Skip to content

GitLab

Unverified Commit ec959387 authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into ck_tile/fmha_receipt_aiter

parents c1e2fef7 0e5e29c4
Hide whitespace changes
Inline Side-by-side
Showing with 1064 additions and 168 deletions
library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp
View file @ ec959387
......@@ -39,12 +39,13 @@ template <GemmSpecialization GemmSpec,
typename DsLayout = ck::Tuple<>,
typename DsDataType = ck::Tuple<>>
using device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| DsLayout| CLayout|AData| BData| DsData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffleV3R1< Row, Row, DsLayout, Row, F16, F16, DsDataType, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3R1< Row, Row, DsLayout, Row, F16, F16, DsDataType, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3R1< Row, Row, DsLayout, Row, F16, F16, DsDataType, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
......@@ -63,12 +64,13 @@ template <BlockGemmPipelineScheduler BlkGemmPipeSched,
typename DsLayout = ck::Tuple<>,
typename DsDataType = ck::Tuple<>>
using device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| DsLayout| CLayout|AData| BData| DsData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// Latency friendly
DeviceGemm_Xdl_CShuffleV3R1< Row, Row, DsLayout, Row, F16, F16, DsDataType, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3R1< Row, Row, DsLayout, Row, F16, F16, DsDataType, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
......
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp
View file @ ec959387
......@@ -34,12 +34,14 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx950__)
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 128, 16, 2, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v2>
#else
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
......@@ -52,20 +54,23 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances = st
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 32, 8, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 32, 8, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
#endif // defined(__gfx950__)
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
#if defined(__gfx950__)
#else
// Latency friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
......@@ -95,6 +100,7 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances = std
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 8, 4, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 64, 8, 4, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
#endif // defined(__gfx950__)
// clang-format on
>;
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn.hpp
View file @ ec959387
......@@ -34,12 +34,13 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// Compute friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
......@@ -69,13 +70,14 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_instances = st
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_instances = std::tuple<
// clang-format off
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// Latency friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
......
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
View file @ ec959387
......@@ -29,13 +29,14 @@ using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
// clang-format off
//################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F8, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F8, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F8, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
......
library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn.hpp
View file @ ec959387
......@@ -44,11 +44,14 @@ template <typename DsLayout,
typename CDEElementwiseOp,
GemmSpecialization GemmSpec = GemmMNKPadding>
using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances = std::tuple<
// clang-format off
// clang-format off
//###########################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###########################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S<C,D0...,D_N|
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 128, 128, 128, 16, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8,8,8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
......@@ -67,11 +70,13 @@ template <typename DsLayout,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave>
using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances =
std::tuple<
// clang-format off
// clang-format off
//###########################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###########################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###########################################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S<C,D0...,D_N|
#if defined(CK_USE_AMD_MFMA_GFX950)
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// Latency friendly
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 64, 16, 16, 256, 8, 4, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 1, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 128, 16, 32, 256, 8, 4, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
......
library/src/utility/CMakeLists.txt
View file @ ec959387
add_library(utility STATIC
add_library(utility
device_memory.cpp
host_tensor.cpp
convolution_parameter.cpp
......@@ -6,6 +6,10 @@ add_library(utility STATIC
add_library(composable_kernel::utility ALIAS utility)
set_target_properties(utility PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(utility
PROPERTIES
VERSION ${CMAKE_PROJECT_VERSION}
SOVERSION ${CMAKE_PROJECT_VERSION_MAJOR})
target_compile_options(utility PRIVATE ${CMAKE_COMPILER_WARNINGS})
target_include_directories(utility PUBLIC
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
......
