diff options
Diffstat (limited to 'ggml/src/ggml-vulkan.cpp')
| -rw-r--r-- | ggml/src/ggml-vulkan.cpp | 965 |
1 files changed, 669 insertions, 296 deletions
diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index 6bcd81a7..86732837 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -177,24 +177,33 @@ struct vk_device_struct { 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_add_f32; + 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_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_leaky_relu_f32; + vk_pipeline pipeline_tanh_f32; vk_pipeline pipeline_diag_mask_inf_f32; vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16; vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16; vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16; vk_pipeline pipeline_argsort_f32; vk_pipeline pipeline_sum_rows_f32; + vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16; + vk_pipeline pipeline_timestep_embedding_f32; std::vector<vk_pipeline_ref> pipelines; @@ -236,8 +245,8 @@ struct vk_device_struct { }; struct vk_buffer_struct { - vk::Buffer buffer; - vk::DeviceMemory device_memory; + vk::Buffer buffer = VK_NULL_HANDLE; + vk::DeviceMemory device_memory = VK_NULL_HANDLE; vk::MemoryPropertyFlags memory_property_flags; void * ptr; size_t size = 0; @@ -259,6 +268,10 @@ struct vk_subbuffer { vk_buffer buffer; uint64_t offset; uint64_t size; + + operator vk::DescriptorBufferInfo() const { + return { buffer->buffer, offset, size }; + } }; struct vk_semaphore { @@ -320,7 +333,7 @@ struct vk_op_binary_push_constants { 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; - float param1; float param2; + float param1; float param2; int32_t param3; }; struct vk_op_diag_mask_push_constants { @@ -358,6 +371,25 @@ struct vk_op_argsort_push_constants { int32_t order; }; +struct vk_op_im2col_push_constants { + uint32_t batch_offset; uint32_t offset_delta; + uint32_t IC; + uint32_t IW; uint32_t IH; + uint32_t OW; uint32_t OH; + uint32_t KW; uint32_t KH; + uint32_t pelements; + uint32_t CHW; + int32_t s0; int32_t s1; + int32_t p0; int32_t p1; + int32_t d0; int32_t d1; +}; + +struct vk_op_timestep_embedding_push_constants { + uint32_t nb1; + uint32_t dim; + 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) {} @@ -367,28 +399,32 @@ struct vk_staging_memcpy { size_t n; }; -struct vk_context { - size_t idx; +struct vk_op_upscale_push_constants { + uint32_t ne; 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; +}; +struct vk_context_struct { vk_submission * s; std::vector<vk_sequence> seqs; - ggml_tensor * exit_tensor; + int exit_tensor_idx; std::vector<vk_staging_memcpy> in_memcpys; std::vector<vk_staging_memcpy> out_memcpys; vk_queue * q; }; +typedef std::shared_ptr<vk_context_struct> vk_context; +typedef std::weak_ptr<vk_context_struct> vk_context_ref; struct ggml_tensor_extra_gpu { - size_t ctx_idx; - vk_buffer_ref buffer_gpu; uint64_t offset; void reset() { - ctx_idx = 0; buffer_gpu.reset(); offset = 0; } @@ -459,8 +495,10 @@ struct ggml_backend_vk_context { vk_buffer buffer_pool[MAX_VK_BUFFERS]; - vk_context * compute_ctx; - vk_context * transfer_ctx; + vk_context_ref compute_ctx; + vk_context_ref transfer_ctx; + + std::vector<vk_context_ref> tensor_ctxs; }; #ifdef GGML_VULKAN_MEMORY_DEBUG @@ -510,12 +548,12 @@ static vk_instance_t vk_instance; static size_t vk_skip_checks; static size_t vk_output_tensor; -static void ggml_vk_print_tensor(ggml_backend * ctx, const ggml_tensor * tensor, const char * name); -static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor); -static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor); +static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name); +static void ggml_vk_check_results_0(ggml_tensor * tensor); +static void ggml_vk_check_results_1(ggml_tensor * tensor); #endif -typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); +typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend); @@ -708,11 +746,11 @@ static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, s return s; } -static void ggml_vk_submit(vk_context * ctx, vk::Fence fence) { - VK_LOG_DEBUG("ggml_vk_submit(" << ctx->seqs.size() << ", " << fence << ")"); +static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) { if (ctx->seqs.empty()) { return; } + VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")"); std::vector<std::vector<uint64_t>> tl_wait_vals; std::vector<std::vector<uint64_t>> tl_signal_vals; @@ -844,21 +882,17 @@ static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_ q.stage_flags = stage_flags; } -static vk_context * ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) { - VK_LOG_DEBUG("ggml_vk_create_context()"); - ctx->gc.contexts.emplace_back(); - vk_context * result = &ctx->gc.contexts[ctx->gc.contexts.size() - 1]; - memset((void *) result, 0, sizeof(vk_context)); - result->idx = ctx->gc.contexts.size() - 1; +static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) { + 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; return result; } -static vk_context * ggml_vk_create_temporary_context(vk_queue& q) { - VK_LOG_DEBUG("ggml_vk_create_temporary_context()"); - vk_context * result = new vk_context; - memset((void *) result, 0, sizeof(vk_context)); - result->idx = 0; +static vk_context ggml_vk_create_temporary_context(vk_queue& q) { + vk_context result = std::make_shared<vk_context_struct>(); + VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")"); result->q = &q; return result; } @@ -915,6 +949,10 @@ static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_pr static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) { VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")"); + if (size > device->max_memory_allocation_size) { + 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>(); @@ -1027,21 +1065,22 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) { return { buf, 0, VK_WHOLE_SIZE }; } -static void ggml_vk_sync_buffers(vk_context * ctx) { +static void ggml_vk_sync_buffers(vk_context& ctx) { VK_LOG_DEBUG("ggml_vk_sync_buffers()"); - const std::vector<vk::MemoryBarrier> mem_barriers{ { { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite }, { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite } } }; - ctx->s->buffer.pipelineBarrier( ctx->q->stage_flags, ctx->q->stage_flags, {}, - mem_barriers, + { { + {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite}, + {vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite} + } }, {}, {} ); } -static void ggml_vk_wait_events(vk_context * ctx, std::vector<vk::Event>&& events) { +static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events) { VK_LOG_DEBUG("ggml_vk_wait_events()"); if (events.empty()) { return; @@ -1598,6 +1637,7 @@ static void ggml_vk_load_shaders(vk_device& device) { 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_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_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); @@ -1605,20 +1645,31 @@ static void ggml_vk_load_shaders(vk_device& device) { 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_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); 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_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); + ggml_vk_create_pipeline(device, device->pipeline_concat_i32, "concat_i32", concat_i32_len, concat_i32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_upscale_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + 