From 89fadec2bb48a171956ec57f846fb1eeeec3794d Mon Sep 17 00:00:00 2001 From: lalala-sh Date: Mon, 28 Apr 2025 06:30:13 +0000 Subject: [PATCH 01/11] fix moe i4 example bug --- example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp index 3c3ef16198..9e80a2ca35 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp @@ -233,7 +233,7 @@ int main(int argc, char* argv[]) ck::index_t StrideB = K; ck::index_t StrideE = N; constexpr ck::index_t NumDTensor = DsDataType::Size(); - constexpr auto StrideDs = std::array{0, 0, 0}; + constexpr auto StrideDs = std::array{1, 1, 1}; ck::index_t KBatch = 1; @@ -266,7 +266,8 @@ int main(int argc, char* argv[]) Tensor b0_e_n_k(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K})); Tensor b0_preshuffled(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K})); Tensor d0_t_n(HostTensorDescriptor({tokens, N}, {StrideDs[0], 0})); - Tensor d1_e_n(HostTensorDescriptor({experts, N * 2}, {1, StrideDs[1]})); + Tensor d1_e_n( + HostTensorDescriptor({experts, N * 2}, {StrideDs[1] * N * 2, StrideDs[1]})); Tensor d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0})); Tensor e_t_n_host_result(HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1})); Tensor e_t_n_device_result( From cb27e7c77fe807dbdc763feb128bbd127f49b4c8 Mon Sep 17 00:00:00 2001 From: Andriy Roshchenko <107577548+andriy-ca@users.noreply.github.com> Date: Thu, 8 May 2025 13:26:03 -0600 Subject: [PATCH 02/11] Ensure MX GEMM Instances can be Cross-Compiled for Multiple Architectures (#2171) * Re-enable MX GEMM instances * Fix compilation error when building MX GEMM for multiple architectures --- .../gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp | 2 +- .../gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp | 4 ++-- library/src/tensor_operation_instance/gpu/CMakeLists.txt | 5 ++++- .../device_gemm_mx_xdl_bf8_f8_f16_mk_kn_mn.hpp | 4 +--- .../device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn.hpp | 4 +--- .../device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp | 7 +------ .../device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp | 7 +------ 7 files changed, 11 insertions(+), 22 deletions(-) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp index c37af49387..2c34be9007 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp @@ -714,7 +714,7 @@ struct DeviceGemmMX_Xdl_CShuffleV3 : public DeviceGemmMX using device_gemm_mx_xdl_bf8_f8_f16_mk_kn_mn_instances = std::tuple< -// clang-format off + // clang-format off //#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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| //#########################| | | | | Type| | Type| | | | 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__) DeviceGemmMX_Xdl_CShuffleV3< Row, Row, Row, BF8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 64, 16, 128, 16, 4, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Row, Row, BF8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 256, 16, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, false, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Row, Row, BF8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 64, 64, 256, 16, 4, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Row, Row, BF8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 16, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Row, Row, BF8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 16, 32, 512, 16, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<64, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> -#endif // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn.hpp index 5b0c5137b3..d3f74b2907 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn.hpp @@ -39,12 +39,11 @@ static constexpr auto ScaleBlockSize = 32; template using device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn_instances = std::tuple< -// clang-format off + // clang-format off //#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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| //#########################| | | | | Type| | Type| | | | 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__) DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 4, 16, 32, 32, 2, 2, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 16, 256, 128, 4, 16, 16, 16, 1, 4, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, @@ -52,7 +51,6 @@ using device_gemm_mx_xdl_f8_f8_bf16_km_nk_mn_instances = std::tuple< DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 8, 16, 16, 16, 8, 8, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 64, 4, 16, 32, 32, 4, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Col, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 128, 128, 128, 4, 16, 16, 16, 4, 8, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> -#endif // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp index 8e25bcc25f..ac09df7ea2 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp @@ -39,21 +39,16 @@ static constexpr auto ScaleBlockSize = 32; template using device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_instances = std::tuple< -// clang-format off + // clang-format off //#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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| //#########################| | | | | Type| | Type| | | | 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__) DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 256, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> - -//Require verification - //DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> -#endif // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp index 5fefb57257..68363de523 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp @@ -39,21 +39,16 @@ static constexpr auto ScaleBlockSize = 32; template using device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_instances = std::tuple< -// clang-format off + // clang-format off //#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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| //#########################| | | | | Type| | Type| | | | 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__) DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 256, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> - - //Require verification - //DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, -#endif // clang-format on >; From 3448e12609f9c8a623e31e3eadc2617928f2780c Mon Sep 17 00:00:00 2001 From: Illia Silin <98187287+illsilin@users.noreply.github.com> Date: Thu, 8 May 2025 13:29:14 -0700 Subject: [PATCH 03/11] Generate ckProfiler package for gfx942 only. (#2180) * build CI for gfx942 exclusively * run the last stage in a docker with user jenkins * update the image for the last stage * ignore perf_log if not found * archive and store all packages * use ccache for building packages --- Jenkinsfile | 50 ++++++++++++++++++++++++++++++++------------------ 1 file changed, 32 insertions(+), 18 deletions(-) diff --git a/Jenkinsfile b/Jenkinsfile index a9d30d9f71..2ad96ed44b 100644 --- a/Jenkinsfile +++ b/Jenkinsfile @@ -116,7 +116,7 @@ def getDockerImage(Map conf=[:]){ def retimage try { - echo "Pulling down image: ${image}" + echo "Pulling image: ${image}" retimage = docker.image("${image}") withDockerRegistry([ credentialsId: "ck_docker_cred", url: "" ]) { retimage.pull() @@ -335,8 +335,8 @@ def cmake_build(Map conf=[:]){ } } - // Only archive from master or develop - if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) { + // Only archive from develop + if (package_build == true && env.BRANCH_NAME == "develop") { archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true } //check the node gpu architecture @@ -539,13 +539,16 @@ def Build_CK(Map conf=[:]){ """ } dir("build"){ - if (params.RUN_FULL_QA && arch_type == 1 ){ + if (params.RUN_FULL_QA && arch_type == 2 ){ // build deb packages for all gfx9 targets on gfx90a system and prepare to export echo "Build ckProfiler package" sh 'make -j package' - archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb' - sh 'mv composablekernel-ckprofiler_*.deb ckprofiler_0.2.0_amd64.deb' - stash includes: "ckprofiler_0.2.0_amd64.deb", name: "ckprofiler_0.2.0_amd64.deb" + archiveArtifacts artifacts: 'composablekernel*.deb' + sh 'mv composablekernel-ckprofiler_*.deb composablekernel-ckprofiler_1.1.0_amd64.deb' + sh 'mv composablekernel-dev_*.deb composablekernel-dev_1.1.0_amd64.deb' + sh 'mv composablekernel-examples_*.deb composablekernel-examples_1.1.0_amd64.deb' + sh 'mv composablekernel-tests_*.deb composablekernel-tests_1.1.0_amd64.deb' + stash includes: "composablekernel-**.deb", name: "packages" } } // run performance tests, stash the logs, results will be processed on the master node @@ -654,7 +657,8 @@ def Build_CK_and_Reboot(Map conf=[:]){ def process_results(Map conf=[:]){ env.HSA_ENABLE_SDMA=0 checkout scm - def image = getDockerImageName() + //use older image that has user jenkins + def image = "rocm/composable_kernel:ck_ub22.04_rocm6.3" def prefixpath = "/opt/rocm" // Jenkins is complaining about the render group @@ -667,12 +671,17 @@ def process_results(Map conf=[:]){ def retimage gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') { - try { - (retimage, image) = getDockerImage(conf) + try + { + echo "Pulling image: ${image}" + retimage = docker.image("${image}") + withDockerRegistry([ credentialsId: "ck_docker_cred", url: "" ]) { + retimage.pull() + } } - catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){ - echo "The job was cancelled or aborted" - throw e + catch(Exception ex) + { + error "Unable to locate image: ${image}" } } @@ -700,9 +709,14 @@ def process_results(Map conf=[:]){ } if (params.RUN_FULL_QA){ // unstash perf files to master - unstash "ckprofiler_0.2.0_amd64.deb" - sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no ckprofiler_0.2.0_amd64.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/" - unstash "perf_log" + unstash "packages" + sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no composablekernel-*.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/" + try{ + unstash "perf_log" + } + catch(Exception err){ + echo "could not locate perf_log: ${err.getMessage()}." + } try{ unstash "perf_log_gfx11" unstash "perf_log_gfx12" @@ -1114,11 +1128,11 @@ pipeline { agent{ label rocmnode("gfx942") } environment{ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install \ - -DGPU_TARGETS="gfx90a;gfx942" \ + -DGPU_TARGETS="gfx942" \ -DCMAKE_CXX_FLAGS=" -O3 " """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \ cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \ - -DGPU_TARGETS="gfx90a;gfx942" \ + -DGPU_TARGETS="gfx942" \ -DCMAKE_CXX_COMPILER="${build_compiler()}" \ -DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """ } From ef72a4b9bc2e5ddc63d9138cae4e5eba23d35b16 Mon Sep 17 00:00:00 2001 From: Khushbu Agarwal Date: Fri, 9 May 2025 00:18:07 -0700 Subject: [PATCH 04/11] Disable SMFMA for gfx90a (#2182) --- include/ck_tile/ops/gemm/warp/warp_gemm.hpp | 10 +++++++++- tile_engine/ops/gemm/gemm_instance_builder.py | 6 +++++- 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp index c98d46e3a0..5cc5ddc70e 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp @@ -97,12 +97,20 @@ using WarpGemmMfmaF16F16F32M64N4K16 = WarpGemmImpl>; // fp16 2:4 structured sparsity - +#if defined(__gfx94__) || defined(__gfx95__) using WarpGemmSmfmacF16F16F32M32N32K16 = WarpGemmSmfmacImpl>>; using WarpGemmSmfmacF16F16F32M16N16K32 = WarpGemmSmfmacImpl>>; +#else // gfx 90a does not support smfmac +using WarpGemmSmfmacF16F16F32M32N32K16 = WarpGemmImpl, + 2>>; +using WarpGemmSmfmacF16F16F32M16N16K32 = WarpGemmImpl, + 2>>; +#endif // bf16 using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl< diff --git a/tile_engine/ops/gemm/gemm_instance_builder.py b/tile_engine/ops/gemm/gemm_instance_builder.py index 3839523e3d..c00554df8f 100755 --- a/tile_engine/ops/gemm/gemm_instance_builder.py +++ b/tile_engine/ops/gemm/gemm_instance_builder.py @@ -535,7 +535,11 @@ struct GemmDispatcher { ((tile[6] == 32 and tile[7] == 32 and tile[8] == 16) or (tile[6] == 16 and tile[7] == 16 and tile[8] == 32)) content += f""" - run_kernel<{group}::GemmKernel<{tile[0]}, {tile[1]}, {tile[2]}, {tile[3]}, {tile[4]}, {tile[5]}, {tile[6]}, {tile[7]}, {tile[8]}, {BOOL_MAP(sparse)}>>(c_m_n_dev_buf, c_m_n_host_result, c_m_n_dev_result, verify, args, stream);""" +#if defined(__gfx908__) + run_kernel<{group}::GemmKernel<{tile[0]}, {tile[1]}, {tile[2]}, {tile[3]}, {tile[4]}, {tile[5]}, {tile[6]}, {tile[7]}, {tile[8]}, {BOOL_MAP(False)}>>(c_m_n_dev_buf, c_m_n_host_result, c_m_n_dev_result, verify, args, stream); +#else + run_kernel<{group}::GemmKernel<{tile[0]}, {tile[1]}, {tile[2]}, {tile[3]}, {tile[4]}, {tile[5]}, {tile[6]}, {tile[7]}, {tile[8]}, {BOOL_MAP(sparse)}>>(c_m_n_dev_buf, c_m_n_host_result, c_m_n_dev_result, verify, args, stream); +#endif""" content += f""" }} else {{""" for tile in tile_params: From a23390163d604d9f00ea43e920822ac7cfb0884f Mon Sep 17 00:00:00 2001 From: Mingtao Gu <145657261+mtgu0705@users.noreply.github.com> Date: Fri, 9 May 2025 23:25:31 +0800 Subject: [PATCH 05/11] fix moe gemm2 for gfx950 (#2164) Co-authored-by: mtgu0705 --- example/65_gemm_multiply_multiply/CMakeLists.txt | 2 +- example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 5d2a097576..8d51d43c65 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -7,7 +7,7 @@ add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_mul add_example_executable(example_moe_gemm1_xdl_fp8 moe_gemm1_xdl_fp8.cpp) add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp) -list(APPEND gpu_list gfx942) +list(APPEND gpu_list gfx942 gfx950) set(target 0) foreach(gpu IN LISTS GPU_TARGETS) if(gpu IN_LIST gpu_list AND target EQUAL 0) diff --git a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp index b9621cc9b3..3745e3d0af 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp @@ -281,7 +281,7 @@ int main(int argc, char* argv[]) break; case 4: a0_t_k_k.GenerateTensorValue(GeneratorTensor_1{}); - b0_e_n_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_e_n_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); d0_t_n.GenerateTensorValue(GeneratorTensor_1{}); d1_e_n.GenerateTensorValue(GeneratorTensor_1{}); d2_e_n.GenerateTensorValue(GeneratorTensor_1{}); From 6b1a339b6faca7e423fdbce67a40a8fca7445abd Mon Sep 17 00:00:00 2001 From: jefyang1 <146495389+jefyang1@users.noreply.github.com> Date: Fri, 9 May 2025 09:01:06 -0700 Subject: [PATCH 06/11] Fix grouped conv bwd data tests on gfx950 (#2173) --- ...ice_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp index 3028cd7cbc..41f596d160 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp @@ -179,8 +179,7 @@ __global__ void const ComputePtrOffsetOfN compute_ptr_offset_of_n, const index_t num_k_per_block) { -#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__)) // offset base pointer for each work-group const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z); const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / karg.KBatch); @@ -251,8 +250,7 @@ __global__ void const ComputePtrOffsetOfN compute_ptr_offset_of_n, const index_t num_k_per_block) { -#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ - defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__)) const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z); const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / karg.KBatch); const index_t k_idx = From 6fddb5708ca28a84519675ffd3f0ca5c25442706 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= Date: Fri, 9 May 2025 22:52:34 +0200 Subject: [PATCH 07/11] Add grouped conv fwd bias relu instances (#2179) * Add grouped conv fwd bias relu instances * fixes * fix --- ..._conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp | 174 ++++---- ...d_multiple_d_xdl_large_tensor_cshuffle.hpp | 250 ++++++------ .../element/binary_element_wise_operation.hpp | 8 + ...ice_grouped_conv_fwd_xdl_comp_instance.hpp | 131 +++--- .../device_grouped_conv_fwd_xdl_instance.hpp | 383 ++++++++++-------- ...ped_conv_fwd_xdl_large_tensor_instance.hpp | 37 +- ...vice_grouped_conv_fwd_xdl_mem_instance.hpp | 171 ++++---- ...ed_conv_fwd_xdl_merged_groups_instance.hpp | 63 +-- .../grouped_convolution_forward_bias_relu.hpp | 141 +++++++ ...uped_convolution_forward_bias_relu_xdl.inc | 242 +++++++++++ .../CMakeLists.txt | 16 + ...hwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp | 67 +++ ...l_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp | 61 +++ ...c_gkyxc_nhwgk_bf16_comp_part2_instance.cpp | 67 +++ ..._nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp | 60 +++ ...lu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 60 +++ ...tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 41 ++ ...gc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp | 63 +++ ...gc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp | 63 +++ ...groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 80 ++++ .../CMakeLists.txt | 16 + ...dhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp | 127 ++++++ ...hwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp | 58 +++ ...xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 58 +++ ...sor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 41 ++ ..._gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp | 61 +++ ..._gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp | 61 +++ ...ups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 51 +++ ...rofile_grouped_conv_fwd_bias_relu_impl.hpp | 277 +++++++++++++ .../profile_grouped_conv_fwd_impl.hpp | 2 +- test/CMakeLists.txt | 1 + .../CMakeLists.txt | 4 + .../test_grouped_convnd_fwd_bias_relu.cpp | 92 +++++ 33 files changed, 2477 insertions(+), 550 deletions(-) create mode 100644 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu.hpp create mode 100644 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu_xdl.inc create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/CMakeLists.txt create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/CMakeLists.txt create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/comp/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp create mode 100644 profiler/include/profiler/profile_grouped_conv_fwd_bias_relu_impl.hpp create mode 100644 test/grouped_convnd_fwd_bias_relu/CMakeLists.txt create mode 100644 test/grouped_convnd_fwd_bias_relu/test_grouped_convnd_fwd_bias_relu.cpp diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp index a93e6ded96..bebcd72ceb 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp @@ -279,9 +279,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3 static constexpr bool isMultiD = DsDataType::Size() > 0; static constexpr bool isMultiABD = isMultiA || isMultiB || isMultiD; - // multi ABD not supported - static_assert(!isMultiABD, "Multi A, Mutli B and Multi D are not supported"); - static constexpr index_t NumATensor = GetNumABTensors(); static constexpr index_t NumBTensor = GetNumABTensors(); static constexpr index_t NumDTensor = DsDataType::Size(); @@ -1080,91 +1077,96 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3 float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) { float avg_time = 0.f; - - if constexpr(is_NGCHW_GKCYX_NGKHW() || - is_NGCDHW_GKCZYX_NGKDHW()) + if constexpr(!isMultiABD) { - const index_t a_grid_size = - arg.elementwise_block_2_ctile_map_transpose_a_.CalculateGridSize( - arg.a_in_transpose_desc_); - const index_t b_grid_size = - arg.elementwise_block_2_ctile_map_transpose_b_.CalculateGridSize( - arg.b_in_transpose_desc_); + if constexpr(is_NGCHW_GKCYX_NGKHW() || + is_NGCDHW_GKCZYX_NGKDHW()) + { + const index_t a_grid_size = + arg.elementwise_block_2_ctile_map_transpose_a_.CalculateGridSize( + arg.a_in_transpose_desc_); + const index_t b_grid_size = + arg.elementwise_block_2_ctile_map_transpose_b_.CalculateGridSize( + arg.b_in_transpose_desc_); - ADataType* p_a_out_grid = type_convert(arg.p_workspace_); - BDataType* p_b_out_grid = type_convert(arg.p_workspace_) + - arg.GetWorkspaceATensorSizeBytes() / sizeof(BDataType); + ADataType* p_a_out_grid = type_convert(arg.p_workspace_); + BDataType* p_b_out_grid = + type_convert(arg.p_workspace_) + + arg.GetWorkspaceATensorSizeBytes() / sizeof(BDataType); - auto kernel_transpose = kernel_elementwise_dual, - ck::Tuple, - ck::Tuple, - ck::Tuple, - ck::Tuple, - ck::Tuple, - ck::Tuple, - ck::Tuple, - Block2TileMapElementwise, - Block2TileMapElementwise, - element_wise::PassThrough>; + auto kernel_transpose = + kernel_elementwise_dual, + ck::Tuple, + ck::Tuple, + ck::Tuple, + ck::Tuple, + ck::Tuple, + ck::Tuple, + ck::Tuple, + Block2TileMapElementwise, + Block2TileMapElementwise, + element_wise::PassThrough>; - avg_time += launch_and_time_kernel(stream_config, - kernel_transpose, - dim3(a_grid_size + b_grid_size), - dim3(ElementwiseBlocksize), - 0, - make_tuple(arg.a_in_transpose_desc_), - make_tuple(arg.b_in_transpose_desc_), - make_tuple(arg.a_out_transpose_desc_), - make_tuple(arg.b_out_transpose_desc_), - make_tuple(arg.p_a_grid_), - make_tuple(arg.p_b_grid_), - make_tuple(p_a_out_grid), - make_tuple(p_b_out_grid), - arg.elementwise_block_2_ctile_map_transpose_a_, - arg.elementwise_block_2_ctile_map_transpose_b_, - element_wise::PassThrough{}, - a_grid_size); + avg_time += + launch_and_time_kernel(stream_config, + kernel_transpose, + dim3(a_grid_size + b_grid_size), + dim3(ElementwiseBlocksize), + 0, + make_tuple(arg.a_in_transpose_desc_), + make_tuple(arg.b_in_transpose_desc_), + make_tuple(arg.a_out_transpose_desc_), + make_tuple(arg.b_out_transpose_desc_), + make_tuple(arg.p_a_grid_), + make_tuple(arg.p_b_grid_), + make_tuple(p_a_out_grid), + make_tuple(p_b_out_grid), + arg.elementwise_block_2_ctile_map_transpose_a_, + arg.elementwise_block_2_ctile_map_transpose_b_, + element_wise::PassThrough{}, + a_grid_size); + } + + avg_time += RunGemm(arg, stream_config); + + if constexpr(is_NGCHW_GKCYX_NGKHW() || + is_NGCDHW_GKCZYX_NGKDHW()) + { + const index_t grid_size = + arg.elementwise_block_2_ctile_map_transpose_e_.CalculateGridSize( + arg.e_in_transpose_desc_); + + const EDataType* p_e_in_grid = + type_convert(arg.p_workspace_) + + (arg.GetWorkspaceATensorSizeBytes() + arg.GetWorkspaceBTensorSizeBytes()) / + sizeof(EDataType); + + EDataType* p_e_out_grid = arg.p_e_grid_; + + auto kernel_transpose = kernel_elementwise, + ck::Tuple, + ck::Tuple, + ck::Tuple, + Block2TileMapElementwise, + element_wise::PassThrough>; + + avg_time += + launch_and_time_kernel(stream_config, + kernel_transpose, + dim3(grid_size), + dim3(ElementwiseBlocksize), + 0, + make_tuple(arg.e_in_transpose_desc_), + make_tuple(arg.e_out_transpose_desc_), + make_tuple(p_e_in_grid), + make_tuple(p_e_out_grid), + arg.elementwise_block_2_ctile_map_transpose_e_, + element_wise::PassThrough{}); + } } - - avg_time += RunGemm(arg, stream_config); - - if constexpr(is_NGCHW_GKCYX_NGKHW() || - is_NGCDHW_GKCZYX_NGKDHW()) - { - const index_t grid_size = - arg.elementwise_block_2_ctile_map_transpose_e_.CalculateGridSize( - arg.e_in_transpose_desc_); - - const EDataType* p_e_in_grid = - type_convert(arg.p_workspace_) + - (arg.GetWorkspaceATensorSizeBytes() + arg.GetWorkspaceBTensorSizeBytes()) / - sizeof(EDataType); - - EDataType* p_e_out_grid = arg.p_e_grid_; - - auto kernel_transpose = kernel_elementwise, - ck::Tuple, - ck::Tuple, - ck::Tuple, - Block2TileMapElementwise, - element_wise::PassThrough>; - - avg_time += launch_and_time_kernel(stream_config, - kernel_transpose, - dim3(grid_size), - dim3(ElementwiseBlocksize), - 0, - make_tuple(arg.e_in_transpose_desc_), - make_tuple(arg.e_out_transpose_desc_), - make_tuple(p_e_in_grid), - make_tuple(p_e_out_grid), - arg.elementwise_block_2_ctile_map_transpose_e_, - element_wise::PassThrough{}); - } - return avg_time; } @@ -1182,6 +1184,12 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3 const index_t G = arg.b_g_k_c_xs_lengths_[I0]; const index_t K = arg.b_g_k_c_xs_lengths_[I1]; const index_t C = arg.b_g_k_c_xs_lengths_[I2]; + // Move this to runtime check to align Conv instances + // with Conv Multiple D instances + if constexpr(isMultiABD) + { + return false; + } // check device if(get_device_name() == "gfx908") diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp index b2903121b1..3c34d77cc9 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp @@ -1,5 +1,5 @@ // 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 @@ -192,7 +192,6 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor static constexpr index_t NumDTensor = DsDataType::Size(); static constexpr index_t MaxGemmsNum = 32; - static_assert(NumDTensor == 0, "MultiD not supported."); static constexpr auto I0 = Number<0>{}; static constexpr auto I1 = Number<1>{}; @@ -440,89 +439,94 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor input_left_pads_{input_left_pads}, input_right_pads_{input_right_pads} { - // Perform grouped gemm, generate array of tranformer for convolution - Array conv_to_gemm_transformer_arr; - Array a_grid_ptrs; - Array c_grid_ptrs; - - ck::tie(conv_to_gemm_transformer_arr, - a_grid_ptrs, - c_grid_ptrs, - gemms_count_, - is_split_valid_) = - GenerateConvToGemmTransforms( - ConvToGemmFwdTransformerLongIndexT{a_g_n_c_wis_lengths_, - a_g_n_c_wis_strides_, - b_g_k_c_xs_lengths_, - b_g_k_c_xs_strides_, - e_g_n_k_wos_lengths_, - e_g_n_k_wos_strides_, - conv_filter_strides_, - conv_filter_dilations_, - input_left_pads_, - input_right_pads_}, - static_cast(p_a), - static_cast(p_e)); - - grid_size_ = 0; - valid_gemms_count_ = 0; - - if(is_split_valid_) + if constexpr(NumDTensor == 0) { - // Create GemmArg for each gemm(conv) - for(index_t i = 0; i < gemms_count_; i++) + // Perform grouped gemm, generate array of tranformer for convolution + Array conv_to_gemm_transformer_arr; + Array a_grid_ptrs; + Array c_grid_ptrs; + + ck::tie(conv_to_gemm_transformer_arr, + a_grid_ptrs, + c_grid_ptrs, + gemms_count_, + is_split_valid_) = + GenerateConvToGemmTransforms( + ConvToGemmFwdTransformerLongIndexT{a_g_n_c_wis_lengths_, + a_g_n_c_wis_strides_, + b_g_k_c_xs_lengths_, + b_g_k_c_xs_strides_, + e_g_n_k_wos_lengths_, + e_g_n_k_wos_strides_, + conv_filter_strides_, + conv_filter_dilations_, + input_left_pads_, + input_right_pads_}, + static_cast(p_a), + static_cast(p_e)); + + grid_size_ = 0; + valid_gemms_count_ = 0; + + if(is_split_valid_) { - const AGridDesc_M_K a_grid_desc_m_k{DeviceOp::MakeAGridDescriptor_M_K( - conv_to_gemm_transformer_arr[i])}; - const BGridDesc_N_K b_grid_desc_n_k{DeviceOp::MakeBGridDescriptor_N_K( - conv_to_gemm_transformer_arr[i])}; - const auto e_grid_desc_m_n = - DeviceOp::MakeEGridDescriptor_M_N(conv_to_gemm_transformer_arr[i]); - - const auto block_2_etile_map = - GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n); - - const index_t grid_size_grp = - block_2_etile_map.CalculateGridSize(e_grid_desc_m_n); - - const index_t BlockStart = grid_size_; - const index_t BlockEnd = grid_size_ + grid_size_grp; - - grid_size_ += grid_size_grp; - - if(GridwiseGemm::CheckValidity(a_grid_desc_m_k, - b_grid_desc_n_k, - Tuple<>{}, - e_grid_desc_m_n, - block_2_etile_map)) + // Create GemmArg for each gemm(conv) + for(index_t i = 0; i < gemms_count_; i++) { + const AGridDesc_M_K a_grid_desc_m_k{ + DeviceOp::MakeAGridDescriptor_M_K( + conv_to_gemm_transformer_arr[i])}; + const BGridDesc_N_K b_grid_desc_n_k{ + DeviceOp::MakeBGridDescriptor_N_K( + conv_to_gemm_transformer_arr[i])}; + const auto e_grid_desc_m_n = DeviceOp::MakeEGridDescriptor_M_N( + conv_to_gemm_transformer_arr[i]); - gemm_desc_kernel_args_(valid_gemms_count_) = GemmArgs{ - a_grid_ptrs[i], - static_cast(p_b), - c_grid_ptrs[i], - GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k), - GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k), - GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( - e_grid_desc_m_n), - block_2_etile_map, - BlockStart, - BlockEnd}; + const auto block_2_etile_map = + GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n); - valid_gemms_count_++; + const index_t grid_size_grp = + block_2_etile_map.CalculateGridSize(e_grid_desc_m_n); + + const index_t BlockStart = grid_size_; + const index_t BlockEnd = grid_size_ + grid_size_grp; + + grid_size_ += grid_size_grp; + + if(GridwiseGemm::CheckValidity(a_grid_desc_m_k, + b_grid_desc_n_k, + Tuple<>{}, + e_grid_desc_m_n, + block_2_etile_map)) + { + + gemm_desc_kernel_args_(valid_gemms_count_) = GemmArgs{ + a_grid_ptrs[i], + static_cast(p_b), + c_grid_ptrs[i], + GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k), + GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k), + GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + e_grid_desc_m_n), + block_2_etile_map, + BlockStart, + BlockEnd}; + + valid_gemms_count_++; + } } + // N is the same for all convs + conv_N_per_block_ = static_cast(conv_to_gemm_transformer_arr[I0].N_); } - // N is the same for all convs - conv_N_per_block_ = static_cast(conv_to_gemm_transformer_arr[I0].N_); + + // Strides for G and N remain the same + compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0]; + compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0]; + compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0]; + + compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_; + compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_; } - - // Strides for G and N remain the same - compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0]; - compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0]; - compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0]; - - compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_; - compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_; } void Print() const @@ -578,55 +582,63 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor { float Run(const DeviceOp::Argument& arg, const StreamConfig& stream_config = StreamConfig{}) { - if(stream_config.log_level_ > 0) + if constexpr(NumDTensor == 0) { - arg.Print(); - } + if(stream_config.log_level_ > 0) + { + arg.Print(); + } - const index_t num_workgroups_per_Conv_N = - arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_; + const index_t num_workgroups_per_Conv_N = + arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_; - const index_t gdx = arg.grid_size_; - const index_t gdy = arg.num_group_; - const index_t gdz = num_workgroups_per_Conv_N; + const index_t gdx = arg.grid_size_; + const index_t gdy = arg.num_group_; + const index_t gdz = num_workgroups_per_Conv_N; - // K is constant for all gemms - const auto K = arg.gemm_desc_kernel_args_[I0].a_grid_desc_ak0_m_ak1_.GetLength(I0) * - arg.gemm_desc_kernel_args_[I0].a_grid_desc_ak0_m_ak1_.GetLength(I2); + // K is constant for all gemms + const auto K = arg.gemm_desc_kernel_args_[I0].a_grid_desc_ak0_m_ak1_.GetLength(I0) * + arg.gemm_desc_kernel_args_[I0].a_grid_desc_ak0_m_ak1_.GetLength(I2); - auto launch_kernel = [&](auto has_main_k_block_loop) { - constexpr bool has_main_loop = has_main_k_block_loop.value; - const auto kernel = kernel_grouped_conv_fwd_multiple_d_grouped_gemm_xdl_cshuffle< - GridwiseGemm, - MaxGemmsNum, - GemmArgs, - AElementwiseOperation, - BElementwiseOperation, - CDEElementwiseOperation, - ComputePtrOffsetOfStridedBatch, - has_main_loop>; + auto launch_kernel = [&](auto has_main_k_block_loop) { + constexpr bool has_main_loop = has_main_k_block_loop.value; + const auto kernel = + kernel_grouped_conv_fwd_multiple_d_grouped_gemm_xdl_cshuffle< + GridwiseGemm, + MaxGemmsNum, + GemmArgs, + AElementwiseOperation, + BElementwiseOperation, + CDEElementwiseOperation, + ComputePtrOffsetOfStridedBatch, + has_main_loop>; - return launch_and_time_kernel(stream_config, - kernel, - dim3(gdx, gdy, gdz), - dim3(BlockSize), - 0, - arg.gemm_desc_kernel_args_, - arg.gemms_count_, - arg.a_element_op_, - arg.b_element_op_, - arg.cde_element_op_, - arg.compute_ptr_offset_of_groups_, - arg.compute_ptr_offset_of_n_); - }; + return launch_and_time_kernel(stream_config, + kernel, + dim3(gdx, gdy, gdz), + dim3(BlockSize), + 0, + arg.gemm_desc_kernel_args_, + arg.gemms_count_, + arg.a_element_op_, + arg.b_element_op_, + arg.cde_element_op_, + arg.compute_ptr_offset_of_groups_, + arg.compute_ptr_offset_of_n_); + }; - if(GridwiseGemm::CalculateHasMainKBlockLoop(K)) - { - return launch_kernel(integral_constant{}); + if(GridwiseGemm::CalculateHasMainKBlockLoop(K)) + { + return launch_kernel(integral_constant{}); + } + else + { + return launch_kernel(integral_constant{}); + } } else { - return launch_kernel(integral_constant{}); + return 0.f; } } @@ -643,6 +655,12 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor const long_index_t K = arg.b_g_k_c_xs_lengths_[I1]; const long_index_t C = arg.b_g_k_c_xs_lengths_[I2]; + // Move this to runtime check to align Conv instances + // with Conv Multiple D instances + if constexpr(NumDTensor != 0) + { + return false; + } // Check if all descs are valid if(!