profiler/include/profiler/profile_batched_gemm_b_scale_impl.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl_fpAintB_b_scale.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_b_scale.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename BScaleDataType,
typename ComputeDataType,
typename AccDataType,
typename CDataType,
index_t ScaleBlockK,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_batched_gemm_b_scale_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int BatchStrideA,
int BatchStrideB,
int BatchStrideC,
int BatchStrideScaleB,
int BatchSize,
int KBatch,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
}
};
ck::index_t Scale_Stride_BN = ck::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
? ((K + ScaleBlockK - 1) / ScaleBlockK)
: N;
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchSize, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(BatchSize, K, N, StrideB, BatchStrideB, BLayout{}));
Tensor<BDataType> b_g_k_n_permute(
f_host_tensor_descriptor(BatchSize, K, N, StrideB, BatchStrideB, BLayout{}));
Tensor<BScaleDataType> b1_g_k_n(f_host_tensor_descriptor(
BatchSize,
(K + ScaleBlockK - 1) / ScaleBlockK, // K direction group size is ScaleBlockK
N, // N direction group size is 1
Scale_Stride_BN,
BatchStrideScaleB,
BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchSize, M, N, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchSize, M, N, StrideC, BatchStrideC, CLayout{}));
int total_gemm_needed = a_g_m_k.GetElementSpaceSizeInBytes() +
b_g_k_n.GetElementSpaceSizeInBytes() +
b1_g_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1,
std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "b1_g_k_n: " << b1_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_device_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n_permute.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(BScaleDataType) * b1_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b1_device_buf.ToDevice(b1_g_k_n.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmV2BScale<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
BScaleDataType,
CDataType,
1,
ScaleBlockK,
AElementOp,
BElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Run reference GEMM
if(do_verification)
{
Tensor<float> b_g_k_n_dequant({K, N});
float v_b = 0;
for(int bs = 0; bs < BatchSize; bs++)
{
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
ck::pk_i4_t i4x2 = b_g_k_n(bs, k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2.data >> 0) & 0xf;
else
i4 = (i4x2.data >> 4) & 0xf;
i4 = i4 - 8;
v_b = ck::type_convert<float>(i4);
b_g_k_n_dequant(bs, k, n) =
ck::type_convert<float>(v_b) *
ck::type_convert<float>(b1_g_k_n(bs, k / ScaleBlockK, n));
}
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp,
ComputeDataType>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_g_m_k,
b_g_k_n_dequant,
c_g_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
const int KPerBlock = op_ptr->GetKPerBlock();
if(op_ptr->GetPermuteB())
{
int K1 = KPerBlock;
int K0 = K / KPerBlock;
// int K0, N, K1
for(int bs = 0; bs < BatchSize; bs++)
{
for(int j = 0; j < K0; j++)
{
for(int i = 0; i < N; i++)
{
for(int jj = 0; jj < K1; jj++)
{
b_g_k_n_permute(bs * BatchStrideB + j * N * K1 + i * K1 + jj) =
b_g_k_n(bs * BatchStrideB + i * K + (j * K1 + jj));
}
}
}
}
if(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
{
// vector pk_i4x4 permute
for(int bs = 0; bs < BatchSize; bs++)
{
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int input[8];
for(int k = 0; k < 4; k++)
{
int i4x2 = b_g_k_n_permute(bs, j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int hi = input[2];
int lo = input[0];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 0, i) = i4x2;
}
{
int hi = input[6];
int lo = input[4];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 2, i) = i4x2;
}
{
int hi = input[3];
int lo = input[1];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 4, i) = i4x2;
}
{
int hi = input[7];
int lo = input[5];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 6, i) = i4x2;
}
}
}
}
}
}
else
{
b_g_k_n_permute = b_g_k_n;
}
b_device_buf.ToDevice(b_g_k_n_permute.mData.data());
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0)
{
kbatch_list = {KBatch};
}
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
Scale_Stride_BN,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideScaleB,
static_cast<BScaleDataType*>(b1_device_buf.GetDeviceBuffer()),
BatchSize, // Batch count
kbatch_curr, // Split K count
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
// invoker_ptr->Run(argument_ptr.get(),
// StreamConfig{nullptr, false, 0, n_warmup, n_iter});
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0});
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
#if defined CK_ENABLE_FP8
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<CDataType, f8_t>)
{
std::string msg = "Error: Incorrect results!";
double rtol = 1e-1;
double atol = 1e-1;
pass =
pass & ck::utils::check_err(
c_g_m_n_device_result, c_g_m_n_host_result, msg, rtol, atol);
}
else
{
#endif
pass =
pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
#if defined CK_ENABLE_FP8
}
#endif
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
}
}
std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr,
time_kernel,
0,
n_warmup,
n_iter,
rotating_count > 1,
rotating_count});
std::size_t flop = std::size_t(2) * M * N * K * BatchSize;
static constexpr index_t BPackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
return 2;
else
return 1;
}();
std::size_t num_btype = sizeof(ADataType) * M * K +
sizeof(BDataType) * K * N / BPackedSize +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
if constexpr(is_same<CDataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<CDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<CDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<CDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ ec959387
......@@ -59,6 +59,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
......@@ -143,6 +144,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
......
