + const uint col_x = col_x0 + 4*tid; - const uint row_y = col_x; + const uint row_y = col_x; - const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel_y*p.channel_stride_y + row_y; - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + const vec4 av4 = vec4(data_a_v4[ix / 4]); + const vec4 bv4 = vec4(data_b_v4[iy / 4]); - tmp[tid] += xi * FLOAT_TYPE(data_b[iy]); + temp += dot(av4, bv4); + + col_x0 += 4*BLOCK_SIZE; + } else { + const uint col_x = col_x0 + tid; + if (col_x >= p.ncols_x) { + break; + } + + const uint row_y = col_x; + + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel_y*p.channel_stride_y + row_y; + + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + + temp = fma(xi, FLOAT_TYPE(data_b[iy]), temp); + col_x0 += BLOCK_SIZE; + } } + tmp[tid] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp b/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp index 4b1871ca..7aa070ee 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp @@ -2,16 +2,25 @@ #extension GL_EXT_control_flow_attributes : enable #extension GL_EXT_shader_16bit_storage : require +#if USE_SUBGROUP_ADD +#extension GL_KHR_shader_subgroup_arithmetic : enable +#endif -#define BLOCK_SIZE 32 #define FLOAT_TYPE float -layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE dst[];}; +layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + +layout(constant_id = 0) const int BLOCK_SIZE = 32; +// gqa_ratio is in the range [1,8] +layout(constant_id = 1) const uint gqa_ratio = 1; + layout (push_constant) uniform parameter { uint ncols_x; @@ -22,52 +31,124 @@ layout (push_constant) uniform parameter uint d_offset; } p; -shared FLOAT_TYPE tmp[BLOCK_SIZE]; +#if !USE_SUBGROUP_ADD +shared FLOAT_TYPE tmp[8][BLOCK_SIZE]; +#endif void main() { const uint tid = gl_LocalInvocationID.x; const uint row_x = gl_GlobalInvocationID.y; - const uint channel = gl_GlobalInvocationID.z; - const uint channel_x = channel / (p.nchannels_y / p.nchannels_x); + + uint channel, channel_x; + + // When gqa_ratio > 1, each invocation does multiple rows. + // The row in the A matrix is starting from channel / gqa_ratio and the + // rows in the B matrix are [channel, channel+gqa_ratio). + // When gpa_ratio is 1, each invocation does one row. + if (gqa_ratio > 1) { + channel_x = gl_GlobalInvocationID.z; + channel = channel_x * gqa_ratio; + } else { + channel = gl_GlobalInvocationID.z; + channel_x = channel / (p.nchannels_y / p.nchannels_x);; + } const uint nrows_y = p.ncols_x; const uint nrows_dst = p.nrows_x; const uint row_dst = row_x; - tmp[tid] = FLOAT_TYPE(0.0f); + FLOAT_TYPE temp[8]; + [[unroll]] for (uint i = 0; i < 8; ++i) { + temp[i] = FLOAT_TYPE(0.0f); + } + + // Detect alignment for vector loads + bool is_aligned = (p.ncols_x % 4) == 0 && (p.nchannels_x % 4) == 0 && (nrows_y % 4) == 0; for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) { - const uint col_x = col_x0 + tid; - if (col_x >= p.ncols_x) { - break; - } + // Use vec4 loads if aligned + if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) { - // x is transposed and permuted - const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + uint col_x = col_x0 + 4*tid; + const uint row_y = col_x; - const uint row_y = col_x; + // x is transposed and permuted + const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; + const vec4 av4 = vec4(data_a_v4[ix / 4]); - // y is not transposed but permuted - const uint iy = channel*nrows_y + row_y; + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // y is not transposed but permuted + const uint iy = (channel + c)*nrows_y + row_y; - tmp[tid] += xi * FLOAT_TYPE(data_b[iy]); - } + vec4 bv4 = data_b_v4[iy / 4]; + temp[c] += dot(av4, bv4); + } + + col_x0 += 3*BLOCK_SIZE; + } else { + const uint col_x = col_x0 + tid; + + if (col_x >= p.ncols_x) { + break; + } - // dst is not transposed and not permuted - const uint idst = channel*nrows_dst + row_dst; + // x is transposed and permuted + const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + const uint row_y = col_x; + + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // y is not transposed but permuted + const uint iy = (channel + c)*nrows_y + row_y; + + temp[c] = fma(xi, FLOAT_TYPE(data_b[iy]), temp[c]); + } + } + } + +#if USE_SUBGROUP_ADD + // reduce vec4 at a time + vec4 t = vec4(temp[0], temp[1], temp[2], temp[3]); + t = subgroupAdd(t); + temp[0] = t[0]; + temp[1] = t[1]; + temp[2] = t[2]; + temp[3] = t[3]; + if (gqa_ratio > 4) { + t = vec4(temp[4], temp[5], temp[6], temp[7]); + t = subgroupAdd(t); + temp[4] = t[0]; + temp[5] = t[1]; + temp[6] = t[2]; + temp[7] = t[3]; + } +#else + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + tmp[c][tid] = temp[c]; + } // sum up partial sums and write back result barrier(); [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { if (tid < s) { - tmp[tid] += tmp[tid + s]; + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + temp[c] += tmp[c][tid + s]; + tmp[c][tid] = temp[c]; + } } barrier(); } + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + temp[c] = tmp[c][tid]; + } +#endif if (tid == 0) { - dst[idst] = tmp[0]; + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // dst is not transposed and not permuted + const uint idst = (channel + c)*nrows_dst + row_dst; + dst[idst] = temp[c]; + } } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp index 4cd97799..423ceb8a 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -1,73 +1,130 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require #include "mul_mat_vec_base.comp" -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE tmp[32]; +shared FLOAT_TYPE sccache1[2][BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache2[2][BLOCK_SIZE/16][16]; -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { + const uint y_idx = i * QUANT_K + y_offset; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; + + if (!all_threads) { // when we don't have enough blocks to use all threads + if (i < num_blocks_per_row) { + const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + } + barrier(); + + if (i >= num_blocks_per_row) + continue; + } else { + const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + barrier(); + } + + const uint32_t qs_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); + const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303)); + const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303)); + const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303)); + const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303)); + + vec2 d = vec2(data_a[ib0 + i].d); + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]); + vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]); + vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]); + vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]); + vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]); + vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]); + vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]); + vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]); + + FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); + FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); + [[unroll]] for (int l = 0; l < 2; ++l) { + sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[csel][ix][ 8*v_im] * qs_u32_0[l ], + fma(FLOAT_TYPE(b16[l]), sccache1[csel][ix][1 + 8*v_im] * qs_u32_0[l+2], + fma(FLOAT_TYPE(b32[l]), sccache1[csel][ix][2 + 8*v_im] * qs_u32_2[l ], + fma(FLOAT_TYPE(b48[l]), sccache1[csel][ix][3 + 8*v_im] * qs_u32_2[l+2], + fma(FLOAT_TYPE(b64[l]), sccache1[csel][ix][4 + 8*v_im] * qs_u32_4[l ], + fma(FLOAT_TYPE(b80[l]), sccache1[csel][ix][5 + 8*v_im] * qs_u32_4[l+2], + fma(FLOAT_TYPE(b96[l]), sccache1[csel][ix][6 + 8*v_im] * qs_u32_6[l ], + fma(FLOAT_TYPE(b112[l]), sccache1[csel][ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); + sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[csel][ix][ 8*v_im], + fma(FLOAT_TYPE(b16[l]), sccache2[csel][ix][1 + 8*v_im], + fma(FLOAT_TYPE(b32[l]), sccache2[csel][ix][2 + 8*v_im], + fma(FLOAT_TYPE(b48[l]), sccache2[csel][ix][3 + 8*v_im], + fma(FLOAT_TYPE(b64[l]), sccache2[csel][ix][4 + 8*v_im], + fma(FLOAT_TYPE(b80[l]), sccache2[csel][ix][5 + 8*v_im], + fma(FLOAT_TYPE(b96[l]), sccache2[csel][ix][6 + 8*v_im], + fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2)))))))); + } + temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n])); + } + } +} +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; - const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...15 + const uint ix = tid/16; - const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128... + const uint v_in = itid - 8*v_im; // 0...7 - const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const uint v_in = tid - step*v_im; // 0...15 or 0...7 - - const uint l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15 + const uint l0 = 2*v_in; // 0...15 const uint q_offset = 32*v_im + l0; - const uint s_offset = 8*v_im; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp - - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - const uint y_idx = i * QUANT_K + y_offset; - - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); - - FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); - FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum1 += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3) - + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3); - sum2 += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF) - + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF); + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); } - tmp[16 * ix + tid] += dall * sum1 - dmin * sum2; } - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; + const uint nbr_par_th = num_blocks_per_row%it_size; + const uint nbr_all_th = num_blocks_per_row - nbr_par_th; + uint i0 = 0; + [[unroll]] for (; i0 < nbr_all_th; i0 += it_size) + calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true); + calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; } - barrier(); - } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp index a6e430ea..e91724a2 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -1,66 +1,132 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require #include "mul_mat_vec_base.comp" -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE tmp[32]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][2][8]; -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { + const uint y_idx = i * QUANT_K + y_offset; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; + + if (!all_threads) { // when we don't have enough blocks to use all threads + if (i < num_blocks_per_row) + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + barrier(); + + if (i >= num_blocks_per_row) + continue; + } + + const uint32_t hmk = ~(uint32_t(data_a_packed16[ib0 + i].hmask[v_in]) | (uint32_t(data_a_packed16[ib0 + i].hmask[v_in + 8]) << 16)); + const vec4 hmk_0 = vec4(unpack8(((hmk & hm_m[0]) >> ( v_im4)) << 2)); + const vec4 hmk_1 = vec4(unpack8(((hmk & hm_m[1]) >> (1 + v_im4)) << 2)); + const vec4 hmk_2 = vec4(unpack8(((hmk & hm_m[2]) >> (2 + v_im4)) << 2)); + const vec4 hmk_3 = vec4(unpack8(((hmk & hm_m[3]) >> (3 + v_im4)) << 2)); + + // 0, 1, 16, 17 + uint32_t qs_u32 = uint32_t(data_a[ib0 + i].qs[q_offset]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 1]) << 8); + qs_u32 |= (uint32_t(data_a[ib0 + i].qs[q_offset + 16]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 17]) << 8)) << 16; + const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303)); + const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303)); + const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303)); + const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303)); + + if (all_threads) { + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + barrier(); + } + + const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]); + vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]); + vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]); + vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]); + vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]); + vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]); + vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]); + vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]); + + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (int l = 0; l < 2; ++l) { + sum = fma(FLOAT_TYPE( b0[l]) * sccache[csel][ix][v_im][0], qs_u32_0[l ] - hmk_0[l ], + fma(FLOAT_TYPE( b16[l]) * sccache[csel][ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2], + fma(FLOAT_TYPE( b32[l]) * sccache[csel][ix][v_im][2], qs_u32_2[l ] - hmk_1[l ], + fma(FLOAT_TYPE( b48[l]) * sccache[csel][ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2], + fma(FLOAT_TYPE( b64[l]) * sccache[csel][ix][v_im][4], qs_u32_4[l ] - hmk_2[l ], + fma(FLOAT_TYPE( b80[l]) * sccache[csel][ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2], + fma(FLOAT_TYPE( b96[l]) * sccache[csel][ix][v_im][6], qs_u32_6[l ] - hmk_3[l ], + fma(FLOAT_TYPE(b112[l]) * sccache[csel][ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum)))))))); + } + temp[j][n] = fma(d, sum, temp[j][n]); + } + } +} +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; - const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - - const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 - const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...15 + const uint ix = tid/16; + const uint itid8 = itid%8; - const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const uint v_in = tid - step*v_im; // 0...15 or 0...7 + const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128... + const uint v_im4 = v_im*4; + const uint v_in = itid - 8*v_im; // 0...7 - const uint8_t m = uint8_t(1 << (4 * v_im)); + const uint32_t m = 0x01010101 << (4 * v_im); + uint32_t hm_m[4]; + [[unroll]] for (uint j = 0; j < 4; ++j) + hm_m[j] = m << j; - const uint l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15 + const uint l0 = 2*v_in; // 0...15 const uint q_offset = 32*v_im + l0; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp - - const uint s_shift = 4 * v_im; + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - const uint y_idx = i * QUANT_K + y_offset; + const uint s_shift = v_im4 + 2*(itid8/4); - const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); + const uint nbr_par_th = num_blocks_per_row%it_size; + const uint nbr_all_th = num_blocks_per_row - nbr_par_th; + uint i0 = 0; + [[unroll]] for (; i0 < nbr_all_th; i0 += it_size) + calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, true); + calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, false); - FLOAT_TYPE sum = FLOAT_TYPE(0.0); - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)) - + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)); - } - tmp[16 * ix + tid] += d * sum; - } + reduce_result(temp, d_offset, first_row, num_rows, tid); +} - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; } - barrier(); - } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp index 75569363..f9cde064 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp @@ -1,28 +1,106 @@ #version 450 -#include "mul_mat_vec_base.comp" - -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require -shared FLOAT_TYPE tmp[32]; +#include "mul_mat_vec_base.comp" -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y1_idx = i * QUANT_K + y_offset; + const uint y2_idx = y1_idx + 128; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + vec2 d = vec2(data_a[ib0 + i].d); + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + + const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; + const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; + const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); + const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); + + const FLOAT_TYPE sc0 = scale_0_4_l_f.x; + const FLOAT_TYPE sc1 = scale_0_4_l_f.y; + const FLOAT_TYPE sc2 = scale_0_4_l_f.z; + const FLOAT_TYPE sc3 = scale_0_4_l_f.w; + const FLOAT_TYPE sc4 = scale8_f.x; + const FLOAT_TYPE sc5 = scale8_f.y; + const FLOAT_TYPE sc6 = scale8_f.z; + const FLOAT_TYPE sc7 = scale8_f.w; + + const uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; + const uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; + + const uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; + const uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; + const uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F; + const uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F; + + const vec4 qs0_lo4 = vec4(unpack8(qs0_u32_lo4)); + const vec4 qs64_lo4 = vec4(unpack8(qs64_u32_lo4)); + const vec4 qs0_hi4 = vec4(unpack8(qs0_u32_hi4)); + const vec4 qs64_hi4 = vec4(unpack8(qs64_u32_hi4)); + + const FLOAT_TYPE q4_0 = qs0_lo4.x; + const FLOAT_TYPE q4_1 = qs0_lo4.y; + const FLOAT_TYPE q4_2 = qs0_lo4.z; + const FLOAT_TYPE q4_3 = qs0_lo4.w; + const FLOAT_TYPE q4_4 = qs0_hi4.x; + const FLOAT_TYPE q4_5 = qs0_hi4.y; + const FLOAT_TYPE q4_6 = qs0_hi4.z; + const FLOAT_TYPE q4_7 = qs0_hi4.w; + const FLOAT_TYPE q4_8 = qs64_lo4.x; + const FLOAT_TYPE q4_9 = qs64_lo4.y; + const FLOAT_TYPE q4_10 = qs64_lo4.z; + const FLOAT_TYPE q4_11 = qs64_lo4.w; + const FLOAT_TYPE q4_12 = qs64_hi4.x; + const FLOAT_TYPE q4_13 = qs64_hi4.y; + const FLOAT_TYPE q4_14 = qs64_hi4.z; + const FLOAT_TYPE q4_15 = qs64_hi4.w; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 by10 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 ]); + vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]); + vec4 by20 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 ]); + vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]); + + const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3))); + const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7))); + const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11))); + const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15))); + const FLOAT_TYPE smin = + fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7, + fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7, + fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7, + fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7))))))))))))))); + temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); + } + } +} +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; - const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...15 + const uint ix = tid/16; - const uint step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 - - const uint il = tid/step; // 0...3 - const uint ir = tid - step*il; // 0...7 or 0...3 - const uint n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 + const uint il = itid/4; // 0...3 + const uint ir = itid - 4*il; // 0...3 + const uint n = 4; const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 const uint v_in = il % 2; @@ -31,85 +109,28 @@ void main() { const uint q_offset = 32*v_im + l0; const uint y_offset = 64*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp - - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - const uint y1_idx = i * QUANT_K + y_offset; - const uint y2_idx = y1_idx + 128; - - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); - - const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f); - const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f); - const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f); - const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f); - const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2)); - const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2)); - const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2)); - const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2)); - -#if K_QUANTS_PER_ITERATION == 2 - const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); - const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); - const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] & 0xf); - const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] & 0xf); - const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); - const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); - const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] >> 4); - const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] >> 4); - const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); - const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); - const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf); - const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf); - const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); - const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); - const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] >> 4); - const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] >> 4); - - const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3); - const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7); - const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11); - const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15); - const FLOAT_TYPE smin = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y1_idx ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7 - + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7 - + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7 - + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7 - ); - tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin); -#else - const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); - const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); - const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); - const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); - const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); - const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); - const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); - const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); - - const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx ]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1); - const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3); - const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx ]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_5); - const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7); - const FLOAT_TYPE smin = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7 - + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7 - ); - - tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin); -#endif + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } } - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) + calc_superblock(a_offset, b_offset, v_im, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; } - barrier(); - } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp index 9be3645b..6c84ef3c 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp @@ -1,25 +1,137 @@ #version 450 -#include "mul_mat_vec_base.comp" - -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require -shared FLOAT_TYPE tmp[32]; +#include "mul_mat_vec_base.comp" -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint l0, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y1_idx = i * QUANT_K + y_offset; + const uint y2_idx = y1_idx + 128; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + vec2 d = vec2(data_a[ib0 + i].d); + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + + const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; + const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; + const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); + const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); + + const FLOAT_TYPE sc0 = scale_0_4_l_f.x; + const FLOAT_TYPE sc1 = scale_0_4_l_f.y; + const FLOAT_TYPE sc2 = scale_0_4_l_f.z; + const FLOAT_TYPE sc3 = scale_0_4_l_f.w; + const FLOAT_TYPE sc4 = scale8_f.x; + const FLOAT_TYPE sc5 = scale8_f.y; + const FLOAT_TYPE sc6 = scale8_f.z; + const FLOAT_TYPE sc7 = scale8_f.w; + + const uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); + const uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16); + + uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F; + uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F; + uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F; + uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F; + + const uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8])); + + const uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4; + const uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3; + const uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010); + const uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1; + + qs0_16_u32_lo4 += qs0_16_lo4_offset16; + qs0_16_u32_hi4 += qs0_16_hi4_offset16; + qs64_80_u32_lo4 += qs64_80_lo4_offset16; + qs64_80_u32_hi4 += qs64_80_hi4_offset16; + + const vec4 qs0_16_lo4 = vec4(unpack8(qs0_16_u32_lo4)); + const vec4 qs64_80_lo4 = vec4(unpack8(qs64_80_u32_lo4)); + const vec4 qs0_16_hi4 = vec4(unpack8(qs0_16_u32_hi4)); + const vec4 qs64_80_hi4 = vec4(unpack8(qs64_80_u32_hi4)); + + const FLOAT_TYPE q4_0 = qs0_16_lo4.x; + const FLOAT_TYPE q4_1 = qs0_16_lo4.y; + const FLOAT_TYPE q4_2 = qs0_16_lo4.z; + const FLOAT_TYPE q4_3 = qs0_16_lo4.w; + const FLOAT_TYPE q4_4 = qs0_16_hi4.x; + const FLOAT_TYPE q4_5 = qs0_16_hi4.y; + const FLOAT_TYPE q4_6 = qs0_16_hi4.z; + const FLOAT_TYPE q4_7 = qs0_16_hi4.w; + const FLOAT_TYPE q4_8 = qs64_80_lo4.x; + const FLOAT_TYPE q4_9 = qs64_80_lo4.y; + const FLOAT_TYPE q4_10 = qs64_80_lo4.z; + const FLOAT_TYPE q4_11 = qs64_80_lo4.w; + const FLOAT_TYPE q4_12 = qs64_80_hi4.x; + const FLOAT_TYPE q4_13 = qs64_80_hi4.y; + const FLOAT_TYPE q4_14 = qs64_80_hi4.z; + const FLOAT_TYPE q4_15 = qs64_80_hi4.w; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec2 by10 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 ]); + vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8]); + vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]); + vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]); + vec2 by20 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 ]); + vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8]); + vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]); + vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]); + + const FLOAT_TYPE sx = + fma(FLOAT_TYPE(by10.x), q4_0, + fma(FLOAT_TYPE(by10.y), q4_1, + fma(FLOAT_TYPE(by116.x), q4_2, + FLOAT_TYPE(by116.y) * q4_3))); + const FLOAT_TYPE sy = + fma(FLOAT_TYPE(by132.x), q4_4, + fma(FLOAT_TYPE(by132.y), q4_5, + fma(FLOAT_TYPE(by148.x), q4_6, + FLOAT_TYPE(by148.y) * q4_7))); + const FLOAT_TYPE sz = + fma(FLOAT_TYPE(by20.x), q4_8, + fma(FLOAT_TYPE(by20.y), q4_9, + fma(FLOAT_TYPE(by216.x), q4_10, + FLOAT_TYPE(by216.y) * q4_11))); + const FLOAT_TYPE sw = + fma(FLOAT_TYPE(by232.x), q4_12, + fma(FLOAT_TYPE(by232.y), q4_13, + fma(FLOAT_TYPE(by248.x), q4_14, + FLOAT_TYPE(by248.y) * q4_15))); + const FLOAT_TYPE smin = + fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2, + fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3, + fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6, + (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7))); + temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); + } + } +} +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; - const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - const uint tid = gl_LocalInvocationID.x/2; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%2; // 0 or 0, 1 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...15 + const uint ix = tid/16; - const uint il = tid/4; // 