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_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_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); @@ -1634,6 +1685,11 @@ static void ggml_vk_load_shaders(vk_device& device) { 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_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_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); } static vk_device ggml_vk_get_device(size_t idx) { @@ -1961,7 +2017,7 @@ void ggml_vk_instance_init() { // Make sure at least one device exists if (devices.empty()) { std::cerr << "ggml_vulkan: Error: No devices found." << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } // Default to using all dedicated GPUs @@ -2057,9 +2113,9 @@ void ggml_vk_instance_init() { } static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { - GGML_ASSERT(idx < vk_instance.device_indices.size()); VK_LOG_DEBUG("ggml_vk_init(" << ctx->name << ", " << idx << ")"); ggml_vk_instance_init(); + GGML_ASSERT(idx < vk_instance.device_indices.size()); ctx->name = GGML_VK_NAME + std::to_string(idx); @@ -2077,9 +2133,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { ctx->staging_size = 0; ctx->staging_offset = 0; - ctx->compute_ctx = nullptr; - ctx->transfer_ctx = nullptr; - #ifdef GGML_VULKAN_CHECK_RESULTS const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS"); vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks)); @@ -2112,7 +2165,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_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()"); + VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_f32; } @@ -2126,7 +2179,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte return ctx->device->pipeline_matmul_f16; } - GGML_ASSERT(src1_type == GGML_TYPE_F32); + if (src1_type != GGML_TYPE_F32) { + return nullptr; + } switch (src0_type) { case GGML_TYPE_Q4_0: @@ -2370,28 +2425,23 @@ 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::vector<vk_subbuffer>&& buffers, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) { + + +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 : buffers) { - std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.size << "), "; + for (auto& buffer : descriptor_buffer_infos) { + std::cerr << "(" << buffer << ", " << buffer.offset << ", " << buffer.size << "), "; } std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))"); - std::vector<vk::DescriptorBufferInfo> descriptor_buffer_infos; - std::vector<vk::WriteDescriptorSet> write_descriptor_sets; GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size()); - GGML_ASSERT(buffers.size() == pipeline->parameter_count); - vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++]; - for (uint32_t i = 0; i < pipeline->parameter_count; i++) { - descriptor_buffer_infos.push_back({buffers[i].buffer->buffer, buffers[i].offset, buffers[i].size}); - } - for (uint32_t i = 0; i < pipeline->parameter_count; i++) { - write_descriptor_sets.push_back({descriptor_set, i, 0, 1, vk::DescriptorType::eStorageBuffer, nullptr, &descriptor_buffer_infos[i]}); - } + GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count); - ctx->device->device.updateDescriptorSets(write_descriptor_sets, {}); + vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->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 }, {}); subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants); subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline); @@ -2410,7 +2460,7 @@ static void ggml_vk_end_submission(vk_submission& s, std::vector<vk_semaphore> w s.signal_semaphores = std::move(signal_semaphores); } -static void ggml_vk_ctx_end(vk_context * ctx) { +static void ggml_vk_ctx_end(vk_context& ctx) { VK_LOG_DEBUG("ggml_vk_ctx_end(" << ctx << ", " << ctx->seqs.size() << ")"); if (ctx->s == nullptr) { return; @@ -2420,7 +2470,7 @@ static void ggml_vk_ctx_end(vk_context * ctx) { ctx->s = nullptr; } -static void ggml_vk_ctx_begin(vk_device& device, vk_context * subctx) { +static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) { VK_LOG_DEBUG("ggml_vk_ctx_begin(" << device->name << ")"); if (subctx->s != nullptr) { ggml_vk_ctx_end(subctx); @@ -2453,13 +2503,13 @@ static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) { } } -static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context * subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) { +static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context& subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) { VK_LOG_DEBUG("ggml_vk_buffer_write_nc_async(" << tensor << ")"); GGML_ASSERT(!ggml_is_contiguous(tensor)); // Buffer is already mapped if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) { std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } // Check if src is pinned memory vk_buffer buf; @@ -2527,7 +2577,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont staging = ctx->device->sync_staging; staging_offset = 0; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } @@ -2558,12 +2608,12 @@ 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, vk_buffer staging_buffer, size_t staging_offset, 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) { std::cerr << "ggml_vulkan: buffer_write_async dst buffer is host_visible. Use synchronous write." << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } // Check if src is pinned memory vk_buffer buf = nullptr; @@ -2602,7 +2652,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s staging_buffer = dst->device->sync_staging; staging_offset = 0; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } @@ -2623,7 +2673,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s } } -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, vk_buffer staging_buffer, size_t staging_offset, 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); } @@ -2638,7 +2688,7 @@ 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); + vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue); 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_ctx_end(subctx); @@ -2650,8 +2700,6 @@ 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 }); - - delete subctx; } } @@ -2660,12 +2708,14 @@ 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, vk_buffer staging_buffer, size_t staging_offset, 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); GGML_ASSERT(src != nullptr); + // TODO: staging_offset is not used + // Check if dst is pinned memory vk_buffer buf = nullptr; size_t buf_offset; @@ -2704,7 +2754,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, si staging_buffer = src->device->sync_staging; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } @@ -2714,18 +2764,18 @@ static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, si 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) { +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(vk_buffer& src, size_t offset, void * dst, size_t size) { - VK_LOG_DEBUG("ggml_vk_buffer_read(" << offset << ", " << size << ")"); + VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")"); if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) { 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); + vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue); ggml_vk_ctx_begin(src->device, subctx); ggml_vk_buffer_read_async(subctx, src, offset, dst, size, nullptr, 