profiler/src/profile_batched_gemm_b_scale.cpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "profiler/profile_batched_gemm_b_scale_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
F16_F16_F16_F8, // 6
F8_F8_BF16, // 7
F16_I4_F16, // 8
};
enum struct BScaleBlockTile
{
K_64, // 0
K_128, // 1
};
#define OP_NAME "batched_gemm_b_scale"
#define OP_DESC "Int4-dequant batched GEMM"
int profile_batched_gemm_b_scale(int argc, char* argv[])
{
if(argc != 17 && argc != 20)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, "
"comp f8; 8: f16@i4)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: B scale block tile (0: 64, 1: 128):\n");
printf("arg5: verification (0: no; 1: yes)\n");
printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg7: print tensor value (0: no; 1: yes)\n");
printf("arg8: time kernel (0=no, 1=yes)\n");
printf("arg9 to 15: M, N, K, StrideA, StrideB, StrideC, BatachCount\n");
printf("arg16: split k into mulitiple batch\n");
printf("optional:\n");
printf("arg17: number of warm-up cycles (default 1)\n");
printf("arg18: number of iterations (default 10)\n");
printf("arg19: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
printf("Start profiling\n");
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const auto B_scale_block = static_cast<BScaleBlockTile>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const int init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const int M = std::stoi(argv[9]);
const int N = std::stoi(argv[10]);
const int K = std::stoi(argv[11]);
const int StrideA = std::stoi(argv[12]);
const int StrideB = std::stoi(argv[13]);
const int StrideC = std::stoi(argv[14]);
const int BatchStrideA = M * N;
const int BatchStrideB = N * K;
const int BatchStrideC = M * N;
const int BatchStrideScaleB =
(K + static_cast<int>(B_scale_block) - 1) / static_cast<int>(B_scale_block) * N;
const int BatchSize = std::stoi(argv[15]);
const int KBatch = std::stoi(argv[16]);
printf("M:%d, N:%d, K:%d, StrideA:%d, StrideB:%d, StrideC:%d, BatchStrideA:%d, "
"BatchStrideB:%d, BatchStrideC:%d, BatchStrideScaleB:%d, BatchSize:%d, KBatch:%d,\n",
M,
N,
K,
StrideA,
StrideB,
StrideC,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideScaleB,
BatchSize,
KBatch);
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 20)
{
n_warmup = std::stoi(argv[17]);
n_iter = std::stoi(argv[18]);
rotating = std::stoull(argv[19]) * 1024 * 1024;
printf("n_warmup:%d, n_iter:%d, rotating:%lu\n", n_warmup, n_iter, rotating);
}
using F32 = float;
using F16 = ck::half_t;
using I4 = ck::pk_i4_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto b_scale_type,
auto comp_type,
auto acc_type,
auto c_type,
auto scale_block_k,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using BScaleDataType = decltype(b_scale_type);
using ComputeDataType = decltype(comp_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass = ck::profiler::profile_batched_gemm_b_scale_impl<ADataType,
BDataType,
BScaleDataType,
ComputeDataType,
AccDataType,
CDataType,
scale_block_k,
ALayout,
BLayout,
CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideScaleB,
BatchSize,
KBatch,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN &&
B_scale_block == BScaleBlockTile::K_128)
{
printf("F16_I4_F16 MK_NK_MN K_128\n");
return profile(
F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<128>{}, Row{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_b_scale);
test/CMakeLists.txt
View file @ ec959387
......@@ -100,7 +100,7 @@ function(add_test_executable TEST_NAME)
if(ARGN MATCHES "_xdl")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
elseif(ARGN MATCHES "_smfmac")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