0...3 - const uint ir = tid - 4*il; // 0...7 or 0...3 + const uint il = itid/4; // 0...3 + const uint ir = itid - 4*il; // 0...3 const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 const uint v_in = il % 2; @@ -28,84 +140,28 @@ void main() { const uint q_offset = 32*v_im + l0; const uint y_offset = 64*v_im + l0; - const uint8_t hm1 = uint8_t(1 << (2*v_im)); - const uint8_t hm2 = uint8_t(hm1 << 4); - - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp - - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) { - const uint y1_idx = i * QUANT_K + y_offset; - const uint y2_idx = y1_idx + 128; - - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); - - const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f); - const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f); - const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f); - const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f); - const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2)); - const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2)); - const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2)); - const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2)); - - const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); - const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); - const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf); - const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf); - const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); - const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); - const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] >> 4); - const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] >> 4); - const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); - const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); - const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf); - const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf); - const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); - const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); - const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] >> 4); - const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] >> 4); - - const FLOAT_TYPE sx = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y1_idx ]) * (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)) - ); - const FLOAT_TYPE sy = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)) - ); - const FLOAT_TYPE sz = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y2_idx ]) * (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)) - ); - const FLOAT_TYPE sw = FLOAT_TYPE( - FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)) - + FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)) - ); - const FLOAT_TYPE smin = FLOAT_TYPE( - (FLOAT_TYPE(data_b[b_offset + y1_idx]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17])) * sc2 + (FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49])) * sc3 - + (FLOAT_TYPE(data_b[b_offset + y2_idx]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17])) * sc6 + (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7 - ); - tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin); + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } } - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) + calc_superblock(a_offset, b_offset, v_im, l0, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; } - barrier(); - } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp index d610cf03..d53d9ee0 100644 --- a/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp +++ b/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp @@ -1,79 +1,130 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + #include "mul_mat_vec_base.comp" -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE tmp[32]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][16]; -void main() { - const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint ix, const uint ql_offset, const uint qh_offset, const uint s_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { + const uint y_idx = i * QUANT_K + y_offset; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; + + if (!all_threads) { // when we don't have enough blocks to use all threads + if (i < num_blocks_per_row) + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + barrier(); + + if (i >= num_blocks_per_row) + continue; + } + + const uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16); + const uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16); + + const uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F; + const uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F; + const uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F; + const uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F; + + const uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16); + const uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4; + const uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2; + const uint32_t qh4_u32 = (qh_u32 & 0x30303030); + const uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2; + + const uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32; + const uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32; + const uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32; + const uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32; + + const vec4 q0 = vec4(unpack8(q0_u32)) - 32; + const vec4 q1 = vec4(unpack8(q1_u32)) - 32; + const vec4 q2 = vec4(unpack8(q2_u32)) - 32; + const vec4 q3 = vec4(unpack8(q3_u32)) - 32; + + if (all_threads) { + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + barrier(); + } + + const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 by0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 ]); + vec4 by32 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8]); + vec4 by64 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16]); + vec4 by96 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24]); + + FLOAT_TYPE sum[4] = {0, 0, 0, 0}; + [[unroll]] for (uint l = 0; l < 4; ++l) { + sum[0] = fma(FLOAT_TYPE(by0[l]), q0[l], sum[0]); + sum[1] = fma(FLOAT_TYPE(by32[l]), q1[l], sum[1]); + sum[2] = fma(FLOAT_TYPE(by64[l]), q2[l], sum[2]); + sum[3] = fma(FLOAT_TYPE(by96[l]), q3[l], sum[3]); + } + temp[j][n] = fma(fma(sum[0], sccache[csel][ix][s_offset], fma(sum[1], sccache[csel][ix][s_offset + 2], fma(sum[2], sccache[csel][ix][s_offset + 4], sum[3] * sccache[csel][ix][s_offset + 6]))), d, temp[j][n]); + } + } +} +void compute_outputs(const uint first_row, const uint num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; - const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...15 + const uint ix = tid/16; - const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128... + const uint v_in = itid - 8*v_im; // 0...7 - const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const uint v_in = tid - step*v_im; // 0...15 or 0...7 - -#if K_QUANTS_PER_ITERATION == 1 - const uint l0 = v_in; // 0...15 - const uint is = 0; -#else const uint l0 = 4 * v_in; // 0, 4, 8, ..., 28 const uint is = v_in / 4; -#endif const uint ql_offset = 64*v_im + l0; const uint qh_offset = 32*v_im + l0; const uint s_offset = 8*v_im + is; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - const uint y_idx = i * QUANT_K + y_offset; + const uint nbr_par_th = num_blocks_per_row%it_size; + const uint nbr_all_th = num_blocks_per_row - nbr_par_th; + uint i0 = 0; + [[unroll]] for (; i0 < nbr_all_th; i0 += it_size) + calc_superblock(a_offset, b_offset, itid, ix, ql_offset, qh_offset, s_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true); + calc_superblock(a_offset, b_offset, itid, ix, ql_offset, qh_offset, s_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false); - const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); + reduce_result(temp, d_offset, first_row, num_rows, tid); +} -#if K_QUANTS_PER_ITERATION == 1 - FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32); - tmp[16 * ix + tid] += sum; -#else - FLOAT_TYPE sum = FLOAT_TYPE(0.0); - [[unroll]] for (int l = 0; l < 4; ++l) { - sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32) - + FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32); +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; } - tmp[16 * ix + tid] += sum; -#endif - } - - // sum up partial sums and write back result - barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; - } - barrier(); - } - if (tid == 0) { - data_d[d_offset + row] = D_TYPE(tmp[0]); + compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/vulkan-shaders/mul_mm.comp b/ggml/src/vulkan-shaders/mul_mm.comp index 5fe9d524..26163b16 100644 --- a/ggml/src/vulkan-shaders/mul_mm.comp +++ b/ggml/src/vulkan-shaders/mul_mm.comp @@ -6,6 +6,19 @@ #ifdef FLOAT16 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif +#if defined(DATA_A_IQ1_M) +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require +#endif + +#if defined(DATA_A_BF16) && defined(COOPMAT) +#extension GL_EXT_bfloat16 : enable +#endif + +#ifdef COOPMAT +#extension GL_KHR_cooperative_matrix : enable +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_shader_subgroup_basic : enable +#endif #ifdef MUL_MAT_ID #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require @@ -20,9 +33,20 @@ #define LOAD_VEC_B 1 #endif +#if !defined(TO_FLOAT_TYPE) +#define TO_FLOAT_TYPE FLOAT_TYPE +#endif + layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +#if defined(A_TYPE_PACKED16) +layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];}; +#endif +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif + layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; @@ -57,6 +81,7 @@ layout (push_constant) uniform parameter #endif } p; +layout (constant_id = 0) const uint BLOCK_SIZE = 64; layout (constant_id = 1) const uint BM = 64; layout (constant_id = 2) const uint BN = 64; layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant @@ -65,16 +90,33 @@ layout (constant_id = 5) const uint WN = 32; layout (constant_id = 6) const uint WMITER = 2; layout (constant_id = 7) const uint TM = 4; layout (constant_id = 8) const uint TN = 2; -layout (constant_id = 9) const uint WARP = 32; +layout (constant_id = 9) const uint TK = 1; // Only needed for coopmat +layout (constant_id = 10) const uint WARP = 32; -shared FLOAT_TYPE buf_a[BM * (BK+1)]; -shared FLOAT_TYPE buf_b[BN * (BK+1)]; +#ifdef COOPMAT +#define SHMEM_STRIDE (BK + 8) +#else +#define SHMEM_STRIDE (BK + 1) +#endif + +shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE]; +shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE]; #ifdef MUL_MAT_ID -shared u16vec2 row_ids[3072]; +shared u16vec2 row_ids[4096]; +#endif // MUL_MAT_ID + +#define NUM_WARPS (BLOCK_SIZE / WARP) + +#ifdef COOPMAT +shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; #endif void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + #ifdef MUL_MAT_ID const uint expert_idx = gl_GlobalInvocationID.z; #else @@ -94,17 +136,32 @@ void main() { const uint ik = gl_WorkGroupID.x / blocks_m; const uint ic = gl_WorkGroupID.y; - const uint warp_i = gl_LocalInvocationID.x / WARP; - const uint warp_r = warp_i % (BM / WM); - const uint warp_c = warp_i / (BM / WM); - const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); const uint WSUBM = WM / WMITER; const uint WSUBN = WN / WNITER; +#ifdef COOPMAT + const uint warp_i = gl_SubgroupID; + + const uint tiw = gl_SubgroupInvocationID; + + const uint cms_per_row = WM / TM; + const uint cms_per_col = WN / TN; + + const uint storestride = WARP / TM; + const uint store_r = tiw % TM; + const uint store_c = tiw / TM; +#else + const uint warp_i = gl_LocalInvocationID.x / WARP; + const uint tiw = gl_LocalInvocationID.x % WARP; + const uint tiwr = tiw % (WSUBM / TM); const uint tiwc = tiw / (WSUBM / TM); +#endif + + const uint warp_r = warp_i % (BM / WM); + const uint warp_c = warp_i / (BM / WM); const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A); const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A); @@ -152,21 +209,31 @@ void main() { uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B; #endif - float sums[WMITER * TM * WNITER * TN]; +#ifdef COOPMAT + coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a; + coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b; + coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col]; + + [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { + sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f); + } +#else + ACC_TYPE sums[WMITER * TM * WNITER * TN]; FLOAT_TYPE cache_a[WMITER * TM]; - FLOAT_TYPE cache_b[WNITER * TN]; + FLOAT_TYPE cache_b[TN]; [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) { - sums[i] = 0.0f; + sums[i] = ACC_TYPE(0.0f); } +#endif - [[unroll]] for (uint block = start_k; block < end_k; block += BK) { + for (uint block = start_k; block < end_k; block += BK) { [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) { #if defined(DATA_A_F32) || defined(DATA_A_F16) #if LOAD_VEC_A == 8 const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx][0].x); buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y); buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z); @@ -177,91 +244,132 @@ void main() { buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w); #elif LOAD_VEC_A == 4 const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx].x); buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y); buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z); buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w); #else if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) { - buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); + } else { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f); + } +#endif +#elif defined(DATA_A_BF16) +#if LOAD_VEC_A == 4 + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + buf_a[buf_idx ] = TO_FLOAT_TYPE(data_a[idx].x); + buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y); + buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z); + buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w); +#else + if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); } else { - buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f); + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0)); } #endif #elif defined(DATA_A_Q4_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; - - const uint ib = idx / 16; - const uint iqs = idx & 0xF; - - const float d = float(data_a[ib].d); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; - - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; + + const uint ib = idx / 4; + const uint iqs = idx & 0x03; + + const float d = float(data_a_packed16[ib].d); + const uint vui = uint(data_a_packed16[ib].qs[2*iqs]) | (uint(data_a_packed16[ib].qs[2*iqs + 1]) << 16); + const vec4 v0 = (vec4(unpack8(vui & 0x0F0F0F0F)) - 8.0f) * d; + const vec4 v1 = (vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) - 8.0f) * d; + + buf_a[buf_idx ] = FLOAT_TYPE(v0.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y); + buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z); + buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w); + buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x); + buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y); + buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z); + buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w); #elif defined(DATA_A_Q4_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; - - const uint ib = idx / 16; - const uint iqs = idx & 0xF; - - const float d = float(data_a[ib].d); - const float m = float(data_a[ib].m); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m; - - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; + + const uint ib = idx / 4; + const uint iqs = idx & 0x03; + + const float d = float(data_a_packed16[ib].d); + const float m = float(data_a_packed16[ib].m); + const uint vui = uint(data_a_packed16[ib].qs[2*iqs]) | (uint(data_a_packed16[ib].qs[2*iqs + 1]) << 16); + const vec4 v0 = vec4(unpack8(vui & 0x0F0F0F0F)) * d + m; + const vec4 v1 = vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) * d + m; + + buf_a[buf_idx ] = FLOAT_TYPE(v0.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y); + buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z); + buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w); + buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x); + buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y); + buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z); + buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w); #elif defined(DATA_A_Q5_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]; - const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d; + const float d = float(data_a_packed16[ib].d); + const uint uint_qh = uint(data_a_packed16[ib].qh[1]) << 16 | uint(data_a_packed16[ib].qh[0]); + const ivec2 qh0 = ivec2(((uint_qh >> 2*iqs) << 4) & 0x10, (uint_qh >> (2*iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (2*iqs + 1)) << 4) & 0x10, (uint_qh >> (2*iqs + 13)) & 0x10); + + const uint vui = uint(data_a_packed16[ib].qs[iqs]); + const vec4 v = (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f) * d; buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z); buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 17] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q5_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const float m = float(data_a[ib].m); - const uint uint_qh = data_a[ib].qh; - const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m; + const float d = float(data_a_packed16[ib].d); + const float m = float(data_a_packed16[ib].m); + const uint uint_qh = data_a_packed16[ib].qh; + const ivec2 qh0 = ivec2(((uint_qh >> 2*iqs) << 4) & 0x10, (uint_qh >> (2*iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (2*iqs + 1)) << 4) & 0x10, (uint_qh >> (2*iqs + 13)) & 0x10); + + const uint vui = uint(data_a_packed16[ib].qs[iqs]); + const vec4 v = vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) * d + m; buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z); buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 17] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q8_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; - const uint ib = idx / 16; - const uint iqs = (idx & 0xF) * 2; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d; + const float d = float(data_a_packed16[ib].d); + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy; + const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d; buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 2] = FLOAT_TYPE(v.z); + buf_a[buf_idx + 3] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q2_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -280,7 +388,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q3_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -294,17 +402,15 @@ void main() { const uint qsshift = halfsplit * 2; // 0,2,4,6 const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128 - const int8_t us = int8_t(is < 4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) : - is < 8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) : - is < 12 ? (data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) : - (data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4)); + const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF) + | (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4)); const float dl = float(data_a[ib].d) * float(us - 32); buf_a[buf_idx ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi ] & m) != 0) ? 0 : 4))); buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4))); #elif defined(DATA_A_Q4_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -316,23 +422,28 @@ void main() { const vec2 loadd = vec2(data_a[ib].d); - uint8_t sc; - uint8_t mbyte; - if (is < 4) { - sc = uint8_t(data_a[ib].scales[is ] & 63); - mbyte = uint8_t(data_a[ib].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4)); - mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4)); - } + const uint scidx0 = (is < 4) ? is : (is + 4); + const uint scidx1 = (is < 4) ? is : (is - 4); + const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint scidxshift1 = (is < 4) ? 0 : 2; + const uint mbidx0 = is + 4; + const uint mbidx1 = (is < 4) ? is + 4 : is; + const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + const uint mbidxshift0 = (is < 4) ? 0 : 4; + const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint mbidxshift1 = (is < 4) ? 0 : 2; + + const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + const uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const float d = loadd.x * sc; - const float m = loadd.y * mbyte; + const float m = -loadd.y * mbyte; - buf_a[buf_idx ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) - m); - buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m); + buf_a[buf_idx ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF), m)); + buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m)); #elif defined(DATA_A_Q5_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -347,23 +458,28 @@ void main() { const vec2 loadd = vec2(data_a[ib].d); - uint8_t sc; - uint8_t mbyte; - if (is < 4) { - sc = uint8_t(data_a[ib].scales[is ] & 63); - mbyte = uint8_t(data_a[ib].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4)); - mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4)); - } + const uint scidx0 = (is < 4) ? is : (is + 4); + const uint scidx1 = (is < 4) ? is : (is - 4); + const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint scidxshift1 = (is < 4) ? 0 : 2; + const uint mbidx0 = is + 4; + const uint mbidx1 = (is < 4) ? is + 4 : is; + const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + const uint mbidxshift0 = (is < 4) ? 0 : 4; + const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint mbidxshift1 = (is < 4) ? 0 : 2; + + const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + const uint8_t mbyte = uint8_t(((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const float d = loadd.x * sc; - const float m = loadd.y * mbyte; + const float m = -loadd.y * mbyte; - buf_a[buf_idx ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi ] & hm) != 0 ? 16 : 0)) - m); - buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m); + buf_a[buf_idx ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi ] & hm) != 0 ? 16 : 0), m)); + buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m)); #elif defined(DATA_A_Q6_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -380,19 +496,201 @@ void main() { buf_a[buf_idx ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32)); buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32)); -#elif defined(DATA_A_IQ4_NL) +#elif defined(DATA_A_IQ1_S) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 128; // 2 values per idx + const uint ib32 = (idx % 128) / 16; // 0..7 + const uint ib8 = (idx % 128) / 4; + const int i8 = 2 * int(idx % 4); const float d = float(data_a[ib].d); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d; + const uint qh = data_a[ib].qh[ib32]; + const uint qs = data_a[ib].qs[ib8]; + const float dl = d * (2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]); + + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + const vec2 v = dl * (vec2(gvec) + delta); - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ1_M) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib8 = (idx % 128) / 4; + const uint ib16 = ib8 / 2; + const int i8 = 2 * int(idx % 4); + + const uint16_t[4] scales = data_a[ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + const uint sc = scales[ib8 / 8]; + const uint qs = data_a[ib].qs[ib8]; + const uint qh = data_a[ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1); + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + const vec2 v = dl * (vec2(gvec) + delta); + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ2_XXS) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib32 = (idx % 128) / 16; // 0..7 + const uint ib8 = (idx / 4) % 4; + + const float d = float(data_a[ib].d); + const uint qs = data_a[ib].qs[8 * ib32 + ib8]; + const uint signs = pack32(u8vec4( + data_a[ib].qs[8*ib32 + 4], + data_a[ib].qs[8*ib32 + 5], + data_a[ib].qs[8*ib32 + 6], + data_a[ib].qs[8*ib32 + 7] + )); + const float db = d * 0.25 * (0.5 + (signs >> 28)); + const uint32_t sign7 = bitfieldExtract(signs, 7 * int(ib8), 7); + const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); + const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); + const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ2_XS) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib32 = (idx % 128) / 16; // 0..7 + const uint ib8 = (idx / 4) % 4; // 0..3 + + const float d = float(data_a[ib].d); + const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf; + const float db = d * 0.25 * (0.5 + scale); + const uint qs = data_a[ib].qs[4 * ib32 + ib8]; + const uint sign7 = qs >> 9; + const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); + const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); + const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ2_S) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib8 = (idx % 128) / 4; // 0..31 + const uint ib32 = ib8 / 4; // 0..7 + + const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf; + const uint qs = data_a[ib].qs[ib8]; + const uint qh = data_a[ib].qh[ib32]; + const uint qhshift = 2 * (ib8 % 4); + const uint sign = data_a[ib].qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4)); + + const float d = float(data_a[ib].d); + const float db = d * 0.25 * (0.5 + scale); + const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); + const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1]; + const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147 + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ3_XXS) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint iqs = (idx % 128) / 2; // 0..63 + const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values + + const float d = float(data_a[ib].d); + const uint qs = data_a[ib].qs[iqs]; + const uint signs = pack32(u8vec4( + data_a[ib].qs[is+0], + data_a[ib].qs[is+1], + data_a[ib].qs[is+2], + data_a[ib].qs[is+3] + )); + const float db = d * 0.5 * (0.5 + (signs >> 28)); + const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7); + const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); + const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); + const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ3_S) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint iqs = (idx % 128) / 2; // 0..63 + const uint iqh = iqs / 8; + + const float d = float(data_a[ib].d); + const uint qs = data_a[ib].qs[iqs]; + const uint qh = data_a[ib].qh[iqh]; + const int8_t sign = int8_t(data_a[ib].signs[iqs / 2] >> (2 * (idx % 4))); + const uint scale = data_a[ib].scales[iqs / 16]; + const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign))); + const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf)); + const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ4_XS) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib32 = (idx % 128) / 16; // 0..7 + const uint iq = 16 * ib32 + 2 * (idx % 8); + + const uint sl = (data_a[ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const uint sh = ((data_a[ib].scales_h) >> (2 * ib32)) & 3; + const uint qshift = (idx & 8) >> 1; + u8vec2 qs = u8vec2(data_a[ib].qs[iq], data_a[ib].qs[iq + 1]); + qs = (qs >> qshift) & uint8_t(0xF); + + const float d = float(data_a[ib].d); + const vec2 v = d * float(int(sl | (sh << 4)) - 32) * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]); + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ4_NL) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; + + const uint ib = idx / 8; + const uint iqs = idx & 0x07; + + const FLOAT_TYPE d = FLOAT_TYPE(data_a_packed16[ib].d); + const uint vui = uint(data_a_packed16[ib].qs[iqs]); + + buf_a[buf_idx ] = FLOAT_TYPE(kvalues_iq4nl[vui & 0xF]) * d; + buf_a[buf_idx + 1 ] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]) * d; + buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)]) * d; + buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_iq4nl[vui >> 12]) * d; #endif } [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) { @@ -403,7 +701,7 @@ void main() { #else const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif - const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B; + const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x); buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y); buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z); @@ -419,24 +717,24 @@ void main() { #else const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif - const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B; - buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x); - buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y); - buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z); - buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w); + const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; + buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x); + buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y); + buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z); + buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w); #elif !MUL_MAT_ID if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); } else { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } #else const uint row_i = ic * BN + loadc_b + l; if (row_i < _ne1) { const u16vec2 row_idx = row_ids[row_i]; - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); } else { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } #endif } @@ -446,29 +744,43 @@ void main() { pos_a += BK / LOAD_VEC_A; pos_b += BK / LOAD_VEC_B; - for (uint i = 0; i < BK; i++) { +#ifdef COOPMAT + [[unroll]] for (uint i = 0; i < BK; i += TK) { + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + // Load from shared into cache + coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor); + + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor); + + sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]); + } + } + } +#else + [[unroll]] for (uint i = 0; i < BK; i++) { // Load from shared into cache [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { [[unroll]] for (uint j = 0; j < TM; j++) { - cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i]; + cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i]; } } [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint j = 0; j < TN; j++) { - cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i]; + cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i]; } - } - [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { [[unroll]] for (uint cc = 0; cc < TN; cc++) { [[unroll]] for (uint cr = 0; cr < TM; cr++) { - sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]); + const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr; + sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[cc]), sums[sums_idx]); } } } } } +#endif barrier(); } @@ -480,6 +792,54 @@ void main() { const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; #endif +#ifdef COOPMAT +#ifdef MUL_MAT_ID + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + const uint row_i = dc + cm_col * TN + col + store_c; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; + + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } +#else + const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float + + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N; + + if (is_aligned && is_in_bounds) { + // Full coopMat is within bounds and stride_d is aligned with 16B + coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]); + coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor); + } else if (is_in_bounds) { + // Full coopMat is within bounds, but stride_d is not aligned + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) { + // Partial coopMat is within bounds + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } + } + } +#endif // MUL_MAT_ID +#else [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { @@ -491,7 +851,7 @@ void main() { if (row_i >= _ne1) break; const u16vec2 row_idx = row_ids[row_i]; -#endif +#endif // MUL_MAT_ID [[unroll]] for (uint cr = 0; cr < TM; cr++) { #ifdef MUL_MAT_ID data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); @@ -499,9 +859,10 @@ void main() { if (dr_warp + cr < p.M && dc_warp + cc < p.N) { data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); } -#endif +#endif // MUL_MAT_ID } } } } +#endif // COOPMAT } diff --git a/ggml/src/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/vulkan-shaders/mul_mm_cm2.comp new file mode 100644 index 00000000..91846575 --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mm_cm2.comp @@ -0,0 +1,441 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require + +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_cooperative_matrix : enable +#extension GL_NV_cooperative_matrix2 : enable +#extension GL_EXT_buffer_reference : enable +#extension GL_KHR_shader_subgroup_ballot : enable +#extension GL_KHR_shader_subgroup_vote : enable +#ifdef DATA_A_BF16 +#extension GL_EXT_bfloat16 : enable +#endif + +#include "types.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +#define IS_MUL_MM2 1 + +layout (constant_id = 0) const uint BLOCK_SIZE = 256; +layout (constant_id = 1) const uint BM = 64; +layout (constant_id = 2) const uint BN = 64; +layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant + +layout (constant_id = 4) const bool enable_smaller_matrices = false; +const uint BNover2 = enable_smaller_matrices ? (BN / 2) : BN; +const uint BNover4 = enable_smaller_matrices ? (BN / 4) : BN; + +layout (push_constant) uniform parameter +{ + uint M; + uint N; + uint K; + uint stride_a; + uint stride_b; + uint stride_d; + + uint batch_stride_a; + uint batch_stride_b; + uint batch_stride_d; + +#ifdef MUL_MAT_ID + uint nei0; + uint nei1; + uint nbi1; + uint ne11; +#else + uint k_split; + uint ne02; + uint ne12; + uint broadcast2; + uint broadcast3; +#endif + // N dimension for the B matrix can be >= p.N + uint padded_N; +} p; + + +layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +#if QUANT_K > 1 +#define DECODEFUNCA , dequantFuncA + +#include "dequant_funcs_cm2.comp" + +#else +#define DECODEFUNCA +#endif + +#if !defined(fetch_scales) +#define fetch_scales(a, b, c, d, e, f) +#endif +#if !defined(store_scales) +#define store_scales(a) +#endif + +#if defined(DATA_A_BF16) +#define MAT_TYPE bfloat16_t +#else +#define MAT_TYPE FLOAT_TYPE +#endif + +#ifdef MUL_MAT_ID +layout (binding = 3) readonly buffer IDS {int data_ids[];}; + +shared u16vec4 row_ids[4096]; + +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB { + B_TYPE b[]; +}; + +uint _ne1; +shared uint _ne1_sh; + +B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const uint row_i = blockCoords[0]; + + if (row_i >= _ne1) { + return B_TYPE(0.0); + } + + const u16vec4 row_idx = row_ids[row_i]; + B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]]; + + return ret; +} + +D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t ir, const in uint32_t ic) +{ + uint dr = ir * BM + r; + uint dc = ic * BN + c; + + if (dr < p.M && dc < _ne1) { + uint row_i = dc; + const u16vec4 row_idx = row_ids[row_i]; + data_d[row_idx.y * p.batch_stride_d + row_idx.z * p.stride_d + dr] = elem; + } + return elem; +} + +#endif + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + const uint tid = gl_LocalInvocationIndex; + +#ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; +#else + const uint batch_idx = gl_GlobalInvocationID.z; + + const uint i13 = batch_idx / p.ne12; + const uint i12 = batch_idx % p.ne12; + + const uint i03 = i13 / p.broadcast3; + const uint i02 = i12 / p.broadcast2; + + const uint batch_idx_a = i03 * p.ne02 + i02; +#endif + + const uint blocks_m = (p.M + BM - 1) / BM; + const uint ir = gl_WorkGroupID.x % blocks_m; + const uint ik = gl_WorkGroupID.x / blocks_m; + const uint ic = gl_WorkGroupID.y; + +#ifdef MUL_MAT_ID + // Spread the search across all elements in the first subgroup + if (gl_SubgroupID == 0) { + _ne1 = 0; + uint num_elements = p.nei1 * p.nei0; + + for (uint i = gl_SubgroupInvocationID; subgroupAny(i < num_elements); i += gl_SubgroupSize) { + bool in_range = i < num_elements; + uint ii0 = i % p.nei0; + uint ii1 = i / p.nei0; + uint id = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0; + uvec4 ballot = subgroupBallot(in_range && id == expert_idx); + uint idx = subgroupBallotExclusiveBitCount(ballot); + if (in_range && id == expert_idx) { + row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0); + } + _ne1 += subgroupBallotBitCount(ballot); + } + _ne1_sh = _ne1; + } + + barrier(); + + _ne1 = _ne1_sh; + + // Workgroup has no work + if (ic * BN >= _ne1) return; +#endif + +#ifdef MUL_MAT_ID + uint start_k = 0; + const uint end_k = p.K; +#else + uint start_k = ik * p.k_split; + const uint end_k = min(p.K, (ik + 1) * p.k_split); +#endif + +#ifdef MUL_MAT_ID + uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K; + uint pos_b = 0; +#else + uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K; + uint pos_b = batch_idx * p.batch_stride_b; + uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; +#endif + + uint stride_a = p.stride_a / QUANT_K; + uint stride_b = p.stride_b; + + // Hint to the compiler that values are aligned (want 16B alignment). + // Quants are always block-aligned, no alignment needed. +#if ALIGNED +#if QUANT_K == 1 + stride_a &= ~7; +#endif + stride_b &= ~7; +#endif + + // Create layouts for both clamped and unclamped accesses + tensorLayoutNV<2> tensorLayoutA = createTensorLayoutNV(2); + tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutAClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutNV<2> tensorLayoutB = createTensorLayoutNV(2); + tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutBClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1); + +#if QUANT_K > 1 + tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K); + tensorLayoutAClamp = setTensorLayoutBlockSizeNV(tensorLayoutAClamp, 1, QUANT_K); +#endif + + // Use end_k rather than p.K as the dimension because that's what + // we need to bound check against when using split_k. + // Bounds check B against padded_N, but bounds check D against N. + tensorLayoutA = setTensorLayoutDimensionNV(tensorLayoutA, p.M, end_k); + tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.padded_N, end_k); + tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.N, p.M); + tensorLayoutAClamp = setTensorLayoutDimensionNV(tensorLayoutAClamp, p.M, end_k); + tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.padded_N, end_k); + + tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0); + +#if !defined(MUL_MAT_ID) + + const uint START_ALIGN_K = 256; + // For Qi_K (block size 256), unroll whole 256 element tiles. + // For legacy quants (block size 32), unroll 8x. + const uint UNROLL_K = (QUANT_K == 256) ? 256 : (BK * 8); + const uint unroll_count = UNROLL_K / BK; + + // Detect a fast path where all loads are entirely in bounds and no clamping is required + if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.padded_N && (start_k % START_ALIGN_K) == 0 && (end_k % BK) == 0 && +#if QUANT_K == 1 + (stride_a % 8) == 0 && +#endif + (stride_b % 8) == 0) { + // Hint to the compiler that values are aligned (want 16B alignment) + start_k &= ~(START_ALIGN_K-1); + stride_b &= ~7; +#if QUANT_K == 1 + stride_a &= ~7; +#endif + + tensorLayoutA = setTensorLayoutStrideNV(tensorLayoutA, stride_a, 1); + tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1); + + uint k_iters = (end_k - start_k) / UNROLL_K; + uint block_k = start_k; + + // fetch scale values for a tile of quants. These will be copied into shared memory. + // The fetches and stores are pipelined to hide the latency. + fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, true); + + if (enable_smaller_matrices && ic * BN + BNover4 >= p.N) { + coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator>(0.0); + for (uint i = 0; i < k_iters; ++i) { + + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } + + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator>(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover4, ir * BM, BM), tensorViewTranspose); + return; + } else if (enable_smaller_matrices && ic * BN + BNover2 >= p.N) { + coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(0.0); + for (uint i = 0; i < k_iters; ++i) { + + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } + + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover2, ir * BM, BM), tensorViewTranspose); + return; + } else { + coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0); + + for (uint i = 0; i < k_iters; ++i) { + + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } + + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); + return; + } + } else +#endif // !defined(MUL_MAT_ID) + { + tensorLayoutA = setTensorLayoutStrideNV(tensorLayoutA, stride_a, 1); + + tensorLayoutAClamp = setTensorLayoutStrideNV(tensorLayoutAClamp, stride_a, 1); + + tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1); + + tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1); + + coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum; + sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0); + + uint k_iters = (end_k - start_k + BK - 1) / BK; + + fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, false); + + [[dont_unroll]] + for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + + store_scales(tid); + if (block_k + BK < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false); + } + + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a; + coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); +#ifdef MUL_MAT_ID + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); +#else + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); +#endif + + sum = coopMatMulAdd(mat_a, mat_b, sum); + } + + // Convert from ACC_TYPE to D_TYPE + coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> mat_d; + mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(sum); + +#ifdef MUL_MAT_ID + // Call callback to store each element, remapping row through shared memory + coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic); +#else + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); +#endif + } +} diff --git a/ggml/src/vulkan-shaders/mul_mmq.comp b/ggml/src/vulkan-shaders/mul_mmq.comp new file mode 100644 index 00000000..83de90eb --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mmq.comp @@ -0,0 +1,442 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require + +#extension GL_EXT_integer_dot_product : require + +#ifdef FLOAT16 +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#endif + +#ifdef COOPMAT +#extension GL_KHR_cooperative_matrix : enable +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_shader_subgroup_basic : enable +#endif + +#ifdef MUL_MAT_ID +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require +#endif + +#include "types.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];}; +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif +layout (binding = 1) readonly buffer B {block_q8_1_packed32 data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +#ifdef MUL_MAT_ID +layout (binding = 3) readonly buffer IDS {int data_ids[];}; +#endif + +layout (push_constant) uniform parameter +{ + uint M; + uint N; + uint K; + uint stride_a; + uint stride_b; + uint stride_d; + + uint batch_stride_a; + uint batch_stride_b; + uint batch_stride_d; + +#ifdef MUL_MAT_ID + uint nei0; + uint nei1; + uint nbi1; + uint ne11; +#else + uint k_split; + uint ne02; + uint ne12; + uint broadcast2; + uint broadcast3; +#endif +} p; + +layout (constant_id = 0) const uint BLOCK_SIZE = 64; +layout (constant_id = 1) const uint BM = 64; +layout (constant_id = 2) const uint BN = 64; +// layout (constant_id = 3) const uint BK = 32; +layout (constant_id = 4) const uint WM = 32; +layout (constant_id = 5) const uint WN = 32; +layout (constant_id = 6) const uint WMITER = 2; +layout (constant_id = 7) const uint TM = 4; +layout (constant_id = 8) const uint TN = 2; +layout (constant_id = 9) const uint TK = 1; // Only needed for coopmat +layout (constant_id = 10) const uint WARP = 32; + +#define BK 32 + +#ifdef COOPMAT +#define SHMEM_STRIDE (BK / 4 + 4) +#else +#define SHMEM_STRIDE (BK / 4 + 1) +#endif + +shared int32_t buf_a_qs[BM * SHMEM_STRIDE]; + +#ifndef COOPMAT +#if QUANT_AUXF == 1 +shared FLOAT_TYPE buf_a_dm[BM]; +#else +shared FLOAT_TYPE_VEC2 buf_a_dm[BM]; +#endif +#endif + +shared int32_t buf_b_qs[BN * SHMEM_STRIDE]; +#ifndef COOPMAT +shared FLOAT_TYPE_VEC2 buf_b_ds[BN]; +#endif + +#define LOAD_VEC_A (4 * QUANT_R) +#define LOAD_VEC_B 4 + +#ifdef MUL_MAT_ID +shared u16vec2 row_ids[4096]; +#endif // MUL_MAT_ID + +#define NUM_WARPS (BLOCK_SIZE / WARP) + +#ifdef COOPMAT +shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; +#endif + +#include "mul_mmq_funcs.comp" + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + +#ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; +#else + const uint batch_idx = gl_GlobalInvocationID.z; + + const uint i13 = batch_idx / p.ne12; + const uint i12 = batch_idx % p.ne12; + + const uint i03 = i13 / p.broadcast3; + const uint i02 = i12 / p.broadcast2; + + const uint batch_idx_a = i03 * p.ne02 + i02; +#endif + + const uint blocks_m = (p.M + BM - 1) / BM; + const uint ir = gl_WorkGroupID.x % blocks_m; + const uint ik = gl_WorkGroupID.x / blocks_m; + const uint ic = gl_WorkGroupID.y; + + const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); + const uint WSUBM = WM / WMITER; + const uint WSUBN = WN / WNITER; + +#ifdef COOPMAT + const uint warp_i = gl_SubgroupID; + + const uint tiw = gl_SubgroupInvocationID; + + const uint cms_per_row = WM / TM; + const uint cms_per_col = WN / TN; + + const uint storestride = WARP / TM; + const uint store_r = tiw % TM; + const uint store_c = tiw / TM; +#else + const uint warp_i = gl_LocalInvocationID.x / WARP; + + const uint tiw = gl_LocalInvocationID.x % WARP; + + const uint tiwr = tiw % (WSUBM / TM); + const uint tiwc = tiw / (WSUBM / TM); +#endif + + const uint warp_r = warp_i % (BM / WM); + const uint warp_c = warp_i / (BM / WM); + + const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A); + const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A); + const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B); + const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B); + + const uint loadstride_a = BLOCK_SIZE * LOAD_VEC_A / BK; + const uint loadstride_b = BLOCK_SIZE * LOAD_VEC_B / BK; + +#ifdef MUL_MAT_ID + uint _ne1 = 0; + for (uint ii1 = 0; ii1 < p.nei1; ii1++) { + for (uint ii0 = 0; ii0 < p.nei0; ii0++) { + if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) { + row_ids[_ne1] = u16vec2(ii0, ii1); + _ne1++; + } + } + } + + barrier(); + + // Workgroup has no work + if (ic * BN >= _ne1) return; +#endif + +#ifdef MUL_MAT_ID + const uint start_k = 0; + const uint end_k = p.K; +#else + const uint start_k = ik * p.k_split; + const uint end_k = min(p.K, (ik + 1) * p.k_split); +#endif + + uint pos_a_ib = ( +#ifdef MUL_MAT_ID + expert_idx * p.batch_stride_a + +#else + batch_idx_a * p.batch_stride_a + +#endif + ir * BM * p.stride_a + start_k) / BK; +#ifdef MUL_MAT_ID + uint pos_b_ib = 0; +#else + uint pos_b_ib = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / BK; +#endif + +#ifdef COOPMAT + coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a; + coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b; + coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_result; + + coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> factors[cms_per_row * cms_per_col]; + + coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col]; + + [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { + sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f); + } +#else + int32_t cache_a_qs[WMITER * TM * BK / 4]; + + int32_t cache_b_qs[TN * BK / 4]; + + ACC_TYPE sums[WMITER * TM * WNITER * TN]; + + [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) { + sums[i] = ACC_TYPE(0.0f); + } +#endif + +#if QUANT_AUXF == 1 + FLOAT_TYPE cache_a_dm[WMITER * TM]; +#else + FLOAT_TYPE_VEC2 cache_a_dm[WMITER * TM]; +#endif + + FLOAT_TYPE_VEC2 cache_b_ds[TN]; + + for (uint block = start_k; block < end_k; block += BK) { + [[unroll]] for (uint l = 0; loadc_a + l < BM; l += loadstride_a) { + const uint ib = pos_a_ib + (loadc_a + l) * p.stride_a / BK; + const uint iqs = loadr_a; + const uint buf_ib = loadc_a + l; + + if (iqs == 0) { +#if QUANT_AUXF == 1 + buf_a_dm[buf_ib] = get_d(ib); +#else + buf_a_dm[buf_ib] = get_dm(ib); +#endif + } +#if QUANT_R == 1 + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs] = repack(ib, iqs); +#else + const i32vec2 vals = repack(ib, iqs); + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs ] = vals.x; + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs + 4] = vals.y; +#endif + } + [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) { +#ifdef MUL_MAT_ID + const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l]; + const uint idx = pos_b_ib + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b; + const uint ib = idx / 8; + const uint iqs = idx & 0x7; +#else + const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK; + const uint iqs = loadr_b; +#endif + + const uint buf_ib = loadc_b + l; + + if (iqs == 0) { + buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds); + } + buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs]; + } + + barrier(); + + pos_a_ib += 1; + pos_b_ib += 1; + +#ifdef COOPMAT + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + const uint ib_a = warp_r * WM + cm_row * TM; + // Load from shared into cache + coopMatLoad(cache_a, buf_a_qs, ib_a * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor); + + // TODO: only cache values that are actually needed + [[unroll]] for (uint t_idx = 0; t_idx < TM; t_idx++) { + cache_a_dm[t_idx] = buf_a_dm[ib_a + t_idx]; + } + + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const uint ib_b = warp_c * WN + cm_col * TN; + coopMatLoad(cache_b, buf_b_qs, ib_b * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor); + + // TODO: only cache values that are actually needed + [[unroll]] for (uint t_idx = 0; t_idx < TN; t_idx++) { + cache_b_dm[t_idx] = buf_b_d[ib_b + t_idx]; + } + + cm_result = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0); + cm_result = coopMatMulAdd(cache_a, cache_b, cm_result); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + coopmat_stage[warp_i * TM * TN + (store_c + col) * TM + store_r] = ACC_TYPE(float(cache_a_d[store_r]) * float(cache_b_d[store_c + col])); + } + + coopMatLoad(factors, coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + sums[cm_col * cms_per_row + cm_row] += factors * coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(cm_result); + } + } +#else + // Load from shared into cache + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + [[unroll]] for (uint cr = 0; cr < TM; cr++) { + const uint ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr; + cache_a_dm[wsir * TM + cr] = buf_a_dm[ib]; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + cache_a_qs[(wsir * TM + cr) * (BK / 4) + idx_k] = buf_a_qs[ib * SHMEM_STRIDE + idx_k]; + } + } + } + + [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { + [[unroll]] for (uint cc = 0; cc < TN; cc++) { + const uint ib = warp_c * WN + wsic * WSUBN + tiwc * TN + cc; + cache_b_ds[cc] = buf_b_ds[ib]; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + cache_b_qs[cc * (BK / 4) + idx_k] = buf_b_qs[ib * SHMEM_STRIDE + idx_k]; + } + } + + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + [[unroll]] for (uint cc = 0; cc < TN; cc++) { + [[unroll]] for (uint cr = 0; cr < TM; cr++) { + const uint cache_a_idx = wsir * TM + cr; + const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr; + int32_t q_sum = 0; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + q_sum += dotPacked4x8EXT(cache_a_qs[cache_a_idx * (BK / 4) + idx_k], + cache_b_qs[cc * (BK / 4) + idx_k]); + } + + sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]); + } + } + } + } +#endif + + barrier(); + } + + const uint dr = ir * BM + warp_r * WM; + const uint dc = ic * BN + warp_c * WN; + +#ifndef MUL_MAT_ID + const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; +#endif + +#ifdef COOPMAT +#ifdef MUL_MAT_ID + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < BN; col += storestride) { + const uint row_i = dc + cm_col * TN + col + store_c; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; + + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } +#else + const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float + + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N; + + if (is_aligned && is_in_bounds) { + // Full coopMat is within bounds and stride_d is aligned with 16B + coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]); + coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor); + } else if (is_in_bounds) { + // Full coopMat is within bounds, but stride_d is not aligned + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) { + // Partial coopMat is within bounds + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } + } + } +#endif // MUL_MAT_ID +#else + [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + + const uint dr_warp = dr + wsir * WSUBM + tiwr * TM; + const uint dc_warp = dc + wsic * WSUBN + tiwc * TN; + [[unroll]] for (uint cc = 0; cc < TN; cc++) { +#ifdef MUL_MAT_ID + const uint row_i = dc_warp + cc; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; +#endif // MUL_MAT_ID + [[unroll]] for (uint cr = 0; cr < TM; cr++) { +#ifdef MUL_MAT_ID + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); +#else + if (dr_warp + cr < p.M && dc_warp + cc < p.N) { + data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); + } +#endif // MUL_MAT_ID + } + } + } + } +#endif // COOPMAT +} diff --git a/ggml/src/vulkan-shaders/mul_mmq_funcs.comp b/ggml/src/vulkan-shaders/mul_mmq_funcs.comp new file mode 100644 index 00000000..63b15471 --- /dev/null +++ b/ggml/src/vulkan-shaders/mul_mmq_funcs.comp @@ -0,0 +1,99 @@ +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require + +#include "types.comp" + +// Each iqs value maps to a 32-bit integer + +#if defined(DATA_A_Q4_0) +i32vec2 repack(uint ib, uint iqs) { + // Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4 + const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1]); + const uint32_t vui = pack32(quants); + return i32vec2( vui & 0x0F0F0F0F, + (vui >> 4) & 0x0F0F0F0F); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y)); +} +#endif + +#if defined(DATA_A_Q4_1) +i32vec2 repack(uint ib, uint iqs) { + // Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4 + const uint32_t vui = data_a_packed32[ib].qs[iqs]; + return i32vec2( vui & 0x0F0F0F0F, + (vui >> 4) & 0x0F0F0F0F); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) { + return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y); +} +#endif + +#if defined(DATA_A_Q5_0) +i32vec2 repack(uint ib, uint iqs) { + // Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4 + const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1]); + const uint32_t vui = pack32(quants); + const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs)); + const int32_t v0 = int32_t(vui & 0x0F0F0F0F) + | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28) + + const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F) + | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28) + + return i32vec2(v0, v1); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y)); +} +#endif + +#if defined(DATA_A_Q5_1) +i32vec2 repack(uint ib, uint iqs) { + // Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4 + const uint32_t vui = data_a_packed32[ib].qs[iqs]; + const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs)); + const int32_t v0 = int32_t(vui & 0x0F0F0F0F) + | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28) + + const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F) + | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28) + + return i32vec2(v0, v1); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) { + return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y); +} +#endif + +#if defined(DATA_A_Q8_0) +int32_t repack(uint ib, uint iqs) { + // Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4 + return pack32(i16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1])); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(float(q_sum) * da * dsb.x); +} +#endif + +#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +FLOAT_TYPE get_d(uint ib) { + return FLOAT_TYPE(data_a[ib].d); +} +#endif + +#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1) +FLOAT_TYPE_VEC2 get_dm(uint ib) { + return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm); +} +#endif diff --git a/ggml/src/vulkan-shaders/opt_step_adamw.comp b/ggml/src/vulkan-shaders/opt_step_adamw.comp new file mode 100644 index 00000000..e0214fe7 --- /dev/null +++ b/ggml/src/vulkan-shaders/opt_step_adamw.comp @@ -0,0 +1,42 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) buffer X {A_TYPE x[];}; +layout (binding = 1) readonly buffer G {A_TYPE grad[];}; +layout (binding = 2) buffer GM {A_TYPE gradm[];}; +layout (binding = 3) buffer GV {A_TYPE gradv[];}; +layout (binding = 4) readonly buffer P {float params[7];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + + const float alpha = params[0]; + const float beta1 = params[1]; + const float beta2 = params[2]; + const float eps = params[3]; + const float wd = params[4]; + const float beta1h = params[5]; + const float beta2h = params[6]; + + const float gi = grad[i]; + const float gmi = gradm[i]*beta1 + gi*(1.0f - beta1); + const float gvi = gradv[i]*beta2 + gi*gi*(1.0f - beta2); + + gradm[i] = gmi; + gradv[i] = gvi; + + const float mh = gmi*beta1h; + const float vh = sqrt(gvi*beta2h) + eps; + + x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh; +} diff --git a/ggml/src/vulkan-shaders/pad.comp b/ggml/src/vulkan-shaders/pad.comp index a465cd52..450b67fc 100644 --- a/ggml/src/vulkan-shaders/pad.comp +++ b/ggml/src/vulkan-shaders/pad.