summaryrefslogtreecommitdiff
path: root/ggml/src/ggml-cuda
diff options
context:
space:
mode:
Diffstat (limited to 'ggml/src/ggml-cuda')
-rw-r--r--ggml/src/ggml-cuda/argsort.cu2
-rw-r--r--ggml/src/ggml-cuda/binbcast.cu2
-rw-r--r--ggml/src/ggml-cuda/common.cuh208
-rw-r--r--ggml/src/ggml-cuda/convert.cu1
-rw-r--r--ggml/src/ggml-cuda/cpy.cu4
-rw-r--r--ggml/src/ggml-cuda/dmmv.cu23
-rw-r--r--ggml/src/ggml-cuda/dmmv.cuh2
-rw-r--r--ggml/src/ggml-cuda/fattn-common.cuh6
-rw-r--r--ggml/src/ggml-cuda/fattn-tile-f16.cu2
-rw-r--r--ggml/src/ggml-cuda/fattn-tile-f32.cu2
-rw-r--r--ggml/src/ggml-cuda/fattn.cu10
-rw-r--r--ggml/src/ggml-cuda/getrows.cu3
-rw-r--r--ggml/src/ggml-cuda/iqk_mmvq.cu1
-rw-r--r--ggml/src/ggml-cuda/mmq.cu2
-rw-r--r--ggml/src/ggml-cuda/mmq.cuh4
-rw-r--r--ggml/src/ggml-cuda/mmvq.cu6
-rw-r--r--ggml/src/ggml-cuda/norm.cu9
-rw-r--r--ggml/src/ggml-cuda/quantize.cu2
-rw-r--r--ggml/src/ggml-cuda/rope.cu4
-rw-r--r--ggml/src/ggml-cuda/vendors/cuda.h14
-rw-r--r--ggml/src/ggml-cuda/vendors/hip.h177
-rw-r--r--ggml/src/ggml-cuda/vendors/musa.h171
22 files changed, 423 insertions, 232 deletions
diff --git a/ggml/src/ggml-cuda/argsort.cu b/ggml/src/ggml-cuda/argsort.cu
index 15757ca1..607ded85 100644
--- a/ggml/src/ggml-cuda/argsort.cu
+++ b/ggml/src/ggml-cuda/argsort.cu
@@ -81,7 +81,7 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
} else if (order == GGML_SORT_ORDER_DESC) {
k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
} else {
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
diff --git a/ggml/src/ggml-cuda/binbcast.cu b/ggml/src/ggml-cuda/binbcast.cu
index 76cc01b2..62d115f1 100644
--- a/ggml/src/ggml-cuda/binbcast.cu
+++ b/ggml/src/ggml-cuda/binbcast.cu
@@ -259,7 +259,7 @@ static void ggml_cuda_op_bin_bcast(
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 07a53bcd..9aff6c13 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -12,6 +12,10 @@
#else
#define GGML_COMMON_DECL_CUDA
#define GGML_COMMON_IMPL_CUDA
+#if defined(GGML_USE_MUSA)
+#define GGML_COMMON_DECL_MUSA
+#define GGML_COMMON_IMPL_MUSA
+#endif
#endif
#include "ggml-common.h"
@@ -23,111 +27,11 @@
#include <vector>
#if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
-#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_OP_N HIPBLAS_OP_N
-#define CUBLAS_OP_T HIPBLAS_OP_T
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH 0
-#define CUDA_R_16F HIPBLAS_R_16F
-#define CUDA_R_32F HIPBLAS_R_32F
-#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
-#define cublasCreate hipblasCreate
-#define cublasDestroy hipblasDestroy
-#define cublasGemmEx hipblasGemmEx
-#define cublasGemmBatchedEx hipblasGemmBatchedEx
-#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-#define cublasHandle_t hipblasHandle_t
-#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
-#define cublasSetStream hipblasSetStream
-#define cublasSgemm hipblasSgemm
-#define cublasStatus_t hipblasStatus_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
-#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
-#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
-#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
-#define cudaDeviceProp hipDeviceProp_t
-#define cudaDeviceSynchronize hipDeviceSynchronize
-#define cudaError_t hipError_t
-#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
-#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
-#define cudaEventCreateWithFlags hipEventCreateWithFlags
-#define cudaEventDisableTiming hipEventDisableTiming
-#define cudaEventRecord hipEventRecord
-#define cudaEventSynchronize hipEventSynchronize
-#define cudaEvent_t hipEvent_t
-#define cudaEventDestroy hipEventDestroy
-#define cudaFree hipFree
-#define cudaFreeHost hipHostFree
-#define cudaGetDevice hipGetDevice
-#define cudaGetDeviceCount hipGetDeviceCount
-#define cudaGetDeviceProperties hipGetDeviceProperties
-#define cudaGetErrorString hipGetErrorString
-#define cudaGetLastError hipGetLastError
-#define cudaHostRegister hipHostRegister
-#define cudaHostRegisterPortable hipHostRegisterPortable
-#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
-#define cudaHostUnregister hipHostUnregister
-#define cudaLaunchHostFunc hipLaunchHostFunc
-#define cudaMalloc hipMalloc
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#define cudaMemcpy hipMemcpy
-#define cudaMemcpyAsync hipMemcpyAsync
-#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
-#define cudaMemcpy2DAsync hipMemcpy2DAsync
-#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
-#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
-#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
-#define cudaMemcpyKind hipMemcpyKind
-#define cudaMemset hipMemset