......@@ -169,26 +169,38 @@ function(add_gtest_executable TEST_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT TEST_TARGETS MATCHES "gfx9" AND source MATCHES "xdl")
message("removing xdl test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT TEST_TARGETS MATCHES "gfx95" AND source MATCHES "mx_")
message("removing microscaling test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT TEST_TARGETS MATCHES "gfx11" AND NOT TEST_TARGETS MATCHES "gfx12" AND source MATCHES "wmma")
message("removing wmma test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(ARGN)
if(ARGN MATCHES "_xdl")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
elseif(ARGN MATCHES "_smfmac")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(ARGN MATCHES "_mx") #only build mx example for gfx950
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
set_source_files_properties(${ARGN} PROPERTIES LANGUAGE HIP)
add_executable(${TEST_NAME} ${ARGN})
......@@ -258,8 +270,11 @@ add_subdirectory(wrapper)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11")
add_subdirectory(wmma_op)
endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL 6 AND CK_HIP_VERSION_MINOR GREATER_EQUAL 2) # smfmac needs ROCm6.2
if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" OR SUPPORTED_GPU_TARGETS MATCHES "gfx950") # smfmac needs ROCm6.2
add_subdirectory(smfmac_op)
endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx950")
add_subdirectory(mx_mfma_op)
endif()
add_subdirectory(position_embedding)
add_subdirectory(scatter_gather)
test/ck_tile/CMakeLists.txt
View file @ ec959387
......@@ -2,3 +2,4 @@ add_subdirectory(image_to_column)
add_subdirectory(gemm)
add_subdirectory(batched_gemm)
add_subdirectory(grouped_gemm)
add_subdirectory(data_type)
test/ck_tile/batched_gemm/test_batched_gemm.cpp
View file @ ec959387
......@@ -17,7 +17,7 @@ using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
// clang-format off
using KernelTypes = ::testing::Types<
// ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType
std::tuple< Row, Row, Row, F16, F16, F32, F16>,
// std::tuple< Row, Row, Row, F16, F16, F32, F16>,
//std::tuple< Col, Row, Row, F16, F16, F32, F16>,
std::tuple< Row, Col, Row, F16, F16, F32, F16>//,
//std::tuple< Col, Col, Row, F16, F16, F32, F16>
......
test/ck_tile/batched_gemm/test_batched_gemm_util.hpp
View file @ ec959387
......@@ -29,12 +29,9 @@ class TestCkTileBatchedGemm : public ::testing::Test
const ck_tile::stream_config& s)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
......@@ -51,32 +48,12 @@ class TestCkTileBatchedGemm : public ::testing::Test
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
......@@ -88,6 +65,20 @@ class TestCkTileBatchedGemm : public ::testing::Test
CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
CodegenGemmPipeline::BlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
using Kernel =
ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
......
test/ck_tile/data_type/CMakeLists.txt 0 → 100644
View file @ ec959387
# Currently ck_tile is only built on gfx9
if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_ck_tile_pk_int4 test_pk_int4.cpp)
endif()
test/ck_tile/data_type/test_pk_int4.cpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include <hip/hip_runtime.h>
#include "ck_tile/core.hpp"
using ck_tile::bf16_t;
using ck_tile::bf16x2_t;
using ck_tile::fp16x2_t;
using ck_tile::fp32x2_t;
using ck_tile::half_t;
using ck_tile::pk_int4_t;
TEST(PackedInt4, ConvertToFloat)
{
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
constexpr float first_input_val = 7.f;
constexpr float second_input_val = -1.f;
#else
constexpr float first_input_val = -1.f;
constexpr float second_input_val = 7.f;
#endif
uint8_t data = 0b11110111; // {-1, 7}
pk_int4_t in = ck_tile::bit_cast<int8_t>(data);