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_unary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; @@ -22,5 +24,5 @@ void main() { const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03; - data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f); + data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : 0.0f); } diff --git a/ggml/src/vulkan-shaders/pool2d.comp b/ggml/src/vulkan-shaders/pool2d.comp new file mode 100644 index 00000000..b6124411 --- /dev/null +++ b/ggml/src/vulkan-shaders/pool2d.comp @@ -0,0 +1,74 @@ +#version 450 + +#include "types.comp" + +#extension GL_EXT_shader_16bit_storage : require + +layout(push_constant) uniform parameter { + uint IW; uint IH; + uint OW; uint OH; + uint OC; + uint pelements; + uint op; + int k0; int k1; + int s0; int s1; + int p0; int p1; +} p; + +#define BLOCK_SIZE 512 +#define FLT_MAX 3.402823466e+38F +#define OP_POOL_MAX 0u +#define OP_POOL_AVG 1u + +layout (local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout(binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout(binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +void main() { + const uint idx = gl_GlobalInvocationID.x; + if (idx >= p.pelements) { + return; + } + + const uint O_HW = p.OW * p.OH; + + const uint nc = idx / O_HW; + const uint cur_oh = (idx % O_HW) / p.OW; + const uint cur_ow = (idx % O_HW) % p.OW; + + const int start_h = int(cur_oh) * p.s0 - p.p0; + const uint bh = max(start_h, 0); + const uint eh = min(start_h + p.k0, p.IH); + + const int start_w = int(cur_ow) * p.s1 - p.p1; + const uint bw = max(start_w, 0); + const uint ew = min(start_w + p.k1, p.IW); + + const float scale = 1.0 / float(p.k0 * p.k1); + float res; + + if (p.op == OP_POOL_AVG) { + res = 0.0; + } else if (p.op == OP_POOL_MAX) { + res = -FLT_MAX; + } else { + return; + } + + #pragma unroll + for (uint i = bh; i < eh; i++) { + #pragma unroll + for (uint j = bw; j < ew; j++) { + const float cur = D_TYPE(data_a[nc * p.IH * p.IW + i * p.IW + j]); + + if (p.op == OP_POOL_AVG) { + res += cur * scale; + } else if (p.op == OP_POOL_MAX) { + res = max(res, cur); + } + } + } + + data_d[nc * O_HW + cur_oh * p.OW + cur_ow] = res; +} diff --git a/ggml/src/vulkan-shaders/quantize_q8_1.comp b/ggml/src/vulkan-shaders/quantize_q8_1.comp new file mode 100644 index 00000000..e2e020fe --- /dev/null +++ b/ggml/src/vulkan-shaders/quantize_q8_1.comp @@ -0,0 +1,77 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : require +#extension GL_EXT_shader_16bit_storage : require + +layout (push_constant) uniform parameter +{ + uint ne; +} p; + +#include "types.comp" + +layout(constant_id = 0) const uint GROUP_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {vec4 data_a[];}; +layout (binding = 1) writeonly buffer D {block_q8_1_packed32 data_b[];}; + +shared float shmem[GROUP_SIZE]; + +void quantize() { + const uint wgid = gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + // Each thread handles a vec4, so 8 threads handle a block + const uint blocks_per_group = GROUP_SIZE / 8; + + const uint block_in_wg = tid / 8; + + const uint ib = wgid * blocks_per_group + block_in_wg; + const uint iqs = tid % 8; + + if (ib >= gl_NumWorkGroups.x * blocks_per_group) { + return; + } + + const uint a_idx = ib * 8 + iqs; + + vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f); + const vec4 abs_vals = abs(vals); + + // Find absolute max for each block + shmem[tid] = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w)); + barrier(); + [[unroll]] for (uint s = 4; s > 0; s >>= 1) { + if (iqs < s) { + shmem[tid] = max(shmem[tid], shmem[tid + s]); + } + barrier(); + } + + const float amax = shmem[block_in_wg * 8]; + const float d = amax / 127.0; + const float d_inv = d != 0.0 ? 1.0 / d : 0.0; + vals = round(vals * d_inv); + data_b[ib].qs[iqs] = pack32(i8vec4(round(vals))); + barrier(); + + // Calculate the sum for each block + shmem[tid] = vals.x + vals.y + vals.z + vals.w; + barrier(); + [[unroll]] for (uint s = 4; s > 0; s >>= 1) { + if (iqs < s) { + shmem[tid] += shmem[tid + s]; + } + barrier(); + } + if (iqs == 0) { + const float sum = shmem[tid]; + + data_b[ib].ds = f16vec2(vec2(d, sum * d)); + } +} + +void main() { + quantize(); +} diff --git a/ggml/src/vulkan-shaders/relu.comp b/ggml/src/vulkan-shaders/relu.comp index 52a19b62..4f806270 100644 --- a/ggml/src/vulkan-shaders/relu.comp +++ b/ggml/src/vulkan-shaders/relu.comp @@ -17,5 +17,5 @@ void main() { return; } - data_d[i] = max(float(data_a[i]), 0); + data_d[i] = D_TYPE(max(float(data_a[i]), 0)); } diff --git a/ggml/src/vulkan-shaders/repeat.comp b/ggml/src/vulkan-shaders/repeat.comp new file mode 100644 index 00000000..1568b141 --- /dev/null +++ b/ggml/src/vulkan-shaders/repeat.comp @@ -0,0 +1,26 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +uint src0_idx_mod(uint idx) { + const uint i13 = idx / (p.ne12*p.ne11*p.ne10); + const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; + const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10); + const uint i12_offset = i12*p.ne11*p.ne10; + const uint i11 = (idx - i13_offset - i12_offset) / p.ne10; + const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; + return (i13 % p.ne03)*p.nb03 + (i12 % p.ne02)*p.nb02 + (i11 % p.ne01)*p.nb01 + (i10 % p.ne00)*p.nb00; +} + +void main() { + const uint idx = get_idx(); + + if (idx >= p.ne) { + return; + } + + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx_mod(idx)]); +} diff --git a/ggml/src/vulkan-shaders/repeat_back.comp b/ggml/src/vulkan-shaders/repeat_back.comp new file mode 100644 index 00000000..d8627993 --- /dev/null +++ b/ggml/src/vulkan-shaders/repeat_back.comp @@ -0,0 +1,37 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +void main() { + const uint idx = get_idx(); + + if (idx >= p.ne) { + return; + } + + // Destination multi-index (inlined dst_idx) + const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L); + const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; + const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L); + const uint i12_offset = i12*p.ne11*p.ne10; + const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L); + const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; + const uint d_idx = i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10; + + // Accumulate from sources + A_TYPE acc = A_TYPE(0); + for (uint i3 = i13; i3 < p.ne03; i3 += p.ne13) { + for (uint i2 = i12; i2 < p.ne02; i2 += p.ne12) { + for (uint i1 = i11; i1 < p.ne01; i1 += p.ne11) { + for (uint i0 = i10; i0 < p.ne00; i0 += p.ne10) { + acc += data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00]; + } + } + } + } + + data_d[get_doffset() + d_idx] = D_TYPE(acc); +} diff --git a/ggml/src/vulkan-shaders/rms_norm.comp b/ggml/src/vulkan-shaders/rms_norm.comp index b554400b..deb8ee99 100644 --- a/ggml/src/vulkan-shaders/rms_norm.comp +++ b/ggml/src/vulkan-shaders/rms_norm.comp @@ -1,6 +1,6 @@ #version 450 -#include "generic_head.comp" +#include "generic_unary_head.comp" #include "types.comp" #extension GL_EXT_control_flow_attributes : enable @@ -8,19 +8,29 @@ layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; -layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; -layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; - shared FLOAT_TYPE sum[BLOCK_SIZE]; void main() { - const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; - const uint tid = gl_LocalInvocationID.x; + const uint ncols = p.ne00; + const uint nrows = gl_NumWorkGroups.x; + const uint nchannels = gl_NumWorkGroups.y; + + const uint row = gl_WorkGroupID.x; + const uint channel = gl_WorkGroupID.y; + const uint samp = gl_WorkGroupID.z; + const uint tid = gl_LocalInvocationID.x; + + const uint stride_row = p.nb01; + const uint stride_channel = p.nb02; + const uint stride_sample = p.nb03; + + uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset(); + uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset(); sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp - [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]); + [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) { + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]); sum[tid] += xi * xi; } @@ -33,10 +43,10 @@ void main() { barrier(); } - const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX); + const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols); const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1)); - [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { - data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col])); + [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) { + data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col])); } } diff --git a/ggml/src/vulkan-shaders/rms_norm_back.comp b/ggml/src/vulkan-shaders/rms_norm_back.comp new file mode 100644 index 00000000..76009f3d --- /dev/null +++ b/ggml/src/vulkan-shaders/rms_norm_back.comp @@ -0,0 +1,55 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable +#define BLOCK_SIZE 512 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer G {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer X {B_TYPE data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE sum_xx[BLOCK_SIZE]; +shared FLOAT_TYPE sum_xg[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + // Compute derivative of x[i]/norm(x) = g[i]/norm(x) - x[i] dot(x,g)/KX / norm(x)^1.5 + + // partial sums for thread in warp + sum_xx[tid] = FLOAT_TYPE(0.0f); + sum_xg[tid] = FLOAT_TYPE(0.0f); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + const FLOAT_TYPE gi = FLOAT_TYPE(data_a[row*p.KX + col]); + const FLOAT_TYPE xi = FLOAT_TYPE(data_b[row*p.KX + col]); + sum_xx[tid] += xi * xi; + sum_xg[tid] += xi * gi; + } + + // sum up partial sums and write back result + barrier(); + [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + sum_xx[tid] += sum_xx[tid + s]; + sum_xg[tid] += sum_xg[tid + s]; + } + barrier(); + } + + const FLOAT_TYPE eps = FLOAT_TYPE(p.param1); + const FLOAT_TYPE mean = sum_xx[0] / FLOAT_TYPE(p.KX); + const FLOAT_TYPE scale_g = inversesqrt(mean + eps); + const FLOAT_TYPE scale_x = -scale_g * sum_xg[0] / (sum_xx[0] + FLOAT_TYPE(p.KX) * eps); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + data_d[row*p.KX + col] = D_TYPE( + scale_g * FLOAT_TYPE(data_a[row*p.KX + col]) + + scale_x * FLOAT_TYPE(data_b[row*p.KX + col])); + } +} diff --git a/ggml/src/vulkan-shaders/rope_head.comp b/ggml/src/vulkan-shaders/rope_head.comp index ea895422..96c9c4cb 100644 --- a/ggml/src/vulkan-shaders/rope_head.comp +++ b/ggml/src/vulkan-shaders/rope_head.comp @@ -1,6 +1,11 @@ #include "types.comp" #extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_spirv_intrinsics: enable + +#if RTE16 +spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits +#endif layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in; @@ -20,6 +25,11 @@ layout (push_constant) uniform parameter { float corr_dims[2]; float theta_scale; uint has_ff; + uint ne02; + uint s1; + uint s2; + int sections[4]; + uint is_back; } p; float rope_yarn_ramp(const float low, const float high, const uint i0) { @@ -39,6 +49,10 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out // Get n-d magnitude scaling corrected for interpolation mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale); } + // Backprogagation uses inverted rotation + if (p.is_back != 0) { + theta = -theta; + } cos_theta = cos(theta) * mscale; sin_theta = sin(theta) * mscale; } diff --git a/ggml/src/vulkan-shaders/rope_multi.comp b/ggml/src/vulkan-shaders/rope_multi.comp new file mode 100644 index 00000000..4f5b1a0e --- /dev/null +++ b/ggml/src/vulkan-shaders/rope_multi.comp @@ -0,0 +1,60 @@ +#version 450 + +#include "rope_head.comp" + +void main() { + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; + uint ne2 = p.ne02; + + if (i0 >= ne0) { + return; + } + + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; + + data_d[i + 0] = data_a[i + 0]; + data_d[i + 1] = data_a[i + 1]; + + return; + } + + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; + + const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3]; + const int sec_w = p.sections[1] + p.sections[0]; + const uint sector = (i0 / 2) % sect_dims; + + float theta_base = 0.0; + if (sector < p.sections[0]) { + theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= p.sections[0] && sector < sec_w) { + theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= sec_w && sector < sec_w + p.sections[2]) { + theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= sec_w + p.sections[2]) { + theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f); + } + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; + + float cos_theta, sin_theta; + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); + + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims/2]); + + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); +} diff --git a/ggml/src/vulkan-shaders/rope_neox.comp b/ggml/src/vulkan-shaders/rope_neox.comp index 83b46b69..db775c45 100644 --- a/ggml/src/vulkan-shaders/rope_neox.comp +++ b/ggml/src/vulkan-shaders/rope_neox.comp @@ -3,15 +3,18 @@ #include "rope_head.comp" void main() { - const uint col = gl_GlobalInvocationID.y * 2; - const uint row = gl_GlobalInvocationID.x; + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; - if (col >= p.ncols) { + if (i0 >= ne0) { return; } - if (col >= p.n_dims) { - const uint i = row*p.ncols + col; + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; data_d[i + 0] = data_a[i + 0]; data_d[i + 1] = data_a[i + 1]; @@ -19,19 +22,22 @@ void main() { return; } - const uint i = row*p.ncols + col/2; - const uint i2 = row/p.p_delta_rows; + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; - const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f); + const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f); - const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f; + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; float cos_theta, sin_theta; - rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta); + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); - const float x0 = float(data_a[i + 0]); - const float x1 = float(data_a[i + p.n_dims/2]); + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims/2]); - data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); - data_d[i + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); } diff --git a/ggml/src/vulkan-shaders/rope_norm.comp b/ggml/src/vulkan-shaders/rope_norm.comp index e416ad93..4ad35e54 100644 --- a/ggml/src/vulkan-shaders/rope_norm.comp +++ b/ggml/src/vulkan-shaders/rope_norm.comp @@ -3,15 +3,18 @@ #include "rope_head.comp" void main() { - const uint col = gl_GlobalInvocationID.y * 2; - const uint row = gl_GlobalInvocationID.x; + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; - if (col >= p.ncols) { + if (i0 >= ne0) { return; } - if (col >= p.n_dims) { - const uint i = row*p.ncols + col; + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; data_d[i + 0] = data_a[i + 0]; data_d[i + 1] = data_a[i + 1]; @@ -19,19 +22,22 @@ void main() { return; } - const uint i = row*p.ncols + col; - const uint i2 = row/p.p_delta_rows; + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0; - const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f); + const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f); - const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f; + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; float cos_theta, sin_theta; - rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta); + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); - const float x0 = float(data_a[i + 0]); - const float x1 = float(data_a[i + 1]); + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + 1]); - data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); - data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta); + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + 1] = D_TYPE(x0*sin_theta + x1*cos_theta); } diff --git a/ggml/src/vulkan-shaders/rope_vision.comp b/ggml/src/vulkan-shaders/rope_vision.comp new file mode 100644 index 00000000..cedacc4d --- /dev/null +++ b/ggml/src/vulkan-shaders/rope_vision.comp @@ -0,0 +1,47 @@ +#version 450 + +#include "rope_head.comp" + +void main() { + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; + uint ne2 = p.ne02; + + if (i0 >= ne0) { + return; + } + + const uint row_dst = gl_GlobalInvocationID.x; + + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; + + const int sect_dims = p.sections[0] + p.sections[1]; + const int sec_w = p.sections[1] + p.sections[0]; + const uint sector = (i0 / 2) % sect_dims; + + float theta_base = 0.0; + if (sector < p.sections[0]) { + const uint p0 = sector; + theta_base = data_pos[channel_x]*pow(p.theta_scale, p0); + } + else if (sector >= p.sections[0] && sector < sec_w) { + const uint p0 = sector - p.sections[0]; + theta_base = data_pos[channel_x + ne2]*pow(p.theta_scale, p0); + } + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; + + float cos_theta, sin_theta; + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); + + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims]); + + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims] = D_TYPE(x0*sin_theta + x1*cos_theta); +} diff --git a/ggml/src/vulkan-shaders/scale.comp b/ggml/src/vulkan-shaders/scale.comp index 5cd2f668..4663428d 100644 --- a/ggml/src/vulkan-shaders/scale.comp +++ b/ggml/src/vulkan-shaders/scale.comp @@ -3,12 +3,22 @@ #include "types.comp" #include "generic_unary_head.comp" +const uint num_threads = 128; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + void main() { - const uint idx = get_idx(); + uint idx = get_idx(); - if (idx >= p.ne) { - return; - } + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 4; - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1)); + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + + data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) * FLOAT_TYPE(p.param1)); + idx += num_threads; + } } diff --git a/ggml/src/vulkan-shaders/sigmoid.comp b/ggml/src/vulkan-shaders/sigmoid.comp new file mode 100644 index 00000000..5c9e5c35 --- /dev/null +++ b/ggml/src/vulkan-shaders/sigmoid.comp @@ -0,0 +1,20 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + data_d[i] = D_TYPE(1. / (1 + exp(-1. * float(data_a[i])))); +} diff --git a/ggml/src/vulkan-shaders/silu_back.comp b/ggml/src/vulkan-shaders/silu_back.comp new file mode 100644 index 00000000..f9afa9b1 --- /dev/null +++ b/ggml/src/vulkan-shaders/silu_back.comp @@ -0,0 +1,26 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer G {A_TYPE data_g[];}; +layout (binding = 1) readonly buffer X {B_TYPE data_x[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + + // Compute derivative of SiLU(x): 1/(1+exp(-x)) - x*exp(-x)/(1+exp(-x))^2 + + const float xi = float(data_x[i]); + const float s = 1.0f / (1.0f + exp(-xi)); + data_d[i] = D_TYPE(data_g[i] * (s + xi * s * (1 - s))); +} diff --git a/ggml/src/vulkan-shaders/sin.comp b/ggml/src/vulkan-shaders/sin.comp new file mode 100644 index 00000000..d7c15a16 --- /dev/null +++ b/ggml/src/vulkan-shaders/sin.comp @@ -0,0 +1,17 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +void main() { + const uint idx = get_idx(); + + if (idx >= p.ne) { + return; + } + + const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sin(val)); +} diff --git a/ggml/src/vulkan-shaders/soft_max.comp b/ggml/src/vulkan-shaders/soft_max.comp index 0bd51eca..51fc2dc7 100644 --- a/ggml/src/vulkan-shaders/soft_max.comp +++ b/ggml/src/vulkan-shaders/soft_max.comp @@ -1,6 +1,6 @@ #version 450 -#extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_control_flow_attributes : enable layout (push_constant) uniform parameter { @@ -11,14 +11,13 @@ layout (push_constant) uniform parameter float m0; float m1; uint n_head_log2; + uint nrows_x; } p; #include "types.comp" -#extension GL_EXT_control_flow_attributes : enable -#define BLOCK_SIZE 512 - -layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; +layout(constant_id = 0) const uint BLOCK_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; layout (binding = 1) readonly buffer Y {B_TYPE data_b[];}; @@ -26,10 +25,17 @@ layout (binding = 2) buffer D {D_TYPE data_d[];}; shared FLOAT_TYPE vals[BLOCK_SIZE]; -void main() { +// num_iters is the number of BLOCK_SIZE loop iterations we need to iterate +// over all the columns. The main function tries to pass a constant here, +// as if it were a template function, to allow unrolling. +void soft_max(uint num_iters) { const uint tid = gl_LocalInvocationID.x; const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; - const uint rowy = rowx % p.KY; + const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0; + + if (rowx >= p.nrows_x) { + return; + } float slope = 1.0f; @@ -46,19 +52,39 @@ void main() { // Find max FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000); - [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) { + // Cache values while we compute the max, so we don't need to read them + // again when we're ready to compute exp(x-max). + const uint DATA_CACHE_SIZE = 16; + FLOAT_TYPE data_cache[DATA_CACHE_SIZE]; + + [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) { const uint col = col0 + tid; - if (col >= p.KX) { - break; + FLOAT_TYPE a = FLOAT_TYPE(0); + if (col < p.KX) { + a = data_a[rowx * p.KX + col]; + } + + FLOAT_TYPE b = FLOAT_TYPE(0); + if (p.KY > 0 && col < p.KX) { + b = data_b[rowy * p.KX + col]; } - max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f))); + FLOAT_TYPE v = a * p.scale + slope * b; + + if (col < p.KX) { + max_val = max(max_val, v); + } + + if (idx < DATA_CACHE_SIZE) { + data_cache[idx] = v; + } } - vals[tid] = max_val; + // reduce across the workgroup + vals[tid] = max_val; barrier(); - [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { if (tid < s) { vals[tid] = max(vals[tid], vals[tid + s]); } @@ -68,39 +94,80 @@ void main() { max_val = vals[0]; barrier(); - // Sum up values - vals[tid] = FLOAT_TYPE(0.0f); + FLOAT_TYPE sum = FLOAT_TYPE(0.0f); - [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) { + // Compute sum{exp(x - max)} + [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) { const uint col = col0 + tid; if (col >= p.KX) { break; } + // compute exp(a*scale+b*slope), add it to sum, and cache the new value + // in data_cache if possible. const uint i = rowx * p.KX + col; - const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val); - vals[tid] += val; - data_d[i] = D_TYPE(val); + FLOAT_TYPE val; + if (idx < DATA_CACHE_SIZE) { + val = exp(data_cache[idx] - max_val); + } else { + val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val); + } + sum += val; + if (idx < DATA_CACHE_SIZE) { + data_cache[idx] = val; + } else { + data_d[i] = D_TYPE(val); + } } + // reduce across the workgroup + vals[tid] = sum; barrier(); - [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { if (tid < s) { vals[tid] += vals[tid + s]; } barrier(); } + sum = vals[0]; - const D_TYPE divisor = D_TYPE(vals[0]); + FLOAT_TYPE rcpdivisor = 1.0/sum; - [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) { + [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) { const uint col = col0 + tid; if (col >= p.KX) { - break; + continue; + } + + if (idx < DATA_CACHE_SIZE) { + data_d[rowx*p.KX + col] = D_TYPE(data_cache[idx] * rcpdivisor); + } else { + data_d[rowx*p.KX + col] *= D_TYPE(rcpdivisor); } + } +} - data_d[rowx*p.KX + col] /= divisor; +void main() { + // instantiate the soft_max function for several different + // dimensions, to allow loop unrolling + uint num_blocks = (p.KX + BLOCK_SIZE - 1) / BLOCK_SIZE; + if (num_blocks > 32) { + soft_max(num_blocks); + } else if (num_blocks > 16) { + soft_max(32); + } else if (num_blocks > 8) { + soft_max(16); + } else if (num_blocks > 4) { + soft_max(8); + } else if (num_blocks == 4) { + soft_max(4); + } else if (num_blocks == 3) { + soft_max(3); + } else if (num_blocks == 2) { + soft_max(2); + } else if (num_blocks == 1) { + soft_max(1); } } diff --git a/ggml/src/vulkan-shaders/soft_max_back.comp b/ggml/src/vulkan-shaders/soft_max_back.comp new file mode 100644 index 00000000..29bd77d7 --- /dev/null +++ b/ggml/src/vulkan-shaders/soft_max_back.comp @@ -0,0 +1,50 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#include "generic_head.comp" +#include "types.comp" + +layout(constant_id = 0) const uint BLOCK_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +// In this shader Y = softmax(X) and X is not provided as input. + +layout (binding = 0) readonly buffer G {A_TYPE data_g[];}; +layout (binding = 1) readonly buffer Y {B_TYPE data_y[];}; +layout (binding = 2) buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE sum_yg[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + FLOAT_TYPE scale = p.param1; + + // partial sums for thread in warp + sum_yg[tid] = FLOAT_TYPE(0.0f); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + const FLOAT_TYPE gi = FLOAT_TYPE(data_g[row*p.KX + col]); + const FLOAT_TYPE yi = FLOAT_TYPE(data_y[row*p.KX + col]); + sum_yg[tid] += yi * gi; + } + + // sum up partial sums and write back result + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + sum_yg[tid] += sum_yg[tid + s]; + } + barrier(); + } + + const FLOAT_TYPE dot_yg = sum_yg[0]; + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + data_d[row*p.KX + col] = D_TYPE(scale + * (FLOAT_TYPE(data_g[row*p.KX + col]) - dot_yg) + * FLOAT_TYPE(data_y[row*p.KX + col])); + } +} diff --git a/ggml/src/vulkan-shaders/square.comp b/ggml/src/vulkan-shaders/square.comp index 1fa118c9..ef43598b 100644 --- a/ggml/src/vulkan-shaders/square.comp +++ b/ggml/src/vulkan-shaders/square.comp @@ -3,6 +3,8 @@ #include "types.comp" #include "generic_unary_head.comp" +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + void main() { const uint idx = get_idx(); @@ -10,6 +12,6 @@ void main() { return; } - const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]); - data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val); + const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val * val); } diff --git a/ggml/src/vulkan-shaders/sub.comp b/ggml/src/vulkan-shaders/sub.comp new file mode 100644 index 00000000..72353cc3 --- /dev/null +++ b/ggml/src/vulkan-shaders/sub.comp @@ -0,0 +1,29 @@ +#version 450 + +#extension GL_EXT_shader_16bit_storage : require + +#include "types.comp" +#include "generic_binary_head.comp" + +const uint num_threads = 256; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + +void main() { + uint idx = get_idx(); + + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 2; + + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) - FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)])); + + idx += num_threads; + } +} diff --git a/ggml/src/vulkan-shaders/tanh.comp b/ggml/src/vulkan-shaders/tanh.comp index 74630dc7..8a6f868f 100644 --- a/ggml/src/vulkan-shaders/tanh.comp +++ b/ggml/src/vulkan-shaders/tanh.comp @@ -16,6 +16,5 @@ void main() { if (i >= p.KX) { return; } - - data_d[i] = D_TYPE(tanh(data_a[i])); + data_d[i] = D_TYPE(1. - 2. / (exp(2.