ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-07 08:15:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

f8 mfma issue

Browse Source
This commit is contained in:
aska-0096
2025-04-22 10:59:03 +00:00
parent 4bef60aa57
commit 5366d3415b
5 changed files with 23 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp
View File

@@ -178,7 +178,7 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto mfma_cycle = NPerXDL == 16 ? 32 : 64;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =

8
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View File

@@ -1120,7 +1120,11 @@ struct MfmaSelector
template <>
constexpr auto GetMfma<f8_t, 32, 32>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_32x32x64f8f6f4;
#else
return MfmaInstr::mfma_f32_32x32x16f8f8;
#endif
}
template <>
@@ -1132,7 +1136,11 @@ struct MfmaSelector
template <>
constexpr auto GetMfma<f8_t, 16, 16>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_16x16x128f8f6f4;
#else
return MfmaInstr::mfma_f32_16x16x32f8f8;
#endif
}
template <>