fp32x2_t out = ck_tile::pk_int4_t_to_fp32x2_t(in);
EXPECT_EQ(out.x, first_input_val);
EXPECT_EQ(out.y, second_input_val);
}
TEST(PackedInt4, ConvertToHalf)
{
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
const half_t first_input_val = ck_tile::type_convert<half_t>(7.f);
const half_t second_input_val = ck_tile::type_convert<half_t>(-1.f);
#else
const half_t first_input_val = ck_tile::type_convert<half_t>(-1.f);
const half_t second_input_val = ck_tile::type_convert<half_t>(7.f);
#endif
uint8_t data = 0b11110111; // {-1, 7}
pk_int4_t in = ck_tile::bit_cast<int8_t>(data);
fp16x2_t out = ck_tile::pk_int4_t_to_halfx2_t(in);
EXPECT_EQ(out.x, first_input_val);
EXPECT_EQ(out.y, second_input_val);
}
TEST(PackedInt4, ConvertToBHalf)
{
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
const bf16_t first_input_val = ck_tile::type_convert<bf16_t>(7.f);
const bf16_t second_input_val = ck_tile::type_convert<bf16_t>(-1.f);
#else
const bf16_t first_input_val = ck_tile::type_convert<bf16_t>(-1.f);
const bf16_t second_input_val = ck_tile::type_convert<bf16_t>(7.f);
#endif
uint8_t data = 0b11110111; // {-1, 7}
pk_int4_t in = ck_tile::bit_cast<int8_t>(data);
bf16x2_t out = ck_tile::pk_int4_t_to_bfloat16x2_t(in);
EXPECT_EQ(out.x, first_input_val);
EXPECT_EQ(out.y, second_input_val);
}
test/ck_tile/gemm/test_gemm_pipeline.cpp
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
......@@ -17,22 +17,27 @@ using Intrawave = ck_tile::integral_constant<ck_tile::GemmPipelineScheduler,
using Interwave = ck_tile::integral_constant<ck_tile::GemmPipelineScheduler,
ck_tile::GemmPipelineScheduler::Interwave>;
using Mem = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::Mem>;
using Comp = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::Comp>;
using CompV3 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV3>;
using CompV4 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV4>;
// clang-format off
using KernelTypes = ::testing::Types<
// ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType, GemmPipelineScheduler, PipelineType
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, CompV3>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, CompV4>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, CompV3>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, CompV4>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, CompV3>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, CompV4>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, CompV3>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, CompV4>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Interwave, Mem>
>;
// clang-format on
......
test/ck_tile/gemm/test_gemm_pipeline_ut_cases.inc
View file @ ec959387
......@@ -10,22 +10,43 @@ TYPED_TEST(TestCkTileGemmPipeline, SmallM)
constexpr int K = 320;
for(int M : Ms)
this->Run(M, N, K);
{
if constexpr(std::is_same_v<typename TestFixture::ALayout,
ck_tile::tensor_layout::gemm::ColumnMajor>)
EXPECT_THROW((this->Run(M, N, K)), std::runtime_error);
else
this->Run(M, N, K);
}
}
TYPED_TEST(TestCkTileGemmPipeline, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 1024;
constexpr int K = 320;
constexpr int N = 1024;
constexpr int K = 320;
constexpr int VecLoadSize = 8;
for(int M : Ms)
this->Run(M, N, K);
{
if constexpr(std::is_same_v<typename TestFixture::ALayout,
ck_tile::tensor_layout::gemm::ColumnMajor>)
{
// TODO: Can we anyhow deduce used vector load size?
if(M % VecLoadSize == 0)
this->Run(M, N, K);
else
EXPECT_THROW((this->Run(M, N, K)), std::runtime_error);
}
else
{
this->Run(M, N, K);
}
}
}
TYPED_TEST(TestCkTileGemmPipeline, PaddK)
{
std::vector<int> Ms{127};
std::vector<int> Ms{128};
constexpr int N = 1024;
constexpr int K = 432;
......