*float(data_a[i])) + 1.)); } diff --git a/ggml/src/vulkan-shaders/test_bfloat16_support.comp b/ggml/src/vulkan-shaders/test_bfloat16_support.comp new file mode 100644 index 00000000..fd0ba401 --- /dev/null +++ b/ggml/src/vulkan-shaders/test_bfloat16_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_EXT_bfloat16 : require + +void main() +{ +} diff --git a/ggml/src/vulkan-shaders/test_coopmat2_support.comp b/ggml/src/vulkan-shaders/test_coopmat2_support.comp new file mode 100644 index 00000000..28eb24e1 --- /dev/null +++ b/ggml/src/vulkan-shaders/test_coopmat2_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_NV_cooperative_matrix2 : require + +void main() +{ +} diff --git a/ggml/src/vulkan-shaders/test_coopmat_support.comp b/ggml/src/vulkan-shaders/test_coopmat_support.comp new file mode 100644 index 00000000..8c5dd1bd --- /dev/null +++ b/ggml/src/vulkan-shaders/test_coopmat_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_KHR_cooperative_matrix : require + +void main() +{ +} diff --git a/ggml/src/vulkan-shaders/test_integer_dot_support.comp b/ggml/src/vulkan-shaders/test_integer_dot_support.comp new file mode 100644 index 00000000..470e3074 --- /dev/null +++ b/ggml/src/vulkan-shaders/test_integer_dot_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_EXT_integer_dot_product : require + +void main() +{ +} diff --git a/ggml/src/vulkan-shaders/types.comp b/ggml/src/vulkan-shaders/types.comp index 21dce72f..3bde7178 100644 --- a/ggml/src/vulkan-shaders/types.comp +++ b/ggml/src/vulkan-shaders/types.comp @@ -1,6 +1,11 @@ -#if !defined(DATA_A_F32) && !defined(DATA_A_F16) +#if !defined(GGML_TYPES_COMP) +#define GGML_TYPES_COMP + +#extension GL_EXT_shader_explicit_arithmetic_types_int64 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require -#endif +#extension GL_EXT_shader_16bit_storage : require #if defined(DATA_A_F32) #define QUANT_K 1 @@ -28,24 +33,43 @@ #endif #endif -#if defined(DATA_A_Q4_0) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#if defined(DATA_A_BF16) +#define QUANT_K 1 +#define QUANT_R 1 + +#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1 +#define A_TYPE uint16_t +#elif LOAD_VEC_A == 4 +#define A_TYPE u16vec4 +#elif LOAD_VEC_A == 8 +#error unsupported +#endif +#endif + +#define QUANT_K_Q4_0 32 +#define QUANT_R_Q4_0 2 struct block_q4_0 { float16_t d; uint8_t qs[16]; }; +struct block_q4_0_packed16 +{ + float16_t d; + uint16_t qs[16/2]; +}; +#if defined(DATA_A_Q4_0) +#define QUANT_K QUANT_K_Q4_0 +#define QUANT_R QUANT_R_Q4_0 +#define QUANT_AUXF 1 #define A_TYPE block_q4_0 +#define A_TYPE_PACKED16 block_q4_0_packed16 #endif -#if defined(DATA_A_Q4_1) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q4_1 32 +#define QUANT_R_Q4_1 2 struct block_q4_1 { @@ -54,14 +78,30 @@ struct block_q4_1 uint8_t qs[16]; }; +struct block_q4_1_packed16 +{ + float16_t d; + float16_t m; + uint16_t qs[16/2]; +}; + +struct block_q4_1_packed32 +{ + f16vec2 dm; + uint32_t qs[16/4]; +}; + +#if defined(DATA_A_Q4_1) +#define QUANT_K QUANT_K_Q4_1 +#define QUANT_R QUANT_R_Q4_1 +#define QUANT_AUXF 2 #define A_TYPE block_q4_1 +#define A_TYPE_PACKED16 block_q4_1_packed16 +#define A_TYPE_PACKED32 block_q4_1_packed32 #endif -#if defined(DATA_A_Q5_0) -#extension GL_EXT_shader_16bit_storage : require -#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q5_0 32 +#define QUANT_R_Q5_0 2 struct block_q5_0 { @@ -70,14 +110,23 @@ struct block_q5_0 uint8_t qs[16]; }; +struct block_q5_0_packed16 +{ + float16_t d; + uint16_t qh[2]; + uint16_t qs[16/2]; +}; + +#if defined(DATA_A_Q5_0) +#define QUANT_K QUANT_K_Q5_0 +#define QUANT_R QUANT_R_Q5_0 +#define QUANT_AUXF 1 #define A_TYPE block_q5_0 +#define A_TYPE_PACKED16 block_q5_0_packed16 #endif -#if defined(DATA_A_Q5_1) -#extension GL_EXT_shader_16bit_storage : require -#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q5_1 32 +#define QUANT_R_Q5_1 2 struct block_q5_1 { @@ -87,114 +136,1238 @@ struct block_q5_1 uint8_t qs[16]; }; +struct block_q5_1_packed16 +{ + float16_t d; + float16_t m; + uint qh; + uint16_t qs[16/2]; +}; + +struct block_q5_1_packed32 +{ + f16vec2 dm; + uint qh; + uint32_t qs[16/4]; +}; + +#if defined(DATA_A_Q5_1) +#define QUANT_K QUANT_K_Q5_1 +#define QUANT_R QUANT_R_Q5_1 +#define QUANT_AUXF 2 #define A_TYPE block_q5_1 +#define A_TYPE_PACKED16 block_q5_1_packed16 +#define A_TYPE_PACKED32 block_q5_1_packed32 #endif -#if defined(DATA_A_Q8_0) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 1 +#define QUANT_K_Q8_0 32 +#define QUANT_R_Q8_0 1 struct block_q8_0 { float16_t d; int8_t qs[32]; }; +struct block_q8_0_packed16 +{ + float16_t d; + int16_t qs[32/2]; +}; +struct block_q8_0_packed32 +{ + float16_t d; + int32_t qs[32/4]; +}; +#if defined(DATA_A_Q8_0) +#define QUANT_K QUANT_K_Q8_0 +#define QUANT_R QUANT_R_Q8_0 +#define QUANT_AUXF 1 #define A_TYPE block_q8_0 +#define A_TYPE_PACKED16 block_q8_0_packed16 +#define A_TYPE_PACKED32 block_q8_0_packed32 #endif +#define QUANT_K_Q8_1 32 +#define QUANT_R_Q8_1 1 + +struct block_q8_1 +{ + f16vec2 ds; + int8_t qs[32]; +}; +struct block_q8_1_packed16 +{ + f16vec2 ds; + int16_t qs[16]; +}; +struct block_q8_1_packed32 +{ + f16vec2 ds; + int32_t qs[8]; +}; + // K-quants -#if defined(DATA_A_Q2_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q2_K 256 struct block_q2_K { - uint8_t scales[QUANT_K/16]; - uint8_t qs[QUANT_K/4]; + uint8_t scales[QUANT_K_Q2_K/16]; + uint8_t qs[QUANT_K_Q2_K/4]; + f16vec2 d; +}; + +struct block_q2_K_packed16 +{ + uint16_t scales[QUANT_K_Q2_K/16/2]; + uint16_t qs[QUANT_K_Q2_K/4/2]; f16vec2 d; }; +struct block_q2_K_packed32 +{ + uint32_t scales[QUANT_K_Q2_K/16/4]; + uint32_t qs[QUANT_K_Q2_K/4/4]; + f16vec2 d; +}; + +#if defined(DATA_A_Q2_K) +#define QUANT_K QUANT_K_Q2_K #define A_TYPE block_q2_K +#define A_TYPE_PACKED16 block_q2_K_packed16 +#define A_TYPE_PACKED32 block_q2_K_packed32 #endif -#if defined(DATA_A_Q3_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q3_K 256 struct block_q3_K { - uint8_t hmask[QUANT_K/8]; - uint8_t qs[QUANT_K/4]; + uint8_t hmask[QUANT_K_Q3_K/8]; + uint8_t qs[QUANT_K_Q3_K/4]; uint8_t scales[12]; float16_t d; }; +struct block_q3_K_packed16 +{ + uint16_t hmask[QUANT_K_Q3_K/8/2]; + uint16_t qs[QUANT_K_Q3_K/4/2]; + uint16_t scales[12/2]; + float16_t d; +}; + +#if defined(DATA_A_Q3_K) +#define QUANT_K QUANT_K_Q3_K #define A_TYPE block_q3_K +#define A_TYPE_PACKED16 block_q3_K_packed16 #endif -#if defined(DATA_A_Q4_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q4_K 256 struct block_q4_K { f16vec2 d; - uint8_t scales[3*QUANT_K/64]; - uint8_t qs[QUANT_K/2]; + uint8_t scales[3*QUANT_K_Q4_K/64]; + uint8_t qs[QUANT_K_Q4_K/2]; +}; + +struct block_q4_K_packed16 +{ + f16vec2 d; + uint16_t scales[3*QUANT_K_Q4_K/64/2]; + uint16_t qs[QUANT_K_Q4_K/2/2]; +}; + +struct block_q4_K_packed32 +{ + f16vec2 d; + uint32_t scales[3*QUANT_K_Q4_K/64/4]; + uint32_t qs[QUANT_K_Q4_K/2/4]; +}; + +struct block_q4_K_packed128 +{ + uvec4 q4k[9]; }; +#if defined(DATA_A_Q4_K) +#define QUANT_K QUANT_K_Q4_K #define A_TYPE block_q4_K +#define A_TYPE_PACKED16 block_q4_K_packed16 +#define A_TYPE_PACKED32 block_q4_K_packed32 #endif -#if defined(DATA_A_Q5_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q5_K 256 struct block_q5_K { f16vec2 d; uint8_t scales[12]; - uint8_t qh[QUANT_K/8]; - uint8_t qs[QUANT_K/2]; + uint8_t qh[QUANT_K_Q5_K/8]; + uint8_t qs[QUANT_K_Q5_K/2]; +}; + +struct block_q5_K_packed16 +{ + f16vec2 d; + uint16_t scales[12/2]; + uint16_t qh[QUANT_K_Q5_K/8/2]; + uint16_t qs[QUANT_K_Q5_K/2/2]; }; +struct block_q5_K_packed128 +{ + uvec4 q5k[11]; +}; + +#if defined(DATA_A_Q5_K) +#define QUANT_K QUANT_K_Q5_K #define A_TYPE block_q5_K +#define A_TYPE_PACKED16 block_q5_K_packed16 #endif -#if defined(DATA_A_Q6_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q6_K 256 struct block_q6_K { - uint8_t ql[QUANT_K/2]; - uint8_t qh[QUANT_K/4]; - int8_t scales[QUANT_K/16]; + uint8_t ql[QUANT_K_Q6_K/2]; + uint8_t qh[QUANT_K_Q6_K/4]; + int8_t scales[QUANT_K_Q6_K/16]; float16_t d; }; +struct block_q6_K_packed16 +{ + uint16_t ql[QUANT_K_Q6_K/2/2]; + uint16_t qh[QUANT_K_Q6_K/4/2]; + int8_t scales[QUANT_K_Q6_K/16]; + float16_t d; +}; + +#if defined(DATA_A_Q6_K) +#define QUANT_K QUANT_K_Q6_K #define A_TYPE block_q6_K +#define A_TYPE_PACKED16 block_q6_K_packed16 #endif // IQuants -#if defined(DATA_A_IQ4_NL) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_IQ1_S 256 +#define QUANT_R_IQ1_S 1 + +struct block_iq1_s { + float16_t d; + uint8_t qs[QUANT_K_IQ1_S/8]; + uint16_t qh[QUANT_K_IQ1_S/32]; +}; + +#define QUANT_K_IQ1_M 256 +#define QUANT_R_IQ1_M 1 + +struct block_iq1_m { + uint8_t qs[QUANT_K_IQ1_M/8]; + uint8_t qh[QUANT_K_IQ1_M/16]; + uint16_t scales[QUANT_K_IQ1_M/64]; +}; + +struct block_iq1_m_packed64 { + uint64_t qs[QUANT_K_IQ1_M/8/8]; + uint64_t qh[QUANT_K_IQ1_M/16/8]; + uint64_t scales; +}; + +#if defined(DATA_A_IQ1_S) +#define QUANT_K QUANT_K_IQ1_S +#define QUANT_R QUANT_R_IQ1_S +#define A_TYPE block_iq1_s +#endif + +#if defined(DATA_A_IQ1_M) +#define QUANT_K QUANT_K_IQ1_M +#define QUANT_R QUANT_R_IQ1_M +#define A_TYPE block_iq1_m +#endif + +#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) +#define IQ1S_DELTA 0.125f +#define IQ1M_DELTA 0.125f + +// Packed IQ1S grid where every 2 vec8 are encoded on 32 bits (2 bits per coordinate). +const uint[1024] iq1s_grid_const = { + 0xfffdffff, 0xfff7fff0, 0xffccfff5, 0xffdfffc0, 0xffd7ffdd, 0xff30ffd5, 0xff03ff0c, 0xff10ff01, + 0xff7dff7f, 0xff75ff77, 0xff5fff40, 0xff57ff5d, 0xfcf3ff55, 0xfcccfcf0, 0xfcc1fcc3, 0xfcc5fcc4, + 0xfc3cfcd0, 0xfc34fc31, 0xfc00fc0d, 0xfc1cfc05, 0xfc11fc13, 0xfc70fc17, 0xfc43fc4c, 0xfc50fc41, + 0xfdfdfdff, 0xfdf5fdf7, 0xfddffdc0, 0xfdd7fddd, 0xfd30fdd5, 0xfd04fd0c, 0xfd14fd13, 0xfd7dfd7f, + 0xfd75fd77, 0xfd40fd4c, 0xfd5ffd44, 0xfd57fd5d, 0xf3ccfd55, 0xf3c1f3c3, 0xf33cf3d0, 0xf300f334, + 0xf313f305, 0xf34cf310, 0xf350f344, 0xf0f3f0fc, 0xf0f1f0f0, 0xf0c7f0c0, 0xf0d4f0c5, 0xf030f03f, + 0xf00ff035, 0xf003f00c, 0xf001f000, 0xf01ff004, 0xf010f01d, 0xf015f017, 0xf04cf07c, 0xf047f040, + 0xf05cf045, 0xf050f053, 0xf054f051, 0xf1c4f1c3, 0xf133f13c, 0xf10df10f, 0xf107f100, 0xf11cf11f, + 0xf114f111, 0xf14cf170, 0xf144f143, 0xf7fdf7ff, 0xf7f5f7f7, 0xf7dff7c0, 0xf7d7f7dd, 0xf730f7d5, + 0xf701f70c, 0xf77ff710, 0xf777f77d, 0xf740f775, 0xf75df75f, 0xf755f757, 0xf4ccf4f0, 0xf4c4f4c3, + 0xf4d0f4d3, 0xf40ff43c, 0xf400f40c, 0xf413f41c, 0xf44cf414, 0xf441f443, 0xf450f444, 0xf5fdf5ff, + 0xf5f5f5f7, 0xf5dff5c0, 0xf5d7f5dd, 0xf530f5d5, 0xf504f50c, 0xf510f51c, 0xf57df57f, 0xf577f570, + 0xf540f575, 0xf55df55f, 0xf555f557, 0xcfcccfcf, 0xcfc4cfc3, 0xcfd0cfd3, 0xcf33cf3c, 0xcf00cf0f, + 0xcf1ccf07, 0xcf10cf13, 0xcf4ccf14, 0xcf41cf43, 0xcf50cf5c, 0xccf3ccfc, 0xccf4ccf1, 0xcccdcccf, + 0xccc7ccc0, 0xccd3ccdc, 0xcc30ccd4, 0xcc0fcc35, 0xcc0dcc0c, 0xcc00cc03, 0xcc04cc01, 0xcc10cc1f, + 0xcc4dcc73, 0xcc5ccc40, 0xcdcccc53, 0xcdc1cdc3, 0xcd3fcdd0, 0xcd34cd31, 0xcd00cd0d, 0xcd05cd07, + 0xcd11cd13, 0xcd4ccd70, 0xcd41cd43, 0xc3fccd50, 0xc3f4c3f1, 0xc3c0c3c3, 0xc3c4c3c7, 0xc3d1c3dc, + 0xc330c33c, 0xc337c331, 0xc30cc335, 0xc300c303, 0xc304c301, 0xc310c31d, 0xc373c317, 0xc34fc374, + 0xc340c343, 0xc344c347, 0xc35cc345, 0xc350c353, 0xc0fdc354, 0xc0f5c0f0, 0xc0c3c0cc, 0xc0c1c0c0, + 0xc0dfc0c4, 0xc0d0c0dd, 0xc0d5c0d7, 0xc033c03c, 0xc031c030, 0xc00dc00c, 0xc000c003, 0xc004c001, + 0xc01cc005, 0xc010c013, 0xc014c011, 0xc07dc07f, 0xc070c073, 0xc075c077, 0xc04cc04f, 0xc040c043, + 0xc044c041, 0xc05fc045, 0xc050c05d, 0xc1f3c1fc, 0xc1f1c1f0, 0xc1c1c1c0, 0xc1c5c1c7, 0xc1d1c1dc, + 0xc13dc13f, 0xc130c133, 0xc135c137, 0xc100c10c, 0xc107c101, 0xc11cc104, 0xc110c113, 0xc114c117, + 0xc171c115, 0xc14dc175, 0xc153c140, 0xc7ccc154, 0xc7d0c7c1, 0xc733c73c, 0xc734c731, 0xc700c70f, + 0xc705c707, 0xc71cc71f, 0xc711c713, 0xc770c714, 0xc743c74c, 0xc4cfc750, 0xc4c0c4cd, 0xc4dcc4c5, + 0xc43dc4d0, 0xc430c433, 0xc40cc437, 0xc400c403, 0xc404c401, 0xc41fc405, 0xc415c410, 0xc44cc474, + 0xc440c44d, 0xc45cc447, 0xc454c451, 0xc5c1c5f4, 0xc5d1c5d3, 0xc531c533, 0xc50fc534, 0xc500c50d, + 0xc51cc507, 0xc514c511, 0xc54cc570, 0xc545c541, 0xdffddfff, 0xdff5dff7, 0xdfdfdfc0, 0xdfd0dfdd, + 0xdfd5dfd7, 0xdf0cdf30, 0xdf1cdf04, 0xdf7fdf10, 0xdf77df7d, 0xdf40df75, 0xdf5ddf5f, 0xdf57df50, + 0xdcf0df55, 0xdcc3dccc, 0xdcd0dcc4, 0xdc33dc3d, 0xdc00dc34, 0xdc05dc07, 0xdc13dc1c, 0xdc11dc10, + 0xdc4fdc70, 0xdc44dc41, 0xddfcdc50, 0xddf5ddf7, 0xddc0ddcc, 0xdddddddf, 0xddd5ddd7, 0xdd0cdd30, + 0xdd04dd01, 0xdd7cdd10, 0xdd75dd77, 0xdd40dd4c, 0xdd5ddd5f, 0xdd55dd57, 0xd3c3d3f0, 0xd3c4d3c1, + 0xd333d3d0, 0xd331d330, 0xd30dd334, 0xd307d300, 0xd311d305, 0xd34cd370, 0xd344d343, 0xd350d35c, + 0xd0c0d0f4, 0xd0d4d0dc, 0xd030d03f, 0xd00cd037, 0xd000d003, 0xd01dd004, 0xd017d010, 0xd04fd074, + 0xd040d043, 0xd045d047, 0xd053d05c, 0xd054d051, 0xd1cfd1f0, 0xd1c4d1cd, 0xd13cd1d0, 0xd100d134, + 0xd11cd11f, 0xd173d114, 0xd14fd171, 0xd7ffd145, 0xd7f7d7fd, 0xd7c0d7f5, 0xd7ddd7df, 0xd7d5d7d7, + 0xd70cd730, 0xd710d703, 0xd77dd77f, 0xd775d777, 0xd75dd75f, 0xd755d757, 0xd4ccd4f4, 0xd4c4d4c3, + 0xd431d4d0, 0xd40dd434, 0xd41cd400, 0xd411d413, 0xd470d414, 0xd441d44f, 0xd453d444, 0xd5ffd450, + 0xd5f7d5fd, 0xd5dfd5f5, 0xd5d7d5dd, 0xd530d5d5, 0xd501d50c, 0xd510d504, 0xd57dd57f, 0xd575d577, + 0xd55fd540, 0xd557d55d, 0x3ff0d555, 0x3fc13fcc, 0x3f343fd0, 0x3f003f0d, 0x3f053f07, 0x3f133f1c, + 0x3f433f11, 0x3f5c3f44, 0x3cff3f51, 0x3cf33cfc, 0x3cf43cf1, 0x3cc03ccd, 0x3cc73cc1, 0x3cdc3cc5, + 0x3cd43cd1, 0x3c373c30, 0x3c0c3c35, 0x3c003c03, 0x3c043c01, 0x3c103c05, 0x3c153c17, 0x3c733c7c, + 0x3c4f3c71, 0x3c403c4d, 0x3c5c3c5f, 0x3df03c5d, 0x3dc33dcc, 0x3dd03dc1, 0x3d0d3d3c, 0x3d053d00, + 0x3d143d13, 0x3d433d74, 0x33fc3d50, 0x33c433c0, 0x333033d4, 0x33353337, 0x3303330c, 0x33013300, + 0x331d331c, 0x33173310, 0x337c3315, 0x33743371, 0x334d334f, 0x335f3340, 0x3354335c, 0x30fd30fc, + 0x30f530f0, 0x30c330cc, 0x30c130c0, 0x30df30c4, 0x30d530d0, 0x3033303c, 0x30313030, 0x300f3034, + 0x3003300c, 0x30013000, 0x30043007, 0x3013301c, 0x30113010, 0x307d3014, 0x30703073, 0x304c3077, + 0x30403043, 0x30443041, 0x30503045, 0x30553057, 0x31f031fc, 0x31c331f4, 0x31c731c0, 0x31dc31c5, + 0x31d431d3, 0x313d313f, 0x31373130, 0x310c310f, 0x3100310d, 0x31043101, 0x3110311d, 0x317c3117, + 0x31753170, 0x31403143, 0x3153315c, 0x37f03151, 0x37c037cc, 0x37d037c5, 0x3734373d, 0x3700370f, + 0x371c3707, 0x37113713, 0x37703714, 0x3743374c, 0x37443741, 0x34fc3750, 0x34f134f0, 0x34cf34f5, + 0x34c034c3, 0x34dc34c7, 0x34d134d3, 0x3430343f, 0x340c3435, 0x3403340d, 0x34013400, 0x341f3404, + 0x3410341d, 0x34153411, 0x34743471, 0x3440344d, 0x34473441, 0x3453345c, 0x34543451, 0x353335c1, + 0x35343531, 0x35073500, 0x35133505, 0x35433514, 0x0ffc3550, 0x0ff00ff3, 0x0ff40ff1, 0x0fc00fcd, + 0x0fdc0fc5, 0x0fd40fd3, 0x0f300f3f, 0x0f0c0f37, 0x0f000f03, 0x0f040f01, 0x0f170f10, 0x0f740f71, + 0x0f470f40, 0x0f5c0f5f, 0x0f540f51, 0x0cf70cf0, 0x0cf50cf4, 0x0cc30ccc, 0x0cc10cc0, 0x0cc40cc7, + 0x0cd00cdf, 0x0cd70cd1, 0x0c3c0cd5, 0x0c300c33, 0x0c340c31, 0x0c0c0c0f, 0x0c030c0d, 0x0c010c00, + 0x0c040c07, 0x0c1c0c05, 0x0c100c13, 0x0c140c11, 0x0c700c7d, 0x0c430c4c, 0x0c410c40, 0x0c5f0c44, + 0x0c550c50, 0x0df10dfc, 0x0dc00dcd, 0x0ddc0dc5, 0x0d3d0dd3, 0x0d350d30, 0x0d030d0c, 0x0d010d00, + 0x0d1d0d04, 0x0d700d10, 0x0d4d0d4f, 0x0d440d40, 0x0d530d45, 0x03f003f3, 0x03c303cc, 0x03c103c0, + 0x03c403c7, 0x03d003dc, 0x03d503d7, 0x0333033c, 0x03310330, 0x03350334, 0x030c030f, 0x03000303, + 0x03070301, 0x03050304, 0x031d031c, 0x03100313, 0x03140311, 0x0377037f, 0x034c0375, 0x03400343, + 0x03440341, 0x0353035c, 0x03550350, 0x00fd00fc, 0x00f000f3, 0x00f400f1, 0x00cc00cf, 0x00c300cd, + 0x00c100c0, 0x00c500c4, 0x00d300dc, 0x00d100d0, 0x003f00d4, 0x003d003c, 0x00300033, 0x00370031, + 0x000f0034, 0x000d000c, 0x00000003, 0x00070001, 0x00050004, 0x001c001f, 0x00100013, 0x00170011, + 0x00150014, 0x0073007c, 0x00740070, 0x004f0075, 0x0043004c, 0x00410040, 0x00440047, 0x0053005c, + 0x00510050, 0x01ff0054, 0x01fd01fc, 0x01f101f3, 0x01f401f7, 0x01c301cc, 0x01c701c0, 0x01df01c4, + 0x01dd01dc, 0x01d001d3, 0x01d701d1, 0x013c01d4, 0x01310130, 0x01340137, 0x010f0135, 0x010d010c, + 0x01000103, 0x01070101, 0x01050104, 0x0113011c, 0x01140110, 0x0170017d, 0x01770171, 0x01750174, + 0x0140014c, 0x015d0145, 0x01510150, 0x01540157, 0x07f007f3, 0x07f407f1, 0x07c007cf, 0x07dc07c7, + 0x073007d5, 0x07350737, 0x0703070c, 0x07010700, 0x07040707, 0x071d071f, 0x07100713, 0x0774077d, + 0x074d074f, 0x07470740, 0x0754075c, 0x04fd04fc, 0x04f504f0, 0x04c304cc, 0x04c104c0, 0x04d004c4, + 0x0433043c, 0x04310430, 0x040f0434, 0x040d040c, 0x04000403, 0x04070401, 0x04050404, 0x0413041c, + 0x04110410, 0x047c0414, 0x04740470, 0x0443044c, 0x04410440, 0x04440447, 0x05f30450, 0x05c005f7, + 0x05df05c5, 0x05d105d0, 0x053005d4, 0x05340537, 0x0500050c, 0x05070501, 0x051d0504, 0x05170510, + 0x057c0515, 0x054d0575, 0x05410540, 0x05450547, 0x1ff0055c, 0x1fc11fc3, 0x1fd01fc4, 0x1f0f1f33, + 0x1f011f00, 0x1f051f07, 0x1f131f1c, 0x1f141f11, 0x1f411f7c, 0x1cfc1f50, 0x1cf11cf3, 0x1ccd1cf4, + 0x1cdc1cc0, 0x1cd11cdd, 0x1c301cd4, 0x1c0c1c34, 0x1c011c00, 0x1c101c04, 0x1c151c11, 0x1c751c73, + 0x1c401c4d, 0x1c511c5c, 0x1dcc1c54, 0x1dc41dc1, 0x1d3c1d3f, 0x1d001d31, 0x1d071d01, 0x1d701d1f, + 0x1d411d4c, 0x13cc1d50, 0x13c013cd, 0x13c513c1, 0x13d113dc, 0x133f13d4, 0x1330133d, 0x13351337, + 0x1303130c, 0x13011300, 0x13051304, 0x131d131f, 0x13731310, 0x13741370, 0x134d134f, 0x13401343, + 0x13471341, 0x135c1345, 0x13541353, 0x10f710f0, 0x10cc10f5, 0x10c110c0, 0x103310c4, 0x10311030, + 0x100f1034, 0x1003100c, 0x10011000, 0x101c1004, 0x10101013, 0x10141011, 0x10741071, 0x104c1075, + 0x10411040, 0x10451044, 0x1050105d, 0x10571051, 0x11f411fd, 0x11df11c0, 0x11d711d1, 0x113f11d4, + 0x11371130, 0x110c1135, 0x11001103, 0x11071101, 0x111f1105, 0x11171110, 0x117d117f, 0x11751170, + 0x11411143, 0x11441147, 0x1153115f, 0x11551151, 0x17c417c1, 0x173c17d0, 0x1700170d, 0x171c1705, + 0x17701714, 0x1747174c, 0x14fc1751, 0x14cf14f3, 0x14dc14c0, 0x14d114d3, 0x143f14d4, 0x1430143c, + 0x14371431, 0x1403140c, 0x14011400, 0x141f1404, 0x14151410, 0x1473147d, 0x14401475, 0x1453145c, + 0x14541450, 0x15c115cc, 0x153c15c7, 0x15341533, 0x1500150f, 0x15051507, 0x15101513, 0x15711514, + 0x15471543, 0x15511545, 0x7ffd7fff, 0x7ff57ff7, 0x7fdd7fdf, 0x7fd57fd7, 0x7f0f7f30, 0x7f037f0c, + 0x7f047f01, 0x7f7f7f10, 0x7f777f7d, 0x7f407f75, 0x7f5d7f5f, 0x7f557f57, 0x7ccc7cf0, 0x7cc17cc3, + 0x7cd07cc4, 0x7c337c3c, 0x7c0f7c34, 0x7c007c0d, 0x7c077c01, 0x7c137c04, 0x7c147c11, 0x7c747c70, + 0x7c417c43, 0x7c507c44, 0x7dfd7dff, 0x7df57df7, 0x7ddf7dc0, 0x7dd77ddd, 0x7d0c7dd5, 0x7d047d03, + 0x7d7f7d10, 0x7d777d7d, 0x7d407d75, 0x7d5d7d5f, 0x7d557d57, 0x73c473c3, 0x7333733c, 0x7300730c, + 0x731c7305, 0x73147313, 0x73447343, 0x70f470fc, 0x70c070cd, 0x70d170c5, 0x703f70d4, 0x7030703c, + 0x700c7037, 0x70007003, 0x70047001, 0x70107005, 0x70177011, 0x707c7015, 0x70717073, 0x704f7074, + 0x7040704d, 0x70517047, 0x71c171cc, 0x71d071c4, 0x7133713c, 0x71357134, 0x7100710f, 0x71057104, + 0x7111711c, 0x71707115, 0x7145714c, 0x77ff7153, 0x77f777fd, 0x77c077f5, 0x77dd77df, 0x77d577d7, + 0x7730773c, 0x7703770c, 0x77107704, 0x777f7714, 0x7777777d, 0x77407775, 0x775d775f, 0x77557757, + 0x74f174f0, 0x74c374cc, 0x74d074c1, 0x7433743c, 0x74347431, 0x740d740f, 0x74057400, 0x7413741c, + 0x74417470, 0x74507444, 0x75fd75ff, 0x75f575f7, 0x75df75c0, 0x75d775dd, 0x753075d5, 0x7503750c, + 0x757f7501, 0x7577757d, 0x75407575, 0x755d755f, 0x75557557, 0x4fcc4ff0, 0x4fc74fc1, 0x4fd04fc4, + 0x4f314f3c, 0x4f004f34, 0x4f054f07, 0x4f154f14, 0x4f4c4f70, 0x4f414f43, 0x4f504f44, 0x4cf34cfc, + 0x4cf44cf1, 0x4cc04ccf, 0x4cc54cc7, 0x4cd34cdc, 0x4cd44cd1, 0x4c304c3f, 0x4c0c4c0f, 0x4c004c03, + 0x4c044c01, 0x4c104c1d, 0x4c714c73, 0x4c404c4d, 0x4c5c4c47, 0x4c514c53, 0x4df04c54, 0x4dc34dcc, + 0x4dd04dc4, 0x4d314d33, 0x4d0f4d34, 0x4d004d0d, 0x4d114d07, 0x4d704d14, 0x4d414d43, 0x43fc4d54, + 0x43f143f3, 0x43c043cf, 0x43d143c7, 0x4335433f, 0x4303430c, 0x43014300, 0x43044307, 0x431c431f, + 0x4310431d, 0x43714373, 0x4343434d, 0x43474340, 0x4354435c, 0x40f040ff, 0x40f540f7, 0x40cc40cf, + 0x40c040c3, 0x40c440c1, 0x40d040dc, 0x40d540d4, 0x4033403c, 0x40314030, 0x400f4034, 0x400d400c, + 0x40004003, 0x40074001, 0x40054004, 0x4013401c, 0x40114010, 0x407c4014, 0x40774070, 0x404d404c, + 0x40404043, 0x40444041, 0x405f4045, 0x4050405d, 0x40554057, 0x41f341fc, 0x41c041cf, 0x41df41c4, + 0x41d441d1, 0x41374130, 0x410c4134, 0x4100410d, 0x41044101, 0x41174110, 0x4173417d, 0x41754174, + 0x4143414d, 0x41534140, 0x41544151, 0x47c147f0, 0x47d047c4, 0x4731473c, 0x470d470f, 0x47014700, + 0x47134705, 0x47704710, 0x4741474c, 0x47504744, 0x44f144f3, 0x44cf44f4, 0x44c044cd, 0x44c544c7, + 0x44dc44df, 0x44d144d3, 0x443d443f, 0x44374430, 0x440c4435, 0x44004403, 0x44044401, 0x4410441d, + 0x44154411, 0x4473447c, 0x444d444f, 0x44454440, 0x4451445c, 0x45c045f0, 0x453345d0, 0x45344531, + 0x4500450f, 0x451c4507, 0x454c4570, 0x45404543, 0x5fff4541, 0x5ff75ffd, 0x5fc05ff5, 0x5fdd5fdf, + 0x5fd55fd7, 0x5f0c5f30, 0x5f015f03, 0x5f7f5f04, 0x5f775f7d, 0x5f405f75, 0x5f5d5f5f, 0x5f555f57, + 0x5cf45cf0, 0x5cc35ccc, 0x5cc45cc1, 0x5c315cc5, 0x5c0c5c34, 0x5c075c00, 0x5c1c5c05, 0x5c705c13, + 0x5c4d5c4f, 0x5c445c41, 0x5df75dfd, 0x5dcf5df5, 0x5ddd5dc4, 0x5dd55dd7, 0x5d0c5d30, 0x5d045d01, + 0x5d7f5d10, 0x5d775d7d, 0x5d405d75, 0x5d5d5d5f, 0x5d555d57, 0x53d053c4, 0x5333533c, 0x5303530f, + 0x53075300, 0x531c5305, 0x53115310, 0x53145317, 0x50f15370, 0x50cf50f4, 0x50c050cd, 0x50d150c7, + 0x503d50d4, 0x500c5030, 0x50005003, 0x50045001, 0x50155010, 0x5073507c, 0x50715070, 0x504d5074, + 0x50475040, 0x51cc51f0, 0x51c551c1, 0x51d051dc, 0x51315133, 0x510d5135, 0x51015100, 0x511f5107, + 0x5171511d, 0x5140514f, 0x51445141, 0x5153515c, 0x57ff5151, 0x57f757fd, 0x57df57f5, 0x57d757dd, + 0x570c57d5, 0x57015703, 0x577f5704, 0x5777577d, 0x57405775, 0x575d575f, 0x57555757, 0x54c354f0, + 0x54dc54c4, 0x543c54d0, 0x5400540f, 0x541c5405, 0x54145411, 0x5441544f, 0x55fd55ff, 0x55f555f7, + 0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557 +}; + +shared uint16_t iq1s_grid[2048]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq1s_grid_const.length(); i += wgsize.x) { + uint idx = i + gl_LocalInvocationIndex.x; + if (iq1s_grid_const.length() % wgsize.x == 0 || idx < iq1s_grid_const.length()) { + u16vec2 g = unpack16(iq1s_grid_const[idx]); + iq1s_grid[2*idx+0] = g.x; + iq1s_grid[2*idx+1] = g.y; + } + } + barrier(); +} +#endif + +#define QUANT_K_IQ2_XXS 256 +#define QUANT_R_IQ2_XXS 1 + +struct block_iq2_xxs +{ + float16_t d; + uint8_t qs[QUANT_K_IQ2_XXS/4]; +}; + +struct block_iq2_xxs_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ2_XXS/8]; +}; + +#if defined(DATA_A_IQ2_XXS) + +const uvec2[256] iq2xxs_grid_const = { + uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808), + uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x082b0808, 0x08080808), + uvec2(0x082b082b, 0x08080808), uvec2(0x082b2b08, 0x08080808), uvec2(0x082b2b2b, 0x08080808), uvec2(0x19080819, 0x08080808), + uvec2(0x19081908, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808), + uvec2(0x192b1908, 0x08080808), uvec2(0x2b080808, 0x08080808), uvec2(0x2b08082b, 0x08080808), uvec2(0x2b082b2b, 0x08080808), + uvec2(0x2b2b082b, 0x08080808), uvec2(0x08080819, 0x08080819), uvec2(0x08081908, 0x08080819), uvec2(0x08190808, 0x08080819), + uvec2(0x08191919, 0x08080819), uvec2(0x19080808, 0x08080819), uvec2(0x2b081908, 0x08080819), uvec2(0x2b192b08, 0x08080819), + uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b), uvec2(0x082b082b, 0x0808082b), uvec2(0x2b08082b, 0x0808082b), + uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x082b0819, 0x08081908), + uvec2(0x082b1908, 0x08081908), uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19082b08, 0x08081908), + uvec2(0x192b0808, 0x08081908), uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b190808, 0x08081908), + uvec2(0x2b2b1908, 0x08081908), uvec2(0x08080808, 0x08081919), uvec2(0x0808082b, 0x08081919), uvec2(0x08082b08, 0x08081919), + uvec2(0x082b0808, 0x08081919), uvec2(0x1908192b, 0x08081919), uvec2(0x192b2b19, 0x08081919), uvec2(0x2b080808, 0x08081919), + uvec2(0x2b190819, 0x08081919), uvec2(0x08082b19, 0x0808192b), uvec2(0x08190808, 0x0808192b), uvec2(0x19080808, 0x0808192b), + uvec2(0x2b081908, 0x0808192b), uvec2(0x2b2b1908, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x08081919, 0x08082b08), + uvec2(0x08082b08, 0x08082b08), uvec2(0x08191908, 0x08082b08), uvec2(0x082b2b08, 0x08082b08), uvec2(0x19080819, 0x08082b08), + uvec2(0x19081908, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x1919082b, 0x08082b08), uvec2(0x2b082b08, 0x08082b08), + uvec2(0x08081908, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x0808082b, 0x08082b2b), uvec2(0x08191908, 0x08082b2b), + uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x082b0819, 0x08190808), + uvec2(0x19080808, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x2b081908, 0x08190808), uvec2(0x2b190808, 0x08190808), + uvec2(0x2b191919, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x08082b08, 0x08190819), uvec2(0x082b0808, 0x08190819), + uvec2(0x19190808, 0x08190819), uvec2(0x19192b2b, 0x08190819), uvec2(0x2b080808, 0x08190819), uvec2(0x082b1908, 0x0819082b), + uvec2(0x19081919, 0x0819082b), uvec2(0x08080808, 