ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Sync the kname with instance name (#1989)

Browse Source 
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
This commit is contained in:
rocking
2025-03-20 00:06:45 +08:00
committed by GitHub
parent 7eaedeb36c
commit b819c217e4
5 changed files with 22 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp
View File

@@ -100,10 +100,10 @@ struct FmhaBwdDQDKDVKernel
"r" + _TS_(gbr4::at(ck_tile::number<0>{})) + "x" + _TS_(gbr4::at(ck_tile::number<1>{})) + "x" + _TS_(gbr4::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt0::at(ck_tile::number<0>{})) + "x" + _TS_(gwt0::at(ck_tile::number<1>{})) + "x" + _TS_(gwt0::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt1::at(ck_tile::number<0>{})) + "x" + _TS_(gwt1::at(ck_tile::number<1>{})) + "x" + _TS_(gwt1::at(ck_tile::number<2>{})) + "_" +
("o" + _TS_(kBlockPerCu) + "_") + _SS_(FmhaPipeline::name) + (pn.empty() ? "" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasBiasGrad ? "_dbias" : "") + (kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kHasDropout ? "_dropout" : "" ) +
(kIsStoreRandval ? "_storerandval" : "" ) + (kIsDeterministic ? "_deterministic" : "" );
("o" + _TS_(kBlockPerCu) + "_") + _SS_(FmhaPipeline::name) + (pn.empty() ? "_npad" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("_nbias") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasBiasGrad ? "_dbias" : "_ndbias") + (kHasMask ? "_" + _SS_(FmhaMask::name) : "_nmask") + (kHasDropout ? "_dropout" : "_ndropout" ) +
(kIsStoreRandval ? "_storerandval" : "" ) + (kIsDeterministic ? "_deterministic" : "_ndeterministic" );
#undef _SS_
#undef _TS_
// clang-format on
@@ -1620,7 +1620,7 @@ struct FmhaBwdOGradDotOKernel
return
_SS_("fmha_bwd_dot_do_o_d") + _TS_(kVHeaddim) + "_" + _SS_(t2s<ODataType>::name) +
"_" + (kIsGroupMode ? "group" : "batch") + "_" +
("o" + _TS_(kBlockPerCu)) + (pn.empty() ? "" : "_" + pn);
("o" + _TS_(kBlockPerCu)) + (pn.empty() ? "_npad" : "_" + pn);
#undef _SS_
#undef _TS_
// clang-format on
@@ -1875,8 +1875,8 @@ struct FmhaBwdConvertQGradKernel
return n.empty() ? n : std::string("p") + n; }();
return
_SS_("fmha_bwd_convert_dq_d") + _TS_(kQKHeaddim) + "_" + _SS_(t2s<QGradDataType>::name) +
"_" + (kIsGroupMode ? "group" : "batch") + (kIsDeterministic ? "_deterministic" : "") + "_" +
("o" + _TS_(kBlockPerCu)) + (pn.empty() ? "" : "_" + pn);
"_" + (kIsGroupMode ? "group" : "batch") + "_" + ("o" + _TS_(kBlockPerCu)) + (pn.empty() ? "_npad" : "_" + pn) +
(kIsDeterministic ? "_deterministic" : "_ndeterministic") ;
#undef _SS_
#undef _TS_
// clang-format on

6
include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp
View File

@@ -93,9 +93,9 @@ struct FmhaFwdKernel
"w" + _TS_(g0wt::at(ck_tile::number<0>{})) + "x" + _TS_(g0wt::at(ck_tile::number<1>{})) + "x" + _TS_(g0wt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g1wt::at(ck_tile::number<0>{})) + "x" + _TS_(g1wt::at(ck_tile::number<1>{})) + "x" + _TS_(g1wt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kStoreLSE ? "_lse" : "" ) + (kHasDropout ? "_dropout" : "" ) + (kDoFp8StaticQuant ? "_squant" : "" );
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "_npad" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("_nbias") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "_nmask") + (kStoreLSE ? "_lse" : "_nlse" ) + (kHasDropout ? "_dropout" : "_ndropout" ) + (kDoFp8StaticQuant ? "_squant" : "_nsquant" );
#undef _SS_
#undef _TS_
// clang-format on

6
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
View File

@@ -54,9 +54,9 @@ struct FmhaFwdSplitKVCombineKernel
"b" + _TS_(FmhaPipeline::kN1) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) +
_SS_(FmhaPipeline::name) +
(pn.empty() ? "" : "_" + pn) +
(kStoreLSE ? "_lse" : "" ) +
(kDoFp8StaticQuant ? "_squant" : "" );
(pn.empty() ? "_npad" : "_" + pn) +
(kStoreLSE ? "_lse" : "_nlse" ) +
(kDoFp8StaticQuant ? "_squant" : "_nsquant" );
#undef _SS_
#undef _TS_
// clang-format on

7
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp
View File

@@ -94,9 +94,10 @@ struct FmhaFwdSplitKVKernel
"w" + _TS_(g0wt::at(ck_tile::number<0>{})) + "x" + _TS_(g0wt::at(ck_tile::number<1>{})) + "x" + _TS_(g0wt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g1wt::at(ck_tile::number<0>{})) + "x" + _TS_(g1wt::at(ck_tile::number<1>{})) + "x" + _TS_(g1wt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kStoreLSE ? "_lse" : "" ) + (kDoFp8StaticQuant ? "_squant" : "") + (kIsPagedKV ? "_pagedkv" : "" );
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "_npad" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("_nbias") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "_nmask") + (kStoreLSE ? "_lse" : "_nlse" ) +
(kDoFp8StaticQuant ? "_squant" : "_nsquant") + (kIsPagedKV ? "_pagedkv" : "_npagedkv" );
#undef _SS_
#undef _TS_
// clang-format on