test/ck_tile/gemm/test_gemm_pipeline_util.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <sstream>
......@@ -14,8 +14,34 @@
enum struct GemmPipelineType
{
Mem,
Comp
CompV3,
CompV4
};
template <GemmPipelineType PT, typename Problem>
struct GemmPipelineTypeSelector;
template <typename Problem>
struct GemmPipelineTypeSelector<GemmPipelineType::Mem, Problem>
{
using base_pipeline = ck_tile::BaseGemmPipelineAgBgCrMem<Problem>;
using pipeline = ck_tile::GemmPipelineAgBgCrMem<Problem>;
};
template <typename Problem>
struct GemmPipelineTypeSelector<GemmPipelineType::CompV3, Problem>
{
using base_pipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<Problem>;
using pipeline = ck_tile::GemmPipelineAgBgCrCompV3<Problem>;
};
template <typename Problem>
struct GemmPipelineTypeSelector<GemmPipelineType::CompV4, Problem>
{
using base_pipeline = ck_tile::BaseGemmPipelineAgBgCrCompV4<Problem>;
using pipeline = ck_tile::GemmPipelineAgBgCrCompV4<Problem>;
};
template <typename Tuple>
class TestCkTileGemmPipeline : public ::testing::Test
{
......@@ -35,8 +61,8 @@ class TestCkTileGemmPipeline : public ::testing::Test
void invoke_gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
// TODO: This should be parameterized in tests
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
......@@ -51,7 +77,14 @@ class TestCkTileGemmPipeline : public ::testing::Test
constexpr bool kPadN = PadN;
constexpr bool kPadK = PadK;
constexpr int kBlockPerCu = 1;
constexpr bool DoubleSmemBuffer = (PipelineType == GemmPipelineType::CompV4) ? true : false;
// TODO: For now - but this should also be a test parameter
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 1;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
// ===============================================
......@@ -59,20 +92,24 @@ class TestCkTileGemmPipeline : public ::testing::Test
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
kPadN,
kPadK,
DoubleSmemBuffer,
ALayout,
BLayout,
CLayout,
TransposeC>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
using BaseGemmPipeline = std::conditional_t<
PipelineType == GemmPipelineType::Mem,
ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>,
ck_tile::BaseGemmPipelineAgBgCrCompV3<
ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>>;
using BaseGemmPipeline =
typename GemmPipelineTypeSelector<PipelineType, GemmPipelineProblem>::base_pipeline;
const ck_tile::index_t k_grain = args.k_batch * K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * K_Tile;
......@@ -84,26 +121,32 @@ class TestCkTileGemmPipeline : public ::testing::Test
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
Scheduler,
has_hot_loop_v,
tail_number_v>;
using GemmPipeline =
std::conditional_t<PipelineType == GemmPipelineType::Mem,
ck_tile::GemmPipelineAgBgCrMem<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
Scheduler,
has_hot_loop_v,
tail_number_v>>,
ck_tile::GemmPipelineAgBgCrCompV3<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
Scheduler,
has_hot_loop_v,
tail_number_v>>>;
typename GemmPipelineTypeSelector<PipelineType, UniversalGemmProblem>::pipeline;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
GemmPipeline::BlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
UniversalGemmProblem::TransposeC>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
......@@ -129,70 +172,110 @@ class TestCkTileGemmPipeline : public ::testing::Test
if(has_hot_loop)
{
// Tail pipeline One to Seven
if(tail_num == ck_tile::TailNumber::One)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::One>{});
}
else if(tail_num == ck_tile::TailNumber::Full)
if constexpr(PipelineType == GemmPipelineType::CompV3)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
if constexpr(BaseGemmPipeline::PrefetchStages > 2)
{
if(tail_num == ck_tile::TailNumber::Two)
if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Two>{});
ck_tile::TailNumber::Full>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 3)
{
if(tail_num == ck_tile::TailNumber::Three)
else
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Three>{});
std::ostringstream err;
err << "For compute pipeline tail number should always be Full, but have \""
<< tail_num << "\" which is not supported! PrefetchStages: "
<< BaseGemmPipeline::PrefetchStages << "\n File: " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 4)
if constexpr(PipelineType == GemmPipelineType::Mem)
{
if(tail_num == ck_tile::TailNumber::Four)
// Tail pipeline One to Seven
if(tail_num == ck_tile::TailNumber::One)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Four>{});
ck_tile::TailNumber::One>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 5)
{
if(tail_num == ck_tile::TailNumber::Five)
else if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Five>{});