0, true); ggml_vk_ctx_end(subctx); @@ -2737,12 +2787,10 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_ for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - - delete subctx; } } -static void ggml_vk_buffer_copy_async(vk_context * ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) { +static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) { VK_LOG_DEBUG("ggml_vk_buffer_copy_async(" << size << ")"); // Make sure both buffers are on same device GGML_ASSERT(src->device == dst->device); @@ -2756,15 +2804,13 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr if (src->device == dst->device) { 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); 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 }); - - delete subctx; } else { VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")"); // Copy device to device @@ -2783,7 +2829,7 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr 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); + vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue); ggml_vk_ctx_begin(dst->device, subctx); subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c); ggml_vk_ctx_end(subctx); @@ -2791,8 +2837,6 @@ 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 }); - - delete subctx; } static uint32_t ggml_vk_guess_split_k(int m, int n, int k) { @@ -2855,7 +2899,7 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct } static void ggml_vk_matmul( - ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline, + ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline, 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, @@ -2879,7 +2923,7 @@ static void ggml_vk_matmul( } static void ggml_vk_matmul_id( - ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline, + 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, @@ -2913,10 +2957,10 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_ } std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } -static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) { +static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) { VK_LOG_DEBUG("ggml_vk_cpy_to_contiguous((" << tensor << ", 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] << "), "; std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")"); const int tensor_type_size = ggml_type_size(tensor->type); @@ -2934,7 +2978,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { 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) { +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) { 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] << "),)"); @@ -3079,7 +3123,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su } else if (qx_needs_dequant) { const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); } 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 }); @@ -3107,7 +3151,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su ); // 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) { 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] << "),)"); @@ -3268,11 +3312,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context }; ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, - { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23} }, + { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} }, 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) { 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] << "),)"); @@ -3340,10 +3384,10 @@ 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)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { 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, { 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 }); } -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) { 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] << "),)"); @@ -3415,10 +3459,11 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con // 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)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); + 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) { 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); @@ -3431,7 +3476,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, } } -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) { 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]; @@ -3499,7 +3544,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; if (mmp == nullptr) { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } // Not implemented @@ -3590,7 +3635,8 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * } else if (qx_needs_dequant) { const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, + { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); } 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 }); @@ -3618,7 +3664,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context * ); // 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) { 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]; @@ -3790,11 +3836,12 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte }; ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, - { { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne22 * ne23}, { d_ids, ids_buf_offset, ids_sz } }, + { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, + vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } }, 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) { 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); @@ -3803,8 +3850,8 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subct } } -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) { - // guaranteed to be an integer due to the check in ggml_can_repeat +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]; @@ -3825,6 +3872,7 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx const uint64_t nb02 = src0->nb[2]; const uint64_t nb03 = src0->nb[3]; + // 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; @@ -3852,8 +3900,8 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx for (uint64_t k1 = 0; k1 < ne01; k1++) { for (uint64_t i0 = 0; i0 < nr0; i0++) { copies.push_back({ - src_offset + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0, - dst_offset + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01, + 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, }); } @@ -3874,11 +3922,6 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx 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) { - 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; - } - return nullptr; case GGML_OP_GET_ROWS: GGML_ASSERT(src1->type == GGML_TYPE_I32); if (dst->type == GGML_TYPE_F16) { @@ -3888,6 +3931,14 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_get_rows_f32[src0->type]; } 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; + } + if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { + return ctx->device->pipeline_add_f16_f32_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; @@ -3898,6 +3949,22 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_div_f32; } return nullptr; + case GGML_OP_CONCAT: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_concat_f32; + } + if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + return ctx->device->pipeline_concat_f16; + } + if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) { + return ctx->device->pipeline_concat_i32; + } + return nullptr; + case GGML_OP_UPSCALE: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_upscale_f32; + } + return