0x08191908), uvec2(0x08082b08, 0x08191908), uvec2(0x082b0808, 0x08191908), + uvec2(0x082b1919, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x08192b08, 0x08191919), + uvec2(0x192b082b, 0x08191919), uvec2(0x08080808, 0x0819192b), uvec2(0x0819192b, 0x0819192b), uvec2(0x08080819, 0x08192b08), + uvec2(0x08081908, 0x08192b08), uvec2(0x08190808, 0x08192b08), uvec2(0x19080808, 0x08192b08), uvec2(0x2b080819, 0x08192b08), + uvec2(0x08080808, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x2b2b0808, 0x08192b19), uvec2(0x19190819, 0x08192b2b), + uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08082b2b, 0x082b0808), uvec2(0x19081908, 0x082b0808), + uvec2(0x192b0819, 0x082b0808), uvec2(0x2b080808, 0x082b0808), uvec2(0x2b08082b, 0x082b0808), uvec2(0x082b2b19, 0x082b0819), + uvec2(0x19082b08, 0x082b0819), uvec2(0x08080808, 0x082b082b), uvec2(0x0808082b, 0x082b082b), uvec2(0x08080819, 0x082b1908), + uvec2(0x08081908, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x19080808, 0x082b1908), uvec2(0x1919192b, 0x082b1908), + uvec2(0x08080808, 0x082b1919), uvec2(0x19080819, 0x082b1919), uvec2(0x192b1908, 0x082b1919), uvec2(0x2b190808, 0x082b192b), + uvec2(0x08082b08, 0x082b2b08), uvec2(0x082b0808, 0x082b2b08), uvec2(0x2b191908, 0x082b2b08), uvec2(0x19081908, 0x082b2b2b), + uvec2(0x08080819, 0x19080808), uvec2(0x08081908, 0x19080808), uvec2(0x08190808, 0x19080808), uvec2(0x08192b08, 0x19080808), + uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808), uvec2(0x19080808, 0x19080808), uvec2(0x19082b08, 0x19080808), + uvec2(0x1919192b, 0x19080808), uvec2(0x192b0808, 0x19080808), uvec2(0x2b080819, 0x19080808), uvec2(0x2b081908, 0x19080808), + uvec2(0x2b190808, 0x19080808), uvec2(0x08080808, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x192b0819, 0x19080819), + uvec2(0x2b080808, 0x19080819), uvec2(0x2b081919, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08190808, 0x1908082b), + uvec2(0x19082b08, 0x1908082b), uvec2(0x1919192b, 0x1908082b), uvec2(0x192b2b08, 0x1908082b), uvec2(0x08080808, 0x19081908), + uvec2(0x08082b08, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x2b080808, 0x19081908), uvec2(0x2b192b19, 0x19081908), + uvec2(0x0819082b, 0x19081919), uvec2(0x082b1908, 0x19081919), uvec2(0x08080808, 0x1908192b), uvec2(0x08080819, 0x19082b08), + uvec2(0x08081908, 0x19082b08), uvec2(0x08190808, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x19081919, 0x19082b08), + uvec2(0x08080808, 0x19082b19), uvec2(0x19192b08, 0x19082b19), uvec2(0x192b0819, 0x19082b19), uvec2(0x2b08082b, 0x19082b19), + uvec2(0x19081919, 0x19082b2b), uvec2(0x2b190808, 0x19082b2b), uvec2(0x08080808, 0x19190808), uvec2(0x08082b08, 0x19190808), + uvec2(0x08190819, 0x19190808), uvec2(0x08192b19, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x2b080808, 0x19190808), + uvec2(0x2b082b08, 0x19190808), uvec2(0x08081908, 0x19190819), uvec2(0x1908082b, 0x19190819), uvec2(0x2b2b1908, 0x19190819), + uvec2(0x2b190819, 0x1919082b), uvec2(0x2b190808, 0x19191908), uvec2(0x2b19082b, 0x19191908), uvec2(0x08082b2b, 0x19191919), + uvec2(0x08080819, 0x1919192b), uvec2(0x19191908, 0x1919192b), uvec2(0x08080808, 0x19192b08), uvec2(0x08190819, 0x19192b08), + uvec2(0x08192b19, 0x19192b08), uvec2(0x192b1908, 0x19192b08), uvec2(0x19080808, 0x19192b19), uvec2(0x08082b08, 0x19192b2b), + uvec2(0x08081908, 0x192b0808), uvec2(0x08190808, 0x192b0808), uvec2(0x19080808, 0x192b0808), uvec2(0x192b2b08, 0x192b0808), + uvec2(0x08080808, 0x192b0819), uvec2(0x19191919, 0x192b0819), uvec2(0x08192b08, 0x192b082b), uvec2(0x192b0808, 0x192b082b), + uvec2(0x08080808, 0x192b1908), uvec2(0x08081919, 0x192b1908), uvec2(0x08190808, 0x192b1919), uvec2(0x0819082b, 0x192b1919), + uvec2(0x2b081908, 0x192b1919), uvec2(0x1908082b, 0x192b2b08), uvec2(0x08080808, 0x2b080808), uvec2(0x0808082b, 0x2b080808), + uvec2(0x08082b2b, 0x2b080808), uvec2(0x19080819, 0x2b080808), uvec2(0x2b08082b, 0x2b080808), uvec2(0x08081908, 0x2b080819), + uvec2(0x08192b08, 0x2b080819), uvec2(0x19080808, 0x2b080819), uvec2(0x08190819, 0x2b08082b), uvec2(0x08080819, 0x2b081908), + uvec2(0x08081908, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x08191919, 0x2b081908), uvec2(0x19080808, 0x2b081908), + uvec2(0x192b0808, 0x2b081908), uvec2(0x08080808, 0x2b081919), uvec2(0x1908192b, 0x2b081919), uvec2(0x2b191908, 0x2b081919), + uvec2(0x08082b19, 0x2b08192b), uvec2(0x19080808, 0x2b08192b), uvec2(0x192b0808, 0x2b08192b), uvec2(0x0808082b, 0x2b082b08), + uvec2(0x08081908, 0x2b082b19), uvec2(0x08190819, 0x2b082b2b), uvec2(0x08081908, 0x2b190808), uvec2(0x08190808, 0x2b190808), + uvec2(0x082b1908, 0x2b190808), uvec2(0x19080808, 0x2b190808), uvec2(0x2b2b0819, 0x2b190808), uvec2(0x0819192b, 0x2b190819), + uvec2(0x2b080808, 0x2b190819), uvec2(0x19081919, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x082b082b, 0x2b191908), + uvec2(0x19081908, 0x2b191908), uvec2(0x19190819, 0x2b191919), uvec2(0x2b080819, 0x2b192b08), uvec2(0x082b0808, 0x2b192b19), + uvec2(0x0808082b, 0x2b2b0808), uvec2(0x19190808, 0x2b2b0808), uvec2(0x2b081919, 0x2b2b0808), uvec2(0x08082b19, 0x2b2b0819), + uvec2(0x08080808, 0x2b2b082b), uvec2(0x08192b08, 0x2b2b1908), uvec2(0x19190808, 0x2b2b2b08), uvec2(0x08081908, 0x2b2b2b19) +}; + +shared uvec2 iq2xxs_grid[256]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq2xxs_grid.length(); i += wgsize.x) { + if (iq2xxs_grid_const.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xxs_grid_const.length()) { + iq2xxs_grid[i + gl_LocalInvocationIndex.x] = iq2xxs_grid_const[i + gl_LocalInvocationIndex.x]; + } + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ2_XXS +#define QUANT_R QUANT_R_IQ2_XXS +#define A_TYPE block_iq2_xxs +#define A_TYPE_PACKED16 block_iq2_xxs_packed16 +#endif + +#define QUANT_K_IQ2_XS 256 +#define QUANT_R_IQ2_XS 1 + +struct block_iq2_xs +{ + float16_t d; + uint16_t qs[QUANT_K_IQ2_XS/8]; + uint8_t scales[QUANT_K_IQ2_XS/32]; +}; + +struct block_iq2_xs_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ2_XS/8]; + uint16_t scales[QUANT_K_IQ2_XS/64]; +}; + +#if defined(DATA_A_IQ2_XS) + +const uvec2 iq2xs_grid_const[512] = { + uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808), + uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x0819192b, 0x08080808), + uvec2(0x08192b19, 0x08080808), uvec2(0x082b0808, 0x08080808), uvec2(0x082b082b, 0x08080808), uvec2(0x082b1919, 0x08080808), + uvec2(0x082b2b08, 0x08080808), uvec2(0x19080819, 0x08080808), uvec2(0x19081908, 0x08080808), uvec2(0x1908192b, 0x08080808), + uvec2(0x19082b19, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x1919082b, 0x08080808), uvec2(0x19191919, 0x08080808), + uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808), uvec2(0x192b1908, 0x08080808), uvec2(0x2b080808, 0x08080808), + uvec2(0x2b08082b, 0x08080808), uvec2(0x2b081919, 0x08080808), uvec2(0x2b082b08, 0x08080808), uvec2(0x2b190819, 0x08080808), + uvec2(0x2b191908, 0x08080808), uvec2(0x2b192b19, 0x08080808), uvec2(0x2b2b0808, 0x08080808), uvec2(0x08080819, 0x08080819), + uvec2(0x08081908, 0x08080819), uvec2(0x0808192b, 0x08080819), uvec2(0x08082b19, 0x08080819), uvec2(0x08190808, 0x08080819), + uvec2(0x0819082b, 0x08080819), uvec2(0x08191919, 0x08080819), uvec2(0x08192b08, 0x08080819), uvec2(0x08192b2b, 0x08080819), + uvec2(0x082b0819, 0x08080819), uvec2(0x082b1908, 0x08080819), uvec2(0x19080808, 0x08080819), uvec2(0x1908082b, 0x08080819), + uvec2(0x19081919, 0x08080819), uvec2(0x19082b08, 0x08080819), uvec2(0x19190819, 0x08080819), uvec2(0x19191908, 0x08080819), + uvec2(0x192b0808, 0x08080819), uvec2(0x192b2b08, 0x08080819), uvec2(0x2b080819, 0x08080819), uvec2(0x2b081908, 0x08080819), + uvec2(0x2b190808, 0x08080819), uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b), uvec2(0x08081919, 0x0808082b), + uvec2(0x08082b08, 0x0808082b), uvec2(0x08190819, 0x0808082b), uvec2(0x08191908, 0x0808082b), uvec2(0x082b0808, 0x0808082b), + uvec2(0x19080819, 0x0808082b), uvec2(0x19081908, 0x0808082b), uvec2(0x19190808, 0x0808082b), uvec2(0x19191919, 0x0808082b), + uvec2(0x2b080808, 0x0808082b), uvec2(0x2b082b2b, 0x0808082b), uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908), + uvec2(0x0808192b, 0x08081908), uvec2(0x08082b19, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x0819082b, 0x08081908), + uvec2(0x08191919, 0x08081908), uvec2(0x08192b08, 0x08081908), uvec2(0x082b0819, 0x08081908), uvec2(0x082b1908, 0x08081908), + uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19081919, 0x08081908), uvec2(0x19082b08, 0x08081908), + uvec2(0x19190819, 0x08081908), uvec2(0x19191908, 0x08081908), uvec2(0x1919192b, 0x08081908), uvec2(0x192b0808, 0x08081908), + uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b190808, 0x08081908), uvec2(0x08080808, 0x08081919), + uvec2(0x0808082b, 0x08081919), uvec2(0x08081919, 0x08081919), uvec2(0x08082b08, 0x08081919), uvec2(0x08190819, 0x08081919), + uvec2(0x08191908, 0x08081919), uvec2(0x082b0808, 0x08081919), uvec2(0x19080819, 0x08081919), uvec2(0x19081908, 0x08081919), + uvec2(0x19190808, 0x08081919), uvec2(0x192b0819, 0x08081919), uvec2(0x2b080808, 0x08081919), uvec2(0x08080819, 0x0808192b), + uvec2(0x08081908, 0x0808192b), uvec2(0x08190808, 0x0808192b), uvec2(0x082b192b, 0x0808192b), uvec2(0x19080808, 0x0808192b), + uvec2(0x1908082b, 0x0808192b), uvec2(0x2b081908, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x0808082b, 0x08082b08), + uvec2(0x08081919, 0x08082b08), uvec2(0x08082b08, 0x08082b08), uvec2(0x08082b2b, 0x08082b08), uvec2(0x08190819, 0x08082b08), + uvec2(0x08191908, 0x08082b08), uvec2(0x082b0808, 0x08082b08), uvec2(0x082b1919, 0x08082b08), uvec2(0x19080819, 0x08082b08), + uvec2(0x19081908, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x19192b08, 0x08082b08), uvec2(0x2b080808, 0x08082b08), + uvec2(0x2b2b0808, 0x08082b08), uvec2(0x2b2b2b2b, 0x08082b08), uvec2(0x08080819, 0x08082b19), uvec2(0x08081908, 0x08082b19), + uvec2(0x08190808, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x2b080819, 0x08082b19), uvec2(0x2b082b19, 0x08082b19), + uvec2(0x08080808, 0x08082b2b), uvec2(0x082b0808, 0x08082b2b), uvec2(0x082b2b08, 0x08082b2b), uvec2(0x2b19192b, 0x08082b2b), + uvec2(0x2b2b0808, 0x08082b2b), uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808), uvec2(0x0808192b, 0x08190808), + uvec2(0x08082b19, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x0819082b, 0x08190808), uvec2(0x08191919, 0x08190808), + uvec2(0x08192b08, 0x08190808), uvec2(0x082b0819, 0x08190808), uvec2(0x082b1908, 0x08190808), uvec2(0x19080808, 0x08190808), + uvec2(0x1908082b, 0x08190808), uvec2(0x19081919, 0x08190808), uvec2(0x19082b08, 0x08190808), uvec2(0x19190819, 0x08190808), + uvec2(0x19191908, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x192b2b2b, 0x08190808), uvec2(0x2b080819, 0x08190808), + uvec2(0x2b081908, 0x08190808), uvec2(0x2b190808, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x0808082b, 0x08190819), + uvec2(0x08081919, 0x08190819), uvec2(0x08082b08, 0x08190819), uvec2(0x08190819, 0x08190819), uvec2(0x08191908, 0x08190819), + uvec2(0x082b0808, 0x08190819), uvec2(0x19080819, 0x08190819), uvec2(0x19081908, 0x08190819), uvec2(0x19190808, 0x08190819), + uvec2(0x2b080808, 0x08190819), uvec2(0x2b191908, 0x08190819), uvec2(0x2b19192b, 0x08190819), uvec2(0x08080819, 0x0819082b), + uvec2(0x08081908, 0x0819082b), uvec2(0x0808192b, 0x0819082b), uvec2(0x08190808, 0x0819082b), uvec2(0x19080808, 0x0819082b), + uvec2(0x192b0808, 0x0819082b), uvec2(0x08080808, 0x08191908), uvec2(0x0808082b, 0x08191908), uvec2(0x08081919, 0x08191908), + uvec2(0x08082b08, 0x08191908), uvec2(0x08190819, 0x08191908), uvec2(0x08191908, 0x08191908), uvec2(0x082b0808, 0x08191908), + uvec2(0x19080819, 0x08191908), uvec2(0x19081908, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x19190808, 0x08191908), + uvec2(0x192b1908, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x08080819, 0x08191919), uvec2(0x08081908, 0x08191919), + uvec2(0x08190808, 0x08191919), uvec2(0x19080808, 0x08191919), uvec2(0x08080808, 0x0819192b), uvec2(0x08191908, 0x0819192b), + uvec2(0x19082b19, 0x0819192b), uvec2(0x08080819, 0x08192b08), uvec2(0x08081908, 0x08192b08), uvec2(0x08190808, 0x08192b08), + uvec2(0x0819082b, 0x08192b08), uvec2(0x19080808, 0x08192b08), uvec2(0x19191908, 0x08192b08), uvec2(0x2b08192b, 0x08192b08), + uvec2(0x08080808, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x192b192b, 0x08192b19), uvec2(0x19190819, 0x08192b2b), + uvec2(0x2b2b2b19, 0x08192b2b), uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08081919, 0x082b0808), + uvec2(0x08082b08, 0x082b0808), uvec2(0x08082b2b, 0x082b0808), uvec2(0x08190819, 0x082b0808), uvec2(0x08191908, 0x082b0808), + uvec2(0x082b0808, 0x082b0808), uvec2(0x19080819, 0x082b0808), uvec2(0x19081908, 0x082b0808), uvec2(0x19190808, 0x082b0808), + uvec2(0x2b080808, 0x082b0808), uvec2(0x2b2b0808, 0x082b0808), uvec2(0x08080819, 0x082b0819), uvec2(0x08081908, 0x082b0819), + uvec2(0x08190808, 0x082b0819), uvec2(0x19080808, 0x082b0819), uvec2(0x19082b08, 0x082b0819), uvec2(0x192b1919, 0x082b0819), + uvec2(0x08080808, 0x082b082b), uvec2(0x082b082b, 0x082b082b), uvec2(0x2b080808, 0x082b082b), uvec2(0x2b2b2b08, 0x082b082b), + uvec2(0x08080819, 0x082b1908), uvec2(0x08081908, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x082b2b19, 0x082b1908), + uvec2(0x19080808, 0x082b1908), uvec2(0x08080808, 0x082b1919), uvec2(0x19080819, 0x082b1919), uvec2(0x1919082b, 0x082b1919), + uvec2(0x2b192b19, 0x082b1919), uvec2(0x08080819, 0x082b192b), uvec2(0x08192b2b, 0x082b192b), uvec2(0x2b2b192b, 0x082b192b), + uvec2(0x08080808, 0x082b2b08), uvec2(0x08082b08, 0x082b2b08), uvec2(0x08082b2b, 0x082b2b08), uvec2(0x082b0808, 0x082b2b08), + uvec2(0x19191919, 0x082b2b08), uvec2(0x2b082b08, 0x082b2b08), uvec2(0x2b2b082b, 0x082b2b08), uvec2(0x192b2b08, 0x082b2b19), + uvec2(0x2b190808, 0x082b2b19), uvec2(0x08082b08, 0x082b2b2b), uvec2(0x082b0808, 0x082b2b2b), uvec2(0x2b08082b, 0x082b2b2b), + uvec2(0x2b082b08, 0x082b2b2b), uvec2(0x2b082b2b, 0x082b2b2b), uvec2(0x08080819, 0x19080808), uvec2(0x08081908, 0x19080808), + uvec2(0x0808192b, 0x19080808), uvec2(0x08082b19, 0x19080808), uvec2(0x08190808, 0x19080808), uvec2(0x0819082b, 0x19080808), + uvec2(0x08191919, 0x19080808), uvec2(0x08192b08, 0x19080808), uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808), + uvec2(0x19080808, 0x19080808), uvec2(0x1908082b, 0x19080808), uvec2(0x19081919, 0x19080808), uvec2(0x19082b08, 0x19080808), + uvec2(0x19082b2b, 0x19080808), uvec2(0x19190819, 0x19080808), uvec2(0x19191908, 0x19080808), uvec2(0x192b0808, 0x19080808), + uvec2(0x192b1919, 0x19080808), uvec2(0x2b080819, 0x19080808), uvec2(0x2b081908, 0x19080808), uvec2(0x2b190808, 0x19080808), + uvec2(0x08080808, 0x19080819), uvec2(0x0808082b, 0x19080819), uvec2(0x08081919, 0x19080819), uvec2(0x08082b08, 0x19080819), + uvec2(0x08190819, 0x19080819), uvec2(0x08191908, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x19080819, 0x19080819), + uvec2(0x19081908, 0x19080819), uvec2(0x19190808, 0x19080819), uvec2(0x2b080808, 0x19080819), uvec2(0x2b081919, 0x19080819), + uvec2(0x2b2b082b, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08081908, 0x1908082b), uvec2(0x08190808, 0x1908082b), + uvec2(0x0819082b, 0x1908082b), uvec2(0x082b2b19, 0x1908082b), uvec2(0x19080808, 0x1908082b), uvec2(0x08080808, 0x19081908), + uvec2(0x0808082b, 0x19081908), uvec2(0x08081919, 0x19081908), uvec2(0x08082b08, 0x19081908), uvec2(0x08190819, 0x19081908), + uvec2(0x08191908, 0x19081908), uvec2(0x08192b19, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x19080819, 0x19081908), + uvec2(0x19081908, 0x19081908), uvec2(0x19190808, 0x19081908), uvec2(0x2b080808, 0x19081908), uvec2(0x2b191908, 0x19081908), + uvec2(0x08080819, 0x19081919), uvec2(0x08081908, 0x19081919), uvec2(0x08190808, 0x19081919), uvec2(0x082b1908, 0x19081919), + uvec2(0x19080808, 0x19081919), uvec2(0x2b192b2b, 0x19081919), uvec2(0x08080808, 0x1908192b), uvec2(0x08082b2b, 0x1908192b), + uvec2(0x19081908, 0x1908192b), uvec2(0x19190808, 0x1908192b), uvec2(0x08080819, 0x19082b08), uvec2(0x08081908, 0x19082b08), + uvec2(0x08190808, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x19081919, 0x19082b08), uvec2(0x19191908, 0x19082b08), + uvec2(0x192b082b, 0x19082b08), uvec2(0x08080808, 0x19082b19), uvec2(0x08190819, 0x19082b19), uvec2(0x19081908, 0x19082b19), + uvec2(0x19190808, 0x19082b19), uvec2(0x192b2b19, 0x19082b19), uvec2(0x08081908, 0x19082b2b), uvec2(0x08080808, 0x19190808), + uvec2(0x0808082b, 0x19190808), uvec2(0x08081919, 0x19190808), uvec2(0x08082b08, 0x19190808), uvec2(0x08190819, 0x19190808), + uvec2(0x08191908, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x082b2b08, 0x19190808), uvec2(0x19080819, 0x19190808), + uvec2(0x19081908, 0x19190808), uvec2(0x19190808, 0x19190808), uvec2(0x2b080808, 0x19190808), uvec2(0x08080819, 0x19190819), + uvec2(0x08081908, 0x19190819), uvec2(0x08190808, 0x19190819), uvec2(0x08191919, 0x19190819), uvec2(0x19080808, 0x19190819), + uvec2(0x1908082b, 0x19190819), uvec2(0x08080808, 0x1919082b), uvec2(0x19081908, 0x1919082b), uvec2(0x2b2b2b2b, 0x1919082b), + uvec2(0x08080819, 0x19191908), uvec2(0x08081908, 0x19191908), uvec2(0x08190808, 0x19191908), uvec2(0x082b0819, 0x19191908), + uvec2(0x19080808, 0x19191908), uvec2(0x192b0808, 0x19191908), uvec2(0x2b080819, 0x19191908), uvec2(0x2b2b0819, 0x19191908), + uvec2(0x08080808, 0x19191919), uvec2(0x08082b08, 0x19191919), uvec2(0x2b080808, 0x19191919), uvec2(0x2b082b08, 0x19191919), + uvec2(0x082b0819, 0x1919192b), uvec2(0x192b2b08, 0x1919192b), uvec2(0x2b2b0819, 0x1919192b), uvec2(0x08080808, 0x19192b08), + uvec2(0x08191908, 0x19192b08), uvec2(0x19080819, 0x19192b08), uvec2(0x19190808, 0x19192b08), uvec2(0x2b192b19, 0x19192b08), + uvec2(0x08192b2b, 0x19192b19), uvec2(0x19080808, 0x19192b19), uvec2(0x1908082b, 0x19192b19), uvec2(0x2b081919, 0x19192b2b), + uvec2(0x08080819, 0x192b0808), uvec2(0x08081908, 0x192b0808), uvec2(0x08190808, 0x192b0808), uvec2(0x19080808, 0x192b0808), + uvec2(0x19191908, 0x192b0808), uvec2(0x192b082b, 0x192b0808), uvec2(0x2b08192b, 0x192b0808), uvec2(0x2b2b2b19, 0x192b0808), + uvec2(0x08080808, 0x192b0819), uvec2(0x082b1908, 0x192b082b), uvec2(0x19082b2b, 0x192b082b), uvec2(0x2b19082b, 0x192b082b), + uvec2(0x08080808, 0x192b1908), uvec2(0x0819192b, 0x192b1908), uvec2(0x08190808, 0x192b1919), uvec2(0x19080808, 0x192b1919), + uvec2(0x19081919, 0x192b1919), uvec2(0x2b2b1908, 0x192b1919), uvec2(0x08080819, 0x192b2b08), uvec2(0x192b2b2b, 0x192b2b08), + uvec2(0x082b1919, 0x192b2b19), uvec2(0x0808192b, 0x192b2b2b), uvec2(0x19191908, 0x192b2b2b), uvec2(0x192b082b, 0x192b2b2b), + uvec2(0x08080808, 0x2b080808), uvec2(0x0808082b, 0x2b080808), uvec2(0x08081919, 0x2b080808), uvec2(0x08082b08, 0x2b080808), + uvec2(0x08190819, 0x2b080808), uvec2(0x08191908, 0x2b080808), uvec2(0x082b0808, 0x2b080808), uvec2(0x082b2b2b, 0x2b080808), + uvec2(0x19080819, 0x2b080808), uvec2(0x19081908, 0x2b080808), uvec2(0x19190808, 0x2b080808), uvec2(0x2b080808, 0x2b080808), + uvec2(0x2b08082b, 0x2b080808), uvec2(0x2b2b2b08, 0x2b080808), uvec2(0x2b2b2b2b, 0x2b080808), uvec2(0x08080819, 0x2b080819), + uvec2(0x08081908, 0x2b080819), uvec2(0x0808192b, 0x2b080819), uvec2(0x08190808, 0x2b080819), uvec2(0x19080808, 0x2b080819), + uvec2(0x19190819, 0x2b080819), uvec2(0x19192b19, 0x2b080819), uvec2(0x08080808, 0x2b08082b), uvec2(0x082b0808, 0x2b08082b), + uvec2(0x2b080808, 0x2b08082b), uvec2(0x2b08082b, 0x2b08082b), uvec2(0x2b2b0808, 0x2b08082b), uvec2(0x2b2b2b08, 0x2b08082b), + uvec2(0x08080819, 0x2b081908), uvec2(0x08081908, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x0819082b, 0x2b081908), + uvec2(0x08191919, 0x2b081908), uvec2(0x19080808, 0x2b081908), uvec2(0x192b0808, 0x2b081908), uvec2(0x2b082b19, 0x2b081908), + uvec2(0x08080808, 0x2b081919), uvec2(0x19081908, 0x2b081919), uvec2(0x2b2b1919, 0x2b081919), uvec2(0x08192b08, 0x2b08192b), + uvec2(0x192b2b2b, 0x2b08192b), uvec2(0x08080808, 0x2b082b08), uvec2(0x08082b08, 0x2b082b08), uvec2(0x082b1919, 0x2b082b08), + uvec2(0x19192b2b, 0x2b082b08), uvec2(0x2b080808, 0x2b082b08), uvec2(0x2b08082b, 0x2b082b08), uvec2(0x2b2b2b08, 0x2b082b08), + uvec2(0x0808192b, 0x2b082b19), uvec2(0x082b082b, 0x2b082b2b), uvec2(0x2b080808, 0x2b082b2b), uvec2(0x2b082b08, 0x2b082b2b), + uvec2(0x2b19192b, 0x2b082b2b), uvec2(0x2b2b2b08, 0x2b082b2b), uvec2(0x08080819, 0x2b190808), uvec2(0x08081908, 0x2b190808), + uvec2(0x08190808, 0x2b190808), uvec2(0x19080808, 0x2b190808), uvec2(0x1919192b, 0x2b190808), uvec2(0x2b081908, 0x2b190808), + uvec2(0x08080808, 0x2b190819), uvec2(0x082b082b, 0x2b190819), uvec2(0x192b1908, 0x2b190819), uvec2(0x1919192b, 0x2b19082b), + uvec2(0x2b082b19, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x08081919, 0x2b191908), uvec2(0x19081908, 0x2b191908), + uvec2(0x19190808, 0x2b191908), uvec2(0x19192b08, 0x2b191908), uvec2(0x082b2b19, 0x2b191919), uvec2(0x2b190808, 0x2b191919), + uvec2(0x2b19082b, 0x2b191919), uvec2(0x19080819, 0x2b19192b), uvec2(0x19190819, 0x2b192b08), uvec2(0x2b2b192b, 0x2b192b08), + uvec2(0x19082b19, 0x2b192b19), uvec2(0x08191919, 0x2b192b2b), uvec2(0x192b0808, 0x2b192b2b), uvec2(0x08080808, 0x2b2b0808), + uvec2(0x0808082b, 0x2b2b0808), uvec2(0x08082b08, 0x2b2b0808), uvec2(0x08082b2b, 0x2b2b0808), uvec2(0x082b0808, 0x2b2b0808), + uvec2(0x082b2b2b, 0x2b2b0808), uvec2(0x2b2b0808, 0x2b2b0808), uvec2(0x19190819, 0x2b2b0819), uvec2(0x19192b19, 0x2b2b0819), + uvec2(0x2b2b192b, 0x2b2b0819), uvec2(0x08080808, 0x2b2b082b), uvec2(0x0808082b, 0x2b2b082b), uvec2(0x08082b08, 0x2b2b082b), + uvec2(0x082b2b2b, 0x2b2b082b), uvec2(0x2b080808, 0x2b2b082b), uvec2(0x2b2b0808, 0x2b2b082b), uvec2(0x19080808, 0x2b2b1908), + uvec2(0x2b191919, 0x2b2b1908), uvec2(0x192b1919, 0x2b2b192b), uvec2(0x2b192b08, 0x2b2b192b), uvec2(0x08082b2b, 0x2b2b2b08), + uvec2(0x082b0808, 0x2b2b2b08), uvec2(0x082b082b, 0x2b2b2b08), uvec2(0x082b2b08, 0x2b2b2b08), uvec2(0x2b2b0808, 0x2b2b2b08), + uvec2(0x2b2b2b08, 0x2b2b2b08), uvec2(0x08081908, 0x2b2b2b19), uvec2(0x2b081908, 0x2b2b2b19), uvec2(0x2b08192b, 0x2b2b2b19), + uvec2(0x082b2b08, 0x2b2b2b2b), uvec2(0x082b2b2b, 0x2b2b2b2b), uvec2(0x2b190819, 0x2b2b2b2b), uvec2(0x2b2b2b2b, 0x2b2b2b2b), +}; + +shared uvec2 iq2xs_grid[512]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq2xs_grid.length(); i += wgsize.x) { + if (iq2xs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xs_grid_const.length()) { + iq2xs_grid[i + gl_LocalInvocationIndex.x] = iq2xs_grid_const[i + gl_LocalInvocationIndex.x]; + } + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ2_XS +#define QUANT_R QUANT_R_IQ2_XS +#define A_TYPE block_iq2_xs +#define A_TYPE_PACKED16 block_iq2_xs_packed16 +#endif + +#define QUANT_K_IQ2_S 256 +#define QUANT_R_IQ2_S 1 + +struct block_iq2_s +{ + float16_t d; + uint8_t qs[QUANT_K_IQ2_S/4]; + uint8_t qh[QUANT_K_IQ2_S/32]; + uint8_t scales[QUANT_K_IQ2_S/32]; +}; + +struct block_iq2_s_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ2_S/8]; + uint16_t qh[QUANT_K_IQ2_S/64]; + uint16_t scales[QUANT_K_IQ2_S/64]; +}; + +#if defined(DATA_A_IQ2_S) + +const uvec2 iq2s_grid_const[1024] = { + uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808), + uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x0819192b, 0x08080808), + uvec2(0x08192b19, 0x08080808), uvec2(0x082b0808, 0x08080808), uvec2(0x082b082b, 0x08080808), uvec2(0x082b1919, 0x08080808), + uvec2(0x082b2b08, 0x08080808), uvec2(0x19080819, 0x08080808), uvec2(0x19081908, 0x08080808), uvec2(0x1908192b, 0x08080808), + uvec2(0x19082b19, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x1919082b, 0x08080808), uvec2(0x19191919, 0x08080808), + uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808), uvec2(0x192b1908, 0x08080808), uvec2(0x192b192b, 0x08080808), + uvec2(0x192b2b19, 0x08080808), uvec2(0x2b080808, 0x08080808), uvec2(0x2b08082b, 0x08080808), uvec2(0x2b081919, 0x08080808), + uvec2(0x2b082b08, 0x08080808), uvec2(0x2b190819, 0x08080808), uvec2(0x2b191908, 0x08080808), uvec2(0x2b2b0808, 0x08080808), + uvec2(0x2b2b1919, 0x08080808), uvec2(0x2b2b2b2b, 0x08080808), uvec2(0x08080819, 0x08080819), uvec2(0x08081908, 0x08080819), + uvec2(0x0808192b, 0x08080819), uvec2(0x08082b19, 0x08080819), uvec2(0x08190808, 0x08080819), uvec2(0x0819082b, 0x08080819), + uvec2(0x08191919, 0x08080819), uvec2(0x08192b08, 0x08080819), uvec2(0x082b0819, 0x08080819), uvec2(0x082b1908, 0x08080819), + uvec2(0x19080808, 0x08080819), uvec2(0x1908082b, 0x08080819), uvec2(0x19081919, 0x08080819), uvec2(0x19082b08, 0x08080819), + uvec2(0x19190819, 0x08080819), uvec2(0x19191908, 0x08080819), uvec2(0x1919192b, 0x08080819), uvec2(0x19192b19, 0x08080819), + uvec2(0x192b0808, 0x08080819), uvec2(0x192b1919, 0x08080819), uvec2(0x192b2b08, 0x08080819), uvec2(0x2b080819, 0x08080819), + uvec2(0x2b081908, 0x08080819), uvec2(0x2b190808, 0x08080819), uvec2(0x2b19082b, 0x08080819), uvec2(0x2b191919, 0x08080819), + uvec2(0x2b2b0819, 0x08080819), uvec2(0x2b2b1908, 0x08080819), uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b), + uvec2(0x08081919, 0x0808082b), uvec2(0x08082b08, 0x0808082b), uvec2(0x08190819, 0x0808082b), uvec2(0x08191908, 0x0808082b), + uvec2(0x082b0808, 0x0808082b), uvec2(0x082b2b2b, 0x0808082b), uvec2(0x19080819, 0x0808082b), uvec2(0x19081908, 0x0808082b), + uvec2(0x1908192b, 0x0808082b), uvec2(0x19082b19, 0x0808082b), uvec2(0x19190808, 0x0808082b), uvec2(0x19191919, 0x0808082b), + uvec2(0x2b080808, 0x0808082b), uvec2(0x2b081919, 0x0808082b), uvec2(0x2b082b2b, 0x0808082b), uvec2(0x2b191908, 0x0808082b), + uvec2(0x2b2b082b, 0x0808082b), uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908), uvec2(0x0808192b, 0x08081908), + uvec2(0x08082b19, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x0819082b, 0x08081908), uvec2(0x08191919, 0x08081908), + uvec2(0x08192b08, 0x08081908), uvec2(0x082b0819, 0x08081908), uvec2(0x082b1908, 0x08081908), uvec2(0x082b192b, 0x08081908), + uvec2(0x082b2b19, 0x08081908), uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19081919, 0x08081908), + uvec2(0x19082b08, 0x08081908), uvec2(0x19082b2b, 0x08081908), uvec2(0x19190819, 0x08081908), uvec2(0x19191908, 0x08081908), + uvec2(0x1919192b, 0x08081908), uvec2(0x19192b19, 0x08081908), uvec2(0x192b0808, 0x08081908), uvec2(0x192b082b, 0x08081908), + uvec2(0x192b1919, 0x08081908), uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b08192b, 0x08081908), + uvec2(0x2b082b19, 0x08081908), uvec2(0x2b190808, 0x08081908), uvec2(0x2b191919, 0x08081908), uvec2(0x2b192b08, 0x08081908), + uvec2(0x2b2b0819, 0x08081908), uvec2(0x2b2b1908, 0x08081908), uvec2(0x08080808, 0x08081919), uvec2(0x0808082b, 0x08081919), + uvec2(0x08081919, 0x08081919), uvec2(0x08082b08, 0x08081919), uvec2(0x08082b2b, 0x08081919), uvec2(0x08190819, 0x08081919), + uvec2(0x08191908, 0x08081919), uvec2(0x0819192b, 0x08081919), uvec2(0x08192b19, 0x08081919), uvec2(0x082b0808, 0x08081919), + uvec2(0x082b1919, 0x08081919), uvec2(0x082b2b08, 0x08081919), uvec2(0x19080819, 0x08081919), uvec2(0x19081908, 0x08081919), + uvec2(0x1908192b, 0x08081919), uvec2(0x19082b19, 0x08081919), uvec2(0x19190808, 0x08081919), uvec2(0x1919082b, 0x08081919), + uvec2(0x19191919, 0x08081919), uvec2(0x19192b08, 0x08081919), uvec2(0x192b0819, 0x08081919), uvec2(0x192b1908, 0x08081919), + uvec2(0x2b080808, 0x08081919), uvec2(0x2b08082b, 0x08081919), uvec2(0x2b081919, 0x08081919), uvec2(0x2b082b08, 0x08081919), + uvec2(0x2b190819, 0x08081919), uvec2(0x2b191908, 0x08081919), uvec2(0x2b2b0808, 0x08081919), uvec2(0x08080819, 0x0808192b), + uvec2(0x08081908, 0x0808192b), uvec2(0x0808192b, 0x0808192b), uvec2(0x08082b19, 0x0808192b), uvec2(0x08190808, 0x0808192b), + uvec2(0x08191919, 0x0808192b), uvec2(0x19080808, 0x0808192b), uvec2(0x19081919, 0x0808192b), uvec2(0x19082b08, 