(arg.is_split_valid_ && arg.gemms_count_ == arg.valid_gemms_count_)) diff --git a/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp index 530876650e..0e58d5acb4 100644 --- a/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp +++ b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp @@ -404,6 +404,14 @@ struct AddRelu y = a > type_convert(0.0f) ? a : type_convert(0.0f); }; + template <> + __host__ __device__ constexpr void + operator()(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const + { + const float a = type_convert(x0) + type_convert(x1); + y = a > type_convert(0.0f) ? a : type_convert(0.0f); + }; + template <> __host__ __device__ constexpr void operator()(int& y, const int& x0, const int8_t& x1) const diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp index 158ed26ec4..17ffa65d1c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp @@ -33,6 +33,7 @@ using Empty_Tuple = ck::Tuple<>; using namespace ck::tensor_layout::convolution; using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddRelu = ck::tensor_operation::element_wise::AddRelu; static constexpr auto ConvFwdDefault = ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; @@ -55,14 +56,16 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_comp_instances_2x = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -71,7 +74,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -79,17 +84,17 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Compute friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3> // clang-format on >; @@ -99,15 +104,17 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_comp_instances_part2 = std::tuple< // clang-format off - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, // AGPR Spill when use permuted lds layout. so, use padding for these two. - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5> // clang-format on >; @@ -117,14 +124,16 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_comp_instances_2x = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -133,14 +142,16 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 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> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 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> // clang-format on >; @@ -150,22 +161,24 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_comp_instances_part2 = std::tuple< // clang-format off - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, // AGPR Spill when use permuted lds layout. so, use padding for these two. - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -174,17 +187,19 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f32_comp_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<4, 64, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<4, 64, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -194,14 +209,16 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_comp_instances_2x = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -210,14 +227,16 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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> // clang-format on >; @@ -227,18 +246,20 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_comp_instances_part2 = std::tuple< // clang-format off - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v5>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, // AGPR Spill when use permuted lds layout. so, use padding for these two. - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::Interwave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; } // namespace instance diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp index f5397308dc..df24b4cbcb 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp @@ -33,6 +33,7 @@ using Empty_Tuple = ck::Tuple<>; using namespace ck::tensor_layout::convolution; using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddRelu = ck::tensor_operation::element_wise::AddRelu; static constexpr auto ConvFwdDefault = ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; @@ -51,7 +52,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_generic_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -59,7 +62,7 @@ using device_grouped_conv_fwd_xdl_bf16_generic_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -68,7 +71,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -76,24 +81,24 @@ using device_grouped_conv_fwd_xdl_bf16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -102,17 +107,19 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_16x16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8> // clang-format on >; @@ -121,7 +128,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_generic_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -129,7 +138,7 @@ using device_grouped_conv_fwd_xdl_f16_generic_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -138,7 +147,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -146,24 +157,24 @@ using device_grouped_conv_fwd_xdl_f16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -172,17 +183,19 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_16x16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8> // clang-format on >; @@ -191,7 +204,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f32_generic_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -199,7 +214,7 @@ using device_grouped_conv_fwd_xdl_f32_generic_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1> // clang-format on >; @@ -208,7 +223,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f32_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -216,24 +233,24 @@ using device_grouped_conv_fwd_xdl_f32_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> // clang-format on >; @@ -242,7 +259,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f32_16x16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -250,9 +269,9 @@ using device_grouped_conv_fwd_xdl_f32_16x16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4> // clang-format on >; @@ -261,7 +280,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_generic_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -269,7 +290,7 @@ using device_grouped_conv_fwd_xdl_int8_generic_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -278,7 +299,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -286,24 +309,24 @@ using device_grouped_conv_fwd_xdl_int8_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -312,7 +335,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_comp_f8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| ComputeType| @@ -321,24 +346,24 @@ using device_grouped_conv_fwd_xdl_f16_comp_f8_instances = std::tuple< //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | #ifdef CK_ENABLE_FP8 // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8> #endif // clang-format on >; @@ -348,7 +373,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| ComputeType| @@ -357,24 +384,24 @@ using device_grouped_conv_fwd_xdl_f8_instances = std::tuple< //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | #ifdef CK_ENABLE_FP8 // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8> #endif // clang-format on >; @@ -384,7 +411,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| ComputeType| @@ -393,24 +422,24 @@ using device_grouped_conv_fwd_xdl_bf8_instances = std::tuple< //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | #ifdef CK_ENABLE_BF8 // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8> #endif // clang-format on >; @@ -420,7 +449,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f8_bf8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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|AComputeType|BComputeType| @@ -429,24 +460,24 @@ using device_grouped_conv_fwd_xdl_f8_bf8_instances = std::tuple< //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | #if(defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8)) // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, BF8>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, BF8> #endif // clang-format on >; @@ -456,7 +487,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf8_f8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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|AComputeType|BComputeType| @@ -465,24 +498,24 @@ using device_grouped_conv_fwd_xdl_bf8_f8_instances = std::tuple< //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | #if(defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8)) // generic instance - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>, // instances for small conv.K and conv.C - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> #endif // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp index 0a85cde3bc..6bb6d255f3 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp @@ -25,6 +25,7 @@ using Empty_Tuple = ck::Tuple<>; using namespace ck::tensor_layout::convolution; using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddRelu = ck::tensor_operation::element_wise::AddRelu; static constexpr auto ConvFwdDefault = ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; @@ -36,7 +37,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_large_tensor_bf16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -44,10 +47,10 @@ using device_grouped_conv_fwd_xdl_large_tensor_bf16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> // clang-format on >; @@ -56,7 +59,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_large_tensor_f16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -64,10 +69,10 @@ using device_grouped_conv_fwd_xdl_large_tensor_f16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> // clang-format on >; @@ -76,7 +81,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_large_tensor_f32_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -84,9 +91,9 @@ using device_grouped_conv_fwd_xdl_large_tensor_f32_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4> // clang-format on >; @@ -95,7 +102,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_large_tensor_int8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -103,9 +112,9 @@ using device_grouped_conv_fwd_xdl_large_tensor_int8_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> // clang-format on >; } // namespace instance diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp index 1f381af08c..195367ffd7 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" @@ -33,6 +33,7 @@ using Empty_Tuple = ck::Tuple<>; using namespace ck::tensor_layout::convolution; using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddRelu = ck::tensor_operation::element_wise::AddRelu; static constexpr auto ConvFwdDefault = ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; @@ -52,7 +53,9 @@ template + BlockGemmPipelineScheduler BlkGemmPipeSched, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_bf16_mem_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -60,27 +63,27 @@ using device_grouped_conv_fwd_xdl_bf16_mem_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, // Memory friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -90,34 +93,36 @@ template + BlockGemmPipelineScheduler BlkGemmPipeSched, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f16_mem_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, // Memory friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -127,30 +132,32 @@ template + BlockGemmPipelineScheduler BlkGemmPipeSched, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_f32_mem_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, // Memory friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -160,34 +167,36 @@ template + BlockGemmPipelineScheduler BlkGemmPipeSched, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_int8_mem_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, // Memory friendly - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, 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, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp index 153cc61b09..182c785978 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp @@ -1,5 +1,5 @@ // 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 "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" @@ -25,6 +25,7 @@ using Empty_Tuple = ck::Tuple<>; using namespace ck::tensor_layout::convolution; using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using AddRelu = ck::tensor_operation::element_wise::AddRelu; static constexpr auto ConvFwdDefault = ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; @@ -38,7 +39,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| ACompute| BCompute| BlockGemm| NumGroups| @@ -46,9 +49,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| | | Scheduler| | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Instances with NumGroupsPerBatch > 1 - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32> // clang-format on >; @@ -58,16 +61,18 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances_2x = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| ACompute| BCompute| BlockGemm| NumGroups| //########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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| Type| Type| Pipeline| ToMerge| //########################################| | | | | | | | | | | | 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| | | Scheduler| | //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32> // clang-format on >; @@ -76,7 +81,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -84,9 +91,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Instances with NumGroupsPerBatch > 1 - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32> // clang-format on >; @@ -96,7 +103,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_f16_instances_2x = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -104,9 +113,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_f16_instances_2x = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Instances with NumGroupsPerBatch > 1 - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32> // clang-format on >; @@ -115,7 +124,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -123,9 +134,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Instances with NumGroupsPerBatch > 1 - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 32> // clang-format on >; @@ -134,7 +145,9 @@ template + ConvolutionForwardSpecialization ConvSpec, + typename DsDataTypes = Tuple<>, + typename OutElementOp = PassThrough> using device_grouped_conv_fwd_xdl_merged_groups_int8_instances = std::tuple< // clang-format off //########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| 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| @@ -142,9 +155,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_int8_instances = std::tuple< //########################################| | | | | | | | | | | | 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| //########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Instances with NumGroupsPerBatch > 1 - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 8>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 16>, - DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 32> + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 8>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 16>, + DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 32> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu.hpp new file mode 100644 index 0000000000..d873edadba --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu.hpp @@ -0,0 +1,141 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.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/device_operation_instance_factory.hpp" + +#ifdef CK_USE_XDL +#include "grouped_convolution_forward_bias_relu_xdl.inc" +#endif + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +struct DeviceOperationInstanceFactory> +{ + using DeviceOp = + DeviceGroupedConvFwdMultipleABD; + + static auto GetInstances() + { + std::vector> op_ptrs; + +#ifdef CK_USE_XDL + // layout NHWGC/GKYXC/NHWGK + if constexpr(NumDimSpatial == 2 && is_same_v && + is_same_v && is_same_v) + { +#ifdef CK_ENABLE_BF16 + if constexpr(is_same_v && + is_same_v && + is_same_v && + is_same_v && + is_same_v) + { + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances( + op_ptrs); + add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances( + op_ptrs); + } +#endif + } + // layout NDHWGC/GKZYXC/NDHWGK + if constexpr(NumDimSpatial == 3 && is_same_v && + is_same_v && is_same_v) + { +#ifdef CK_ENABLE_BF16 + if constexpr(is_same_v && + is_same_v && + is_same_v && + is_same_v && + is_same_v) + { + add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances( + op_ptrs); + add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances( + op_ptrs); + } +#endif + } +#endif // CK_USE_XDL + + return op_ptrs; + } +}; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu_xdl.inc b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu_xdl.inc new file mode 100644 index 0000000000..1935f123a8 --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_relu_xdl.inc @@ -0,0 +1,242 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +#ifdef CK_ENABLE_BF16 + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances); + +#endif + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/CMakeLists.txt new file mode 100644 index 0000000000..98b0b1c4cb --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/CMakeLists.txt @@ -0,0 +1,16 @@ +# ONLY XDL_KERNELS +add_instance_library(device_grouped_conv2d_fwd_bias_relu_instance + xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp + xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp + + xdl/large_tensor/device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp + + xdl/merged_groups/device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp + + xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp + xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp + + xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp + xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp + xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp +) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp new file mode 100644 index 0000000000..75acd604ee --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp @@ -0,0 +1,67 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + if(ck::get_device_name() == "gfx950") + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); + } +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp new file mode 100644 index 0000000000..69a8a4bd9d --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp @@ -0,0 +1,61 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp new file mode 100644 index 0000000000..043c724e4a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/comp/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp @@ -0,0 +1,67 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + if(ck::get_device_name() != "gfx950") + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); + } +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp new file mode 100644 index 0000000000..c58631e169 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instance.cpp @@ -0,0 +1,60 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_16x16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..cd80f2875f --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -0,0 +1,60 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..a6286b55e8 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp new file mode 100644 index 0000000000..0736325b05 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp @@ -0,0 +1,63 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Interwave, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Interwave, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Interwave, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp new file mode 100644 index 0000000000..0d35ab1b05 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/mem/device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp @@ -0,0 +1,63 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Intrawave, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1P0, + Intrawave, + Tuple, + AddRelu>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd1x1S1P0, + Intrawave, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..253e8b196e --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -0,0 +1,80 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_bias_relu_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances( + std::vector, + NHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + if(ck::get_device_name() == "gfx950") + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances_2x<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances_2x<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd3x3, + Tuple, + AddRelu>{}); + } + else + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2, + NHWGC, + GKYXC, + Tuple, + NHWGK, + ConvFwd3x3, + Tuple, + AddRelu>{}); + } +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/CMakeLists.txt new file mode 100644 index 0000000000..afdddfec70 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/CMakeLists.txt @@ -0,0 +1,16 @@ +# ONLY XDL_KERNELS +set(GROUPED_CONV3D_FWD + xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp + xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp + + xdl/large_tensor/device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp + + xdl/merged_groups/device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp + + xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp + xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp + + xdl/comp/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp +) + +add_instance_library(device_grouped_conv3d_fwd_bias_relu_instance ${GROUPED_CONV3D_FWD}) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/comp/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/comp/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp new file mode 100644 index 0000000000..9819f0ea0b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/comp/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instance.cpp @@ -0,0 +1,127 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); + + if(ck::get_device_name() != "gfx950") + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); + } + + if(ck::get_device_name() == "gfx950") + { + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); + } +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp new file mode 100644 index 0000000000..dc3fc7a4bf --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instance.cpp @@ -0,0 +1,58 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_16x16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..a9a8ff8459 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -0,0 +1,58 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..e58e879973 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/large_tensor/device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp new file mode 100644 index 0000000000..e76052c6e0 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp @@ -0,0 +1,61 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Interwave, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Interwave, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Interwave, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp new file mode 100644 index 0000000000..0593f3f46a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/mem/device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instance.cpp @@ -0,0 +1,61 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Intrawave, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1P0, + Intrawave, + Tuple, + AddRelu>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd1x1S1P0, + Intrawave, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..6552f26f88 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bias_relu/xdl/merged_groups/device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -0,0 +1,51 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_grouped_conv3d_fwd_bias_relu_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances( + std::vector, + NDHWGK, + BF16, + BF16, + Tuple, + BF16, + PassThrough, + PassThrough, + AddRelu>>>& instances) +{ + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwdDefault, + Tuple, + AddRelu>{}); + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3, + NDHWGC, + GKZYXC, + Tuple, + NDHWGK, + ConvFwd3x3, + Tuple, + AddRelu>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/include/profiler/profile_grouped_conv_fwd_bias_relu_impl.hpp b/profiler/include/profiler/profile_grouped_conv_fwd_bias_relu_impl.hpp new file mode 100644 index 0000000000..9d38263d4e --- /dev/null +++ b/profiler/include/profiler/profile_grouped_conv_fwd_bias_relu_impl.hpp @@ -0,0 +1,277 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include + +#include "ck/ck.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/grouped_convolution_forward_bias_relu.hpp" + +#include "ck/library/utility/algorithm.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/convolution_parameter.hpp" +#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp" + +namespace ck { +namespace profiler { + +template +bool profile_grouped_conv_fwd_bias_relu_impl(int do_verification, + int init_method, + bool do_log, + bool time_kernel, + const ck::utils::conv::ConvParam& conv_param) +{ + using InElementOp = ck::tensor_operation::element_wise::PassThrough; + using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; + using OutElementOp = ck::tensor_operation::element_wise::AddRelu; + + const auto in_element_op = InElementOp{}; + const auto wei_element_op = WeiElementOp{}; + const auto out_element_op = OutElementOp{}; + + const auto in_g_n_c_wis_desc = + ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed(conv_param); + + const auto wei_g_k_c_xs_desc = + ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed(conv_param); + + const auto out_g_n_k_wos_desc = + ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed(conv_param); + + std::array a_g_n_c_wis_lengths{}; + std::array a_g_n_c_wis_strides{}; + std::array b_g_k_c_xs_lengths{}; + std::array b_g_k_c_xs_strides{}; + std::array e_g_n_k_wos_lengths{}; + std::array e_g_n_k_wos_strides{}; + std::array conv_filter_strides{}; + std::array conv_filter_dilations{}; + std::array input_left_pads{}; + std::array input_right_pads{}; + + auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); }; + + copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths); + copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides); + copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths); + copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides); + copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths); + copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides); + copy(conv_param.conv_filter_strides_, conv_filter_strides); + copy(conv_param.conv_filter_dilations_, conv_filter_dilations); + copy(conv_param.input_left_pads_, input_left_pads); + copy(conv_param.input_right_pads_, input_right_pads); + + Tensor input(in_g_n_c_wis_desc); + Tensor weight(wei_g_k_c_xs_desc); + Tensor host_output(out_g_n_k_wos_desc); + Tensor device_output(out_g_n_k_wos_desc); + Tensor bias(out_g_n_k_wos_desc); + + std::cout << "input: " << input.mDesc << std::endl; + std::cout << "weight: " << weight.mDesc << std::endl; + std::cout << "output: " << host_output.mDesc << std::endl; + std::cout << "bias: " << bias.mDesc << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + input.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + weight.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + bias.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + break; + default: + input.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + weight.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + bias.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + } + + DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize()); + DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize()); + DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize()); + DeviceMem bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpaceSize()); + + in_device_buf.ToDevice(input.mData.data()); + wei_device_buf.ToDevice(weight.mData.data()); + bias_device_buf.ToDevice(bias.mData.data()); + + // run reference op + if(do_verification) + { + auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd{}; + + std::array, 1> d_tensors = {bias}; + auto ref_invoker = ref_conv.MakeInvoker(); + auto ref_argument = ref_conv.MakeArgument(input, + weight, + host_output, + conv_param.conv_filter_strides_, + conv_param.conv_filter_dilations_, + conv_param.input_left_pads_, + conv_param.input_right_pads_, + in_element_op, + wei_element_op, + out_element_op, + {}, + {}, + d_tensors); + + // init host output to zero + host_output.SetZero(); + + ref_invoker.Run(ref_argument); + } + + std::string best_op_name; + float best_avg_time = 0; + float best_tflops = 0; + float best_gb_per_sec = 0; + + // profile device op instances + bool pass = true; + + auto run_impl = [&](auto& op_ptr, auto& argument_ptr) { + // workspace_sz will be equal to 0 for other layout than NGCHW + const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get()); + DeviceMem workspace_dev(workspace_sz); + op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer()); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + // re-init output to zero before profiling next kernel + out_device_buf.SetZero(); + + std::string op_name = op_ptr->GetTypeString(); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + float avg_time = + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); + + std::size_t flop = conv_param.GetFlops(); + std::size_t num_btype = conv_param.GetByte(); + + float tflops = static_cast(flop) / 1.E9 / avg_time; + + float gb_per_sec = num_btype / 1.E6 / avg_time; + + std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, " + << gb_per_sec << " GB/s, " << op_name << std::endl; + + if(tflops > best_tflops) + { + best_op_name = op_name; + best_tflops = tflops; + best_avg_time = avg_time; + best_gb_per_sec = gb_per_sec; + } + + if(do_verification) + { + out_device_buf.FromDevice(device_output.mData.data()); + + pass = pass & ck::utils::check_err(device_output, host_output); + + if(do_log) + { + LogRangeAsType(std::cout << "input : ", input.mData, ",") << std::endl; + LogRangeAsType(std::cout << "weight: ", weight.mData, ",") << std::endl; + LogRangeAsType(std::cout << "host_output : ", host_output.mData, ",") + << std::endl; + LogRangeAsType(std::cout << "device_output: ", device_output.mData, ",") + << std::endl; + } + } + } + else + { + std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl; + } + }; + + using DeviceOp = + ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD, + OutLayout, + InDataType, + WeiDataType, + ck::Tuple, + OutDataType, + InElementOp, + WeiElementOp, + OutElementOp, + AComputeType, + BComputeType>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl; + + for(auto& op_ptr : op_ptrs) + { + auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(), + wei_device_buf.GetDeviceBuffer(), + {bias_device_buf.GetDeviceBuffer()}, + out_device_buf.GetDeviceBuffer(), + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + {e_g_n_k_wos_lengths}, + {e_g_n_k_wos_strides}, + e_g_n_k_wos_lengths, + e_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + in_element_op, + wei_element_op, + out_element_op); + + run_impl(op_ptr, argument_ptr); + } + + std::cout << "Best configuration parameters:" + << "\nname: " << best_op_name << "\navg_time: " << best_avg_time + << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl; + + return pass; +} + +} // namespace profiler +} // namespace ck diff --git a/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp index 4bfbca5437..dfa6bc1edd 100644 --- a/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp +++ b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 6bde1140d9..69ffb94488 100755 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -255,6 +255,7 @@ add_subdirectory(reduce) add_subdirectory(convnd_fwd) add_subdirectory(convnd_bwd_data) add_subdirectory(grouped_convnd_fwd) +add_subdirectory(grouped_convnd_fwd_bias_relu) add_subdirectory(grouped_convnd_bwd_weight) add_subdirectory(block_to_ctile_map) add_subdirectory(softmax) diff --git a/test/grouped_convnd_fwd_bias_relu/CMakeLists.txt b/test/grouped_convnd_fwd_bias_relu/CMakeLists.txt new file mode 100644 index 0000000000..680a92b19c --- /dev/null +++ b/test/grouped_convnd_fwd_bias_relu/CMakeLists.txt @@ -0,0 +1,4 @@ +if(GPU_TARGETS MATCHES "gfx9") + add_gtest_executable(test_grouped_convnd_fwd_bias_relu test_grouped_convnd_fwd_bias_relu.cpp) + target_link_libraries(test_grouped_convnd_fwd_bias_relu PRIVATE utility device_grouped_conv2d_fwd_bias_relu_instance device_grouped_conv3d_fwd_bias_relu_instance) +endif() diff --git a/test/grouped_convnd_fwd_bias_relu/test_grouped_convnd_fwd_bias_relu.cpp b/test/grouped_convnd_fwd_bias_relu/test_grouped_convnd_fwd_bias_relu.cpp new file mode 100644 index 0000000000..c508235d9c --- /dev/null +++ b/test/grouped_convnd_fwd_bias_relu/test_grouped_convnd_fwd_bias_relu.cpp @@ -0,0 +1,92 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include +#include + +#include "profiler/profile_grouped_conv_fwd_bias_relu_impl.hpp" + +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +using AddRelu = ck::tensor_operation::element_wise::AddRelu; + +template +class TestGroupedConvndFwd : public ::testing::Test +{ + protected: + using DataType = std::tuple_element_t<0, Tuple>; + using InLayout = std::tuple_element_t<1, Tuple>; + using WeiLayout = std::tuple_element_t<2, Tuple>; + using OutLayout = std::tuple_element_t<3, Tuple>; + using IndexType = ck::index_t; + + std::vector conv_params; + + template + void Run() + { + EXPECT_FALSE(conv_params.empty()); + bool pass = true; + for(auto& param : conv_params) + { + pass = pass && ck::profiler::profile_grouped_conv_fwd_bias_relu_impl( + true, // do_verification + 1, // init_method: integer value + false, // do_log + false, // time_kernel + param); + } + EXPECT_TRUE(pass); + } +}; + +using namespace ck::tensor_layout::convolution; + +using KernelTypes2d = ::testing::Types>; + +using KernelTypes3d = ::testing::Types>; + +template +class TestGroupedConvndFwd2d : public TestGroupedConvndFwd +{ +}; + +template +class TestGroupedConvndFwd3d : public TestGroupedConvndFwd +{ +}; + +TYPED_TEST_SUITE(TestGroupedConvndFwd2d, KernelTypes2d); +TYPED_TEST_SUITE(TestGroupedConvndFwd3d, KernelTypes3d); + +TYPED_TEST(TestGroupedConvndFwd2d, Test2D) +{ + this->conv_params.clear(); + this->conv_params.push_back( + {2, 2, 32, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}}); + this->conv_params.push_back( + {2, 2, 32, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); + this->template Run<2>(); +} + +TYPED_TEST(TestGroupedConvndFwd3d, Test3D) +{ + this->conv_params.clear(); + this->conv_params.push_back( + {3, 2, 32, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}}); + this->conv_params.push_back( + {3, 2, 32, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); + this->template Run<3>(); +} From d8faf1c6a161ddcee98e9dfca3cc00941eec9f61 Mon Sep 17 00:00:00 2001 From: Khushbu Agarwal Date: Sat, 10 May 2025 22:40:05 -0700 Subject: [PATCH 08/11] Support for swizzle and transpose for MFMA_16x16x32_F16/BF16 (#2172) * Changes for updating tile distribution for shuffle and transpose * Fixed swizzle and transpose, removed comments * clang formatted * Adding support for bf16 type * Addressing review comments --- include/ck_tile/ops/gemm/warp/warp_gemm.hpp | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp index 5cc5ddc70e..5ed97dc05c 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp @@ -77,6 +77,18 @@ using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution = 2>>; #endif +#if defined(__gfx950__) +using WarpGemmMfmaF16F16F32M16N16K32SwizzleBTransposedCDistribution = + WarpGemmImpl, + 1>>; + +using WarpGemmMfmaBf16Bf16F32M16N16K32SwizzleBTransposedCDistribution = + WarpGemmImpl, + 1>>; +#endif + #if defined(__gfx950__) using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution = WarpGemmImpl Date: Mon, 12 May 2025 05:08:58 +0000 Subject: [PATCH 09/11] update moe --- ...peline_xdlops_b_preshuffle_gufusion_v3.hpp | 973 ++++++++++++++++++ ...e_gemm_pipeline_xdlops_b_preshuffle_v3.hpp | 787 +++++++------- .../gpu/device/impl/device_moe_gemm.hpp | 175 +++- .../gpu/grid/gridwise_moe_gemm.hpp | 317 +++++- 4 files changed, 1765 insertions(+), 487 deletions(-) create mode 100644 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp new file mode 100644 index 0000000000..c2fad6132f --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp @@ -0,0 +1,973 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp" + +namespace ck { + +// Compute optimized pipeline +// GlobalPrefetchStages: 2 +// LocalPreFillStages: 1 +// LocalPreFetchStages: 1 +// LocalSharedMemoryBuffer: 1 + +template +struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3 + : BlockwiseGemmXdlops_pipeline_base + +{ + using Base = BlockwiseGemmXdlops_pipeline_base; + using Base::A_K1; + using Base::B_K1; + using Base::I0; + using Base::I1; + using Base::I2; + using Base::KGroup; + using Base::KRepeat; + using Base::xdlops_gemm; + using typename Base::HotLoopInstList; + + using Base::a_block_desc_m0_m1_m2_k; + using Base::CalculateCThreadOriginDataIndex; + using Base::CalculateCThreadOriginDataIndex8D; + using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2; + using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2; + using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4; + using Base::GetCThreadBuffer; + using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2; + using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2; + using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4; + using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2; + using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2; + + using Base::AMmaKStride; + using Base::BMmaKStride; + + using Base::MWaves; + + static constexpr index_t PrefetchStages = 2; + static constexpr index_t PrefillStages = 1; + static constexpr index_t GlobalBufferNum = 1; + static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1; + + template + __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&) + { + constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{}); + constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{}); + constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{}); + constexpr index_t K2 = KPack / KGroup; + constexpr index_t K1 = 64 / NPerXDL; + constexpr index_t K0 = KRepeat * KGroup; + + return transform_tensor_descriptor( + TileDesc_M0_M1_M2_K{}, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_unmerge_transform(make_tuple(Number{}, Number{}, Number{}))), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{})); + } + + static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 = + MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k); + + __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop) + { + return num_loop > PrefetchStages; + } + + __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop) + { + return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd; + } + + __device__ static constexpr auto HotLoopScheduler() + { + // A/B split schedule + // compiler is likely to use ds_read2 when instruction width smaller than 16bytes + constexpr auto num_ds_read_inst_a = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 + ? HotLoopInstList::A_LDS_Read_Inst_Num + : HotLoopInstList::A_LDS_Read_Inst_Num / 2; + + constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; + + constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num; + constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num * 2; + + static_assert(num_buffer_load_inst_a == num_ds_write_inst_a); + + constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num * 2; + constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle; + + constexpr auto ds_read_a_issue_cycle = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4; + constexpr auto ds_read_a_mfma_rate = + math::integer_divide_ceil(mfma_cycle - 4, 2 * ds_read_a_issue_cycle); + + // constexpr auto num_dsread_a_mfma = + // (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate; + + constexpr auto num_total_stages = MRepeat; + + // Group num_mfma_perstage num_ds_read_a_perstage + // since we want to reuse a local register buffer + constexpr auto num_mfma_perstage = num_mfma_inst / num_total_stages; + constexpr auto num_ds_read_a_perstage = num_ds_read_inst_a / num_total_stages; + + constexpr auto num_ds_read_a_mfma_perstage = + math::integer_divide_ceil(num_ds_read_a_perstage, ds_read_a_mfma_rate); + + constexpr auto num_ds_read_a_prefetch_stages = 2; + + constexpr auto buffer_load_perstage_more = math::integer_divide_ceil( + (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2)); + constexpr auto buffer_load_perstage_less = math::integer_divide_floor( + (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2)); + + constexpr auto buffer_load_stages_more = + (num_buffer_load_inst_a + num_buffer_load_inst_b) - + math::integer_divide_floor((num_buffer_load_inst_a + num_buffer_load_inst_b), + (num_total_stages - 2)) * + ((num_total_stages - 2)); + + constexpr auto buffer_load_b_stages = + buffer_load_perstage_more * buffer_load_stages_more > num_buffer_load_inst_b + ? num_buffer_load_inst_b / buffer_load_perstage_more + : (buffer_load_stages_more + + (num_buffer_load_inst_b - buffer_load_perstage_more * buffer_load_stages_more) / + buffer_load_perstage_less); + + constexpr auto buffer_load_a_stages = + num_total_stages - num_ds_read_a_prefetch_stages - buffer_load_b_stages; + + constexpr auto buffer_load_issue_point_b = 0; + constexpr auto buffer_load_issue_point_interval_more = + num_mfma_perstage / buffer_load_perstage_more; + constexpr auto buffer_load_issue_point_interval_less = + num_mfma_perstage / buffer_load_perstage_less; + constexpr auto ds_write_issue_point = 0; + constexpr auto buffer_load_issue_point_a = num_mfma_perstage >= 3 ? 1 : 0; + + // B global read + static_for<0, buffer_load_b_stages, 1>{}([&](auto i) { + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + if constexpr(((i < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + buffer_load_issue_point_b)) || + ((i >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + buffer_load_issue_point_b))) + { + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + } + + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + }); + }); + + // A global read + A local write + static_for<0, buffer_load_a_stages, 1>{}([&](auto i) { + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + ds_write_issue_point)) || + (((i + buffer_load_b_stages) >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + ds_write_issue_point))) + { + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + } + if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + buffer_load_issue_point_a)) || + (((i + buffer_load_b_stages) >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + buffer_load_issue_point_a))) + { + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + } + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + }); + }); + + // lds synchronization, prefetch next loop local A + static_for<0, num_ds_read_a_prefetch_stages, 1>{}([&](auto i) { + ignore = i; + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + }); + }); + } + + template + __device__ static constexpr auto EpilogueScheduler_1(Stage stage) + { + constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num; + constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; + constexpr auto num_buffer_load_inst_b = + MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num * 2; + + constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num * 2; + + constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat; + constexpr auto staged_num_mfma = num_mfma / MRepeat; + + constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a; + + if constexpr(stage.value == 0) + { + constexpr auto staged_num_buffer_load_b_per_ds_read_a = + num_buffer_load_inst_b / staged_num_ds_read_inst_a; + constexpr auto staged_num_mfma_per_buffer_load_b = + staged_num_mfma / num_buffer_load_inst_b; + // B global + static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { + ignore = i_inst; + + static_for<0, staged_num_buffer_load_b_per_ds_read_a, 1>{}([&](auto ibuf_inst) { + ignore = ibuf_inst; + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_buffer_load_b, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + }); + + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + }); + + __builtin_amdgcn_sched_barrier(0); + } + else if constexpr(stage.value == 1) + { +#if 0 + constexpr auto staged_num_ds_write_a_per_ds_read_a = + num_ds_write_inst_a / staged_num_ds_read_inst_a; + constexpr auto staged_num_mfma_per_ds_write_a = staged_num_mfma / num_ds_write_inst_a; + // A local write + static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { + ignore = i_inst; + + static_for<0, staged_num_ds_write_a_per_ds_read_a, 1>{}([&](auto idswrite_inst) { + ignore = idswrite_inst; + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write + }); + + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_ds_write_a_per_ds_read_a, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + }); +#elif 1 + constexpr auto staged_num_mfma_per_ds_write_a = + math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a); + + constexpr auto stage_more_mfma = + staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a; + + // A local write + static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) { + if constexpr(i_inst.value < stage_more_mfma) + { + if(i_inst.value < staged_num_ds_read_inst_a) + { + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + } + else + { + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_write_a, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write + } + } + else + { + if(i_inst.value < staged_num_ds_read_inst_a) + { + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + } + else + { + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write + } + } + }); +#endif + __builtin_amdgcn_sched_barrier(0); + } + else + { + // A local Read + static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { + ignore = i_inst; + __builtin_amdgcn_sched_group_barrier( + 0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + }); + + __builtin_amdgcn_sched_barrier(0); + } + } + + __device__ static constexpr auto EpilogueScheduler_2() + { + constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num; + + constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num * 2; + + constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat; + constexpr auto staged_num_mfma = num_mfma / MRepeat; + + constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a; + + // A local Read + static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { + ignore = i_inst; + __builtin_amdgcn_sched_group_barrier(0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + }); + + __builtin_amdgcn_sched_barrier(0); + } + + template + __device__ void Run(const AGridDesc& a_grid_desc, + const ABlockDesc& a_block_desc, + ABlockTransfer& a_blockwise_copy, + const AGridBuffer& a_grid_buf, + ABlockBuffer& a_block_buf, + const ABlockTransferStep& a_block_copy_step, + const BGridDesc& b_grid_desc, + BBlockTransfer& b_blockwise_copy, + BBlockTransfer& b_blockwise_copy_up, + const BGridBuffer& b_grid_buf, + const BGridBuffer& b_grid_buf_up, + BBlockBuffer& b_block_buf, + const BBlockTransferStep& b_block_copy_step, + CThreadBuffer& c_thread_buf, + CThreadBuffer& c_thread_buf_up, + index_t num_loop) const + { + ignore = b_block_buf; + __builtin_amdgcn_sched_barrier(0); + auto a_thread_buf = make_static_buffer( + a_thread_desc_.GetElementSpaceSize()); + auto b_thread_buf = make_static_buffer( + b_thread_desc_.GetElementSpaceSize()); + + StaticallyIndexedArray{}> b_thread_bufs; + StaticallyIndexedArray{}> b_thread_bufs_up; + constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0); + + // Global prefetch A1 B1 + b_blockwise_copy.Run(b_grid_desc, + b_grid_buf, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs(I0)); + + b_blockwise_copy_up.Run(b_grid_desc, + b_grid_buf_up, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs_up(I0)); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + __builtin_amdgcn_sched_barrier(0); + + // // Local prefill A1 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); + + // // Global prefetch A2 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + // Local prefetch A1 + block_sync_lds(); + static_for<0, 2, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(m0, I0, I0, Number{}, I0, I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + }); + + // Initialize C + c_thread_buf.Clear(); + c_thread_buf_up.Clear(); + + __builtin_amdgcn_sched_barrier(0); + + // main body + if constexpr(HasMainLoop) + { + index_t i = 0; + do + { + auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) { + b_blockwise_copy.Run(b_grid_desc, + b_grid_buf, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs(local_read_buf)); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + b_blockwise_copy_up.Run(b_grid_desc, + b_grid_buf_up, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs_up(local_read_buf)); + b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf)); + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + vector_type b_thread_vec_up; + + static_for<0, KPack, 1>{}([&](auto ik) { + a_thread_vec.template AsType()(ik) = + a_thread_buf[Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[mfma_reg_buf] + [Number{}]; + + b_thread_vec_up.template AsType()(ik) = + b_thread_bufs_up[mfma_reg_buf] + [Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + xdlops_gemm.Run( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + + xdlops_gemm.Run( + a_thread_vec.template AsType(), + b_thread_vec_up.template AsType(), + c_thread_buf_up.GetVectorTypeReference(Number{})); + }); + }); + + if constexpr(m0.value == MRepeat - 2) + { + block_sync_lds(); + + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(local_read_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); + }); + } + else if constexpr(m0.value == (MRepeat - 1)) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(local_read_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); + }); + } + else + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(mfma_reg_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); + }); + } + }); + HotLoopScheduler(); + }; + + LoopFunc(I0, I1); + LoopFunc(I1, I0); + + i += 2; + } while(i < (num_loop - 2)); + } + // tail + if constexpr(TailNum == TailNumber::Even) + { + b_blockwise_copy.Run(b_grid_desc, + b_grid_buf, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs(I1)); + + b_blockwise_copy_up.Run(b_grid_desc, + b_grid_buf_up, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs_up(I1)); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1)); + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + vector_type b_thread_vec_up; + + static_for<0, KPack, 1>{}([&](auto ik) { + a_thread_vec.template AsType()(ik) = + a_thread_buf[Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[I0][Number{}]; + + b_thread_vec_up.template AsType()(ik) = + b_thread_bufs_up[I0][Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec_up.template AsType(), + c_thread_buf_up.GetVectorTypeReference(Number{})); + }); + }); + if constexpr(m0.value == (MRepeat - 2)) + { + block_sync_lds(); + + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + } + else if constexpr(m0.value == MRepeat - 1) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + } + else + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + } + }); + + HotLoopScheduler(); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + vector_type b_thread_vec_up; + + static_for<0, KPack, 1>{}([&](auto ik) { + a_thread_vec.template AsType()(ik) = + a_thread_buf[Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[I1][Number{}]; + b_thread_vec_up.template AsType()(ik) = + b_thread_bufs_up[I1][Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec_up.template AsType(), + c_thread_buf_up.GetVectorTypeReference(Number{})); + }); + }); + + if constexpr(m0.value < (MRepeat - 2)) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple( + Number{}, I0, I0, Number{}, I0, I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple(Number<(m0 + 2 + HotloopLocalBufSwitch) % 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); + }); + } + }); + + HotLoopScheduler(); + // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle + // latency + } + else if constexpr(TailNum == TailNumber::Odd) + { + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + vector_type b_thread_vec_up; + + static_for<0, KPack, 1>{}([&](auto ik) { + a_thread_vec.template AsType()(ik) = + a_thread_buf[Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[I0][Number{}]; + b_thread_vec_up.template AsType()(ik) = + b_thread_bufs_up[I0][Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + xdlops_gemm.Run(a_thread_vec.template AsType(), + b_thread_vec_up.template AsType(), + c_thread_buf_up.GetVectorTypeReference(Number{})); + }); + }); + + if constexpr(m0.value < (MRepeat - 2)) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple( + Number{}, I0, I0, Number{}, I0, I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + } + }); + } + } + + protected: + // MRepeat MWave MLane KRepeat KLane KPack + // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack + // Reduce the vgpr usage here. + static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed( + make_tuple(I2, I1, I1, Number{}, I1, Number{})); + + using AThreadCopy = ThreadwiseTensorSliceTransfer_v4, + Sequence<0, 1, 2, 3, 4, 5>, + 5, + A_K1, + A_K1>; + + AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()}; + + static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, I1, Number{}, Number{})); + + static constexpr BTileDesc b_block_desc_n0_n1_k0_k1; + + using Base::c_thread_desc_; +}; + +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp index 6d115e7620..02da036d10 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -123,6 +123,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}); constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{}); constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{}); - constexpr index_t K2 = KPack; + constexpr index_t K2 = KPack / KGroup; constexpr index_t K1 = 64 / NPerXDL; - constexpr index_t K0 = KRepeat; + constexpr index_t K0 = KRepeat * KGroup; return transform_tensor_descriptor( TileDesc_M0_M1_M2_K{}, @@ -184,298 +185,230 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3 - __device__ static constexpr auto HotLoopScheduler(Stage stage) + __device__ static constexpr auto HotLoopScheduler() { - constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num; - constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; - constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num; - constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num; - - constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num; - - constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat; - constexpr auto staged_num_mfma = num_mfma / MRepeat; - - constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a; - - if constexpr(stage.value == 0) - { - constexpr auto staged_num_buffer_load_b_per_ds_read_a = - num_buffer_load_inst_b / staged_num_ds_read_inst_a; - constexpr auto staged_num_mfma_per_buffer_load_b = - staged_num_mfma / num_buffer_load_inst_b; - // B global - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; - - static_for<0, staged_num_buffer_load_b_per_ds_read_a - 1, 1>{}([&](auto ibuf_inst) { - ignore = ibuf_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_b, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - }); - - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - }); - - __builtin_amdgcn_sched_barrier(0); - } - else if constexpr(stage.value == 1) - { - constexpr auto staged_num_mfma_per_ds_write_a = - math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a); - - constexpr auto stage_more_mfma = - staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a; - - // A local write - static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) { - if constexpr(i_inst.value < stage_more_mfma) - { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - } - } - else - { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - } - } - }); - - __builtin_amdgcn_sched_barrier(0); - } - else if constexpr(stage.value == 2) - { - constexpr auto staged_num_mfma_per_buffer_load_a = - math::integer_divide_ceil(staged_num_mfma, num_buffer_load_inst_a); - - constexpr auto stage_more_mfma = - staged_num_mfma - (staged_num_mfma_per_buffer_load_a - 1) * num_buffer_load_inst_a; - - // A global - static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i_inst) { - if constexpr(i_inst.value < stage_more_mfma) - { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - } - } - else - { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_a - 2, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - } - } - }); - - __builtin_amdgcn_sched_barrier(0); - } - else - { - // A local Read - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - }); - - __builtin_amdgcn_sched_barrier(0); - } - } - - template - __device__ static constexpr auto EpilogueScheduler_1(Stage stage) - { - constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num; - constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; - constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num; - - constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num; - - constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat; - constexpr auto staged_num_mfma = num_mfma / MRepeat; - - constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a; - - if constexpr(stage.value == 0) - { - constexpr auto staged_num_buffer_load_b_per_ds_read_a = - num_buffer_load_inst_b / staged_num_ds_read_inst_a; - constexpr auto staged_num_mfma_per_buffer_load_b = - staged_num_mfma / num_buffer_load_inst_b; - // B global - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; - - static_for<0, staged_num_buffer_load_b_per_ds_read_a, 1>{}([&](auto ibuf_inst) { - ignore = ibuf_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_b, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - }); - - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - }); - - __builtin_amdgcn_sched_barrier(0); - } - else if constexpr(stage.value == 1) - { #if 0 - constexpr auto staged_num_ds_write_a_per_ds_read_a = - num_ds_write_inst_a / staged_num_ds_read_inst_a; - constexpr auto staged_num_mfma_per_ds_write_a = staged_num_mfma / num_ds_write_inst_a; - // A local write - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; + // A/B split schedule + // compiler is likely to use ds_read2 when instruction width smaller than 16bytes + constexpr auto num_ds_read_inst_a = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 + ? HotLoopInstList::A_LDS_Read_Inst_Num + : HotLoopInstList::A_LDS_Read_Inst_Num / 2; - static_for<0, staged_num_ds_write_a_per_ds_read_a, 1>{}([&](auto idswrite_inst) { - ignore = idswrite_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - }); + constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_ds_write_a_per_ds_read_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - }); + constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num; + constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num; + + constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num; + constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle; + + constexpr auto ds_read_a_issue_cycle = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4; + constexpr auto ds_read_a_mfma_rate = + math::integer_divide_ceil(mfma_cycle - 4, 2 * ds_read_a_issue_cycle); + + // constexpr auto num_dsread_a_mfma = + // (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate; + + constexpr auto num_stages = MRepeat; + + // Group num_mfma_perstage num_ds_read_a_perstage + // since we want to reuse a local register buffer + constexpr auto num_mfma_perstage = num_mfma_inst / num_stages; + constexpr auto num_ds_read_a_perstage = num_ds_read_inst_a / num_stages; + + constexpr auto num_ds_read_a_mfma_perstage = + math::integer_divide_ceil(num_ds_read_a_perstage, ds_read_a_mfma_rate); + + constexpr auto num_mfma_per_issue_more = math::integer_divide_ceil( + num_mfma_inst, num_buffer_load_inst_a + num_buffer_load_inst_b); + constexpr auto num_mfma_per_issue_less = math::integer_divide_floor( + num_mfma_inst, num_buffer_load_inst_a + num_buffer_load_inst_b); + // Insert more mfmas between bufferloads + constexpr auto num_stage1_bufferloads = + num_mfma_inst - + (num_buffer_load_inst_a + num_buffer_load_inst_b) * num_mfma_per_issue_less; + constexpr auto num_stage1_mfma = num_mfma_per_issue_more * num_stage1_bufferloads; + // Insert less mfmas between bufferloads + // constexpr auto num_stage2_mfma = num_mfma_inst - num_stage1_mfma; + + constexpr auto buffer_load_issue_point = 0; + constexpr auto ds_write_issue_point_stage1 = num_mfma_per_issue_more >= 3 ? 1 : 0; + constexpr auto ds_write_issue_point_stage2 = num_mfma_per_issue_less >= 3 ? 1 : 0; + + static_for<0, num_mfma_inst, 1>{}([&](auto i) { + constexpr auto current_buffer_load_issue = + i < num_stage1_mfma + ? (i / num_mfma_per_issue_more) + : (num_stage1_bufferloads + (i - num_stage1_mfma) / num_mfma_per_issue_less); + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + // Group num_mfma_perstage num_ds_read_a_perstage + // Hide A lds rd issue latency at begining of each stage + if constexpr((i % num_mfma_perstage) >= + (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + + // Schedule VMEM access instruction distributed evenly in the loop + // Hide B/A global rd issue latency + if constexpr(((i < num_stage1_mfma) && + (i % num_mfma_per_issue_more == buffer_load_issue_point)) || + ((i >= num_stage1_mfma) && + ((i - num_stage1_mfma) % num_mfma_per_issue_less == + buffer_load_issue_point))) + { + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + } + + // Hide A lds wr issue latency + if constexpr((current_buffer_load_issue >= num_buffer_load_inst_b) && + ((((i < num_stage1_mfma) && + (i % num_mfma_per_issue_more == ds_write_issue_point_stage1)) || + ((i >= num_stage1_mfma) && + ((i - num_stage1_mfma) % num_mfma_per_issue_less == + ds_write_issue_point_stage2))) && + (((i < num_stage1_mfma) && + ((i / num_mfma_per_issue_more - num_buffer_load_inst_b) < num_ds_write_inst_a)) || + ((i >= num_stage1_mfma) && + ((i - num_stage1_mfma) / num_mfma_per_issue_less + + num_stage1_bufferloads - num_buffer_load_inst_b) < num_ds_write_inst_a)))) + { + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + } + }); #elif 1 - constexpr auto staged_num_mfma_per_ds_write_a = - math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a); + // A/B split schedule + // compiler is likely to use ds_read2 when instruction width smaller than 16bytes + constexpr auto num_ds_read_inst_a = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 + ? HotLoopInstList::A_LDS_Read_Inst_Num + : HotLoopInstList::A_LDS_Read_Inst_Num / 2; - constexpr auto stage_more_mfma = - staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a; + constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num; - // A local write - static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) { - if constexpr(i_inst.value < stage_more_mfma) + constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num; + constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num; + + static_assert(num_buffer_load_inst_a == num_ds_write_inst_a); + + constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num; + constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle; + + constexpr auto ds_read_a_issue_cycle = + HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4; + constexpr auto ds_read_a_mfma_rate = + math::integer_divide_ceil(mfma_cycle - 4, 2 * ds_read_a_issue_cycle); + + // constexpr auto num_dsread_a_mfma = + // (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate; + + constexpr auto num_total_stages = MRepeat; + + // Group num_mfma_perstage num_ds_read_a_perstage + // since we want to reuse a local register buffer + constexpr auto num_mfma_perstage = num_mfma_inst / num_total_stages; + constexpr auto num_ds_read_a_perstage = num_ds_read_inst_a / num_total_stages; + + constexpr auto num_ds_read_a_mfma_perstage = + math::integer_divide_ceil(num_ds_read_a_perstage, ds_read_a_mfma_rate); + + constexpr auto num_ds_read_a_prefetch_stages = 2; + + constexpr auto buffer_load_perstage_more = math::integer_divide_ceil( + (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2)); + constexpr auto buffer_load_perstage_less = math::integer_divide_floor( + (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2)); + + constexpr auto buffer_load_stages_more = + (num_buffer_load_inst_a + num_buffer_load_inst_b) - + math::integer_divide_floor((num_buffer_load_inst_a + num_buffer_load_inst_b), + (num_total_stages - 2)) * + ((num_total_stages - 2)); + + constexpr auto buffer_load_b_stages = + buffer_load_perstage_more * buffer_load_stages_more > num_buffer_load_inst_b + ? num_buffer_load_inst_b / buffer_load_perstage_more + : (buffer_load_stages_more + + (num_buffer_load_inst_b - buffer_load_perstage_more * buffer_load_stages_more) / + buffer_load_perstage_less); + + constexpr auto buffer_load_a_stages = + num_total_stages - num_ds_read_a_prefetch_stages - buffer_load_b_stages; + + constexpr auto buffer_load_issue_point_b = 0; + constexpr auto buffer_load_issue_point_interval_more = + num_mfma_perstage / buffer_load_perstage_more; + constexpr auto buffer_load_issue_point_interval_less = + num_mfma_perstage / buffer_load_perstage_less; + constexpr auto ds_write_issue_point = 0; + constexpr auto buffer_load_issue_point_a = num_mfma_perstage >= 3 ? 1 : 0; + + // B global read + static_for<0, buffer_load_b_stages, 1>{}([&](auto i) { + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + if constexpr(((i < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + buffer_load_issue_point_b)) || + ((i >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + buffer_load_issue_point_b))) { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - } + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read } - else + + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) { - if(i_inst.value < staged_num_ds_read_inst_a) - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - } - else - { - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write - } + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read } }); -#endif - __builtin_amdgcn_sched_barrier(0); - } - else - { - // A local Read - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - }); - - __builtin_amdgcn_sched_barrier(0); - } - } - - __device__ static constexpr auto EpilogueScheduler_2() - { - constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num; - - constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num; - - constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat; - constexpr auto staged_num_mfma = num_mfma / MRepeat; - - constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a; - - // A local Read - static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) { - ignore = i_inst; - __builtin_amdgcn_sched_group_barrier(0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read }); - __builtin_amdgcn_sched_barrier(0); + // A global read + A local write + static_for<0, buffer_load_a_stages, 1>{}([&](auto i) { + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + ds_write_issue_point)) || + (((i + buffer_load_b_stages) >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + ds_write_issue_point))) + { + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + } + if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_more == + buffer_load_issue_point_a)) || + (((i + buffer_load_b_stages) >= buffer_load_stages_more) && + (imfma % buffer_load_issue_point_interval_less == + buffer_load_issue_point_a))) + { + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + } + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + }); + }); + + // lds synchronization, prefetch next loop local A + static_for<0, num_ds_read_a_prefetch_stages, 1>{}([&](auto i) { + ignore = i; + static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage)) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + }); + }); +#endif } template {}([&](auto k0) { - a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(I0, I0, I0, k0, I0, I0), - a_block_buf.At(I0), - a_thread_desc_, - make_tuple(I0, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, 2, 1>{}([&](auto m0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(m0, I0, I0, Number{}, I0, I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); }); // Initialize C @@ -558,26 +495,18 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto m0) { - if constexpr(m0.value == 0) - { - b_blockwise_copy.Run(b_grid_desc, - b_grid_buf, - b_block_desc_n0_n1_k0_k1, - b_block_origin_idx, - b_thread_bufs(local_read_buf)); - b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); - } - else if constexpr(m0.value == 1) - { - a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf)); - } - else if constexpr(m0.value == 2) - { - a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); - a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); - } + b_blockwise_copy.Run(b_grid_desc, + b_grid_buf, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs(local_read_buf)); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf)); + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { static_for<0, NRepeat, 1>{}([&](auto n0) { vector_type a_thread_vec; @@ -613,49 +542,88 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto k0) { - a_thread_copy_.Run( - a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0), - a_block_buf.At(local_read_buf), - a_thread_desc_, - make_tuple( - Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) % - 2>{}, - I0, - I0, - k0, - I0, - I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(local_read_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); + }); + } + else if constexpr(m0.value == (MRepeat - 1)) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(local_read_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); }); } else { static_for<0, KRepeat, 1>{}([&](auto k0) { - a_thread_copy_.Run( - a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0), - a_block_buf.At(mfma_reg_buf), - a_thread_desc_, - make_tuple( - Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) % - 2>{}, - I0, - I0, - k0, - I0, - I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(mfma_reg_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); }); } - - HotLoopScheduler(m0); }); + HotLoopScheduler(); }; LoopFunc(I0, I1); @@ -667,20 +635,14 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto m0) { - if constexpr(m0.value == 0) - { - b_blockwise_copy.Run(b_grid_desc, - b_grid_buf, - b_block_desc_n0_n1_k0_k1, - b_block_origin_idx, - b_thread_bufs(I1)); - } - else if constexpr(m0.value == MRepeat - 1) - { - a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1)); - } + b_blockwise_copy.Run(b_grid_desc, + b_grid_buf, + b_block_desc_n0_n1_k0_k1, + b_block_origin_idx, + b_thread_bufs(I1)); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1)); + static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { static_for<0, NRepeat, 1>{}([&](auto n0) { vector_type a_thread_vec; @@ -707,36 +669,72 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto k0) { - a_thread_copy_.Run( - a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0), - a_block_buf.At(I1), - a_thread_desc_, - make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); + }); + } + else if constexpr(m0.value == (MRepeat - 1)) + { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); } else { static_for<0, KRepeat, 1>{}([&](auto k0) { - a_thread_copy_.Run( - a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0), - a_block_buf.At(I0), - a_thread_desc_, - make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(Number<(m0 + 2) % MRepeat>{}, + I0, + I0, + Number{}, + I0, + I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); } - - EpilogueScheduler_1(m0); }); + HotLoopScheduler(); + static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { static_for<0, NRepeat, 1>{}([&](auto n0) { @@ -764,25 +762,31 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto k0) { - a_thread_copy_.Run( - a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number{}, I0, I0, k0, I0, I0), - a_block_buf.At(I1), - a_thread_desc_, - make_tuple( - Number<(m0 + 1 + HotloopLocalBufSwitch) % 2>{}, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple( + Number{}, I0, I0, Number{}, I0, I0), + a_block_buf.At(I1), + a_thread_desc_, + make_tuple(Number<(m0 + 2 + HotloopLocalBufSwitch) % 2>{}, + I0, + I0, + k0, + I0, + Number{}), + a_thread_buf); + }); }); - - EpilogueScheduler_2(); } }); + + HotLoopScheduler(); // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle // latency - // __builtin_amdgcn_sched_barrier(0); } else if constexpr(TailNum == TailNumber::Odd) { @@ -813,18 +817,21 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}([&](auto k0) { - a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(Number{}, I0, I0, k0, I0, I0), - a_block_buf.At(I0), - a_thread_desc_, - make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple( + Number{}, I0, I0, Number{}, I0, I0), + a_block_buf.At(I0), + a_thread_desc_, + make_tuple( + Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); - - EpilogueScheduler_2(); } }); } @@ -841,7 +848,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3, + Sequence<1, 1, 1, 1, 1, KPack / KGroup>, Sequence<0, 1, 2, 3, 4, 5>, 5, A_K1, diff --git a/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp b/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp index 08d177035e..a6110d2bfc 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp @@ -264,77 +264,152 @@ struct DeviceMoeGemm : public DeviceGemmMultipleDSplitKBPreShuffle= 256) ? 1 : 2; - constexpr auto MemoryDataOp = - IsInputGemm ? InMemoryDataOperationEnum::Set : InMemoryDataOperationEnum::AtomicAdd; - if(has_main_k_block_loop) + if(IsInputGemm || arg.TopK == 1) { - // Tail number always full - if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + constexpr auto MemoryDataOp = InMemoryDataOperationEnum::Set; + + if(has_main_k_block_loop) { + // Tail number always full + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_moe_gemm; + RunKernel(kernel); + } + else + { + const auto kernel = kernel_moe_gemm; + RunKernel(kernel); + } + } + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || + BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) { if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) { - const auto kernel = kernel_moe_gemm; + const auto kernel = kernel_moe_gemm_2lds; RunKernel(kernel); } else { - const auto kernel = kernel_moe_gemm; + const auto kernel = kernel_moe_gemm_2lds; RunKernel(kernel); } } - } - else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || - BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + else { - const auto kernel = kernel_moe_gemm_2lds; + throw std::runtime_error("todo: only v1 & v2 support now"); + } + } +#if 1 + else + { + // Tail number always 1 + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + const auto kernel = kernel_moe_gemm; RunKernel(kernel); } + } +#endif + } + else + { + constexpr auto MemoryDataOp = InMemoryDataOperationEnum::AtomicAdd; + + if(has_main_k_block_loop) + { + // Tail number always full + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_moe_gemm; + RunKernel(kernel); + } + else + { + const auto kernel = kernel_moe_gemm; + RunKernel(kernel); + } + } + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || + BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_moe_gemm_2lds; + RunKernel(kernel); + } + else + { + const auto kernel = kernel_moe_gemm_2lds; + RunKernel(kernel); + } + } else { - const auto kernel = kernel_moe_gemm_2lds; + throw std::runtime_error("todo: only v1 & v2 support now"); + } + } +#if 1 + else + { + // Tail number always 1 + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + const auto kernel = kernel_moe_gemm; RunKernel(kernel); } } - else - { - throw std::runtime_error("todo: only v1 & v2 support now"); - } - } -#if 1 - else - { - // Tail number always 1 - if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) - { - const auto kernel = kernel_moe_gemm; - RunKernel(kernel); - } - } #endif - + } return ave_time; } diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp index 255fb8cff4..54bc1a76df 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp @@ -188,7 +188,10 @@ struct GridwiseMoeGemm math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk); static constexpr index_t KLane = mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops(); - static constexpr index_t KRepeat = KPerBlock / KLane / KPack; + + static constexpr index_t KGroup = mfma_selector::selected_mfma.k_per_blk == 32 ? 2 : 1; + // static_assert(KGroup == 2, ""); + static constexpr index_t KRepeat = KPerBlock / KLane / (KPack / KGroup); static constexpr index_t NLane = NPerXdl; static constexpr index_t NWave = NPerBlock / NPerXdl / NXdlPerWave; // static constexpr index_t NumTokens = 1; @@ -249,7 +252,7 @@ struct GridwiseMoeGemm } __host__ __device__ static auto CalculateBK0Shuffled(index_t K) { - return math::integer_divide_ceil(K, KLane * KPack); + return math::integer_divide_ceil(K, KLane * KPack / KGroup); } __host__ __device__ static auto CalculateKPadded(index_t K) @@ -391,7 +394,7 @@ struct GridwiseMoeGemm __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0) { - constexpr index_t NkSwizzleNumber = Number{}; + constexpr index_t NkSwizzleNumber = Number{}; return make_naive_tensor_descriptor( make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber), make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1)); @@ -1301,7 +1304,7 @@ struct GridwiseMoeGemm make_multi_index(n_block_data_idx_on_grid, get_warp_local_1d_id() % NWave, 0, - KPack * (get_thread_local_1d_id() % warpSize))); + KPack / KGroup * (get_thread_local_1d_id() % warpSize))); // LDS allocation for A and B: be careful of alignment // Cast after lds @@ -1347,7 +1350,7 @@ struct GridwiseMoeGemm make_multi_index(n_block_data_idx_on_grid, get_warp_local_1d_id() % NWave, 0, - KPack * (get_thread_local_1d_id() % warpSize))); + KPack / KGroup * (get_thread_local_1d_id() % warpSize))); blockwise_gemm_pipeline.template Run( a_grid_desc_ak0_m_ak1, a_block_desc_ak0_m_ak1, @@ -1886,7 +1889,8 @@ struct GridwiseMoeGemm const auto c_grid_desc_mblock_mperblock_nblock_nperblock = MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( c_grid_desc_m_n, problem.MBlock, problem.NBlock); - const index_t max_token_id = __builtin_amdgcn_readfirstlane(p_max_token_id[0]); + const index_t max_token_id = __builtin_amdgcn_readfirstlane(p_max_token_id[0]); + // static_assert(NSwizzle == false, "to do fix: need another pr in sorting merged"); const index_t expert_block_id = NSwizzle ? blockIdx.x / problem.NBlock : blockIdx.y; if(expert_block_id * MPerBlock >= max_token_id) return; @@ -1895,12 +1899,13 @@ struct GridwiseMoeGemm const auto block_mn = [&]() -> std::pair { if constexpr(NSwizzle) { - const index_t ecnt_prefix = p_max_token_id[1 + expert_id]; - const index_t prefix_block = ecnt_prefix * problem.NBlock; - const index_t ecnt = p_max_token_id[2 + expert_id] - ecnt_prefix; - const index_t expert_swizzle = ecnt > 0 ? ecnt : 1; - const index_t bid_new = blockIdx.x - prefix_block; - const index_t nid = __builtin_amdgcn_readfirstlane( + const index_t ecnt_prefix = p_max_token_id[1 + expert_id]; + const index_t prefix_block = ecnt_prefix * problem.NBlock; + const index_t ecnt = p_max_token_id[2 + expert_id] - ecnt_prefix; + const index_t expert_swizzle = + ecnt > 0 ? ecnt : 1; // p_max_token_id[expert_id + 1]; // 2 + const index_t bid_new = blockIdx.x - prefix_block; + const index_t nid = __builtin_amdgcn_readfirstlane( bid_new % 8 + bid_new / (8 * expert_swizzle) * 8); const index_t mid = __builtin_amdgcn_readfirstlane(ecnt_prefix + bid_new / 8 % expert_swizzle); @@ -1911,9 +1916,9 @@ struct GridwiseMoeGemm return {blockIdx.x, blockIdx.y}; } }(); + const index_t block_n_id = block_mn.first; const index_t block_m_id = block_mn.second; - const index_t token0 = __builtin_amdgcn_readfirstlane(p_sorted_token_ids[block_m_id * MPerBlock] & 0xffffff); @@ -1925,11 +1930,9 @@ struct GridwiseMoeGemm constexpr auto AMRepeats = MPerBlock / AMThreads; const index_t token_pos = block_m_id * MPerBlock + threadIdx.x / AKThreads * AMRepeats; - if(token_pos >= max_token_id || expert_block_id * MPerBlock >= max_token_id || - token0 >= problem.NumTokens) + if(token_pos >= max_token_id || token0 >= problem.NumTokens) return; - StaticallyIndexedArray - gather_offsets; //= p_sorted_token_ids[token_pos]; + StaticallyIndexedArray gather_offsets; static_for<0, AMRepeats, 1>{}([&](auto m0) { const index_t fused_token = p_sorted_token_ids[token_pos + m0]; index_t token_offset = fused_token & 0xffffff; @@ -1939,7 +1942,8 @@ struct GridwiseMoeGemm } gather_offsets(m0) = static_cast(token_offset) * problem.K; }); - const index_t expert_stride = __builtin_amdgcn_readfirstlane(problem.N * problem.K); + const index_t expert_stride = + __builtin_amdgcn_readfirstlane(problem.N * problem.K * (IsInputGemm ? 2 : 1)); // N0, K0, Blocksize*KPack const index_t n_block_data_idx_on_grid = @@ -1950,7 +1954,6 @@ struct GridwiseMoeGemm const auto b_grid_buf = make_dynamic_buffer( p_b_grid + expert_id * expert_stride / BPackedSize, b_grid_desc_bpreshuffled.GetElementSpaceSize()); - // A matrix in LDS memory, dst of blockwise copy constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); @@ -2012,7 +2015,7 @@ struct GridwiseMoeGemm make_multi_index(n_block_data_idx_on_grid, get_warp_local_1d_id() % NWave, 0, - KPack * (get_thread_local_1d_id() % warpSize))); + KPack / KGroup * (get_thread_local_1d_id() % warpSize))); // LDS allocation for A and B: be careful of alignment // Cast after lds @@ -2029,24 +2032,76 @@ struct GridwiseMoeGemm static_assert(std::is_default_constructible_v); auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); + decltype(c_thread_buf) c_thread_buf_up; + + StaticBufferTupleOfVector + c_thread_buf_fp32; const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane( (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / KPerBlock); - blockwise_gemm_pipeline.template Run(a_grid_desc_ak0_m_ak1, - a_block_desc_ak0_m_ak1, - a_blockwise_copy, - a_grid_buf, - a_block_bufs, - a_block_slice_copy_step, - b_grid_desc_bpreshuffled, - b_blockwise_copy, - b_grid_buf, - b_block_bufs, - b_block_slice_copy_step, - c_thread_buf, - num_k_block_main_loop); + if constexpr(IsInputGemm) + { + const BDataType* p_b_grid_up = p_b_grid + expert_stride / 2 / BPackedSize; + const auto b_grid_buf_up = make_dynamic_buffer( + p_b_grid_up + expert_id * expert_stride / BPackedSize, + b_grid_desc_bpreshuffled.GetElementSpaceSize()); + auto b_blockwise_copy_up = ThreadwiseTensorSliceTransfer_v2< + BDataType, + BDataType, + decltype(b_grid_desc_bpreshuffled), + decltype(b_block_desc_bk0_n_bk1), + Sequence{}, I1, Number{}, Number{}>, + Sequence<1, 2, 0, 3>, + 3, + BBlockTransferSrcScalarPerVector, + BThreadTransferSrcResetCoordinateAfterRun, + true>(b_grid_desc_bpreshuffled, + make_multi_index(n_block_data_idx_on_grid, + get_warp_local_1d_id() % NWave, + 0, + KPack / KGroup * (get_thread_local_1d_id() % warpSize))); + blockwise_gemm_pipeline.template Run( + a_grid_desc_ak0_m_ak1, + a_block_desc_ak0_m_ak1, + a_blockwise_copy, + a_grid_buf, + a_block_bufs, + a_block_slice_copy_step, + b_grid_desc_bpreshuffled, + b_blockwise_copy, + b_blockwise_copy_up, + b_grid_buf, + b_grid_buf_up, + b_block_bufs, + b_block_slice_copy_step, + c_thread_buf, + c_thread_buf_up, + num_k_block_main_loop); + } + else + { + + blockwise_gemm_pipeline.template Run( + a_grid_desc_ak0_m_ak1, + a_block_desc_ak0_m_ak1, + a_blockwise_copy, + a_grid_buf, + a_block_bufs, + a_block_slice_copy_step, + b_grid_desc_bpreshuffled, + b_blockwise_copy, + b_grid_buf, + b_block_bufs, + b_block_slice_copy_step, + c_thread_buf, + num_k_block_main_loop); + } // shuffle C and write out { @@ -2074,6 +2129,185 @@ struct GridwiseMoeGemm constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6); constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7); + // mul scales + const float* p_sorted_weights_0 = p_ds_grid[I0]; + const float* p_scale_b = p_ds_grid[I1]; + + static_assert(M0 * M1 * M2 * M3 * M4 == MPerBlock); + static_assert(M4 == 4); + const index_t m1 = get_warp_local_1d_id() / NWave; + const index_t m3 = threadIdx.x % get_warp_size() / MPerXdl; + + if(p_sorted_weights_0 != nullptr && p_scale_b != nullptr) + { + if constexpr(PerTokenQuant) + { + constexpr index_t scale_stride = (IsInputGemm ? 2 : 1); + p_scale_b += expert_id * problem.N * scale_stride + block_n_id * NPerBlock + + get_warp_local_1d_id() % NWave * NPerXdl + threadIdx.x % NPerXdl; + } + else + { + p_scale_b += expert_id; + } + + vector_type scale_token_ids; + vector_type topk_weights; + static_for<0, NXdlPerWave, 1>{}([&](auto n0) { + const float scale_b = p_scale_b[n0 * NWave * NPerXdl * PerTokenQuant]; + static_for<0, MXdlPerWave, 1>{}([&](auto m0) { // MXDLPerWave + static_for<0, M2, 1>{}([&](auto m2) { // m_inst_num_groups_per_blk + const index_t m_pos = block_m_id * MPerBlock + m0 * M1 * M2 * M3 * M4 + + m1 * M2 * M3 * M4 + m2 * M3 * M4 + m3 * M4; + if constexpr(PerTokenQuant) + { + scale_token_ids = + *c_style_pointer_cast*>( + p_sorted_token_ids + m_pos); + } + if constexpr(MulRoutedWeight) + { + topk_weights = *c_style_pointer_cast*>( + p_ds_grid[I2] + m_pos); + } + static_for<0, M4, 1>{}([&](auto m4) { // m_inst_group_size + float scale_a = [&]() { + if constexpr(PerTokenQuant) + { + index_t fused_token = scale_token_ids.AsType()[m4]; + const index_t token_offset = fused_token & 0xffffff; + return token_offset < problem.NumTokens + ? p_sorted_weights_0[token_offset] + : 0.0; + } + else + { + return p_sorted_weights_0[0]; + } + }(); + constexpr index_t c_offset = + blockwise_gemm_pipeline.GetCThreadDesc().CalculateOffset( + make_tuple(m0, n0, m2 * M4 + m4)); + constexpr auto cidx = Number{}; + if constexpr(IsInputGemm) // gu fusion + { + if constexpr(ActivationOperation == Activation::silu_and_mul) + { + const float scale_up = + p_scale_b[(n0 * NWave * NPerXdl + problem.N) * + PerTokenQuant]; + float gate = scale_a * scale_b * c_thread_buf[cidx]; + float up = scale_a * scale_up * c_thread_buf_up[cidx]; + if constexpr(MulRoutedWeight) + { + gate = gate * topk_weights.AsType()[m4]; + up = up * topk_weights.AsType()[m4]; + } + if constexpr(is_same_v, pk_i4_t>) + { + gate *= 16; + up *= 16; + } + tensor_operation::element_wise::Silu{}(gate, gate); + c_thread_buf_fp32(cidx) = gate * up; + } + else if(ActivationOperation == Activation::gelu_and_mul) + { + const float scale_up = + p_scale_b[(n0 * NWave * NPerXdl + problem.N) * + PerTokenQuant]; + float gate = scale_a * scale_b * c_thread_buf[cidx]; + float up = scale_a * scale_up * c_thread_buf_up[cidx]; + if constexpr(MulRoutedWeight) + { + gate = gate * topk_weights.AsType()[m4]; + up = up * topk_weights.AsType()[m4]; + } + if constexpr(is_same_v, pk_i4_t>) + { + gate *= 16; + up *= 16; + } + tensor_operation::element_wise::Gelu{}(gate, gate); + c_thread_buf_fp32(cidx) = gate * up; + } + } + else + { + c_thread_buf_fp32(cidx) = + scale_a * scale_b * c_thread_buf[cidx]; + if constexpr(MulRoutedWeight) + { + c_thread_buf_fp32(cidx) = c_thread_buf_fp32(cidx) * + topk_weights.AsType()[m4]; + } + } + }); + }); + }); + }); + } + else + { + vector_type topk_weights; // for gemm2 only + static_for<0, NXdlPerWave, 1>{}([&](auto n0) { + static_for<0, MXdlPerWave, 1>{}([&](auto m0) { // MXDLPerWave + static_for<0, M2, 1>{}([&](auto m2) { // m_inst_num_groups_per_blk + const index_t m_pos = block_m_id * MPerBlock + m0 * M1 * M2 * M3 * M4 + + m1 * M2 * M3 * M4 + m2 * M3 * M4 + m3 * M4; + if constexpr(MulRoutedWeight) + { + topk_weights = *c_style_pointer_cast*>( + p_ds_grid[I2] + m_pos); + } + static_for<0, M4, 1>{}([&](auto m4) { // m_inst_group_size + constexpr index_t c_offset = + blockwise_gemm_pipeline.GetCThreadDesc().CalculateOffset( + make_tuple(m0, n0, m2 * M4 + m4)); + constexpr auto cidx = Number{}; + + if constexpr(IsInputGemm) // gu fusion + { + if constexpr(ActivationOperation == Activation::silu_and_mul) + { + float gate = c_thread_buf[cidx]; + float up = c_thread_buf_up[cidx]; + if constexpr(MulRoutedWeight) + { + gate = gate * topk_weights.AsType()[m4]; + up = up * topk_weights.AsType()[m4]; + } + tensor_operation::element_wise::Silu{}(gate, gate); + c_thread_buf_fp32(cidx) = gate * up; + } + else if(ActivationOperation == Activation::gelu_and_mul) + { + float gate = c_thread_buf[cidx]; + float up = c_thread_buf_up[cidx]; + if constexpr(MulRoutedWeight) + { + gate = gate * topk_weights.AsType()[m4]; + up = up * topk_weights.AsType()[m4]; + } + tensor_operation::element_wise::Gelu{}(gate, gate); + c_thread_buf_fp32(cidx) = gate * up; + } + } + else + { + c_thread_buf_fp32(cidx) = c_thread_buf[cidx]; + if constexpr(MulRoutedWeight) + { + c_thread_buf_fp32(cidx) = topk_weights.AsType()[m4] * + c_thread_buf_fp32[cidx]; + } + } + }); + }); + }); + }); + } + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); @@ -2171,18 +2405,8 @@ struct GridwiseMoeGemm const auto ds_grid_buf = generate_tuple( [&](auto i) { - using DDataType = remove_cvref_t>; - const DDataType* ptr_ = p_ds_grid[i]; - // hack logic here to support different kind of strides. todo fix it. - // ascale t, 1; bscale E, N, 1, move ptr to E - // if(i.value == 1) - // { - // ptr_ += - // expert_id * (problem.StrideDs[1] ? problem.StrideDs[1] * problem.N : - // 1); - // } return make_dynamic_buffer( - ptr_, ds_grid_desc_m_n[i].GetElementSpaceSize()); + p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize()); }, Number{}); @@ -2258,7 +2482,6 @@ struct GridwiseMoeGemm auto c_grid_buf = make_dynamic_buffer( p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); - // space filling curve for threadwise C in VGPR constexpr auto sfc_c_vgpr = SpaceFillingCurve, Sequence<0, 1, 2, 3, 4, 5, 6, 7>, @@ -2297,7 +2520,7 @@ struct GridwiseMoeGemm block_m_id * MPerBlock + threadIdx.x / ENThreads * EMRepeats + dstidx(I1); static_for<0, EMRepeats, 1>{}([&](auto m0) { const index_t fused_token = p_sorted_token_ids[c_token_pos + m0]; - index_t token_offset = fused_token & 0xffffff; + IndexType token_offset = fused_token & 0xffffff; if constexpr(IsInputGemm) { token_offset = token_offset * problem.TopK + (fused_token >> 24); @@ -2310,7 +2533,7 @@ struct GridwiseMoeGemm // each thread write its data from VGPR to LDS c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2, sfc_c_vgpr.GetIndexTupleOfNumber(access_id), - c_thread_buf, + c_thread_buf_fp32, c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, c_shuffle_block_buf); From 037a5ff76c7e1cf984011b446ec0927acf30783e Mon Sep 17 00:00:00 2001 From: valarLip <340077269@qq.com> Date: Mon, 12 May 2025 06:36:23 +0000 Subject: [PATCH 10/11] update moe --- .../moe_gemm1_xdl_fp8.cpp | 30 ++++---- ..._pipeline_xdlops_b_preshuffle_selector.hpp | 69 +++++++++++++------ .../blockwise_gemm_pipeline_xdlops_base.hpp | 5 ++ include/ck/utility/blkgemmpipe_scheduler.hpp | 14 +++- 4 files changed, 84 insertions(+), 34 deletions(-) diff --git a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp index 3b31460953..159fb21851 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp @@ -158,11 +158,14 @@ using BElementOp = PassThrough; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr ck::index_t MPerBlock = 128; -static constexpr ck::index_t MXDLPerWave = 4; -static constexpr ck::index_t NXDLPerWave = 2; -static constexpr ck::index_t BLOCKSIZE = 256; -static constexpr ck::index_t NPerBlock = 64; +static constexpr ck::index_t NPerBlock = 128; static constexpr ck::index_t MNPerXDL = 16; +static constexpr ck::index_t MXDLPerWave = MPerBlock / (MNPerXDL * 1); +static constexpr ck::index_t NXDLPerWave = NPerBlock / (MNPerXDL * 4); +// static constexpr ck::index_t CShuffleMXDLPerWave = MXDLPerWave; +// static constexpr ck::index_t CShuffleNXDLPerWave = NXDLPerWave; +static constexpr ck::index_t BLOCKSIZE = 256; + static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType); static constexpr ck::index_t Nswizzle = false; static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType); @@ -183,15 +186,15 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceM // mn_perxdl MNPerXDL, MNPerXDL, // mn_xdlperwave - MXDLPerWave, NXDLPerWave, + MXDLPerWave, NXDLPerWave, // a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0, // CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| // MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| // PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| - 2, 2, S<1, 32, 1, 8>, S, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>; + 2, 2, S<1, 32, 1, 8>, S, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>; // clang-format on @@ -205,9 +208,9 @@ int main(int argc, char* argv[]) ck::index_t N = 4096; ck::index_t K = 6144; ck::index_t experts = 8; - ck::index_t sorted_tile_num = 16; - ck::index_t valid_tile_num = 13; - ck::index_t tokens = 64; + ck::index_t sorted_tile_num = 256; + ck::index_t valid_tile_num = 256; + ck::index_t tokens = 16384; ck::index_t topk = 2; if(argc == 1) @@ -263,11 +266,12 @@ int main(int argc, char* argv[]) Tensor sorted_token_ids(HostTensorDescriptor({sorted_size}, {1})); Tensor max_token_id(HostTensorDescriptor({1 + sorted_tile_num})); max_token_id.mData = {valid_size}; - int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3}; + // int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3}; for(int i = 0; i < sorted_tile_num; i++) { - expert_ids.mData[i] = eids[i]; + expert_ids.mData[i] = i / (valid_tile_num / experts); } + int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num; int tokenid = 0; @@ -307,7 +311,7 @@ int main(int argc, char* argv[]) case 0: break; case 1: a0_t_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); - b0_e_n_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + b0_e_n_k.GenerateTensorValue(GeneratorTensor_3{-0.1, 0.1}); d0_t_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); d1_e_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); d2_e_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp index 074b5873ee..df59513a88 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp @@ -8,6 +8,7 @@ #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_dequant_v1.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp" +#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_dequant_v3.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp" @@ -171,26 +172,54 @@ constexpr auto BlockGemmBPreshufflePipeline_Selector() static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3"); if constexpr(std::is_same::value) { - return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}; + if constexpr(GUFusion) + { + return BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3< + BlkGemmPipeSche, + BlockSize, + ADataType, + BDataType, + ComputeDataType, + AccDataType, + ATileDesc, + BTileDesc, + AMmaTileDesc, + BMmaTileDesc, + ABlockTransferSrcScalarPerVector, + BBlockTransferSrcScalarPerVector, + MPerBlock, + NPerBlock, + KPerBlock, + MPerXDL, + NPerXDL, + MRepeat, + NRepeat, + KPack>{}; + } + else + { + + return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}; + } } else { diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp index ce507ca8d3..6c1c5b1c4d 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp @@ -58,6 +58,11 @@ struct BlockwiseGemmXdlops_pipeline_base static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops; static constexpr index_t KRepeat = KPerThread / KPack; static constexpr index_t KPerInnerLoop = KPack; + static constexpr index_t KGroup = + ((MPerXDL == 16 && MPerXDL == 16 && xdlops_gemm.KPerXdlops == 128) || + (MPerXDL == 32 && MPerXDL == 32 && xdlops_gemm.KPerXdlops == 64)) + ? 2 + : 1; static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL); static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL); diff --git a/include/ck/utility/blkgemmpipe_scheduler.hpp b/include/ck/utility/blkgemmpipe_scheduler.hpp index 39407cb8f6..34c874353d 100644 --- a/include/ck/utility/blkgemmpipe_scheduler.hpp +++ b/include/ck/utility/blkgemmpipe_scheduler.hpp @@ -90,6 +90,17 @@ struct BlockwiseGemmXdlops_pipeline_hotloop_inst static constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock / (BlockSize / WaveSize) / (MPerXDL * NPerXDL * KPerXDL); + static constexpr index_t C_MFMA_Inst_Cycle = []() { + if constexpr(NPerXDL == 16) + { + return KPerXDL == 128 ? 32 : 16; + } + else if constexpr(NPerXDL == 32) + { + return KPerXDL == 64 ? 64 : 32; + } + }(); + static constexpr auto Print() { printf(" Blk/Wave Size: %d, %d, M/N/K PerBlk: %d, %d, %d, M/N/K PerXdl: %d, %d, %d\n", @@ -103,7 +114,7 @@ struct BlockwiseGemmXdlops_pipeline_hotloop_inst KPerXDL); printf(" A/B buffer load inst: %d, %d\n A/B LDS write inst: %d, %d\n A/B LDS read inst: " - "%d, %d\n C MFMA inst: %d\n" + "%d, %d\n C MFMA inst: %d C MFMA cycle: %d\n" "A/B LDS read width: %d, %d, A/B LDS write width: %d, %d, A/B buffer load width: " "%d/ %d\n", A_Buffer_Load_Inst_Num, @@ -113,6 +124,7 @@ struct BlockwiseGemmXdlops_pipeline_hotloop_inst A_LDS_Read_Inst_Num, B_LDS_Read_Inst_Num, C_MFMA_Inst_Num, + C_MFMA_Inst_Cycle, A_LDS_Read_Width, B_LDS_Read_Width, ALDSWriteWidth, From 3ea573acd666ff54c28b4fff58aae23802564c88 Mon Sep 17 00:00:00 2001 From: valarLip <340077269@qq.com> Date: Mon, 12 May 2025 08:10:50 +0000 Subject: [PATCH 11/11] update moe --- ...dlops_b_preshuffle_gufusion_dequant_v1.hpp | 51 +++++++++++-------- 1 file changed, 30 insertions(+), 21 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp index 29750b8baa..c042350f3c 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp @@ -122,6 +122,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< using Base::B_K1; using Base::I0; using Base::I1; + using Base::KGroup; using Base::KRepeat; using Base::xdlops_gemm; using typename Base::HotLoopInstList; @@ -153,9 +154,9 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{}); constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{}); constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{}); - constexpr index_t K2 = KPack; + constexpr index_t K2 = KPack / KGroup; constexpr index_t K1 = 64 / NPerXDL; - constexpr index_t K0 = KRepeat; + constexpr index_t K0 = KRepeat * KGroup; return transform_tensor_descriptor( TileDesc_M0_M1_M2_K{}, @@ -290,12 +291,14 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< block_sync_lds(); static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { - a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(m0, I0, I0, k0, I0, I0), - a_block_buf, - a_thread_desc_, - make_tuple(m0, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(m0, I0, I0, Number{}, I0, I0), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); }); // B VGPR->VGPR dequant @@ -388,12 +391,15 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { - a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(m0, I0, I0, k0, I0, I0), - a_block_buf, - a_thread_desc_, - make_tuple(m0, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(m0, I0, I0, Number{}, I0, I0), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); }); // B VGPR->VGPR dequant @@ -477,12 +483,15 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, KRepeat, 1>{}([&](auto k0) { - a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2, - make_tuple(m0, I0, I0, k0, I0, I0), - a_block_buf, - a_thread_desc_, - make_tuple(m0, I0, I0, k0, I0, I0), - a_thread_buf); + static_for<0, KGroup, 1>{}([&](auto kg0) { + a_thread_copy_.Run( + a_block_desc_m0_m1_m2_k0_k1_k2, + make_tuple(m0, I0, I0, Number{}, I0, I0), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, Number{}), + a_thread_buf); + }); }); }); // B VGPR->VGPR dequant @@ -588,7 +597,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_bdequant_v1< ComputeDataType, decltype(a_block_desc_m0_m1_m2_k0_k1_k2), decltype(a_thread_desc_), - Sequence<1, 1, 1, 1, 1, KPack>, + Sequence<1, 1, 1, 1, 1, KPack / KGroup>, Sequence<0, 1, 2, 3, 4, 5>, 5, A_K1,