nullptr; case GGML_OP_SCALE: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_scale_f32; @@ -3913,6 +3980,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_clamp_f32; } return nullptr; + case GGML_OP_PAD: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_pad_f32; + } + return nullptr; case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: @@ -3922,6 +3994,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_norm_f32; } return nullptr; + case GGML_OP_GROUP_NORM: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_group_norm_f32; + } + return nullptr; case GGML_OP_RMS_NORM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_rms_norm_f32; @@ -3939,11 +4016,21 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_gelu_f32; } break; + 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; case GGML_UNARY_OP_RELU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_relu_f32; } break; + case GGML_UNARY_OP_TANH: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_tanh_f32; + } + break; default: break; } @@ -3995,6 +4082,24 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_sum_rows_f32; } return nullptr; + case GGML_OP_IM2COL: + if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_im2col_f32; + } + if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { + return ctx->device->pipeline_im2col_f32_f16; + } + return nullptr; + case GGML_OP_TIMESTEP_EMBEDDING: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_timestep_embedding_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; + } + return nullptr; default: return nullptr; } @@ -4018,9 +4123,12 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { case GGML_OP_ADD: 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: return true; default: return false; @@ -4028,7 +4136,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { } 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, const PC&& pc) { 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]; @@ -4078,7 +4186,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c std::cerr << " and " << ggml_type_name(src1->type); } std::cerr << " to " << ggml_type_name(dst->type) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } op_func(ctx, subctx, src0, src1, dst); @@ -4124,7 +4232,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c vk_buffer d_D = extra->buffer_gpu.lock(); // Workaround for tiny tensor inputs on ROPE - if (use_src1 && y_sz > d_D->size) { + if (op == GGML_OP_ROPE && use_src1 && y_sz > d_D->size) { y_sz = VK_WHOLE_SIZE; } @@ -4173,13 +4281,26 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c if (op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) { ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, 1); - switch (dst->op) { + switch (op) { case GGML_OP_NORM: case GGML_OP_RMS_NORM: case GGML_OP_SOFT_MAX: case GGML_OP_SUM_ROWS: - elements = { (uint32_t)ggml_nrows(src0), 1, 1 }; - break; + { + 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 }; @@ -4190,6 +4311,49 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c 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 IC = src1->ne[is_2D ? 2 : 1]; + + const uint32_t KH = is_2D ? src0->ne[1] : 1; + const uint32_t KW = src0->ne[0]; + + const uint32_t OH = is_2D ? dst->ne[2] : 1; + const uint32_t OW = dst->ne[1]; + + const uint32_t batch = src1->ne[3]; + + 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 }; + } else { + elements = { ne, 1, 1 }; + } + } break; default: elements = { (uint32_t)ggml_nelements(src0), 1, 1 }; break; @@ -4216,31 +4380,35 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c if (use_src1) { subbuf_y = { d_Y, y_buf_offset, y_sz }; } else { - subbuf_y = { d_X, 0, d_X->size }; + subbuf_y = { d_X, 0, x_sz }; } ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + 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, d_X->size }; + subbuf_z = { d_X, 0, x_sz }; } ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, subbuf_z, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + 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, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_Z, z_buf_offset, z_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + 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, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + 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, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + 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); } } else { GGML_ASSERT(op != GGML_OP_SOFT_MAX); @@ -4249,8 +4417,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c ggml_pipeline_allocate_descriptor_sets(ctx->device, pipeline, ne02 * ne03); - switch (dst->op) { + switch (op) { case GGML_OP_NORM: + case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: elements = { (uint32_t)ne01, 1, 1 }; break; @@ -4276,21 +4445,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c if (use_src1) { ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_Y, y_buf_offset + y_offset, y_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); + 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, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); + 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); } } } } } -static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); +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) { 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); @@ -4301,11 +4470,11 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx, (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, + 0.0f, 0.0f, 0, }); } -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) { 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); @@ -4316,11 +4485,11 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, cons (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, + 0.0f, 0.0f, 0, }); } -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) { 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); @@ -4331,11 +4500,11 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, cons (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, + 0.0f, 0.0f, 0, }); } -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) { 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); @@ -4346,11 +4515,44 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context * subctx, cons (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, + 0.0f, 0.0f, 0, }); } -static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, 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) { + int * op_params = (int *)dst->op_params; + + 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_CONCAT, { + (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)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, op_params[0], + }); +} + +static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + + const float sf0 = (float)dst->ne[0] / src0->ne[0]; + const float sf1 = (float)dst->ne[1] / src0->ne[1]; + const float sf2 = (float)dst->ne[2] / src0->ne[2]; + 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)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, + }); +} + +static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { 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); @@ -4364,7 +4566,7 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, co }); } -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) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4377,7 +4579,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, cons }); } -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) { 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); @@ -4391,7 +4593,20 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, co }); } -static void ggml_vk_cpy(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) { + 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_PAD, { + (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, + }); +} + +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; const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4406,27 +4621,37 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons }); } -static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { 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 }); } -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_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { + int * op_params = (int *)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; + + ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }); +} + +static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { 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 }); } -static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { +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_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) { 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] }); } -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) { float * op_params = (float *)dst->op_params; float scale = op_params[0]; @@ -4451,7 +4676,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, }); } -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) { const int n_dims = ((int32_t *) dst->op_params)[1]; // const int mode = ((int32_t *) dst->op_params)[2]; // const int n_ctx = ((int32_t *) dst->op_params)[3]; @@ -4475,7 +4700,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con }); } -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) { int32_t * op_params = (int32_t *)dst->op_params; uint32_t ncols = src0->ne[0]; @@ -4494,10 +4719,59 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx, }); } -static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { +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_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + const int32_t s0 = dst->op_params[0]; + const int32_t s1 = dst->op_params[1]; + const int32_t p0 = dst->op_params[2]; + const int32_t p1 = dst->op_params[3]; + const int32_t d0 = dst->op_params[4]; + const int32_t d1 = dst->op_params[5]; + + const bool is_2D = dst->op_params[6] == 1; + + const uint32_t IC = src1->ne[is_2D ? 2 : 1]; + const uint32_t IH = is_2D ? src1->ne[1] : 1; + const uint32_t IW = src1->ne[0]; + + const uint32_t KH = is_2D ? src0->ne[1] : 1; + const uint32_t KW = src0->ne[0]; + + const uint32_t OH = is_2D ? dst->ne[2] : 1; + 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 pelements = OW * KW * KH; + + ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL, { + batch_offset, offset_delta, + IC, IW, IH, OW, OH, KW, KH, + pelements, + IC * KH * KW, + s0, s1, p0, p1, d0, d1, + }); +} + +static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { + 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, + }); +} + +static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { + 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 }); +} + #ifdef GGML_VULKAN_RUN_TESTS static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0, int ne1, int i0, int i1, int i2) { if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16) { @@ -4521,7 +4795,7 @@ static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0 } else if (type == GGML_TYPE_F16) { val = ggml_fp16_to_fp32(*((const ggml_fp16_t *) data + i2*ne1*ne0 + idx1*ne0 + idx0)); } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } fprintf(stderr, "% 7.2f ", val); } else { @@ -4555,7 +4829,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f16->a_s; shname = "F16_ALIGNED_S"; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } else if (shader_size == 1) { if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) { @@ -4571,7 +4845,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f16->a_m; shname = "F16_ALIGNED_M"; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } else if (shader_size == 2) { if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) { @@ -4587,7 +4861,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t p = ctx->device->pipeline_matmul_f16->a_l; shname = "F16_ALIGNED_L"; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } else { GGML_ASSERT(0); @@ -4668,7 +4942,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t } else if (std::is_same<ggml_fp16_t, X_TYPE>()) { x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } for (size_t i = 0; i < y_ne; i++) { @@ -4679,14 +4953,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t // 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); } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } 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->device->compute_queue); for (size_t i = 0; i < num_it; i++) { ggml_vk_ctx_begin(ctx->device, subctx); ggml_vk_matmul( @@ -4727,14 +5001,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t } else if (std::is_same<ggml_fp16_t, X_TYPE>()) { src0_type = GGML_TYPE_F16; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if (std::is_same<float, Y_TYPE>()) { src1_type = GGML_TYPE_F32; } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) { src1_type = GGML_TYPE_F16; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, src0_type, k, m, batch); @@ -4841,7 +5115,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, int i0, int i1 } else if (tensor->type == GGML_TYPE_F16) { val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0])); } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } fprintf(stderr, "% 7.2f ", val); } else { @@ -4894,7 +5168,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ 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->device->compute_queue); 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}); @@ -5027,7 +5301,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, 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); + 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( @@ -5175,7 +5449,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm const bool y_f32_kernel = use_src1 && src1->type == GGML_TYPE_F32 && !y_non_contig; - bool mmp = (use_src0 && use_src1 && src1_type == GGML_TYPE_F32) ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0_type, y_non_contig ? GGML_TYPE_F16 : src1->type) != nullptr : false; + 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); @@ -5211,24 +5485,33 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm 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; @@ -5236,6 +5519,13 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm 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; } @@ -5391,7 +5681,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { std::cerr << std::endl; } - GGML_ASSERT(false); + GGML_ABORT("fatal error"); #endif if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) { @@ -5430,7 +5720,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { } } -static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){ +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; if (ggml_is_empty(node) || extra == nullptr) { @@ -5457,7 +5747,9 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod 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; @@ -5468,13 +5760,17 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_ADD: 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_NORM: + case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: @@ -5483,109 +5779,147 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_MUL_MAT_ID: case GGML_OP_ARGSORT: case GGML_OP_SUM_ROWS: + case GGML_OP_IM2COL: + case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_LEAKY_RELU: break; default: std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); return; } - if (ctx->compute_ctx == nullptr) { - ctx->compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); - ggml_vk_ctx_begin(ctx->device, ctx->compute_ctx); + 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); + } else { + compute_ctx = ctx->compute_ctx.lock(); } switch (node->op) { case GGML_OP_REPEAT: - ggml_vk_repeat(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_repeat(ctx, compute_ctx, src0, node); break; case GGML_OP_GET_ROWS: - ggml_vk_get_rows(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node); break; case GGML_OP_ADD: - ggml_vk_add(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_add(ctx, compute_ctx, src0, src1, node); break; case GGML_OP_MUL: - ggml_vk_mul(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_mul(ctx, compute_ctx, src0, src1, node); break; case GGML_OP_DIV: - ggml_vk_div(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_div(ctx, compute_ctx, src0, src1, node); + + break; + case GGML_OP_CONCAT: + ggml_vk_concat(ctx, compute_ctx, src0, src1, node); + + break; + case GGML_OP_UPSCALE: + ggml_vk_upscale(ctx, compute_ctx, src0, node); break; case GGML_OP_SCALE: - ggml_vk_scale(ctx, ctx->compute_ctx, src0, node); + ggml_vk_scale(ctx, compute_ctx, src0, node); break; case GGML_OP_SQR: - ggml_vk_sqr(ctx, ctx->compute_ctx, src0, node); + ggml_vk_sqr(ctx, compute_ctx, src0, node); break; case GGML_OP_CLAMP: - ggml_vk_clamp(ctx, ctx->compute_ctx, src0, node); + ggml_vk_clamp(ctx, compute_ctx, src0, node); + + break; + case GGML_OP_PAD: + ggml_vk_pad(ctx, compute_ctx, src0, node); break; case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: - ggml_vk_cpy(ctx, ctx->compute_ctx, src0, node); + ggml_vk_cpy(ctx, compute_ctx, src0, node); break; case GGML_OP_NORM: - ggml_vk_norm(ctx, ctx->compute_ctx, src0, node); + ggml_vk_norm(ctx, compute_ctx, src0, node); + + break; + case GGML_OP_GROUP_NORM: + ggml_vk_group_norm(ctx, compute_ctx, src0, node); break; case GGML_OP_RMS_NORM: - ggml_vk_rms_norm(ctx, ctx->compute_ctx, src0, node); + ggml_vk_rms_norm(ctx, compute_ctx, src0, node); 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: - ggml_vk_unary(ctx, ctx->compute_ctx, src0, node); + case GGML_UNARY_OP_TANH: + ggml_vk_unary(ctx, compute_ctx, src0, node); break; default: return; } break; case GGML_OP_DIAG_MASK_INF: - ggml_vk_diag_mask_inf(ctx, ctx->compute_ctx, src0, node); + ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node); break; case GGML_OP_SOFT_MAX: - ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node); break; case GGML_OP_ROPE: - ggml_vk_rope(ctx, ctx->compute_ctx, src0, src1, src2, node); + ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node); break; case GGML_OP_ARGSORT: - ggml_vk_argsort(ctx, ctx->compute_ctx, src0, node); + ggml_vk_argsort(ctx, compute_ctx, src0, node); break; case GGML_OP_SUM_ROWS: - ggml_vk_sum_rows(ctx, ctx->compute_ctx, src0, node); + ggml_vk_sum_rows(ctx, compute_ctx, src0, node); + + break; + case GGML_OP_IM2COL: + ggml_vk_im2col(ctx, compute_ctx, src0, src1, node); + + break; + case GGML_OP_TIMESTEP_EMBEDDING: + ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node); + + break; + case GGML_OP_LEAKY_RELU: + ggml_vk_leaky_relu(ctx, compute_ctx, src0, node); break; case GGML_OP_MUL_MAT: - ggml_vk_mul_mat(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node); break; case GGML_OP_MUL_MAT_ID: - ggml_vk_mul_mat_id(ctx, ctx->compute_ctx, src0, src1, src2, node); + ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node); break; default: return; } - extra->ctx_idx = ctx->compute_ctx->idx; + ctx->tensor_ctxs[node_idx] = compute_ctx; #ifdef GGML_VULKAN_CHECK_RESULTS // Force context reset on each node so that each tensor ends up in its own context @@ -5594,13 +5928,13 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod #endif if (last_node) { - ggml_vk_ctx_end(ctx->compute_ctx); - ctx->compute_ctx->exit_tensor = node; - ctx->compute_ctx = nullptr; + ggml_vk_ctx_end(compute_ctx); + compute_ctx->exit_tensor_idx = node_idx; + ctx->compute_ctx.reset(); } } -static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor){ +static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx){ ggml_tensor_extra_gpu * extra = nullptr; switch (tensor->op) { @@ -5608,13 +5942,17 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_GET_ROWS: 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_NORM: + case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: @@ -5626,6 +5964,10 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_NONE: case GGML_OP_ARGSORT: case GGML_OP_SUM_ROWS: + case GGML_OP_IM2COL: + case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_LEAKY_RELU: + case GGML_OP_REPEAT: extra = (ggml_tensor_extra_gpu *) tensor->extra; break; @@ -5633,7 +5975,9 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * switch (ggml_get_unary_op(tensor)) { 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: extra = (ggml_tensor_extra_gpu *) tensor->extra; break; default: @@ -5656,31 +6000,31 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * 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(ctx, tensor); + ggml_vk_check_results_0(tensor); #endif - vk_context& subctx = ctx->gc.contexts[extra->ctx_idx]; + vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock(); // Only run if ctx hasn't been submitted yet - if (!subctx.seqs.empty()) { + if (!subctx->seqs.empty()) { // Do staging buffer copies - for (auto& cpy : subctx.in_memcpys) { + for (auto& cpy : subctx->in_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - ggml_vk_submit(&subctx, ctx->fence); + ggml_vk_submit(subctx, ctx->fence); } - if (tensor == subctx.exit_tensor) { + 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) { + for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - subctx.in_memcpys.clear(); - subctx.out_memcpys.clear(); + subctx->in_memcpys.clear(); + subctx->out_memcpys.clear(); } return true; @@ -5725,8 +6069,7 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) { ctx->staging_offset = 0; - ctx->compute_ctx = nullptr; - ctx->transfer_ctx = nullptr; + ctx->tensor_ctxs.clear(); ctx->gc.contexts.clear(); } @@ -6063,15 +6406,20 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - if (ctx->transfer_ctx == nullptr) { + vk_context transfer_ctx; + + if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); - ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx); + transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + 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(); - ggml_vk_buffer_write_async(ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); + ggml_vk_buffer_write_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); } 