-#define cudaMemsetAsync hipMemsetAsync
-#define cudaMemGetInfo hipMemGetInfo
-#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
-#define cudaSetDevice hipSetDevice
-#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
-#define cudaStreamDestroy hipStreamDestroy
-#define cudaStreamFireAndForget hipStreamFireAndForget
-#define cudaStreamNonBlocking hipStreamNonBlocking
-#define cudaStreamPerThread hipStreamPerThread
-#define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
-#define cudaStream_t hipStream_t
-#define cudaSuccess hipSuccess
-#define __trap() do { abort(); __builtin_unreachable(); } while(0)
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
-#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
-#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
-#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
-#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
-#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
-#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
-#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+#include "vendors/hip.h"
+#elif defined(GGML_USE_MUSA)
+#include "vendors/musa.h"
#else
-#include <cuda_runtime.h>
-#include <cuda.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-
-#if CUDART_VERSION < 11020
-#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
-#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
-#define CUBLAS_COMPUTE_16F CUDA_R_16F
-#define CUBLAS_COMPUTE_32F CUDA_R_32F
-#define cublasComputeType_t cudaDataType_t
-#endif // CUDART_VERSION < 11020
-
+#include "vendors/cuda.h"
#endif // defined(GGML_USE_HIPBLAS)
#define STRINGIZE_IMPL(...) #__VA_ARGS__
@@ -168,7 +72,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
-#if CUDART_VERSION >= 12000
+#if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
static const char * cublas_get_error_str(const cublasStatus_t err) {
return cublasGetStatusString(err);
}
@@ -200,7 +104,7 @@ static const char * cu_get_error_str(CUresult err) {
#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
#endif
-#if CUDART_VERSION >= 11100
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
#else
#define GGML_CUDA_ASSUME(x)
@@ -212,93 +116,7 @@ typedef half2 dfloat2;
#else
typedef float dfloat; // dequantize float
typedef float2 dfloat2;
-#endif //GGML_CUDA_F16
-
-#if defined(GGML_USE_HIPBLAS)
-#define __CUDA_ARCH__ 1300
-
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
- defined(__gfx1150__) || defined(__gfx1151__)
-#define RDNA3
-#endif
-
-#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
- defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
-#define RDNA2
-#endif
-
-#if defined(__gfx1010__) || defined(__gfx1012__)
-#define RDNA1
-#endif
-
-#ifndef __has_builtin
- #define __has_builtin(x) 0
-#endif
-
-typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
-typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
-static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
- const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
- const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-#if __has_builtin(__builtin_elementwise_sub_sat)
- const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
- return reinterpret_cast<const int &>(c);
-#else
- int8x4_t c;
- int16_t tmp;
-#pragma unroll
- for (int i = 0; i < 4; i++) {
- tmp = va[i] - vb[i];
- if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
- if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
- c[i] = tmp;
- }
- return reinterpret_cast<int &>(c);
-#endif // __has_builtin(__builtin_elementwise_sub_sat)
-}
-
-static __device__ __forceinline__ int __vsub4(const int a, const int b) {
- return __vsubss4(a, b);
-}
-
-static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
- const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
- const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
- unsigned int c;
- uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
- for (int i = 0; i < 4; ++i) {
- vc[i] = va[i] == vb[i] ? 0xff : 0x00;
- }
- return c;
-}
-
-static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
- const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
- const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
- unsigned int c;
- uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
-#pragma unroll
- for (int i = 0; i < 4; ++i) {
- vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
- }
- return c;
-}
-
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
- typedef union half2_b32 {
- half2 val;
- int b32;
- } half2_b32_t;
- half2_b32_t tmp;
- tmp.val = var;
- tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
- return tmp.val;
-}
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // GGML_CUDA_F16
#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#define FP16_AVAILABLE
@@ -348,7 +166,7 @@ static __device__ void no_device_code(
#ifdef __CUDA_ARCH__
#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
#else
-#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
#endif // __CUDA_ARCH__
static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -455,7 +273,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
return mask_low | mask_high;
}
-#endif // CUDART_VERSION < 12000
+#endif // CUDART_VERSION < CUDART_HMASK
static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index f76c80dc..70305404 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -162,7 +162,6 @@ static __global__ void dequantize_block_iq2_tn(const void * __restrict__ vx, dst
const int64_t tid = threadIdx.x;
const int64_t n = tid/32;
const int64_t l = tid - 32*n;
- const int64_t is = 8*n + l/16;
const uint8_t q = x[i].qs[32*n + l];
dst_t * y = yy + i*QK_K + 128*n;
diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
index 3db57034..aad34bfe 100644
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -451,7 +451,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
} else {
fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
ggml_type_name(src0->type), ggml_type_name(src1->type));
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
@@ -484,6 +484,6 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
} else {
fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
ggml_type_name(src0->type), ggml_type_name(src1->type));
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
diff --git a/ggml/src/ggml-cuda/dmmv.cu b/ggml/src/ggml-cuda/dmmv.cu
index 174489e0..96a5adef 100644
--- a/ggml/src/ggml-cuda/dmmv.cu
+++ b/ggml/src/ggml-cuda/dmmv.cu
@@ -500,7 +500,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
}
static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
// the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
const dim3 block_nums(block_num_y, 1, 1);
@@ -510,7 +510,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
}
static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -519,7 +519,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
}
static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -528,7 +528,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
}
static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -537,7 +537,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
}
static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -588,7 +588,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
}
static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
- GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+ GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -662,7 +662,7 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
@@ -672,3 +672,12 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
GGML_UNUSED(src1_ncols);
GGML_UNUSED(src1_padded_row_size);
}
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
+ return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
+ src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
+ src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
+ src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
+ src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
+ src0_type == GGML_TYPE_F16;
+}
diff --git a/ggml/src/ggml-cuda/dmmv.cuh b/ggml/src/ggml-cuda/dmmv.cuh
index 4c5ebd47..e727eb97 100644
--- a/ggml/src/ggml-cuda/dmmv.cuh
+++ b/ggml/src/ggml-cuda/dmmv.cuh
@@ -16,3 +16,5 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index f24312dd..950fd93d 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -564,7 +564,7 @@ static void on_no_fattn_vec_case(const int D) {
fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
fprintf(stderr, "By default only f16 KV cache is supported.\n");
fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
} else if (D == 128) {
fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
fprintf(stderr, "Supported combinations:\n");
@@ -572,11 +572,11 @@ static void on_no_fattn_vec_case(const int D) {
fprintf(stderr, " - K == q8_0, V == q8_0, 8.50 BPV\n");
fprintf(stderr, " - K == f16, V == f16, 16.00 BPV\n");
fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
} else {
fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
fprintf(stderr, "Only f16 is supported.\n");
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index c6c35134..1b2fd500 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -287,7 +287,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
} break;
default: {
- GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+ GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
}
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index 15e22f49..f3e68dbf 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -284,7 +284,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
} break;
default: {
- GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+ GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
}
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 38d30b21..29f608b0 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -38,7 +38,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, float>(ctx, dst);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
} else {
@@ -63,7 +63,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
// ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
// break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
@@ -86,7 +86,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
return;
@@ -114,7 +114,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
return;
@@ -141,7 +141,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
diff --git a/ggml/src/ggml-cuda/getrows.cu b/ggml/src/ggml-cuda/getrows.cu
index 55af195f..4c370323 100644
--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -171,8 +171,7 @@ void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
break;
default:
// TODO: k-quants
- fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
- GGML_ASSERT(false);
+ GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
break;
}
}
diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cu b/ggml/src/ggml-cuda/iqk_mmvq.cu
index 29721cdd..c567ad1a 100644
--- a/ggml/src/ggml-cuda/iqk_mmvq.cu
+++ b/ggml/src/ggml-cuda/iqk_mmvq.cu
@@ -466,7 +466,6 @@ __device__ __forceinline__ float vec_dot_iq3_k_q8_1(
const int * q8;
int sumi[4] = {0, 0, 0, 0};
- uint16_t v1, v2;
int v;
for (int i = 0; i < 2; ++i) {
uint32_t vl = ql[2*i+0] | (ql[2*i+1] << 16);
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 84f6387e..78d70cd7 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -84,7 +84,7 @@ void ggml_cuda_op_mul_mat_q(
mul_mat_q_case<GGML_TYPE_IQ4_NL>(ctx, args, stream);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index f08a4758..e8a95744 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -75,7 +75,7 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
case GGML_TYPE_IQ4_NL:
return MMQ_Q8_1_DS_LAYOUT_D4;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
@@ -2898,7 +2898,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
break;
default:
fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best);
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 9eb3fa4f..2586ab7e 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -165,7 +165,7 @@ static void mul_mat_vec_q_cuda(
rows_per_cuda_block = 2;
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
@@ -199,7 +199,7 @@ static void mul_mat_vec_q_cuda(
mul_mat_vec_q<type, 8><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
@@ -454,7 +454,7 @@ void ggml_cuda_op_mul_mat_vec_q(
mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
diff --git a/ggml/src/ggml-cuda/norm.cu b/ggml/src/ggml-cuda/norm.cu
index 30866d51..133e219f 100644
--- a/ggml/src/ggml-cuda/norm.cu
+++ b/ggml/src/ggml-cuda/norm.cu
@@ -142,8 +142,7 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
}
}
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
- static const float eps = 1e-6f;
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
if (group_size < 1024) {
const dim3 block_dims(WARP_SIZE, 1, 1);
group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@@ -196,8 +195,12 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
GGML_ASSERT( dst->type == GGML_TYPE_F32);
int num_groups = dst->op_params[0];
+
+ float eps;
+ memcpy(&eps, dst->op_params + 1, sizeof(float));
+
int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
- group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
+ group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
}
void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu
index aa7f1eff..45408ce8 100644
--- a/ggml/src/ggml-cuda/quantize.cu
+++ b/ggml/src/ggml-cuda/quantize.cu
@@ -163,7 +163,7 @@ void quantize_mmq_q8_1_cuda(
<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
break;
default:
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
break;
}
}
diff --git a/ggml/src/ggml-cuda/rope.cu b/ggml/src/ggml-cuda/rope.cu
index 596fb7c1..99ec1dd9 100644
--- a/ggml/src/ggml-cuda/rope.cu
+++ b/ggml/src/ggml-cuda/rope.cu
@@ -251,7 +251,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
attn_factor, corr_dims, freq_factors, stream
);
} else {
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
} else {
if (src0->type == GGML_TYPE_F32) {
@@ -265,7 +265,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
attn_factor, corr_dims, freq_factors, stream
);
} else {
- GGML_ASSERT(false);
+ GGML_ABORT("fatal error");
}
}
}
diff --git a/ggml/src/ggml-cuda/vendors/cuda.h b/ggml/src/ggml-cuda/vendors/cuda.h
new file mode 100644
index 00000000..db9f6a16
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/cuda.h
@@ -0,0 +1,14 @@
+#pragma once
+
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
new file mode 100644
index 00000000..d0c37725
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -0,0 +1,177 @@
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F HIPBLAS_R_16F
+#define CUDA_R_32F HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define __trap() do { abort(); __builtin_unreachable(); } while(0)
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+ defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+ defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#if defined(__gfx1010__) || defined(__gfx1012__)
+#define RDNA1
+#endif
+
+#ifndef __has_builtin
+ #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+ const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+ const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+ const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+ return reinterpret_cast<const int &>(c);
+#else
+ int8x4_t c;
+ int16_t tmp;
+#pragma unroll
+ for (int i = 0; i < 4; i++) {
+ tmp = va[i] - vb[i];
+ if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+ if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+ c[i] = tmp;
+ }
+ return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+ return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+ const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+ const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+ unsigned int c;
+ uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+ }
+ return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
+ const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+ const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+ unsigned int c;
+ uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+ }
+ return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+ typedef union half2_b32 {
+ half2 val;
+ int b32;
+ } half2_b32_t;
+ half2_b32_t tmp;
+ tmp.val = var;
+ tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+ return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
diff --git a/ggml/src/ggml-cuda/vendors/musa.h b/ggml/src/ggml-cuda/vendors/musa.h
new file mode 100644
index 00000000..e50a103a
--- /dev/null
+++ b/ggml/src/ggml-cuda/vendors/musa.h
@@ -0,0 +1,171 @@
+#pragma once
+
+#include <musa_runtime.h>
+#include <musa.h>
+#include <mublas.h>
+#include <musa_fp16.h>
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F MUBLAS_COMPUTE_32F_FAST_16F
+#define CUBLAS_GEMM_DEFAULT MUBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N MUBLAS_OP_N
+#define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUDA_R_16F MUSA_R_16F
+#define CUDA_R_32F MUSA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#define cublasCreate mublasCreate
+#define cublasDestroy mublasDestroy
+#define cublasGemmEx mublasGemmEx
+#define cublasGemmBatchedEx mublasGemmBatchedEx
+#define cublasGemmStridedBatchedEx mublasGemmStridedBatchedEx
+#define cublasHandle_t mublasHandle_t
+#define cublasSetMathMode mublasSetMathMode
+#define cublasSetStream mublasSetStream
+#define cublasSgemm mublasSgemm
+#define cublasStatus_t mublasStatus_t
+#define cublasGetStatusString mublasStatus_to_string
+#define cudaDataType_t musaDataType_t
+#define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceProp musaDeviceProp
+#define cudaDeviceSynchronize musaDeviceSynchronize
+#define cudaError_t musaError_t
+#define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags musaEventCreateWithFlags
+#define cudaEventDisableTiming musaEventDisableTiming
+#define cudaEventRecord musaEventRecord
+#define cudaEventSynchronize musaEventSynchronize
+#define cudaEvent_t musaEvent_t
+#define cudaEventDestroy musaEventDestroy
+#define cudaFree musaFree
+#define cudaFreeHost musaFreeHost
+#define cudaGetDevice musaGetDevice
+#define cudaGetDeviceCount musaGetDeviceCount
+#define cudaGetDeviceProperties musaGetDeviceProperties
+#define cudaGetErrorString musaGetErrorString
+#define cudaGetLastError musaGetLastError
+#define cudaHostRegister musaHostRegister
+#define cudaHostRegisterPortable musaHostRegisterPortable
+#define cudaHostRegisterReadOnly musaHostRegisterReadOnly
+#define cudaHostUnregister musaHostUnregister
+#define cudaLaunchHostFunc musaLaunchHostFunc
+#define cudaMalloc musaMalloc
+#define cudaMallocHost musaMallocHost
+#define cudaMemcpy musaMemcpy
+#define cudaMemcpyAsync musaMemcpyAsync
+#define cudaMemcpyPeerAsync musaMemcpyPeerAsync
+#define cudaMemcpy2DAsync musaMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice musaMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost musaMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice musaMemcpyHostToDevice
+#define cudaMemcpyKind musaMemcpyKind
+#define cudaMemset musaMemset
+#define cudaMemsetAsync musaMemsetAsync
+#define cudaMemGetInfo musaMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize musaOccupancyMaxPotentialBlockSize
+#define cudaSetDevice musaSetDevice
+#define cudaStreamCreateWithFlags musaStreamCreateWithFlags
+#define cudaStreamDestroy musaStreamDestroy
+#define cudaStreamFireAndForget musaStreamFireAndForget
+#define cudaStreamNonBlocking musaStreamNonBlocking
+#define cudaStreamPerThread musaStreamPerThread
+#define cudaStreamSynchronize musaStreamSynchronize
+#define cudaStreamWaitEvent musaStreamWaitEvent
+#define cudaStream_t musaStream_t
+#define cudaSuccess musaSuccess
+
+// Additional mappings for MUSA virtual memory pool
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED MU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE MU_MEM_ACCESS_FLAGS_PROT_READWRITE
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED MU_MEM_ALLOC_GRANULARITY_RECOMMENDED
+#define CU_MEM_ALLOCATION_TYPE_PINNED MU_MEM_ALLOCATION_TYPE_PINNED
+#define CU_MEM_LOCATION_TYPE_DEVICE MU_MEM_LOCATION_TYPE_DEVICE
+#define CUdevice MUdevice
+#define CUdeviceptr MUdeviceptr
+#define CUmemAccessDesc MUmemAccessDesc
+#define CUmemAllocationProp MUmemAllocationProp
+#define CUmemGenericAllocationHandle MUmemGenericAllocationHandle
+#define cuDeviceGet muDeviceGet
+#define cuDeviceGetAttribute muDeviceGetAttribute
+#define cuMemAddressFree muMemAddressFree
+#define cuMemAddressReserve muMemAddressReserve
+#define cuMemCreate muMemCreate
+#define cuMemGetAllocationGranularity muMemGetAllocationGranularity
+#define cuMemMap muMemMap
+#define cuMemRelease muMemRelease
+#define cuMemSetAccess muMemSetAccess
+#define cuMemUnmap muMemUnmap
+#define cudaFuncAttributeMaxDynamicSharedMemorySize musaFuncAttributeMaxDynamicSharedMemorySize
+#define cudaFuncSetAttribute musaFuncSetAttribute
+#define cudaMemcpy3DPeerParms musaMemcpy3DPeerParms
+#define make_cudaExtent make_musaExtent
+#define make_cudaPitchedPtr make_musaPitchedPtr
+
+// Additional mappings for MUSA graphs
+#define CUDA_SUCCESS MUSA_SUCCESS
+#define CUresult MUresult
+#define cuGetErrorString muGetErrorString
+#define cudaErrorGraphExecUpdateFailure musaErrorGraphExecUpdateFailure
+#define cudaErrorInvalidDeviceFunction musaErrorInvalidDeviceFunction
+#define cudaGraphDestroy musaGraphDestroy
+#define cudaGraphExecDestroy musaGraphExecDestroy
+#define cudaGraphExec_t musaGraphExec_t
+#define cudaGraphExecUpdate musaGraphExecUpdate
+#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphGetNodes musaGraphGetNodes
+#define cudaGraphInstantiate musaGraphInstantiate
+#define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
+#define cudaGraphKernelNodeSetParams musaGraphKernelNodeSetParams
+#define cudaGraphLaunch musaGraphLaunch
+#define cudaGraphNodeGetType musaGraphNodeGetType
+#define cudaGraphNode_t musaGraphNode_t
+#define cudaGraphNodeType musaGraphNodeType
+#define cudaGraphNodeTypeKernel musaGraphNodeTypeKernel
+#define cudaGraph_t musaGraph_t
+#define cudaKernelNodeParams musaKernelNodeParams
+#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamEndCapture musaStreamEndCapture
+
+// XXX: Clang builtins mapping
+#define __vsub4 __vsub4_musa
+#define __vcmpeq4 __vcmpeq4_musa
+#define __vcmpne4 __vcmpne4_musa
+
+#ifndef __has_builtin
+ #define __has_builtin(x) 0
+#endif
+
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+
+static __device__ __forceinline__ int __vsub4_musa(const int a, const int b) {
+ return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4_musa(unsigned int a, unsigned int b) {
+ const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+ const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+ unsigned int c;
+ uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+ }
+ return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4_musa(unsigned int a, unsigned int b) {
+ const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+ const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+ unsigned int c;
+ uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+ }
+ return c;
+}
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-16 11:49:47 +0000