EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

HIP: Enable Matrix cores for MMQ Kernels, Enable stream-K for CDNA 3 (#14624)

Browse Source 
This commit adds support for MFMA instructions to MMQ. CDNA1/GFX908 CDNA2/GFX90a and CDNA3/GFX942 are supported by the MFMA-enabled code path added by this commit. The code path and stream-k is only enabled on CDNA3 for now as it fails to outperform blas in all cases on the other devices.
Blas is currently only consistently outperformed on CDNA3 due to issues in the amd-provided blas libraries.
This commit also improves the awareness of MMQ towards different warp sizes and as a side effect improves the performance of all quant formats besides q4_0 and q4_1, which regress slightly, on GCN gpus.
This commit is contained in:
deepsek
2025-07-26 18:28:14 -04:00
committed by GitHub
parent 11dd5a44eb
commit 66906cd82a
6 changed files with 1305 additions and 710 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cuda/mmq.cu
View File

@@ -109,7 +109,8 @@ void ggml_cuda_mul_mat_q(
const int64_t s03 = src0->nb[3] / ts_src0;
const int64_t s3 = dst->nb[3] / ts_dst;
const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
|| (GGML_CUDA_CC_IS_AMD(cc) && GGML_CUDA_CC_IS_CDNA3(cc)));
if (!ids) {
const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
@@ -250,8 +251,9 @@ void ggml_cuda_op_mul_mat_q(
// The stream-k decomposition is only faster for recent NVIDIA GPUs.
// Also its fixup needs to allocate a temporary buffer in the memory pool.
// There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
|| (GGML_CUDA_CC_IS_AMD(cc) && GGML_CUDA_CC_IS_CDNA3(cc)))
&& src1_ncols == ne11;
const mmq_args args = {
src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst,
@@ -304,7 +306,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
return false;
}
if (new_mma_available(cc)) {
if (new_mma_available(cc) || amd_mfma_available(cc)) {
return true;
}