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) { @@ -6081,15 +6429,20 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - if (ctx->transfer_ctx == nullptr) { + vk_context transfer_ctx; + + if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); - ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx); + transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + 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(); - ggml_vk_buffer_read_async(ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); + ggml_vk_buffer_read_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size, ctx->staging, ctx->staging_offset); } GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) { @@ -6099,16 +6452,21 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - if (ctx->transfer_ctx == nullptr) { + vk_context transfer_ctx; + + if (ctx->transfer_ctx.expired()) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); - ggml_vk_ctx_begin(ctx->device, ctx->transfer_ctx); + transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); + 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(); - ggml_vk_buffer_copy_async(ctx->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, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src)); return true; } @@ -6118,25 +6476,27 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) { VK_LOG_DEBUG("ggml_backend_vk_synchronize()"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; - if(ctx->transfer_ctx == nullptr) { + if(ctx->transfer_ctx.expired()) { return; } - ggml_vk_ctx_end(ctx->transfer_ctx); + vk_context transfer_ctx = ctx->transfer_ctx.lock(); + + ggml_vk_ctx_end(transfer_ctx); - for (auto& cpy : ctx->transfer_ctx->in_memcpys) { + for (auto& cpy : transfer_ctx->in_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - ggml_vk_submit(ctx->transfer_ctx, ctx->fence); + 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 }); - for (auto& cpy : ctx->transfer_ctx->out_memcpys) { + for (auto& cpy : transfer_ctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } - ctx->transfer_ctx = nullptr; + ctx->transfer_ctx.reset(); } static bool ggml_vk_is_empty(ggml_tensor * node) { @@ -6159,8 +6519,11 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen last_node -= 1; } + // 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 == last_node); + ggml_vk_build_graph(ctx, cgraph->nodes[i], i, i == last_node); } for (int i = 0; i < cgraph->n_nodes; i++) { @@ -6170,13 +6533,17 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen continue; } - bool ok = ggml_vk_compute_forward(ctx, node); + bool ok = ggml_vk_compute_forward(ctx, node, i); if (!ok) { - fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); + 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; + } } #ifdef GGML_VULKAN_CHECK_RESULTS else { - ggml_vk_check_results_1(ctx, node); + ggml_vk_check_results_1(node); } #endif GGML_ASSERT(ok); @@ -6196,8 +6563,10 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { case GGML_UNARY_OP_GELU: + case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_TANH: return ggml_is_contiguous(op->src[0]); default: return false; @@ -6270,11 +6639,11 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const } 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; + case GGML_OP_REPEAT: + { + ggml_type src0_type = op->src[0]->type; + return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; + } break; case GGML_OP_ROPE: return ggml_is_contiguous(op->src[0]); case GGML_OP_NONE: @@ -6283,18 +6652,25 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NORM: + case GGML_OP_GROUP_NORM: + case GGML_OP_RMS_NORM: case GGML_OP_ADD: case GGML_OP_MUL: case GGML_OP_DIV: - case GGML_OP_RMS_NORM: + 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_ARGSORT: case GGML_OP_SUM_ROWS: + case GGML_OP_IM2COL: + case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_LEAKY_RELU: return true; default: return false; @@ -6498,7 +6874,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d } else if (tensor->type == GGML_TYPE_I32) { val = *(const int32_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]); } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } fprintf(stderr, "% 7.2f ", val); } else { @@ -6509,10 +6885,12 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d } } -static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tensor * tensor, const char * name) { +static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) { void * tensor_data = tensor->data; - if (ggml_backend_buffer_is_vk(tensor->buffer)) { + const bool is_gpu = tensor->buffer != nullptr && ggml_backend_buffer_is_vk(tensor->buffer); + + if (is_gpu) { const size_t tensor_size = ggml_nbytes(tensor); tensor_data = malloc(tensor_size); @@ -6533,13 +6911,10 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso std::cerr << std::endl << "Result:" << std::endl; ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0); std::cerr << std::endl; - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0); - std::cerr << std::endl; std::vector<const ggml_tensor *> done; ggml_vk_print_graph_origin(tensor, done); - if (ggml_backend_buffer_is_vk(tensor->buffer)) { + if (is_gpu) { free(tensor_data); } } @@ -6548,8 +6923,8 @@ void * comp_result; size_t comp_size; size_t comp_nb[GGML_MAX_DIMS]; size_t check_counter = 0; -static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor) { - if (tensor->op == GGML_OP_TRANSPOSE) { +static void ggml_vk_check_results_0(ggml_tensor * tensor) { + if (tensor->op == GGML_OP_TRANSPOSE) { return; } @@ -6565,7 +6940,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * ggml_tensor * src2 = tensor->src[2]; struct ggml_init_params iparams = { - /*.mem_size =*/ 1024*1024*1024, + /*.mem_size =*/ 2ul*1024ul*1024ul*1024ul, /*.mem_buffer =*/ NULL, /*.no_alloc =*/ false, }; @@ -6620,11 +6995,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(ctx, src0, "src0"); + ggml_vk_print_tensor(src0, "src0"); } } if (src1 != nullptr) { @@ -6662,27 +7037,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(ctx, src1, "src1"); - std::cerr << "TENSOR CHECK: " << ggml_op_name(src1_clone->op) << " (check " << check_counter << ")" << std::endl; - std::cerr << "src1_clone=" << tensor << " src1_clone->type: " << ggml_type_name(src1_clone->type) << " ne0=" << src1_clone->ne[0] << " nb0=" << src1_clone->nb[0] << " ne1=" << src1_clone->ne[1] << " nb1=" << src1_clone->nb[1] << " ne2=" << src1_clone->ne[2] << " nb2=" << src1_clone->nb[2] << " ne3=" << src1_clone->ne[3] << " nb3=" << src1_clone->nb[3] << std::endl; - if (src1->src[0] != nullptr) { - std::cerr << "src1->src[0]=" << src1->src[0] << " op=" << ggml_op_name(src1->src[0]->op) << " type=" << ggml_type_name(src1->src[0]->type) << " ne0=" << src1->src[0]->ne[0] << " nb0=" << src1->src[0]->nb[0] << " ne1=" << src1->src[0]->ne[1] << " nb1=" << src1->src[0]->nb[1] << " ne2=" << src1->src[0]->ne[2] << " nb2=" << src1->src[0]->nb[2] << " ne3=" << src1->src[0]->ne[3] << " nb3=" << src1->src[0]->nb[3] << std::endl; - } - if (src1->src[1] != nullptr) { - std::cerr << "src1->src[1]=" << src1->src[1] << " op=" << ggml_op_name(src1->src[1]->op) << " type=" << ggml_type_name(src1->src[1]->type) << " ne0=" << src1->src[1]->ne[0] << " nb0=" << src1->src[1]->nb[0] << " ne1=" << src1->src[1]->ne[1] << " nb1=" << src1->src[1]->nb[1] << " ne2=" << src1->src[1]->ne[2] << " nb2=" << src1->src[1]->nb[2] << " ne3=" << src1->src[1]->ne[3] << " nb3=" << src1->src[1]->nb[3] << std::endl; - } - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 