public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-22 03:48:19 +02:00
Code Issues Packages Projects Releases Wiki Activity

cuda : use SIMT implicit GEMM conv2d on AMD without WMMA (llama/30325)

Browse Source 
* cuda : use SIMT implicit GEMM conv2d on AMD without WMMA

Assisted-by: Claude Opus

* cuda : update conv2d SIMT comment

* cuda : skip SIMT conv2d on AMD for short reductions

Co-authored-by: thelittlefireman <5165783+thelittlefireman@users.noreply.github.com>

---------

Co-authored-by: thelittlefireman <5165783+thelittlefireman@users.noreply.github.com>
This commit is contained in:
An Long
2026-10-19 09:47:11 +02:00
committed by Daniel Bevenius
co-authored by thelittlefireman
parent 900f40e1a2
commit 8242e50cc6
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/conv2d.cu
+2 -2
View File
@@ -378,8 +378,8 @@ void ggml_cuda_op_conv2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const auto & device = ggml_cuda_info().devices[ctx.device];
const bool use_mma =
turing_mma_available(device.cc) || amd_wmma_available(device.cc) || amd_mfma_available(device.cc);
// MUSA can share the tiling without a native fragment implementation in mma.cuh.
const bool use_simt = GGML_CUDA_CC_IS_MTHREADS(device.cc);
// Devices without a native fragment implementation in mma.cuh can share the tiling.
const bool use_simt = GGML_CUDA_CC_IS_MTHREADS(device.cc) || (GGML_CUDA_CC_IS_AMD(device.cc) && !use_mma && IC * KW * KH >= 64);
const bool pointwise = KW == 1 && KH == 1 && ST_X == 1 && ST_Y == 1 && PD_X == 0 && PD_Y == 0;
const bool use_blas = pointwise && fast_fp16_hardware_available(device.cc);
// Short reductions on small maps do not amortize conversion and launch costs.