0x0808192b), + uvec2(0x19190819, 0x0808192b), uvec2(0x19191908, 0x0808192b), uvec2(0x192b0808, 0x0808192b), uvec2(0x2b080819, 0x0808192b), + uvec2(0x2b081908, 0x0808192b), uvec2(0x2b190808, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x0808082b, 0x08082b08), + uvec2(0x08081919, 0x08082b08), uvec2(0x08082b08, 0x08082b08), uvec2(0x08190819, 0x08082b08), uvec2(0x08191908, 0x08082b08), + uvec2(0x0819192b, 0x08082b08), uvec2(0x08192b19, 0x08082b08), uvec2(0x082b0808, 0x08082b08), uvec2(0x082b1919, 0x08082b08), + uvec2(0x082b2b2b, 0x08082b08), uvec2(0x19080819, 0x08082b08), uvec2(0x19081908, 0x08082b08), uvec2(0x1908192b, 0x08082b08), + uvec2(0x19082b19, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x1919082b, 0x08082b08), uvec2(0x19191919, 0x08082b08), + uvec2(0x19192b08, 0x08082b08), uvec2(0x192b0819, 0x08082b08), uvec2(0x192b1908, 0x08082b08), uvec2(0x2b080808, 0x08082b08), + uvec2(0x2b081919, 0x08082b08), uvec2(0x2b191908, 0x08082b08), uvec2(0x2b2b2b2b, 0x08082b08), uvec2(0x08080819, 0x08082b19), + uvec2(0x08081908, 0x08082b19), uvec2(0x08190808, 0x08082b19), uvec2(0x0819082b, 0x08082b19), uvec2(0x08191919, 0x08082b19), + uvec2(0x08192b08, 0x08082b19), uvec2(0x082b0819, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x19081919, 0x08082b19), + uvec2(0x19082b08, 0x08082b19), uvec2(0x19190819, 0x08082b19), uvec2(0x19191908, 0x08082b19), uvec2(0x192b0808, 0x08082b19), + uvec2(0x2b080819, 0x08082b19), uvec2(0x2b190808, 0x08082b19), uvec2(0x08080808, 0x08082b2b), uvec2(0x08190819, 0x08082b2b), + uvec2(0x08191908, 0x08082b2b), uvec2(0x082b082b, 0x08082b2b), uvec2(0x082b2b08, 0x08082b2b), uvec2(0x082b2b2b, 0x08082b2b), + uvec2(0x19190808, 0x08082b2b), uvec2(0x2b192b19, 0x08082b2b), uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808), + uvec2(0x0808192b, 0x08190808), uvec2(0x08082b19, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x0819082b, 0x08190808), + uvec2(0x08191919, 0x08190808), uvec2(0x08192b08, 0x08190808), uvec2(0x082b0819, 0x08190808), uvec2(0x082b1908, 0x08190808), + uvec2(0x082b192b, 0x08190808), uvec2(0x19080808, 0x08190808), uvec2(0x1908082b, 0x08190808), uvec2(0x19081919, 0x08190808), + uvec2(0x19082b08, 0x08190808), uvec2(0x19190819, 0x08190808), uvec2(0x19191908, 0x08190808), uvec2(0x1919192b, 0x08190808), + uvec2(0x19192b19, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x192b082b, 0x08190808), uvec2(0x192b1919, 0x08190808), + uvec2(0x192b2b08, 0x08190808), uvec2(0x2b080819, 0x08190808), uvec2(0x2b081908, 0x08190808), uvec2(0x2b08192b, 0x08190808), + uvec2(0x2b190808, 0x08190808), uvec2(0x2b191919, 0x08190808), uvec2(0x2b192b08, 0x08190808), uvec2(0x2b2b0819, 0x08190808), + uvec2(0x2b2b1908, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x0808082b, 0x08190819), uvec2(0x08081919, 0x08190819), + uvec2(0x08082b08, 0x08190819), uvec2(0x08082b2b, 0x08190819), uvec2(0x08190819, 0x08190819), uvec2(0x08191908, 0x08190819), + uvec2(0x0819192b, 0x08190819), uvec2(0x08192b19, 0x08190819), uvec2(0x082b0808, 0x08190819), uvec2(0x082b082b, 0x08190819), + uvec2(0x082b1919, 0x08190819), uvec2(0x082b2b08, 0x08190819), uvec2(0x19080819, 0x08190819), uvec2(0x19081908, 0x08190819), + uvec2(0x1908192b, 0x08190819), uvec2(0x19082b19, 0x08190819), uvec2(0x19190808, 0x08190819), uvec2(0x1919082b, 0x08190819), + uvec2(0x19191919, 0x08190819), uvec2(0x19192b08, 0x08190819), uvec2(0x192b0819, 0x08190819), uvec2(0x192b1908, 0x08190819), + uvec2(0x2b080808, 0x08190819), uvec2(0x2b08082b, 0x08190819), uvec2(0x2b081919, 0x08190819), uvec2(0x2b082b08, 0x08190819), + uvec2(0x2b190819, 0x08190819), uvec2(0x2b191908, 0x08190819), uvec2(0x08080819, 0x0819082b), uvec2(0x08081908, 0x0819082b), + uvec2(0x08082b19, 0x0819082b), uvec2(0x08190808, 0x0819082b), uvec2(0x08191919, 0x0819082b), uvec2(0x082b0819, 0x0819082b), + uvec2(0x082b1908, 0x0819082b), uvec2(0x19080808, 0x0819082b), uvec2(0x19081919, 0x0819082b), uvec2(0x19190819, 0x0819082b), + uvec2(0x19191908, 0x0819082b), uvec2(0x2b080819, 0x0819082b), uvec2(0x2b081908, 0x0819082b), uvec2(0x2b190808, 0x0819082b), + uvec2(0x08080808, 0x08191908), uvec2(0x0808082b, 0x08191908), uvec2(0x08081919, 0x08191908), uvec2(0x08082b08, 0x08191908), + uvec2(0x08190819, 0x08191908), uvec2(0x08191908, 0x08191908), uvec2(0x0819192b, 0x08191908), uvec2(0x08192b19, 0x08191908), + uvec2(0x082b0808, 0x08191908), uvec2(0x082b1919, 0x08191908), uvec2(0x082b2b08, 0x08191908), uvec2(0x19080819, 0x08191908), + uvec2(0x19081908, 0x08191908), uvec2(0x1908192b, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x19190808, 0x08191908), + uvec2(0x1919082b, 0x08191908), uvec2(0x19191919, 0x08191908), uvec2(0x19192b08, 0x08191908), uvec2(0x192b0819, 0x08191908), + uvec2(0x192b1908, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x2b08082b, 0x08191908), uvec2(0x2b081919, 0x08191908), + uvec2(0x2b082b08, 0x08191908), uvec2(0x2b190819, 0x08191908), uvec2(0x2b191908, 0x08191908), uvec2(0x2b2b0808, 0x08191908), + uvec2(0x08080819, 0x08191919), uvec2(0x08081908, 0x08191919), uvec2(0x0808192b, 0x08191919), uvec2(0x08082b19, 0x08191919), + uvec2(0x08190808, 0x08191919), uvec2(0x0819082b, 0x08191919), uvec2(0x08191919, 0x08191919), uvec2(0x08192b08, 0x08191919), + uvec2(0x082b0819, 0x08191919), uvec2(0x082b1908, 0x08191919), uvec2(0x19080808, 0x08191919), uvec2(0x1908082b, 0x08191919), + uvec2(0x19081919, 0x08191919), uvec2(0x19082b08, 0x08191919), uvec2(0x19190819, 0x08191919), uvec2(0x19191908, 0x08191919), + uvec2(0x192b0808, 0x08191919), uvec2(0x2b080819, 0x08191919), uvec2(0x2b081908, 0x08191919), uvec2(0x2b190808, 0x08191919), + uvec2(0x08080808, 0x0819192b), uvec2(0x08081919, 0x0819192b), uvec2(0x08082b08, 0x0819192b), uvec2(0x08190819, 0x0819192b), + uvec2(0x08191908, 0x0819192b), uvec2(0x082b0808, 0x0819192b), uvec2(0x19080819, 0x0819192b), uvec2(0x19081908, 0x0819192b), + uvec2(0x19190808, 0x0819192b), uvec2(0x2b080808, 0x0819192b), uvec2(0x2b2b2b2b, 0x0819192b), uvec2(0x08080819, 0x08192b08), + uvec2(0x08081908, 0x08192b08), uvec2(0x0808192b, 0x08192b08), uvec2(0x08082b19, 0x08192b08), uvec2(0x08190808, 0x08192b08), + uvec2(0x08191919, 0x08192b08), uvec2(0x08192b08, 0x08192b08), uvec2(0x082b0819, 0x08192b08), uvec2(0x19080808, 0x08192b08), + uvec2(0x1908082b, 0x08192b08), uvec2(0x19081919, 0x08192b08), uvec2(0x19082b08, 0x08192b08), uvec2(0x19190819, 0x08192b08), + uvec2(0x19191908, 0x08192b08), uvec2(0x192b0808, 0x08192b08), uvec2(0x2b080819, 0x08192b08), uvec2(0x2b081908, 0x08192b08), + uvec2(0x08080808, 0x08192b19), uvec2(0x0808082b, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x08082b08, 0x08192b19), + uvec2(0x08190819, 0x08192b19), uvec2(0x08191908, 0x08192b19), uvec2(0x082b0808, 0x08192b19), uvec2(0x19080819, 0x08192b19), + uvec2(0x19081908, 0x08192b19), uvec2(0x19190808, 0x08192b19), uvec2(0x192b2b19, 0x08192b19), uvec2(0x2b2b082b, 0x08192b19), + uvec2(0x08081908, 0x08192b2b), uvec2(0x08190808, 0x08192b2b), uvec2(0x19080808, 0x08192b2b), uvec2(0x1919192b, 0x08192b2b), + uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08081919, 0x082b0808), uvec2(0x08082b08, 0x082b0808), + uvec2(0x08190819, 0x082b0808), uvec2(0x08191908, 0x082b0808), uvec2(0x0819192b, 0x082b0808), uvec2(0x08192b19, 0x082b0808), + uvec2(0x082b0808, 0x082b0808), uvec2(0x082b1919, 0x082b0808), uvec2(0x082b2b2b, 0x082b0808), uvec2(0x19080819, 0x082b0808), + uvec2(0x19081908, 0x082b0808), uvec2(0x19190808, 0x082b0808), uvec2(0x1919082b, 0x082b0808), uvec2(0x19191919, 0x082b0808), + uvec2(0x192b1908, 0x082b0808), uvec2(0x2b080808, 0x082b0808), uvec2(0x2b082b2b, 0x082b0808), uvec2(0x2b191908, 0x082b0808), + uvec2(0x2b2b2b2b, 0x082b0808), uvec2(0x08080819, 0x082b0819), uvec2(0x08081908, 0x082b0819), uvec2(0x08190808, 0x082b0819), + uvec2(0x0819082b, 0x082b0819), uvec2(0x08191919, 0x082b0819), uvec2(0x082b0819, 0x082b0819), uvec2(0x19080808, 0x082b0819), + uvec2(0x1908082b, 0x082b0819), uvec2(0x19081919, 0x082b0819), uvec2(0x19190819, 0x082b0819), uvec2(0x19191908, 0x082b0819), + uvec2(0x192b0808, 0x082b0819), uvec2(0x2b080819, 0x082b0819), uvec2(0x2b081908, 0x082b0819), uvec2(0x2b190808, 0x082b0819), + uvec2(0x08080808, 0x082b082b), uvec2(0x08082b2b, 0x082b082b), uvec2(0x082b082b, 0x082b082b), uvec2(0x082b2b08, 0x082b082b), + uvec2(0x082b2b2b, 0x082b082b), uvec2(0x19081908, 0x082b082b), uvec2(0x19190808, 0x082b082b), uvec2(0x2b082b08, 0x082b082b), + uvec2(0x2b082b2b, 0x082b082b), uvec2(0x2b2b2b08, 0x082b082b), uvec2(0x08080819, 0x082b1908), uvec2(0x08081908, 0x082b1908), + uvec2(0x0808192b, 0x082b1908), uvec2(0x08082b19, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x08191919, 0x082b1908), + uvec2(0x08192b08, 0x082b1908), uvec2(0x082b0819, 0x082b1908), uvec2(0x082b1908, 0x082b1908), uvec2(0x19080808, 0x082b1908), + uvec2(0x1908082b, 0x082b1908), uvec2(0x19081919, 0x082b1908), uvec2(0x19082b08, 0x082b1908), uvec2(0x19190819, 0x082b1908), + uvec2(0x19191908, 0x082b1908), uvec2(0x192b0808, 0x082b1908), uvec2(0x2b080819, 0x082b1908), uvec2(0x2b081908, 0x082b1908), + uvec2(0x2b190808, 0x082b1908), uvec2(0x08080808, 0x082b1919), uvec2(0x08081919, 0x082b1919), uvec2(0x08082b08, 0x082b1919), + uvec2(0x08190819, 0x082b1919), uvec2(0x08191908, 0x082b1919), uvec2(0x082b0808, 0x082b1919), uvec2(0x19080819, 0x082b1919), + uvec2(0x19081908, 0x082b1919), uvec2(0x19190808, 0x082b1919), uvec2(0x192b192b, 0x082b1919), uvec2(0x2b080808, 0x082b1919), + uvec2(0x08080819, 0x082b192b), uvec2(0x08081908, 0x082b192b), uvec2(0x08190808, 0x082b192b), uvec2(0x19080808, 0x082b192b), + uvec2(0x19192b19, 0x082b192b), uvec2(0x08080808, 0x082b2b08), uvec2(0x08081919, 0x082b2b08), uvec2(0x08190819, 0x082b2b08), + uvec2(0x08191908, 0x082b2b08), uvec2(0x19080819, 0x082b2b08), uvec2(0x19081908, 0x082b2b08), uvec2(0x19190808, 0x082b2b08), + uvec2(0x2b082b2b, 0x082b2b08), uvec2(0x2b2b2b2b, 0x082b2b08), uvec2(0x08080819, 0x082b2b19), uvec2(0x08081908, 0x082b2b19), + uvec2(0x08190808, 0x082b2b19), uvec2(0x2b191919, 0x082b2b19), uvec2(0x08082b2b, 0x082b2b2b), uvec2(0x082b082b, 0x082b2b2b), + uvec2(0x192b1908, 0x082b2b2b), uvec2(0x2b082b08, 0x082b2b2b), uvec2(0x2b082b2b, 0x082b2b2b), uvec2(0x08080819, 0x19080808), + uvec2(0x08081908, 0x19080808), uvec2(0x0808192b, 0x19080808), uvec2(0x08082b19, 0x19080808), uvec2(0x08190808, 0x19080808), + uvec2(0x0819082b, 0x19080808), uvec2(0x08191919, 0x19080808), uvec2(0x08192b08, 0x19080808), uvec2(0x08192b2b, 0x19080808), + uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808), uvec2(0x082b192b, 0x19080808), uvec2(0x19080808, 0x19080808), + uvec2(0x1908082b, 0x19080808), uvec2(0x19081919, 0x19080808), uvec2(0x19082b08, 0x19080808), uvec2(0x19082b2b, 0x19080808), + uvec2(0x19190819, 0x19080808), uvec2(0x19191908, 0x19080808), uvec2(0x1919192b, 0x19080808), uvec2(0x19192b19, 0x19080808), + uvec2(0x192b0808, 0x19080808), uvec2(0x192b082b, 0x19080808), uvec2(0x192b1919, 0x19080808), uvec2(0x2b080819, 0x19080808), + uvec2(0x2b081908, 0x19080808), uvec2(0x2b190808, 0x19080808), uvec2(0x2b191919, 0x19080808), uvec2(0x2b192b08, 0x19080808), + uvec2(0x2b2b0819, 0x19080808), uvec2(0x2b2b1908, 0x19080808), uvec2(0x08080808, 0x19080819), uvec2(0x0808082b, 0x19080819), + uvec2(0x08081919, 0x19080819), uvec2(0x08082b08, 0x19080819), uvec2(0x08190819, 0x19080819), uvec2(0x08191908, 0x19080819), + uvec2(0x0819192b, 0x19080819), uvec2(0x08192b19, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x082b082b, 0x19080819), + uvec2(0x082b1919, 0x19080819), uvec2(0x19080819, 0x19080819), uvec2(0x19081908, 0x19080819), uvec2(0x1908192b, 0x19080819), + uvec2(0x19082b19, 0x19080819), uvec2(0x19190808, 0x19080819), uvec2(0x1919082b, 0x19080819), uvec2(0x19191919, 0x19080819), + uvec2(0x19192b08, 0x19080819), uvec2(0x192b0819, 0x19080819), uvec2(0x192b1908, 0x19080819), uvec2(0x2b080808, 0x19080819), + uvec2(0x2b08082b, 0x19080819), uvec2(0x2b081919, 0x19080819), uvec2(0x2b082b08, 0x19080819), uvec2(0x2b190819, 0x19080819), + uvec2(0x2b191908, 0x19080819), uvec2(0x2b2b0808, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08081908, 0x1908082b), + uvec2(0x08190808, 0x1908082b), uvec2(0x0819082b, 0x1908082b), uvec2(0x08191919, 0x1908082b), uvec2(0x08192b08, 0x1908082b), + uvec2(0x082b1908, 0x1908082b), uvec2(0x19080808, 0x1908082b), uvec2(0x19081919, 0x1908082b), uvec2(0x19082b08, 0x1908082b), + uvec2(0x19190819, 0x1908082b), uvec2(0x19191908, 0x1908082b), uvec2(0x192b0808, 0x1908082b), uvec2(0x2b080819, 0x1908082b), + uvec2(0x2b081908, 0x1908082b), uvec2(0x08080808, 0x19081908), uvec2(0x0808082b, 0x19081908), uvec2(0x08081919, 0x19081908), + uvec2(0x08082b08, 0x19081908), uvec2(0x08082b2b, 0x19081908), uvec2(0x08190819, 0x19081908), uvec2(0x08191908, 0x19081908), + uvec2(0x0819192b, 0x19081908), uvec2(0x08192b19, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x082b082b, 0x19081908), + uvec2(0x082b1919, 0x19081908), uvec2(0x082b2b08, 0x19081908), uvec2(0x19080819, 0x19081908), uvec2(0x19081908, 0x19081908), + uvec2(0x1908192b, 0x19081908), uvec2(0x19082b19, 0x19081908), uvec2(0x19190808, 0x19081908), uvec2(0x1919082b, 0x19081908), + uvec2(0x19191919, 0x19081908), uvec2(0x19192b08, 0x19081908), uvec2(0x192b0819, 0x19081908), uvec2(0x192b1908, 0x19081908), + uvec2(0x2b080808, 0x19081908), uvec2(0x2b08082b, 0x19081908), uvec2(0x2b081919, 0x19081908), uvec2(0x2b082b08, 0x19081908), + uvec2(0x2b190819, 0x19081908), uvec2(0x2b191908, 0x19081908), uvec2(0x2b2b0808, 0x19081908), uvec2(0x08080819, 0x19081919), + uvec2(0x08081908, 0x19081919), uvec2(0x0808192b, 0x19081919), uvec2(0x08082b19, 0x19081919), uvec2(0x08190808, 0x19081919), + uvec2(0x0819082b, 0x19081919), uvec2(0x08191919, 0x19081919), uvec2(0x08192b08, 0x19081919), uvec2(0x082b0819, 0x19081919), + uvec2(0x082b1908, 0x19081919), uvec2(0x19080808, 0x19081919), uvec2(0x1908082b, 0x19081919), uvec2(0x19081919, 0x19081919), + uvec2(0x19082b08, 0x19081919), uvec2(0x19190819, 0x19081919), uvec2(0x19191908, 0x19081919), uvec2(0x192b0808, 0x19081919), + uvec2(0x192b2b2b, 0x19081919), uvec2(0x2b080819, 0x19081919), uvec2(0x2b081908, 0x19081919), uvec2(0x2b190808, 0x19081919), + uvec2(0x08080808, 0x1908192b), uvec2(0x0808082b, 0x1908192b), uvec2(0x08081919, 0x1908192b), uvec2(0x08082b08, 0x1908192b), + uvec2(0x08190819, 0x1908192b), uvec2(0x08191908, 0x1908192b), uvec2(0x082b0808, 0x1908192b), uvec2(0x19080819, 0x1908192b), + uvec2(0x19081908, 0x1908192b), uvec2(0x19190808, 0x1908192b), uvec2(0x2b080808, 0x1908192b), uvec2(0x2b2b1919, 0x1908192b), + uvec2(0x08080819, 0x19082b08), uvec2(0x08081908, 0x19082b08), uvec2(0x08082b19, 0x19082b08), uvec2(0x08190808, 0x19082b08), + uvec2(0x0819082b, 0x19082b08), uvec2(0x08191919, 0x19082b08), uvec2(0x08192b08, 0x19082b08), uvec2(0x082b0819, 0x19082b08), + uvec2(0x082b1908, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x1908082b, 0x19082b08), uvec2(0x19081919, 0x19082b08), + uvec2(0x19082b08, 0x19082b08), uvec2(0x19190819, 0x19082b08), uvec2(0x19191908, 0x19082b08), uvec2(0x192b0808, 0x19082b08), + uvec2(0x2b081908, 0x19082b08), uvec2(0x2b190808, 0x19082b08), uvec2(0x08080808, 0x19082b19), uvec2(0x0808082b, 0x19082b19), + uvec2(0x08081919, 0x19082b19), uvec2(0x08082b08, 0x19082b19), uvec2(0x08190819, 0x19082b19), uvec2(0x08191908, 0x19082b19), + uvec2(0x082b0808, 0x19082b19), uvec2(0x19080819, 0x19082b19), uvec2(0x19081908, 0x19082b19), uvec2(0x19190808, 0x19082b19), + uvec2(0x2b080808, 0x19082b19), uvec2(0x2b19192b, 0x19082b19), uvec2(0x08080819, 0x19082b2b), uvec2(0x08081908, 0x19082b2b), + uvec2(0x08190808, 0x19082b2b), uvec2(0x19080808, 0x19082b2b), uvec2(0x08080808, 0x19190808), uvec2(0x0808082b, 0x19190808), + uvec2(0x08081919, 0x19190808), uvec2(0x08082b08, 0x19190808), uvec2(0x08190819, 0x19190808), uvec2(0x08191908, 0x19190808), + uvec2(0x0819192b, 0x19190808), uvec2(0x08192b19, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x082b082b, 0x19190808), + uvec2(0x082b1919, 0x19190808), uvec2(0x082b2b08, 0x19190808), uvec2(0x19080819, 0x19190808), uvec2(0x19081908, 0x19190808), + uvec2(0x1908192b, 0x19190808), uvec2(0x19082b19, 0x19190808), uvec2(0x19190808, 0x19190808), uvec2(0x1919082b, 0x19190808), + uvec2(0x19191919, 0x19190808), uvec2(0x19192b08, 0x19190808), uvec2(0x192b0819, 0x19190808), uvec2(0x192b1908, 0x19190808), + uvec2(0x2b080808, 0x19190808), uvec2(0x2b08082b, 0x19190808), uvec2(0x2b081919, 0x19190808), uvec2(0x2b082b08, 0x19190808), + uvec2(0x2b190819, 0x19190808), uvec2(0x2b191908, 0x19190808), uvec2(0x08080819, 0x19190819), uvec2(0x08081908, 0x19190819), + uvec2(0x0808192b, 0x19190819), uvec2(0x08082b19, 0x19190819), uvec2(0x08190808, 0x19190819), uvec2(0x0819082b, 0x19190819), + uvec2(0x08191919, 0x19190819), uvec2(0x08192b08, 0x19190819), uvec2(0x082b0819, 0x19190819), uvec2(0x082b1908, 0x19190819), + uvec2(0x19080808, 0x19190819), uvec2(0x1908082b, 0x19190819), uvec2(0x19081919, 0x19190819), uvec2(0x19082b08, 0x19190819), + uvec2(0x19190819, 0x19190819), uvec2(0x19191908, 0x19190819), uvec2(0x192b0808, 0x19190819), uvec2(0x2b080819, 0x19190819), + uvec2(0x2b081908, 0x19190819), uvec2(0x2b190808, 0x19190819), uvec2(0x08080808, 0x1919082b), uvec2(0x08081919, 0x1919082b), + uvec2(0x08082b08, 0x1919082b), uvec2(0x08190819, 0x1919082b), uvec2(0x08191908, 0x1919082b), uvec2(0x082b0808, 0x1919082b), + uvec2(0x19080819, 0x1919082b), uvec2(0x19081908, 0x1919082b), uvec2(0x19190808, 0x1919082b), uvec2(0x192b2b19, 0x1919082b), + uvec2(0x2b080808, 0x1919082b), uvec2(0x08080819, 0x19191908), uvec2(0x08081908, 0x19191908), uvec2(0x0808192b, 0x19191908), + uvec2(0x08082b19, 0x19191908), uvec2(0x08190808, 0x19191908), uvec2(0x0819082b, 0x19191908), uvec2(0x08191919, 0x19191908), + uvec2(0x08192b08, 0x19191908), uvec2(0x082b0819, 0x19191908), uvec2(0x082b1908, 0x19191908), uvec2(0x19080808, 0x19191908), + uvec2(0x1908082b, 0x19191908), uvec2(0x19081919, 0x19191908), uvec2(0x19082b08, 0x19191908), uvec2(0x19190819, 0x19191908), + uvec2(0x19191908, 0x19191908), uvec2(0x192b0808, 0x19191908), uvec2(0x2b080819, 0x19191908), uvec2(0x2b081908, 0x19191908), + uvec2(0x2b190808, 0x19191908), uvec2(0x08080808, 0x19191919), uvec2(0x0808082b, 0x19191919), uvec2(0x08081919, 0x19191919), + uvec2(0x08082b08, 0x19191919), uvec2(0x08190819, 0x19191919), uvec2(0x08191908, 0x19191919), uvec2(0x082b0808, 0x19191919), + uvec2(0x19080819, 0x19191919), uvec2(0x19081908, 0x19191919), uvec2(0x19190808, 0x19191919), uvec2(0x2b080808, 0x19191919), + uvec2(0x08080819, 0x1919192b), uvec2(0x08081908, 0x1919192b), uvec2(0x08190808, 0x1919192b), uvec2(0x082b192b, 0x1919192b), + uvec2(0x19080808, 0x1919192b), uvec2(0x08080808, 0x19192b08), uvec2(0x0808082b, 0x19192b08), uvec2(0x08081919, 0x19192b08), + uvec2(0x08082b08, 0x19192b08), uvec2(0x08190819, 0x19192b08), uvec2(0x08191908, 0x19192b08), uvec2(0x082b0808, 0x19192b08), + uvec2(0x19080819, 0x19192b08), uvec2(0x19081908, 0x19192b08), uvec2(0x19190808, 0x19192b08), uvec2(0x19192b2b, 0x19192b08), + uvec2(0x2b080808, 0x19192b08), uvec2(0x08080819, 0x19192b19), uvec2(0x08081908, 0x19192b19), uvec2(0x08190808, 0x19192b19), + uvec2(0x19080808, 0x19192b19), uvec2(0x08080808, 0x19192b2b), uvec2(0x08192b19, 0x19192b2b), uvec2(0x2b081919, 0x19192b2b), + uvec2(0x2b2b2b08, 0x19192b2b), uvec2(0x08080819, 0x192b0808), uvec2(0x08081908, 0x192b0808), uvec2(0x0808192b, 0x192b0808), + uvec2(0x08190808, 0x192b0808), uvec2(0x0819082b, 0x192b0808), uvec2(0x08191919, 0x192b0808), uvec2(0x08192b08, 0x192b0808), + uvec2(0x082b0819, 0x192b0808), uvec2(0x082b1908, 0x192b0808), uvec2(0x19080808, 0x192b0808), uvec2(0x19081919, 0x192b0808), + uvec2(0x19082b08, 0x192b0808), uvec2(0x19190819, 0x192b0808), uvec2(0x19191908, 0x192b0808), uvec2(0x192b0808, 0x192b0808), + uvec2(0x2b081908, 0x192b0808), uvec2(0x2b190808, 0x192b0808), uvec2(0x08080808, 0x192b0819), uvec2(0x0808082b, 0x192b0819), + uvec2(0x08081919, 0x192b0819), uvec2(0x08082b08, 0x192b0819), uvec2(0x08190819, 0x192b0819), uvec2(0x08191908, 0x192b0819), + uvec2(0x082b0808, 0x192b0819), uvec2(0x19080819, 0x192b0819), uvec2(0x19081908, 0x192b0819), uvec2(0x19190808, 0x192b0819), + uvec2(0x2b080808, 0x192b0819), uvec2(0x2b192b19, 0x192b0819), uvec2(0x08081908, 0x192b082b), uvec2(0x08190808, 0x192b082b), + uvec2(0x19080808, 0x192b082b), uvec2(0x1919192b, 0x192b082b), uvec2(0x2b2b0819, 0x192b082b), uvec2(0x08080808, 0x192b1908), + uvec2(0x08081919, 0x192b1908), uvec2(0x08082b08, 0x192b1908), uvec2(0x08190819, 0x192b1908), uvec2(0x08191908, 0x192b1908), + uvec2(0x082b0808, 0x192b1908), uvec2(0x19080819, 0x192b1908), uvec2(0x19081908, 0x192b1908), uvec2(0x19190808, 0x192b1908), + uvec2(0x2b080808, 0x192b1908), uvec2(0x08080819, 0x192b1919), uvec2(0x08081908, 0x192b1919), uvec2(0x08190808, 0x192b1919), + uvec2(0x19080808, 0x192b1919), uvec2(0x19082b2b, 0x192b1919), uvec2(0x192b2b08, 0x192b1919), uvec2(0x2b19082b, 0x192b1919), + uvec2(0x08080808, 0x192b192b), uvec2(0x2b191908, 0x192b192b), uvec2(0x08080819, 0x192b2b08), uvec2(0x08081908, 0x192b2b08), + uvec2(0x08190808, 0x192b2b08), uvec2(0x192b1919, 0x192b2b08), uvec2(0x2b192b08, 0x192b2b08), uvec2(0x08080808, 0x192b2b19), + uvec2(0x082b2b2b, 0x192b2b19), uvec2(0x1908082b, 0x192b2b2b), uvec2(0x2b2b0819, 0x192b2b2b), uvec2(0x08080808, 0x2b080808), + uvec2(0x0808082b, 0x2b080808), uvec2(0x08081919, 0x2b080808), uvec2(0x08082b08, 0x2b080808), uvec2(0x08190819, 0x2b080808), + uvec2(0x08191908, 0x2b080808), uvec2(0x08192b19, 0x2b080808), uvec2(0x082b0808, 0x2b080808), uvec2(0x082b1919, 0x2b080808), + uvec2(0x19080819, 0x2b080808), uvec2(0x19081908, 0x2b080808), uvec2(0x19190808, 0x2b080808), uvec2(0x1919082b, 0x2b080808), + uvec2(0x19191919, 0x2b080808), uvec2(0x19192b08, 0x2b080808), uvec2(0x192b0819, 0x2b080808), uvec2(0x2b080808, 0x2b080808), + uvec2(0x2b081919, 0x2b080808), uvec2(0x2b190819, 0x2b080808), uvec2(0x2b191908, 0x2b080808), uvec2(0x08080819, 0x2b080819), + uvec2(0x08081908, 0x2b080819), uvec2(0x08082b19, 0x2b080819), uvec2(0x08190808, 0x2b080819), uvec2(0x0819082b, 0x2b080819), + uvec2(0x08191919, 0x2b080819), uvec2(0x08192b08, 0x2b080819), uvec2(0x082b0819, 0x2b080819), uvec2(0x082b1908, 0x2b080819), + uvec2(0x19080808, 0x2b080819), uvec2(0x1908082b, 0x2b080819), uvec2(0x19081919, 0x2b080819), uvec2(0x19082b08, 0x2b080819), + uvec2(0x19190819, 0x2b080819), uvec2(0x19191908, 0x2b080819), uvec2(0x2b080819, 0x2b080819), uvec2(0x2b081908, 0x2b080819), + uvec2(0x2b190808, 0x2b080819), uvec2(0x2b2b2b19, 0x2b080819), uvec2(0x08080808, 0x2b08082b), uvec2(0x08081919, 0x2b08082b), + uvec2(0x08082b2b, 0x2b08082b), uvec2(0x08190819, 0x2b08082b), uvec2(0x08191908, 0x2b08082b), uvec2(0x19080819, 0x2b08082b), + uvec2(0x19081908, 0x2b08082b), uvec2(0x19190808, 0x2b08082b), uvec2(0x08080819, 0x2b081908), uvec2(0x08081908, 0x2b081908), + uvec2(0x0808192b, 0x2b081908), uvec2(0x08082b19, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x0819082b, 0x2b081908), + uvec2(0x08191919, 0x2b081908), uvec2(0x08192b08, 0x2b081908), uvec2(0x082b0819, 0x2b081908), uvec2(0x19080808, 0x2b081908), + uvec2(0x1908082b, 0x2b081908), uvec2(0x19081919, 0x2b081908), uvec2(0x19082b08, 0x2b081908), uvec2(0x19190819, 0x2b081908), + uvec2(0x19191908, 0x2b081908), uvec2(0x192b0808, 0x2b081908), uvec2(0x2b080819, 0x2b081908), uvec2(0x2b081908, 0x2b081908), + uvec2(0x2b190808, 0x2b081908), uvec2(0x08080808, 0x2b081919), uvec2(0x0808082b, 0x2b081919), uvec2(0x08081919, 0x2b081919), + uvec2(0x08082b08, 0x2b081919), uvec2(0x08190819, 0x2b081919), uvec2(0x08191908, 0x2b081919), uvec2(0x082b0808, 0x2b081919), + uvec2(0x19080819, 0x2b081919), uvec2(0x19081908, 0x2b081919), uvec2(0x19190808, 0x2b081919), uvec2(0x2b080808, 0x2b081919), + uvec2(0x2b082b2b, 0x2b081919), uvec2(0x08080819, 0x2b08192b), uvec2(0x08081908, 0x2b08192b), uvec2(0x08190808, 0x2b08192b), + uvec2(0x082b2b19, 0x2b08192b), uvec2(0x19080808, 0x2b08192b), uvec2(0x08080808, 0x2b082b08), uvec2(0x08081919, 0x2b082b08), + uvec2(0x08190819, 0x2b082b08), uvec2(0x08191908, 0x2b082b08), uvec2(0x19080819, 0x2b082b08), uvec2(0x19081908, 0x2b082b08), + uvec2(0x19190808, 0x2b082b08), uvec2(0x2b2b082b, 0x2b082b08), uvec2(0x08080819, 0x2b082b19), uvec2(0x08081908, 0x2b082b19), + uvec2(0x19080808, 0x2b082b19), uvec2(0x192b1919, 0x2b082b19), uvec2(0x082b082b, 0x2b082b2b), uvec2(0x19192b08, 0x2b082b2b), + uvec2(0x19192b2b, 0x2b082b2b), uvec2(0x2b08082b, 0x2b082b2b), uvec2(0x2b2b082b, 0x2b082b2b), uvec2(0x08080819, 0x2b190808), + uvec2(0x08081908, 0x2b190808), uvec2(0x08082b19, 0x2b190808), uvec2(0x08190808, 0x2b190808), uvec2(0x0819082b, 0x2b190808), + uvec2(0x08191919, 0x2b190808), uvec2(0x08192b08, 0x2b190808), uvec2(0x082b1908, 0x2b190808), uvec2(0x19080808, 0x2b190808), + uvec2(0x1908082b, 0x2b190808), uvec2(0x19081919, 0x2b190808), uvec2(0x19082b08, 0x2b190808), uvec2(0x19190819, 0x2b190808), + uvec2(0x19191908, 0x2b190808), uvec2(0x192b0808, 0x2b190808), uvec2(0x2b080819, 0x2b190808), uvec2(0x2b081908, 0x2b190808), + uvec2(0x2b190808, 0x2b190808), uvec2(0x08080808, 0x2b190819), uvec2(0x08081919, 0x2b190819), uvec2(0x08190819, 0x2b190819), + uvec2(0x08191908, 0x2b190819), uvec2(0x19080819, 0x2b190819), uvec2(0x19081908, 0x2b190819), uvec2(0x19190808, 0x2b190819), + uvec2(0x19192b2b, 0x2b190819), uvec2(0x08080819, 0x2b19082b), uvec2(0x08081908, 0x2b19082b), uvec2(0x08190808, 0x2b19082b), + uvec2(0x19080808, 0x2b19082b), uvec2(0x2b2b192b, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x0808082b, 0x2b191908), + uvec2(0x08081919, 0x2b191908), uvec2(0x08082b08, 0x2b191908), uvec2(0x08190819, 0x2b191908), uvec2(0x08191908, 0x2b191908), + uvec2(0x082b0808, 0x2b191908), uvec2(0x19080819, 0x2b191908), uvec2(0x19081908, 0x2b191908), uvec2(0x19190808, 0x2b191908), + uvec2(0x2b080808, 0x2b191908), uvec2(0x2b19192b, 0x2b191908), uvec2(0x08080819, 0x2b191919), uvec2(0x08081908, 0x2b191919), + uvec2(0x08190808, 0x2b191919), uvec2(0x19080808, 0x2b191919), uvec2(0x2b192b08, 0x2b191919), uvec2(0x2b2b0819, 0x2b191919), + uvec2(0x08080808, 0x2b19192b), uvec2(0x1908192b, 0x2b19192b), uvec2(0x192b1908, 0x2b19192b), uvec2(0x08080819, 0x2b192b08), + uvec2(0x08081908, 0x2b192b08), uvec2(0x08190808, 0x2b192b08), uvec2(0x082b192b, 0x2b192b08), uvec2(0x19080808, 0x2b192b08), + uvec2(0x2b2b2b19, 0x2b192b08), uvec2(0x08080808, 0x2b192b19), uvec2(0x19082b19, 0x2b192b19), uvec2(0x1919082b, 0x2b192b19), + uvec2(0x2b190808, 0x2b192b2b), uvec2(0x08080808, 0x2b2b0808), uvec2(0x08081919, 0x2b2b0808), uvec2(0x08082b2b, 0x2b2b0808), + uvec2(0x08191908, 0x2b2b0808), uvec2(0x082b082b, 0x2b2b0808), uvec2(0x082b2b2b, 0x2b2b0808), uvec2(0x19080819, 0x2b2b0808), + uvec2(0x19081908, 0x2b2b0808), uvec2(0x19190808, 0x2b2b0808), uvec2(0x2b2b082b, 0x2b2b0808), uvec2(0x2b2b2b2b, 0x2b2b0808), + uvec2(0x19080808, 0x2b2b0819), uvec2(0x192b1919, 0x2b2b0819), uvec2(0x0808082b, 0x2b2b082b), uvec2(0x08082b2b, 0x2b2b082b), + uvec2(0x082b082b, 0x2b2b082b), uvec2(0x082b2b08, 0x2b2b082b), uvec2(0x082b2b2b, 0x2b2b082b), uvec2(0x2b08082b, 0x2b2b082b), + uvec2(0x2b082b08, 0x2b2b082b), uvec2(0x2b082b2b, 0x2b2b082b), uvec2(0x2b2b2b08, 0x2b2b082b), uvec2(0x08080819, 0x2b2b1908), + uvec2(0x08081908, 0x2b2b1908), uvec2(0x08190808, 0x2b2b1908), uvec2(0x19080808, 0x2b2b1908), uvec2(0x2b082b19, 0x2b2b1908), + uvec2(0x2b2b1908, 0x2b2b1908), uvec2(0x08080808, 0x2b2b1919), uvec2(0x08192b19, 0x2b2b1919), uvec2(0x19190819, 0x2b2b192b), + uvec2(0x08082b2b, 0x2b2b2b08), uvec2(0x082b2b08, 0x2b2b2b08), uvec2(0x2b2b082b, 0x2b2b2b08), uvec2(0x19191908, 0x2b2b2b19), + uvec2(0x2b08192b, 0x2b2b2b19), uvec2(0x08082b08, 0x2b2b2b2b), uvec2(0x08082b2b, 0x2b2b2b2b), uvec2(0x082b0808, 0x2b2b2b2b), + uvec2(0x082b082b, 0x2b2b2b2b), uvec2(0x082b2b08, 0x2b2b2b2b), uvec2(0x2b082b08, 0x2b2b2b2b), uvec2(0x2b2b2b2b, 0x2b2b2b2b) +}; + +shared uvec2 iq2s_grid[1024]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq2s_grid.length(); i += wgsize.x) { + if (iq2s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2s_grid_const.length()) { + iq2s_grid[i + gl_LocalInvocationIndex.x] = iq2s_grid_const[i + gl_LocalInvocationIndex.x]; + } + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ2_S +#define QUANT_R QUANT_R_IQ2_S +#define A_TYPE block_iq2_s +#define A_TYPE_PACKED16 block_iq2_s_packed16 +#endif + +#define QUANT_K_IQ3_XXS 256 +#define QUANT_R_IQ3_XXS 1 + +struct block_iq3_xxs +{ + float16_t d; + uint8_t qs[QUANT_K_IQ3_XXS/4 + QUANT_K_IQ3_XXS/8]; +}; + +struct block_iq3_xxs_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ3_XXS/8 + QUANT_K_IQ3_XXS/16]; +}; + +#if defined(DATA_A_IQ3_XXS) + +const uint32_t iq3xxs_grid_const[256] = { + 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, + 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, + 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, + 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, + 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, + 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, + 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, + 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, + 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, + 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, + 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, + 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, + 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, + 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, + 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, + 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, + 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, + 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, + 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, + 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, + 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, + 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, + 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, + 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, + 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, + 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, + 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, + 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, + 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, + 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, + 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, + 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, +}; + +shared uint32_t iq3xxs_grid[256]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq3xxs_grid.length(); i += wgsize.x) { + if (iq3xxs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3xxs_grid.length()) { + iq3xxs_grid[i + gl_LocalInvocationIndex.x] = iq3xxs_grid_const[i + gl_LocalInvocationIndex.x]; + } + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ3_XXS +#define QUANT_R QUANT_R_IQ3_XXS +#define A_TYPE block_iq3_xxs +#define A_TYPE_PACKED16 block_iq3_xxs_packed16 +#endif + +#define QUANT_K_IQ3_S 256 +#define QUANT_R_IQ3_S 1 + +struct block_iq3_s +{ + float16_t d; + uint8_t qs[QUANT_K_IQ3_S/4]; + uint8_t qh[QUANT_K_IQ3_S/32]; + uint8_t signs[QUANT_K_IQ3_S/8]; + uint8_t scales[QUANT_K_IQ3_S/64]; +}; + +struct block_iq3_s_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ3_S/4/2]; + uint16_t qh[QUANT_K_IQ3_S/32/2]; + uint16_t signs[QUANT_K_IQ3_S/8/2]; + uint16_t scales[QUANT_K_IQ3_S/64/2]; +}; + +#if defined(DATA_A_IQ3_S) + +const uint32_t iq3s_grid_const[512] = { + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, +}; + +shared uint32_t iq3s_grid[512]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq3s_grid.length(); i += wgsize.x) { + if (iq3s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3s_grid.length()) { + iq3s_grid[i + gl_LocalInvocationIndex.x] = iq3s_grid_const[i + gl_LocalInvocationIndex.x]; + } + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ3_S +#define QUANT_R QUANT_R_IQ3_S +#define A_TYPE block_iq3_s +#define A_TYPE_PACKED16 block_iq3_s_packed16 +#endif + +#define QUANT_K_IQ4_XS 256 +#define QUANT_R_IQ4_XS 1 + +struct block_iq4_xs +{ + float16_t d; + uint16_t scales_h; + uint8_t scales_l[QUANT_K_IQ4_XS/64]; + uint8_t qs[QUANT_K_IQ4_XS/2]; +}; + +#if defined(DATA_A_IQ4_XS) +#define QUANT_K QUANT_K_IQ4_XS +#define QUANT_R QUANT_R_IQ4_XS +#define A_TYPE block_iq4_xs +#endif + +#define QUANT_K_IQ4_NL 32 +#define QUANT_R_IQ4_NL 2 struct block_iq4_nl { float16_t d; - uint8_t qs[QUANT_K/2]; + uint8_t qs[QUANT_K_IQ4_NL/2]; +}; + +struct block_iq4_nl_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ4_NL/2/2]; }; +#if defined(DATA_A_IQ4_NL) +#define QUANT_K QUANT_K_IQ4_NL +#define QUANT_R QUANT_R_IQ4_NL #define A_TYPE block_iq4_nl +#define A_TYPE_PACKED16 block_iq4_nl_packed16 +#endif -const int8_t kvalues_iq4nl[16] = { +#if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS) +const int8_t kvalues_iq4nl_const[16] = { int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10), int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113) }; + +shared FLOAT_TYPE kvalues_iq4nl[16]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + for (uint i = gl_LocalInvocationIndex.x; i < kvalues_iq4nl.length(); i += wgsize.x) { + kvalues_iq4nl[i] = FLOAT_TYPE(kvalues_iq4nl_const[i]); + } + barrier(); +} #endif + +// returns the bfloat value in the low 16b. +// See ggml_compute_fp32_to_bf16 +uint32_t fp32_to_bf16(float f) +{ + uint32_t u = floatBitsToUint(f); + u = (u + (0x7fff + ((u >> 16) & 1))) >> 16; + return u; +} + +float bf16_to_fp32(uint32_t u) +{ + return uintBitsToFloat(u << 16); +} + +#endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/vulkan-shaders/upscale.comp b/ggml/src/vulkan-shaders/upscale.comp index 511a086e..6f607380 100644 --- a/ggml/src/vulkan-shaders/upscale.comp +++ b/ggml/src/vulkan-shaders/upscale.comp @@ -2,7 +2,7 @@ layout (push_constant) uniform parameter { - uint ne; uint d_offset; + uint ne; uint a_offset; uint d_offset; uint nb00; uint nb01; uint nb02; uint nb03; uint ne10; uint ne11; uint ne12; uint ne13; float sf0; float sf1; float sf2; float sf3; @@ -32,5 +32,5 @@ void main() { const uint i02 = uint(i12 / p.sf2); const uint i03 = uint(i13 / p.sf3); - data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]); + data_d[p.d_offset + idx] = D_TYPE(data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]); } diff --git a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp index a792e203..0f244dea 100644 --- a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp @@ -16,13 +16,14 @@ #include <cstdio> #include <cstring> #include <cstdlib> +#include <cassert> +#include <algorithm> #include <sys/stat.h> #include <sys/types.h> #ifdef _WIN32 #include <windows.h> #include <direct.h> // For _mkdir on Windows - #include <algorithm> // For std::replace on w64devkit #else #include <unistd.h> #include <sys/wait.h> @@ -54,9 +55,19 @@ const std::vector<std::string> type_names = { "q4_k", "q5_k", "q6_k", - "iq4_nl" + "iq1_s", + "iq1_m", + "iq2_xxs", + "iq2_xs", + "iq2_s", + "iq3_xxs", + "iq3_s", + "iq4_xs", + "iq4_nl", + "bf16", }; +namespace { void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) { #ifdef _WIN32 HANDLE stdout_read, stdout_write; @@ -74,7 +85,8 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s } PROCESS_INFORMATION pi; - STARTUPINFOA si = { sizeof(STARTUPINFOA) }; + STARTUPINFOA si = {}; + si.cb = sizeof(STARTUPINFOA); si.dwFlags = STARTF_USESTDHANDLES; si.hStdOutput = stdout_write; si.hStdError = stderr_write; @@ -92,11 +104,11 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s std::array<char, 128> buffer; DWORD bytes_read; - while (ReadFile(stdout_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) { + while (ReadFile(stdout_read, buffer.data(), (DWORD)buffer.size(), &bytes_read, NULL) && bytes_read > 0) { stdout_str.append(buffer.data(), bytes_read); } - while (ReadFile(stderr_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) { + while (ReadFile(stderr_read, buffer.data(), (DWORD)buffer.size(), &bytes_read, NULL) && bytes_read > 0) { stderr_str.append(buffer.data(), bytes_read); } @@ -173,6 +185,13 @@ std::string to_uppercase(const std::string& input) { return result; } +bool string_starts_with(const std::string& str, const std::string& prefix) { + if (prefix.size() > str.size()) { + return false; + } + return std::equal(prefix.begin(), prefix.end(), str.begin()); +} + bool string_ends_with(const std::string& str, const std::string& suffix) { if (suffix.size() > str.size()) { return false; @@ -190,16 +209,31 @@ std::string basename(const std::string &path) { return path.substr(path.find_last_of("/\\") + 1); } -void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) { - std::string name = _name + (fp16 ? "" : "_fp32"); +// variables to track number of compiles in progress +static uint32_t compile_count = 0; +static std::mutex compile_count_mutex; +static std::condition_variable compile_count_cond; + +void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) { + std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")); std::string out_fname = join_paths(output_dir, name + ".spv"); std::string in_path = join_paths(input_dir, in_fname); + std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2"; + + // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734 + std::string opt_level = coopmat ? "" : "-O"; + #ifdef _WIN32 - std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""}; + std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""}; #else - std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname}; + std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o", out_fname}; #endif + + #ifdef GGML_VULKAN_SHADER_DEBUG_INFO + cmd.push_back("-g"); + #endif + for (const auto& define : defines) { cmd.push_back("-D" + define.first + "=" + define.second); } @@ -228,6 +262,12 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const } catch (const std::exception& e) { std::cerr << "Error executing command for " << name << ": " << e.what() << std::endl; } + { + std::lock_guard<std::mutex> guard(compile_count_mutex); + assert(compile_count > 0); + compile_count--; + } + compile_count_cond.notify_all(); } std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b) { @@ -236,12 +276,29 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s return result; } -void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id) { - std::string load_vec = fp16 ? "8" : "4"; - std::string aligned_b_type_f32 = fp16 ? "mat2x4" : "vec4"; - std::string aligned_b_type_f16 = fp16 ? "f16mat2x4" : "f16vec4"; +static std::vector<std::future<void>> compiles; +void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) { + { + // wait until fewer than N compiles are in progress. + // 16 is an arbitrary limit, the goal is to avoid "failed to create pipe" errors. + uint32_t N = 16; + std::unique_lock<std::mutex> guard(compile_count_mutex); + while (compile_count >= N) { + compile_count_cond.wait(guard); + } + compile_count++; + } + compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc)); +} - std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", fp16 ? "float16_t" : "float"}}; +void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) { + std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4"; + std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4"; + std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4"; + + std::map<std::string, std::string> base_dict = { + {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"}, + }; std::string shader_name = "matmul"; if (matmul_id) { @@ -253,225 +310,328 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu base_dict["FLOAT16"] = "1"; } + base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float"; + + if (coopmat) { + base_dict["COOPMAT"] = "1"; + } + + const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; + + auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string { + if (t == "bf16") { + // scalar path promotes to float + if (!coopmat && !coopmat2) { + return "float"; + } + return "bfloat16_t"; + } + if (coopmat2 || fp16) { + return "float16_t"; + } + return "float"; + }; + // Shaders with f16 B_TYPE - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_f32_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_f32_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16); - })); - - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16); - })); + string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + + string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + + // bf16 + { + std::string load_vec_a_unaligned = "1"; + // For aligned matmul loads + std::string load_vec_a = coopmat2 ? "1" : "4"; + + // scalar path promotes to float + std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32"; + + // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader +#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (!(coopmat || coopmat2)) +#endif + { + string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_bf16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); + } + } for (const auto& tname : type_names) { + std::string load_vec_quant = "2"; + if ((tname == "q4_0") || (tname == "q4_1")) + load_vec_quant = "8"; + else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl")) + load_vec_quant = "4"; + + if (tname == "bf16") { + continue; + } + std::string data_a_key = "DATA_A_" + to_uppercase(tname); // For unaligned, load one at a time for f32/f16, or two at a time for quants - std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2"; + std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant; // For aligned matmul loads - std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2"; - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16); - })); + std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant; + + // don't generate f32 variants for coopmat2 + if (!coopmat2) { + string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + } + + if (tname != "f16" && tname != "f32") { + string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + } + +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) { + string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); + } +#endif } } -void process_shaders(std::vector<std::future<void>>& tasks) { +void process_shaders() { std::cout << "ggml_vulkan: Generating and compiling shaders to SPIR-V" << std::endl; std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}}; - for (const auto& fp16 : {false, true}) { - matmul_shaders(tasks, fp16, false); - matmul_shaders(tasks, fp16, true); + // matmul + for (const auto& matmul_id : {false, true}) { + // No coopmats + // fp32 + matmul_shaders(false, matmul_id, false, false, false); + + // fp16, fp32acc and fp16acc + matmul_shaders(true, matmul_id, false, false, false); + matmul_shaders(true, matmul_id, false, false, true); + +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + // Coopmat, fp32acc and fp16acc + matmul_shaders(true, matmul_id, true, false, false); + matmul_shaders(true, matmul_id, true, false, true); +#endif + +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + // Coopmat2, fp32acc and fp16acc + matmul_shaders(true, matmul_id, false, true, false); + matmul_shaders(true, matmul_id, false, true, true); +#endif + } + + // flash attention + for (const auto& f16acc : {false, true}) { + std::string acctype = f16acc ? "float16_t" : "float"; + std::string acctypev4 = f16acc ? "f16vec4" : "vec4"; + + for (const auto& tname : type_names) { + if (tname == "f32") { + continue; + } + if (tname == "bf16") continue; + +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (tname == "f16") { + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", + merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc); + } else { + std::string data_a_key = "DATA_A_" + to_uppercase(tname); + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", + merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc); + } +#endif +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + if (tname == "f16") { + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp", + merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"ACC_TYPEV4", acctypev4}, {"COOPMAT", "1"}}), true, true, false, f16acc); + } else if (tname == "q4_0" || tname == "q8_0") { + std::string data_a_key = "DATA_A_" + to_uppercase(tname); + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp", + merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"ACC_TYPEV4", acctypev4}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc); + } +#endif + if (tname == "f16") { + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp", + merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, false, f16acc); + } else if (tname == "q4_0" || tname == "q8_0") { + std::string data_a_key = "DATA_A_" + to_uppercase(tname); + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp", + merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc); + } + } } for (const auto& tname : type_names) { // mul mat vec std::string data_a_key = "DATA_A_" + to_uppercase(tname); - std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; + std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}})); - })); + string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); - })); + string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); // Dequant shaders - if (tname != "f16") { - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}})); - })); + if (tname != "f16" && tname != "bf16") { + string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}})); } if (!string_ends_with(tname, "_k")) { - shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp"; + shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp"; if (tname == "f16") { - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); - })); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}})); } else { - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}); - })); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}})); } - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}); - })); + string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}})); } } - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - })); + string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}); + string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}); + string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}); // Norms - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}})); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - })); - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); - })); - - tasks.push_back(std::async(std::launch::async, [] { - string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}}); - })); - - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}})); - })); - - tasks.push_back(std::async(std::launch::async, [=] { - string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); - })); + string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); + + string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); + string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("cpy_f16_f32", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); + string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); + string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("contig_cpy_f16_f32", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); + + for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { + string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}}); + string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + } + + auto get_type_str = [](bool f16) { + return f16 ? "float16_t" : "float"; + }; + auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { + std::string s; + s += std::string(src0_f16 ? "_f16" : "_f32"); + s += std::string(src1_f16 ? "_f16" : "_f32"); + s += std::string(dst_f16 ? "_f16" : "_f32"); + return s; + }; + for (std::string op : {"add", "sub", "mul", "div"}) { + for (auto src0_f16 : {false, true}) { + for (auto src1_f16 : {false, true}) { + for (auto dst_f16 : {false, true}) { + auto name = op + get_suffix(src0_f16, src1_f16, dst_f16); + string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}}); + } + } + } + } + + string_to_spv("sub_f32", "sub.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}); + string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {}); + string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {}); + + string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + + string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}); + + string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("gelu_quick_f16", "gelu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("silu_f16", "silu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("relu_f16", "relu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("tanh_f16", "tanh.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("sigmoid_f16", "sigmoid.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("sigmoid_f32", "sigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}})); + string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); + + string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + + string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + + string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + + string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + + string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}}); + + string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}})); + string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}})); + + string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}})); + string_to_spv("im2col_f32_f16_rte", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}})); + + string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + + string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + + string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + + string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + + for (auto &c : compiles) { + c.wait(); + } } void write_output_files() { @@ -481,6 +641,7 @@ void write_output_files() { fprintf(hdr, "#include <cstdint>\n\n"); fprintf(src, "#include \"%s\"\n\n", basename(target_hpp).c_str()); + std::sort(shader_fnames.begin(), shader_fnames.end()); for (const auto& pair : shader_fnames) { const std::string& name = pair.first; #ifdef _WIN32 @@ -522,16 +683,28 @@ void write_output_files() { std::remove(path.c_str()); } } - + for (const char *op : {"add", "sub", "mul", "div"}) { + fprintf(hdr, "extern unsigned char *%s_data[2][2][2];\n", op); + fprintf(hdr, "extern uint64_t %s_len[2][2][2];\n", op); + fprintf(src, "unsigned char *%s_data[2][2][2] = {{{%s_f32_f32_f32_data, %s_f32_f32_f16_data}, {%s_f32_f16_f32_data, %s_f32_f16_f16_data}}, {{%s_f16_f32_f32_data, %s_f16_f32_f16_data}, {%s_f16_f16_f32_data, %s_f16_f16_f16_data}}};\n", op, op, op, op, op, op, op, op, op); + fprintf(src, "uint64_t %s_len[2][2][2] = {{{%s_f32_f32_f32_len, %s_f32_f32_f16_len}, {%s_f32_f16_f32_len, %s_f32_f16_f16_len}}, {{%s_f16_f32_f32_len, %s_f16_f32_f16_len}, {%s_f16_f16_f32_len, %s_f16_f16_f16_len}}};\n", op, op, op, op, op, op, op, op, op); + } fclose(hdr); fclose(src); } +} int main(int argc, char** argv) { std::map<std::string, std::string> args; - for (int i = 1; i < argc; i += 2) { - if (i + 1 < argc) { - args[argv[i]] = argv[i + 1]; + for (int i = 1; i < argc; ++i) { + std::string arg = argv[i]; + if (arg.rfind("--", 0) == 0) { + if (i + 1 < argc && argv[i + 1][0] != '-') { + args[arg] = argv[i + 1]; + ++i; + } else { + args[arg] = ""; + } } } @@ -566,12 +739,7 @@ int main(int argc, char** argv) { } } - std::vector<std::future<void>> tasks; - process_shaders(tasks); - - for (auto& task : tasks) { - task.get(); - } + process_shaders(); write_output_files(); diff --git a/ggml/src/vulkan-shaders/wkv6.comp b/ggml/src/vulkan-shaders/wkv6.comp new file mode 100644 index 00000000..35cc6c45 --- /dev/null +++ b/ggml/src/vulkan-shaders/wkv6.comp @@ -0,0 +1,87 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : require + +#define BLOCK_SIZE 64 +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout(push_constant) uniform Parameters { + uint B; + uint T; + uint C; + uint H; +}; + +layout(binding = 0) readonly buffer KBuf { A_TYPE k[]; }; +layout(binding = 1) readonly buffer VBuf { A_TYPE v[]; }; +layout(binding = 2) readonly buffer RBuf { A_TYPE r[]; }; +layout(binding = 3) readonly buffer TimeFBuf { A_TYPE tf[]; }; +layout(binding = 4) readonly buffer TimeDBuf { A_TYPE td[]; }; +layout(binding = 5) readonly buffer StateBuf { A_TYPE state_in[]; }; +layout(binding = 6) buffer DstBuf { A_TYPE dst[]; }; + +shared A_TYPE _k[BLOCK_SIZE], _r[BLOCK_SIZE], _tf[BLOCK_SIZE], _td[BLOCK_SIZE]; + +void main() { + const uint head_size = BLOCK_SIZE; + const uint batch_id = gl_WorkGroupID.x / H; + const uint head_id = gl_WorkGroupID.x % H; + const uint tid = gl_LocalInvocationID.x; + + const uint state_size = C * head_size; + const uint n_seq_tokens = T / B; + + if (batch_id >= B || head_id >= H) { + return; + } + + A_TYPE state[BLOCK_SIZE]; + [[unroll]] for (uint i = 0; i < head_size; i++) { + state[i] = state_in[batch_id * state_size + head_id * head_size * head_size + + i * head_size + tid]; + } + + barrier(); + _tf[tid] = tf[head_id * head_size + tid]; + barrier(); + + const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid; + const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid; + + for (uint t = start_t; t < end_t; t += C) { + barrier(); + _k[tid] = k[t]; + _r[tid] = r[t]; + _td[tid] = td[t]; + barrier(); + + const A_TYPE v_val = v[t]; + A_TYPE y = 0.0; + + [[unroll]] for (uint j = 0; j < head_size; j += 4) { + vec4 k_vec = vec4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + vec4 r_vec = vec4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + vec4 tf_vec = vec4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]); + vec4 td_vec = vec4(_td[j], _td[j+1], _td[j+2], _td[j+3]); + vec4 s_vec = vec4(state[j], state[j+1], state[j+2], state[j+3]); + + vec4 kv = k_vec * v_val; + + vec4 temp = tf_vec * kv + s_vec; + y += dot(r_vec, temp); + + s_vec = s_vec * td_vec + kv; + state[j] = s_vec.x; + state[j+1] = s_vec.y; + state[j+2] = s_vec.z; + state[j+3] = s_vec.w; + } + + dst[t] = y; + } + + [[unroll]] for (uint i = 0; i < head_size; i++) { + dst[T * C + batch_id * state_size + head_id * head_size * head_size + + i * head_size + tid] = state[i]; + } +} diff --git a/include/llama.h b/include/llama.h index aa0f3980..53adeb94 100644 --- a/include/llama.h +++ b/include/llama.h @@ -117,7 +117,7 @@ extern "C" { enum llama_rope_type { LLAMA_ROPE_TYPE_NONE = -1, LLAMA_ROPE_TYPE_NORM = 0, - LLAMA_ROPE_TYPE_NEOX = 2, + LLAMA_ROPE_TYPE_NEOX = 2, LLAMA_ROPE_TYPE_GLM = 4, }; diff --git a/src/llama.cpp b/src/llama.cpp index 7c10cebd..b06f9d44 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -9595,7 +9595,7 @@ static struct ggml_tensor * llm_build_norm( const llm_build_cb & cb, int il, float scale_eps = 1) { - if (type == LLM_NORM_RMS && mw) { + if (false && type == LLM_NORM_RMS && mw) { cur = ggml_fused_rms_norm(ctx, cur, mw, scale_eps * hparams.f_norm_rms_eps); if (mb) { cb(cur, "fused_norm", il); @@ -9686,7 +9686,7 @@ static struct ggml_tensor * llm_build_ffn( cur = tmp; } - if (type_gate == LLM_FFN_PAR && + if (false && type_gate == LLM_FFN_PAR && (type_op == LLM_FFN_SILU || type_op == LLM_FFN_RELU || (type_op == LLM_FFN_GELU && !act_scales))) { cur = ggml_fused_mul_unary(ctx, cur, tmp, type_op == LLM_FFN_SILU ? GGML_UNARY_OP_SILU : type_op == LLM_FFN_RELU ? GGML_UNARY_OP_RELU : GGML_UNARY_OP_GELU); |