ck_tile::TailNumber::Full>{});
}
if constexpr(BaseGemmPipeline::PrefetchStages > 2)
{
if(tail_num == ck_tile::TailNumber::Two)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Two>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 3)
{
if(tail_num == ck_tile::TailNumber::Three)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Three>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 4)
{
if(tail_num == ck_tile::TailNumber::Four)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Four>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 5)
{
if(tail_num == ck_tile::TailNumber::Five)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Five>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 6)
{
if(tail_num == ck_tile::TailNumber::Six)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Six>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 7)
{
if(tail_num == ck_tile::TailNumber::Seven)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Seven>{});
}
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 6)
if constexpr(PipelineType == GemmPipelineType::CompV4)
{
if(tail_num == ck_tile::TailNumber::Six)
if(tail_num == ck_tile::TailNumber::Three)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Six>{});
ck_tile::TailNumber::Three>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 7)
{
if(tail_num == ck_tile::TailNumber::Seven)
else
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber,
ck_tile::TailNumber::Seven>{});
ck_tile::TailNumber::Two>{});
}
}
}
......@@ -218,7 +301,19 @@ class TestCkTileGemmPipeline : public ::testing::Test
public:
std::vector<int> k_batches_;
void SetUp() override { k_batches_ = {1}; }
void SetUp() override
{
if constexpr(PipelineType == GemmPipelineType::CompV4)
{
// Only do k_batch = 1 when pipeline is CompV4
k_batches_ = {1};
}
else
{
// Otherwise, use k_batch = 1 and 2
k_batches_ = {1, 2};
}
}
template <bool PadM = true, bool PadN = true, bool PadK = true>
void Run(const int M,
......
test/ck_tile/grouped_gemm/test_grouped_gemm.cpp
View file @ ec959387
......@@ -17,7 +17,7 @@ using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
// clang-format off
using KernelTypes = ::testing::Types<
// ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType
std::tuple< Row, Row, Row, F16, F16, F32, F16>,
// std::tuple< Row, Row, Row, F16, F16, F32, F16>,
//std::tuple< Col, Row, Row, F16, F16, F32, F16>,
std::tuple< Row, Col, Row, F16, F16, F32, F16>//,
//std::tuple< Col, Col, Row, F16, F16, F32, F16>
......
test/ck_tile/grouped_gemm/test_grouped_gemm_util.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <sstream>
......@@ -26,12 +26,9 @@ class TestCkTileGroupedGemm : public ::testing::Test
struct GroupedGemKernelParam
{
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const bool kTilePermute = false;
static const ck_tile::index_t kOutputRank = 2;
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const int kBlockPerCu = 1;
static const ck_tile::index_t M_Tile = 128;
......@@ -60,26 +57,6 @@ class TestCkTileGroupedGemm : public ::testing::Test
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
template <typename CLayout>
using GemmEpilogue =
std::conditional_t<std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>,
ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN,
GroupedGemKernelParam::kTilePermute,
GroupedGemKernelParam::kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType,
CDataType,
GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN>>>;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmTraits = ck_tile::TileGemmTraits<GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN,
......@@ -96,17 +73,29 @@ class TestCkTileGroupedGemm : public ::testing::Test
CodegenGemmShape,
CodegenGemmTraits<ALayout, BLayout, CLayout>>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>,
CodegenGemmPolicy>;
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
template <typename ALayout, typename BLayout, typename CLayout>
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
AccDataType,
CDataType,
CLayout,
CodegenGemmPipeline<ALayout, BLayout, CLayout>::BlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GroupedGemKernelParam::M_Warp,
GroupedGemKernelParam::N_Warp,
GroupedGemKernelParam::M_Warp_Tile,
GroupedGemKernelParam::N_Warp_Tile,
GroupedGemKernelParam::K_Warp_Tile,
CodegenPipelineProblem<ALayout, BLayout, CLayout>::TransposeC>>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,
CodegenGemmPipeline<ALayout, BLayout, CLayout>,
GemmEpilogue<CLayout>>;
GemmEpilogue<ALayout, BLayout, CLayout>>;
using grouped_gemm_kargs = ck_tile::GroupedGemmHostArgs;
std::size_t GetWorkspaceSize(const std::vector<grouped_gemm_kargs>& gemm_descs)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment