CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

HIP: enable mfma mmq on gfx908 and gfx90a for select datatypes and shapes (#14949)

Browse Source
This commit is contained in:
uvos
2025-07-30 17:38:06 +02:00
committed by GitHub
parent e32a4ec60e
commit ad4a700117
3 changed files with 25 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-cuda/common.cuh
View File

@@ -227,9 +227,9 @@ typedef float2 dfloat2;
#define FP16_MMA_AVAILABLE
#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
#if defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
#if defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
#define AMD_MFMA_AVAILABLE
#endif // defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
#endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#define NEW_MMA_AVAILABLE
@@ -293,10 +293,9 @@ static bool fp32_mma_hardware_available(const int cc) {
return GGML_CUDA_CC_IS_CDNA(cc);
}
// AMD CDNA3 matrix cores.. Will add support for other CDNA generations later.
static bool amd_mfma_available(const int cc) {
#if !defined(GGML_HIP_NO_MMQ_MFMA)
return GGML_CUDA_CC_IS_CDNA3(cc);
return GGML_CUDA_CC_IS_CDNA(cc);
#else
return false;
#endif //!defined(GGML_HIP_NO_MMQ_MFMA)