0, 0); - std::cerr << std::endl; - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 1, 0); - std::cerr << std::endl; - std::vector<const ggml_tensor *> done; - ggml_vk_print_graph_origin(src1_clone, done); + ggml_vk_print_tensor(src1, "src1"); } } if (src2 != nullptr) { @@ -6720,27 +7079,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(ctx, src2, "src2"); - std::cerr << "TENSOR CHECK: " << ggml_op_name(src2_clone->op) << " (check " << check_counter << ")" << std::endl; - std::cerr << "src2_clone=" << tensor << " src2_clone->type: " << ggml_type_name(src2_clone->type) << " ne0=" << src2_clone->ne[0] << " nb0=" << src2_clone->nb[0] << " ne1=" << src2_clone->ne[1] << " nb1=" << src2_clone->nb[1] << " ne2=" << src2_clone->ne[2] << " nb2=" << src2_clone->nb[2] << " ne3=" << src2_clone->ne[3] << " nb3=" << src2_clone->nb[3] << std::endl; - if (src2->src[0] != nullptr) { - std::cerr << "src2->src[0]=" << src2->src[0] << " op=" << ggml_op_name(src2->src[0]->op) << " type=" << ggml_type_name(src2->src[0]->type) << " ne0=" << src2->src[0]->ne[0] << " nb0=" << src2->src[0]->nb[0] << " ne1=" << src2->src[0]->ne[1] << " nb1=" << src2->src[0]->nb[1] << " ne2=" << src2->src[0]->ne[2] << " nb2=" << src2->src[0]->nb[2] << " ne3=" << src2->src[0]->ne[3] << " nb3=" << src2->src[0]->nb[3] << std::endl; - } - if (src2->src[1] != nullptr) { - std::cerr << "src2->src[1]=" << src2->src[1] << " op=" << ggml_op_name(src2->src[1]->op) << " type=" << ggml_type_name(src2->src[1]->type) << " ne0=" << src2->src[1]->ne[0] << " nb0=" << src2->src[1]->nb[0] << " ne1=" << src2->src[1]->ne[1] << " nb1=" << src2->src[1]->nb[1] << " ne2=" << src2->src[1]->ne[2] << " nb2=" << src2->src[1]->nb[2] << " ne3=" << src2->src[1]->ne[3] << " nb3=" << src2->src[1]->nb[3] << std::endl; - } - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 0, 0); - std::cerr << std::endl; - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 1, 0); - std::cerr << std::endl; - std::vector<const ggml_tensor *> done; - ggml_vk_print_graph_origin(src2_clone, done); + ggml_vk_print_tensor(src2, "src2"); } } @@ -6752,16 +7095,24 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone); } else if (tensor->op == GGML_OP_DIV) { tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone); + } else if (tensor->op == GGML_OP_CONCAT) { + tensor_clone = ggml_concat(ggml_ctx, src0_clone, src1_clone, *(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]); } else if (tensor->op == GGML_OP_SCALE) { tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_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]); + } 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]); } else if (tensor->op == GGML_OP_ADD) { tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone); } else if (tensor->op == GGML_OP_NORM) { tensor_clone = ggml_norm(ggml_ctx, src0_clone, *(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); } else if (tensor->op == GGML_OP_RMS_NORM) { tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); } else if (tensor->op == GGML_OP_SOFT_MAX) { @@ -6777,12 +7128,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * const int mode = ((int32_t *) tensor->op_params)[2]; //const int n_ctx_ggml = ((int32_t *) tensor->op_params)[3]; const int n_ctx_orig_ggml = ((int32_t *) tensor->op_params)[4]; - float freq_base = ((float *) tensor->op_params)[5]; - float freq_scale = ((float *) tensor->op_params)[6]; - float ext_factor = ((float *) tensor->op_params)[7]; - float attn_factor = ((float *) tensor->op_params)[8]; - float beta_fast = ((float *) tensor->op_params)[9]; - float beta_slow = ((float *) tensor->op_params)[10]; + const float freq_base = ((float *) tensor->op_params)[5]; + const float freq_scale = ((float *) tensor->op_params)[6]; + const float ext_factor = ((float *) tensor->op_params)[7]; + 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); } else if (tensor->op == GGML_OP_UNARY) { switch (ggml_get_unary_op(tensor)) { @@ -6792,12 +7143,18 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_UNARY_OP_GELU: tensor_clone = ggml_gelu(ggml_ctx, src0_clone); break; + case GGML_UNARY_OP_GELU_QUICK: + tensor_clone = ggml_gelu_quick(ggml_ctx, src0_clone); + break; case GGML_UNARY_OP_RELU: tensor_clone = ggml_relu(ggml_ctx, src0_clone); break; + case GGML_UNARY_OP_TANH: + tensor_clone = ggml_tanh(ggml_ctx, src0_clone); + break; default: std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) { if (src1 == nullptr) { @@ -6823,9 +7180,26 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params); } else if (tensor->op == GGML_OP_SUM_ROWS) { tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone); + } 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]; + const int32_t p1 = tensor->op_params[3]; + const int32_t d0 = tensor->op_params[4]; + 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); + } 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); + } 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 { std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx); @@ -6834,7 +7208,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8); if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(ctx, tensor_clone, "tensor_clone"); + ggml_vk_print_tensor(tensor_clone, "tensor_clone"); } comp_size = ggml_nbytes(tensor_clone); @@ -6851,9 +7225,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * } ggml_free(ggml_ctx); + + VK_LOG_DEBUG("END ggml_vk_check_results_0(" << tensor->name << ")"); } -static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor) { +static void ggml_vk_check_results_1(ggml_tensor * tensor) { if (tensor->op == GGML_OP_TRANSPOSE) { return; } @@ -6912,7 +7288,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * } } else { std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl; - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) { @@ -6935,7 +7311,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * std::cerr << std::endl; std::vector<const ggml_tensor *> done; ggml_vk_print_graph_origin(tensor, done); - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) { first_error[0] = i0; @@ -6977,11 +7353,6 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * std::cerr << std::endl << "Correct:" << std::endl; ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 0, 0); std::cerr << std::endl; - std::cerr << std::endl << "Result:" << std::endl; - ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0); - std::cerr << std::endl << "Correct:" << std::endl; - ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 1, 0); - std::cerr << std::endl; std::vector<const ggml_tensor *> done; ggml_vk_print_graph_origin(tensor, done); } @@ -7006,7 +7377,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * std::cerr << std::endl; std::vector<const ggml_tensor *> done; ggml_vk_print_graph_origin(tensor, done); - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } else { std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " avg_err=" << avg_err << std::endl; } @@ -7018,5 +7389,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * if (ggml_backend_buffer_is_vk(tensor->buffer)) { free(tensor_data); } + + VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")"); } #endif |