From f29480846cd914038939f9f5ad06588ded244302 Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Fri, 20 Dec 2024 15:42:27 +0800 Subject: [PATCH 01/28] Added two kernel for M=32 problem --- .../gemm_multiply_multiply_xdl_fp8.cpp | 42 +++++++++++-------- 1 file changed, 24 insertions(+), 18 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp index cb4f60764e..285f70ba14 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp @@ -24,9 +24,10 @@ template using S = ck::Sequence; -using F16 = ck::half_t; -using FP8 = ck::f8_t; -using F32 = float; +using F16 = ck::half_t; +using FP8 = ck::f8_t; +using F32 = float; +using BF16 = ck::bhalf_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -38,7 +39,7 @@ using CShuffleDataType = F32; using D0DataType = F32; using D1DataType = F32; using DsDataType = ck::Tuple; -using EDataType = F16; +using EDataType = BF16; using A0Layout = Row; using B0Layout = Col; @@ -47,21 +48,23 @@ using D1Layout = Col; using DsLayout = ck::Tuple; using ELayout = Row; -struct MultiplyMultiply -{ - template - __host__ __device__ constexpr void - operator()(E& e, const C& c, const D0& d0, const D1& d1) const; +// struct MultiplyMultiply +// { +// template +// __host__ __device__ constexpr void +// operator()(E& e, const C& c, const D0& d0, const D1& d1) const; - template <> - __host__ __device__ constexpr void operator()( - ck::half_t& e, const float& c, const float& d0, const float& d1) const - { - const float x0_f = c * d0 * d1; +// template <> +// __host__ __device__ constexpr void operator()( +// ck::half_t& e, const float& c, const float& d0, const float& d1) const +// { +// const float x0_f = c * d0 * d1; - e = ck::type_convert(x0_f); - } -}; +// e = ck::type_convert(x0_f); +// } +// }; + +using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply; using PassThrough = ck::tensor_operation::element_wise::PassThrough; @@ -80,7 +83,10 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu ///###### RRR ///< Row, Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; ///###### RCR - < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; + // kernel 1: 256->32x128x128 + < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; + // kernel 2: 128->32x128x128 + < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on int main(int argc, char* argv[]) From e5bc56a4f9a9ce8cd44c21539e114dddd614f228 Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Fri, 20 Dec 2024 15:45:15 +0800 Subject: [PATCH 02/28] Comment the first one --- .../gemm_multiply_multiply_xdl_fp8.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp index 285f70ba14..18f78851dc 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp @@ -84,7 +84,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu ///< Row, Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; ///###### RCR // kernel 1: 256->32x128x128 - < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; + // < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; // kernel 2: 128->32x128x128 < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on From 1fcd33296705a880b6ff1b37ae499e0306b03c83 Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Mon, 23 Dec 2024 20:35:34 +0800 Subject: [PATCH 03/28] Enable multiply_multiply for Scale_Block_M = 1 for deepseek --- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 19 +++++++-- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 40 ++++++++++++------- ...mm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp | 8 ++-- ..._gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 15 ++++--- 4 files changed, 56 insertions(+), 26 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index 9b7849a654..0faf162bf3 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -26,6 +26,7 @@ using S = ck::Sequence; using BF16 = ck::bhalf_t; using FP8 = ck::f8_t; +using F16 = ck::half_t; using F32 = float; using Row = ck::tensor_layout::gemm::RowMajor; @@ -55,7 +56,7 @@ using CDEElementOp = PassThrough; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; -static constexpr ck::index_t Scale_Block_M = 128; +static constexpr ck::index_t Scale_Block_M = 1; static constexpr ck::index_t Scale_Block_N = 128; static constexpr ck::index_t Scale_Block_K = 128; @@ -67,8 +68,8 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, 128, 128, 128, 16, 16, - 16, 16, - 4, 4, + 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, 2, S<1, 32, 1, 8>, S<8, 8, 1>, @@ -187,6 +188,18 @@ int main(int argc, char* argv[]) a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); break; + case 5: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + case 6: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; default: a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index de542866a6..6654d272ab 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -338,11 +338,18 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0){ + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -350,7 +357,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); @@ -462,11 +468,18 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0){ + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -474,7 +487,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp index 480402b7e1..9ddde91145 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp @@ -363,10 +363,10 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3 return false; } - if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock) - { - return false; - } + // if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock) + // { + // return false; + // } if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding || GemmSpec == GemmSpecialization::NKPadding || diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp index 813acfa656..356113733b 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp @@ -1357,7 +1357,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / KPerBlock); - const index_t ScaleSliceSizeM = 1; + const index_t ScaleSliceSizeM = MXdlPerWave; const index_t ScaleSliceSizeN = 1; const index_t ScaleSliceSizeK = 1; @@ -1365,20 +1365,24 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(Number{}, Number{})); constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( - make_tuple(Number{}, Number{})); + make_tuple(Number{}, Number{})); + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + auto a_thread_offset = + get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) % MWaves * MPerXdl; + auto a_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2, + Sequence<1, ScaleSliceSizeK>, Sequence<0, 1>, 1, 1, 1, false>( - a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM, 0)); + a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); auto b_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2( b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0)); - constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1); + constexpr auto a_scale_thread_slice_copy_step = + make_tuple(make_multi_index(MWaves * MPerXdl, 0), make_multi_index(-MPerBlock, 1)); constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1); const index_t num_k_block_per_scale = ScaleBlockK / KPerBlock; From f728087c61604b0c76dddf557c8e0bc0da97eebb Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Wed, 25 Dec 2024 23:26:17 +0800 Subject: [PATCH 04/28] Modify the a_thread offset since the A data load is different from B. --- .../grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp index 356113733b..a806003297 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp @@ -1368,8 +1368,10 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(Number{}, Number{})); constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); - auto a_thread_offset = - get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) % MWaves * MPerXdl; + // auto a_thread_offset = + // get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) % MWaves * MPerXdl; + + auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 128) * MPerXdl; auto a_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2 Date: Thu, 26 Dec 2024 07:27:49 +0000 Subject: [PATCH 05/28] edit fp8 ab scale for Scale_Block_M=1 --- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 37 ++--- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 54 ++++++-- ..._gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 129 +++++++++--------- 3 files changed, 124 insertions(+), 96 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index 0faf162bf3..f4fa6ab2ab 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -26,7 +26,6 @@ using S = ck::Sequence; using BF16 = ck::bhalf_t; using FP8 = ck::f8_t; -using F16 = ck::half_t; using F32 = float; using Row = ck::tensor_layout::gemm::RowMajor; @@ -68,11 +67,11 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, 128, 128, 128, 16, 16, - 32, 32, - 2, 2, + 16, 16, + 4, 4, 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>, S<8, 8, 1>, + 1, 2, S<1, 32, 1, 8>, S<8>, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; // clang-format on @@ -83,9 +82,9 @@ int main(int argc, char* argv[]) bool time_kernel = false; // GEMM shape - ck::index_t M = 3840; - ck::index_t N = 4096; - ck::index_t K = 4096; + ck::index_t M = 128; + ck::index_t N = 1024; + ck::index_t K = 1024; ck::index_t StrideA = K; ck::index_t StrideB = K; @@ -101,7 +100,7 @@ int main(int argc, char* argv[]) init_method = std::stoi(argv[2]); time_kernel = std::stoi(argv[3]); } - else if(argc == 10) + else if(argc == 7) { do_verification = std::stoi(argv[1]); init_method = std::stoi(argv[2]); @@ -111,9 +110,9 @@ int main(int argc, char* argv[]) N = std::stoi(argv[5]); K = std::stoi(argv[6]); - StrideA = std::stoi(argv[7]); - StrideB = std::stoi(argv[8]); - StrideE = std::stoi(argv[9]); + StrideA = K; + StrideB = K; + StrideE = N; } else { @@ -185,20 +184,10 @@ int main(int argc, char* argv[]) case 4: a0_m_k.GenerateTensorValue(GeneratorTensor_1{}); b0_k_n.GenerateTensorValue(GeneratorTensor_1{}); + // a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - break; - case 5: - a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); - b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); - a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); - b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - break; - case 6: - a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); - b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); - a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + // b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); break; default: a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index 6654d272ab..2f195fa058 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale + KPack, + true> { using Base = BlockwiseGemmXdlops_pipeline_base; + KPack, + true>; using Base::I0; using Base::KRepeat; using Base::xdlops_gemm; @@ -338,18 +340,32 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0){ + // a_scale_thread_copy.Run(a_scale_grid_desc, + // a_scale_grid_buf, + // a_scale_thread_desc, + // make_tuple(I0, I0), + // a_scale_thread_buf); + + static_for<0, MRepeat, 1>{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, a_scale_grid_buf, a_scale_thread_desc, make_tuple(m0, I0), a_scale_thread_buf); - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); }); - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -357,6 +373,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0){ + // a_scale_thread_copy.Run(a_scale_grid_desc, + // a_scale_grid_buf, + // a_scale_thread_desc, + // make_tuple(I0, I0), + // a_scale_thread_buf); + + static_for<0, MRepeat, 1>{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, a_scale_grid_buf, a_scale_thread_desc, make_tuple(m0, I0), a_scale_thread_buf); - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<0>{})); + a_scale_thread_copy_step.At(Number<0>{})); }); - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -487,6 +518,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}, Number{})); + + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); + auto a_thread_offset = + get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl; + // auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 128) * MPerXdl; constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( - make_tuple(Number{}, Number{})); + make_tuple(Number{}, Number{})); - constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); - // auto a_thread_offset = - // get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) % MWaves * MPerXdl; - - auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 128) * MPerXdl; - auto a_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2( - a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); + a_scale_grid_desc_am_ak, + make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); auto b_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2( b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0)); + // constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1); constexpr auto a_scale_thread_slice_copy_step = - make_tuple(make_multi_index(MWaves * MPerXdl, 0), make_multi_index(-MPerBlock, 1)); + make_tuple(make_multi_index(MWaves * MPerXdl, 0), + make_multi_index(-MPerBlock, 0), + make_multi_index(-MPerBlock, 1)); constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1); const index_t num_k_block_per_scale = ScaleBlockK / KPerBlock; @@ -1443,24 +1447,28 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); + + // transposed XDL + // // TODO: hacky, fix it! + constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 = + blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + // // TODO: hacky, fix it! + // only used to get lengths + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp = + blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0); + constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1); + constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2); + constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3); + constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4); + constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5); + constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6); + constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7); - // TODO: hacky, fix it! - constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 = - blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(); - // TODO: hacky, fix it! - // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths - constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp = - blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(); - constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0); - constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1); - constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2); - constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3); - constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4); - constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5); - 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); constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); @@ -1469,24 +1477,24 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 static_cast(p_shared), c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); - constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor( + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor( c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, make_tuple( make_freeze_transform(I0), make_unmerge_transform(make_tuple( Number{}, // M0 (MXdlPerWave) per shuffle M1, // M1 = MWave - M2, // M2 * M3 * M4 = MPerXdl - M3, - M4)), + M2)), // M2 = MPerXdl make_freeze_transform(I0), make_unmerge_transform(make_tuple( Number{}, // N0 (NXdlPerWave) per shuffle N1, // N1 = NWave - N2))), // N2 = NPerXdl + N2, // N2 * N3 * N4 = NPerXdl + N3, + N4))), make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), make_tuple( - Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{})); + Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{})); // calculate origin of thread output tensor on global memory // blockwise GEMM c matrix starting index @@ -1496,57 +1504,57 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0]; const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1]; - const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor = + const auto m_thread_data_on_block_to_m0_m1_m2_adaptor = make_single_stage_tensor_adaptor( - make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))), - make_tuple(Sequence<0, 1, 2, 3, 4>{}), - make_tuple(Sequence<0>{})); - - const auto m_thread_data_on_block_idx = - m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex( - make_multi_index(m_thread_data_on_block)); - - const auto n_thread_data_on_block_to_n0_n1_n2_adaptor = - make_single_stage_tensor_adaptor( - make_tuple(make_merge_transform(make_tuple(N0, N1, N2))), + make_tuple(make_merge_transform(make_tuple(M0, M1, M2))), make_tuple(Sequence<0, 1, 2>{}), make_tuple(Sequence<0>{})); + const auto m_thread_data_on_block_idx = + m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex( + make_multi_index(m_thread_data_on_block)); + + const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))), + make_tuple(Sequence<0, 1, 2, 3, 4>{}), + make_tuple(Sequence<0>{})); + const auto n_thread_data_on_block_idx = - n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex( + n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex( make_multi_index(n_thread_data_on_block)); // shuffle: threadwise copy C from VGPR to LDS auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3, + N2, + I1, + N4>, Sequence<0, 1, 2, 3, 4, 5, 6, 7>, 7, 1, InMemoryDataOperationEnum::Set, 1, true>{ - c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, make_multi_index(0, 0, m_thread_data_on_block_idx[I1], n_thread_data_on_block_idx[I1], m_thread_data_on_block_idx[I2], - m_thread_data_on_block_idx[I3], - m_thread_data_on_block_idx[I4], - n_thread_data_on_block_idx[I2]), - ck::tensor_operation::element_wise::PassThrough{}}; + n_thread_data_on_block_idx[I2], + n_thread_data_on_block_idx[I3], + n_thread_data_on_block_idx[I4]), + tensor_operation::element_wise::PassThrough{}}; using EDataType = CDataType; @@ -1628,18 +1636,17 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)), c_element_op}; - // space filling curve for threadwise C in VGPR constexpr auto sfc_c_vgpr = - SpaceFillingCurve, + SpaceFillingCurve, Sequence<0, 1, 2, 3, 4, 5, 6, 7>, Sequence>{}; + N2, + 1, + N4>>{}; constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess(); @@ -1659,10 +1666,10 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 block_sync_lds(); // 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, + c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4, sfc_c_vgpr.GetIndexTupleOfNumber(access_id), c_thread_buf, - c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, c_shuffle_block_buf); // make sure it's safe to read from LDS From d58d55e5404b1cb7e9c382d0ebfa65cf94d59154 Mon Sep 17 00:00:00 2001 From: chenjun Date: Thu, 26 Dec 2024 07:38:14 +0000 Subject: [PATCH 06/28] edit GemmSpec to MNKPadding --- .../gemm_multiply_multiply_xdl_fp8_ab_scale.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index f4fa6ab2ab..6e103bfedd 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -53,7 +53,7 @@ using AElementOp = PassThrough; using BElementOp = PassThrough; using CDEElementOp = PassThrough; -static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; static constexpr ck::index_t Scale_Block_M = 1; static constexpr ck::index_t Scale_Block_N = 128; From 9dac9713cc4bf82061edc4654d1d6836b5694a67 Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Fri, 10 Jan 2025 14:14:27 +0800 Subject: [PATCH 07/28] enable blockwise pipelie v1 and v2. v1 is work for small K. --- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 2 +- ...kwise_gemm_pipeline_xdlops_v1_ab_scale.hpp | 62 +++++++++---- ...kwise_gemm_pipeline_xdlops_v2_ab_scale.hpp | 93 ++++++++++++++----- 3 files changed, 115 insertions(+), 42 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index 6e103bfedd..55bdb76b9a 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -72,7 +72,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ 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>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp index 821bbb0051..58823214fb 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp @@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale + KPack, + true> { using Base = BlockwiseGemmXdlops_pipeline_base; + KPack, + true>; using Base::I0; using Base::KRepeat; using Base::xdlops_gemm; @@ -231,11 +233,26 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -243,7 +260,6 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); }); - a_scale_thread_copy.Run(a_scale_grid_desc, - a_scale_grid_buf, - a_scale_thread_desc, - make_tuple(I0, I0), - a_scale_thread_buf); + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -336,7 +367,6 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp index 40fa776484..5ed36ac1c0 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp @@ -96,7 +96,8 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale + KPack, + true> { using Base = BlockwiseGemmXdlops_pipeline_base; + KPack, + true>; using Base::I0; using Base::KRepeat; using Base::xdlops_gemm; @@ -270,11 +272,26 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -282,7 +299,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); }); - a_scale_thread_copy.Run(a_scale_grid_desc, - a_scale_grid_buf, - a_scale_thread_desc, - make_tuple(I0, I0), - a_scale_thread_buf); + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -378,8 +409,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); }); - a_scale_thread_copy.Run(a_scale_grid_desc, - a_scale_grid_buf, - a_scale_thread_desc, - make_tuple(I0, I0), - a_scale_thread_buf); + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if(num_loop_per_scale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } b_scale_thread_copy.Run(b_scale_grid_desc, b_scale_grid_buf, @@ -471,7 +515,6 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); @@ -586,7 +629,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{}) += c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[I0]) * + type_convert(a_scale_thread_buf[m0]) * type_convert(b_scale_thread_buf[I0]); }); }); From 363b6744d79bbaaa88edd1d8369c5a914901021a Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Tue, 14 Jan 2025 12:58:44 +0800 Subject: [PATCH 08/28] add instance for gemm_ab_scale --- .../gpu/gemm_ab_scale.hpp | 46 +-- ...le_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp | 42 +-- ...k_mn_128_128_128_comp_default_instance.cpp | 6 +- ..._mn_128_128_128_comp_kpadding_instance.cpp | 6 +- ...n_128_128_128_comp_mnkpadding_instance.cpp | 6 +- ...mn_128_128_128_comp_mnpadding_instance.cpp | 6 +- ...mn_128_128_128_mem_v1_default_instance.cpp | 6 +- ...n_128_128_128_mem_v1_kpadding_instance.cpp | 6 +- ...128_128_128_mem_v1_mnkpadding_instance.cpp | 6 +- profiler/src/CMakeLists.txt | 302 +++++++++--------- profiler/src/profile_gemm_ab_scale.cpp | 23 +- 11 files changed, 237 insertions(+), 218 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp index 7553d5e76e..fb9d6fb078 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp @@ -17,7 +17,7 @@ namespace tensor_operation { namespace device { namespace instance { #if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8)) -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( std::vector, @@ -28,14 +28,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( std::vector, @@ -46,14 +46,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances( std::vector, @@ -64,14 +64,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances( std::vector, @@ -82,14 +82,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( std::vector, @@ -100,14 +100,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( std::vector, @@ -118,14 +118,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances( std::vector, @@ -136,7 +136,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -163,7 +163,7 @@ struct DeviceOperationInstanceFactory, CDataType, - 128, + 1, 128, 128, ck::tensor_operation::element_wise::PassThrough, @@ -180,7 +180,7 @@ struct DeviceOperationInstanceFactory, CDataType, - 128, + 1, 128, 128, ck::tensor_operation::element_wise::PassThrough, @@ -198,20 +198,20 @@ struct DeviceOperationInstanceFactory && is_same_v && is_same_v) { - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( op_ptrs); - add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( + add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances( op_ptrs); } } diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp index 3a7df8d974..d45df83042 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp @@ -34,7 +34,7 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; template -using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances = std::tuple< +using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances = std::tuple< // clang-format off //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| 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| @@ -45,15 +45,15 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances = // Spill in current compiler // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, 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, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> // clang-format on >; template -using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = std::tuple< +using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances = std::tuple< // clang-format off //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| 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| @@ -61,22 +61,22 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = s //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, // Memory friendly - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8> + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8> // clang-format on >; } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp index ab83c7eb3e..aebffc01f2 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances{}); + device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp index dfb1bb6e2d..31fffae080 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances{}); + device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp index d2d3ebe81e..7b61cadea9 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances{}); + device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp index f6ce77a751..aa9b0c828a 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances{}); + device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp index e2205ad728..a466bb23dd 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances{}); } diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp index 5c0a6eb00d..a1722fade8 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances{}); } diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp index cc1a03b060..0f91fdefcc 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( +void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances( std::vector, @@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd F32, Tuple<>, BF16, - 128, + 1, 128, 128, PassThrough, @@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd { add_device_operation_instances( instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances{}); } diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 35e91f8172..c789ac8949 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -1,87 +1,87 @@ # ckProfiler set(PROFILER_SOURCES profiler.cpp - profile_gemm.cpp - profile_reduce.cpp - profile_groupnorm_bwd_data.cpp - profile_groupnorm_fwd.cpp - profile_layernorm_bwd_data.cpp - profile_layernorm_bwd_gamma_beta.cpp - profile_groupnorm_bwd_gamma_beta.cpp - profile_layernorm_fwd.cpp - profile_max_pool2d_fwd.cpp - profile_pool3d_fwd.cpp - profile_avg_pool3d_bwd.cpp - profile_max_pool3d_bwd.cpp - profile_avg_pool2d_bwd.cpp - profile_max_pool2d_bwd.cpp - profile_softmax.cpp - profile_batchnorm_fwd.cpp - profile_batchnorm_bwd.cpp - profile_batchnorm_infer.cpp - profile_conv_tensor_rearrange.cpp - profile_transpose.cpp - profile_permute_scale.cpp + # profile_gemm.cpp + # profile_reduce.cpp + # profile_groupnorm_bwd_data.cpp + # profile_groupnorm_fwd.cpp + # profile_layernorm_bwd_data.cpp + # profile_layernorm_bwd_gamma_beta.cpp + # profile_groupnorm_bwd_gamma_beta.cpp + # profile_layernorm_fwd.cpp + # profile_max_pool2d_fwd.cpp + # profile_pool3d_fwd.cpp + # profile_avg_pool3d_bwd.cpp + # profile_max_pool3d_bwd.cpp + # profile_avg_pool2d_bwd.cpp + # profile_max_pool2d_bwd.cpp + # profile_softmax.cpp + # profile_batchnorm_fwd.cpp + # profile_batchnorm_bwd.cpp + # profile_batchnorm_infer.cpp + # profile_conv_tensor_rearrange.cpp + # profile_transpose.cpp + # profile_permute_scale.cpp ) -if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) - list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) - endif() - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) - endif() - list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") +# if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) +# list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) +# list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) +# endif() +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) +# list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) +# list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) +# endif() +# list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) endif() - list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) - list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) +# list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) +# list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) +# list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) +# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) +# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) +# list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) +# list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) -endif() +# endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) - endif() - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -endif() +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) +# endif() +# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +# endif() -if(DL_KERNELS) - list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -endif() +# if(DL_KERNELS) +# list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +# endif() set(PROFILER_EXECUTABLE ckProfiler) @@ -93,89 +93,89 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress) endif() -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) -if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) - endif() - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) - endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") +# if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) +# endif() +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) +# endif() +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) -endif() +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) +# endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) - endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -endif() +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) +# endif() +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +# endif() -if(DL_KERNELS) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -endif() +# if(DL_KERNELS) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +# endif() rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) diff --git a/profiler/src/profile_gemm_ab_scale.cpp b/profiler/src/profile_gemm_ab_scale.cpp index 56c8b5e7a1..1b46b07864 100644 --- a/profiler/src/profile_gemm_ab_scale.cpp +++ b/profiler/src/profile_gemm_ab_scale.cpp @@ -32,8 +32,10 @@ enum struct GemmDataType enum struct ScaleBlockTile { Tile_128_128_128, // 0 + Tile_1_128_128, // 1 }; + #define OP_NAME "gemm_ab_scale" #define OP_DESC "GEMM_AB_Scale" @@ -154,8 +156,25 @@ int profile_gemm_ab_scale(int argc, char* argv[]) return pass ? 0 : 1; }; + // if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && + // scale_block_tile == ScaleBlockTile::Tile_128_128_128) + // { + // return profile(F8{}, + // F32{}, + // F8{}, + // F32{}, + // F8{}, + // F32{}, + // BF16{}, + // ck::Number<128>{}, + // ck::Number<128>{}, + // ck::Number<128>{}, + // Row{}, + // Col{}, + // Row{}); + // } if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && - scale_block_tile == ScaleBlockTile::Tile_128_128_128) + scale_block_tile == ScaleBlockTile::Tile_1_128_128) { return profile(F8{}, F32{}, @@ -164,7 +183,7 @@ int profile_gemm_ab_scale(int argc, char* argv[]) F8{}, F32{}, BF16{}, - ck::Number<128>{}, + ck::Number<1>{}, ck::Number<128>{}, ck::Number<128>{}, Row{}, From 7ae141faf19ce5f8363d50b43f7419653bbb67d5 Mon Sep 17 00:00:00 2001 From: mtgu0705 Date: Tue, 14 Jan 2025 13:12:25 +0800 Subject: [PATCH 09/28] fix cmakelist of ckProfiler --- profiler/src/CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index c789ac8949..bdd41831c3 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -93,7 +93,7 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress) endif() -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) From 3d4ad53452d967080489ee7312ce3d85d772e822 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Wed, 19 Feb 2025 10:31:21 +0000 Subject: [PATCH 10/28] optimize blockscale gemm. todo: reduce vgpr usage --- CMakeLists.txt | 14 +-- .../65_gemm_multiply_multiply/CMakeLists.txt | 8 +- .../gemm_multiply_multiply_xdl_fp8.cpp | 15 ++- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 10 +- ...kwise_gemm_pipeline_xdlops_v1_ab_scale.hpp | 18 +-- ...kwise_gemm_pipeline_xdlops_v2_ab_scale.hpp | 24 ++-- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 116 ++++++++++++------ ...mm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp | 3 +- .../gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp | 68 ++-------- ..._gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 94 ++++---------- ...mn_128_128_128_mem_v1_default_instance.cpp | 2 +- ...n_128_128_128_mem_v1_kpadding_instance.cpp | 2 +- ...128_128_128_mem_v1_mnkpadding_instance.cpp | 2 +- profiler/src/profile_gemm_ab_scale.cpp | 1 - script/cmake-ck-dev.sh | 2 +- 15 files changed, 174 insertions(+), 205 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index e90f893de0..cd03961477 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -246,13 +246,13 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500500000) add_compile_options("SHELL: -mllvm --lsr-drop-solution=1") endif() endif() -if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) - check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED) - if(HAS_ENABLE_POST_MISCHED) - message("Adding the enable-post-misched=0 compiler flag") - add_compile_options("SHELL: -mllvm -enable-post-misched=0") - endif() -endif() +# if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) +# check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED) +# if(HAS_ENABLE_POST_MISCHED) +# message("Adding the enable-post-misched=0 compiler flag") +# add_compile_options("SHELL: -mllvm -enable-post-misched=0") +# endif() +# endif() set(check-coerce) check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce) if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132) diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 55c884246f..4e916df773 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -1,4 +1,10 @@ add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp) add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp) -add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) \ No newline at end of file +add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) + +list(APPEND TILE_EXAPMLE_BLOCKSCALE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) +list(APPEND TILE_EXAPMLE_BLOCKSCALE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) + +target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${TILE_EXAPMLE_BLOCKSCALE_COMPILE_OPTIONS}) +target_compile_options(example_gemm_multiply_multiply_xdl_fp8 PRIVATE ${TILE_EXAPMLE_BLOCKSCALE_COMPILE_OPTIONS}) \ No newline at end of file diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp index 18f78851dc..3c12b733ac 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp @@ -72,7 +72,7 @@ using AElementOp = PassThrough; using BElementOp = PassThrough; using CDEElementOp = MultiplyMultiply; -static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding; +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3 // clang-format off @@ -86,7 +86,18 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu // kernel 1: 256->32x128x128 // < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; // kernel 2: 128->32x128x128 - < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; + < Row, Col, DsLayout, ELayout, + A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, + AElementOp, BElementOp, CDEElementOp, GemmSpec, + 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>, S<8, 8, 1>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index 55bdb76b9a..76ebe93c36 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -53,7 +53,7 @@ using AElementOp = PassThrough; using BElementOp = PassThrough; using CDEElementOp = PassThrough; -static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr ck::index_t Scale_Block_M = 1; static constexpr ck::index_t Scale_Block_N = 128; @@ -67,12 +67,12 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, 128, 128, 128, 16, 16, - 16, 16, - 4, 4, + 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, 2, S<1, 32, 1, 8>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; + 1, 1, S<1, 32, 1, 8>, S<8>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp index 58823214fb..02bc0a941c 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp @@ -240,18 +240,18 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{})); + a_scale_thread_copy_step.At(Number<0>{})); }); if(num_loop_per_scale == 1) { a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<2>{})); + a_scale_thread_copy_step.At(Number<2>{})); } else { a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + a_scale_thread_copy_step.At(Number<1>{})); } b_scale_thread_copy.Run(b_scale_grid_desc, @@ -346,19 +346,19 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); }); if(num_loop_per_scale == 1) { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<2>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); } else { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); } b_scale_thread_copy.Run(b_scale_grid_desc, diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp index 5ed36ac1c0..c8ad9c5b02 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp @@ -388,19 +388,19 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); }); if(num_loop_per_scale == 1) { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<2>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); } else { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); } b_scale_thread_copy.Run(b_scale_grid_desc, @@ -494,19 +494,19 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); }); if(num_loop_per_scale == 1) { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<2>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); } else { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); } b_scale_thread_copy.Run(b_scale_grid_desc, diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index 2f195fa058..dbbc27d706 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -179,11 +179,11 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale( a_thread_desc_.GetElementSpaceSize()); auto b_thread_buf = make_static_buffer( @@ -332,6 +333,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale( b_scale_thread_desc.GetElementSpaceSize()); + auto c_scale_thread_buf = make_static_buffer( + c_scale_thread_desc.GetElementSpaceSize()); // Global prefetch 1 a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); @@ -340,12 +343,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, a_scale_grid_buf, @@ -356,7 +353,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{})); }); - if(num_loop_per_scale == 1) + if constexpr(NumKBlockPerScale == 1) { a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); @@ -375,6 +372,11 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { + c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0]; + }); + // Local prefill 1 a_blockwise_copy.RunWrite(a_block_desc, a_block_buf); b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); @@ -386,10 +388,42 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + // Initialize C c_thread_buf.Clear(); - auto c_thread_buf_per_scale = remove_cvref_t(); + StaticBufferTupleOfVector + c_thread_buf_per_scale; // Local prefetch 1 block_sync_lds(); @@ -432,7 +466,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - c_thread_buf_per_scale.Clear(); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); static_for<0, KRepeat, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -453,19 +490,23 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale( a_thread_vec.template AsType(), b_thread_vec.template AsType(), - c_thread_buf_per_scale.GetVectorTypeReference(I0)); + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); c_thread_buf(Number{}) += - c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[m0]) * - type_convert(b_scale_thread_buf[I0]); + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); }); }); }); + static_for<0, MRepeat, 1>{}([&](auto m0) { + c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0]; + }); + block_sync_lds(); static_for<0, KRepeat, 1>{}([&](auto k) { static_for<0, MRepeat, 1>{}([&](auto m0) { @@ -485,11 +526,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -497,19 +533,19 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); }); - if(num_loop_per_scale == 1) + if constexpr(NumKBlockPerScale == 1) { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<2>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); } else { - a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, - a_scale_thread_copy_step.At(Number<1>{})); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); } b_scale_thread_copy.Run(b_scale_grid_desc, @@ -518,7 +554,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - c_thread_buf_per_scale.Clear(); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); static_for<0, KRepeat, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -551,15 +589,15 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale( a_thread_vec.template AsType(), b_thread_vec.template AsType(), - c_thread_buf_per_scale.GetVectorTypeReference(I0)); + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); c_thread_buf(Number{}) += - c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[m0]) * - type_convert(b_scale_thread_buf[I0]); + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); }); }); }); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp index 9ddde91145..adc0f749e2 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp @@ -363,7 +363,8 @@ struct DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3 return false; } - // if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock) + // if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != + // KPerBlock) // { // return false; // } diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp index e5a31f8d1f..d1f6cdde22 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp @@ -686,40 +686,19 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 // in some cases. else if constexpr(is_same::value) { - constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1 - ? 1 - : 32 * 4 / KPerBlock / sizeof(LDSTypeA); - constexpr auto a_lds_block_desc = make_naive_tensor_descriptor( - make_tuple( - AK0Number * Number{}, Number{}, AK1Number), - make_tuple(AK1Number, Number{}, I1)); + constexpr auto a_lds_block_desc = + make_naive_tensor_descriptor(make_tuple(AK0Number, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( a_lds_block_desc, - make_tuple(make_xor_with_modulo_transform(make_tuple( - Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), make_pass_through_transform(AK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor( - a_lds_block_desc_permuted, - make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number{})), - make_pass_through_transform(Number{}), - make_pass_through_transform(AK1Number)), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), - make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); - - constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( - a_lds_block_desc_ak0_mldslayer_m_ak1, - make_tuple(make_pass_through_transform(AK0Number), - make_merge_transform_v3_division_mod( - make_tuple(Number{}, Number{})), - make_pass_through_transform(AK1Number)), - make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); - - return a_lds_block_desc_ak0_m_ak1; + return a_lds_block_desc_permuted; } else // ColumnMajor A { @@ -821,42 +800,19 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 } else if constexpr(is_same::value) { - // NLdsLayer * K0 as logical Bank - constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1 - ? 1 - : 32 * 4 / KPerBlock / sizeof(LDSTypeB); - ; - constexpr auto b_lds_block_desc = make_naive_tensor_descriptor( - make_tuple( - BK0Number * Number{}, Number{}, BK1Number), - make_tuple(BK1Number, Number{}, I1)); + constexpr auto b_lds_block_desc = + make_naive_tensor_descriptor(make_tuple(BK0Number, Number{}, BK1Number), + make_tuple(BK1Number, Number{}, I1)); constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor( b_lds_block_desc, - make_tuple(make_xor_with_modulo_transform(make_tuple( - Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), make_pass_through_transform(BK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor( - b_lds_block_desc_permuted, - make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number{})), - make_pass_through_transform(Number{}), - make_pass_through_transform(BK1Number)), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), - make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); - - constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor( - b_lds_block_desc_bk0_nldslayer_n_bk1, - make_tuple(make_pass_through_transform(BK0Number), - make_merge_transform_v3_division_mod( - make_tuple(Number{}, Number{})), - make_pass_through_transform(BK1Number)), - make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); - - return b_lds_block_desc_bk0_n_bk1; + return b_lds_block_desc_permuted; } else // RowMajor B { diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp index a1eb63f401..2710ab7a48 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp @@ -656,40 +656,19 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 // in some cases. else if constexpr(is_same::value) { - constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1 - ? 1 - : 32 * 4 / KPerBlock / sizeof(LDSTypeA); - constexpr auto a_lds_block_desc = make_naive_tensor_descriptor( - make_tuple( - AK0Number * Number{}, Number{}, AK1Number), - make_tuple(AK1Number, Number{}, I1)); + constexpr auto a_lds_block_desc = + make_naive_tensor_descriptor(make_tuple(AK0Number, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( a_lds_block_desc, - make_tuple(make_xor_with_modulo_transform(make_tuple( - Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), make_pass_through_transform(AK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor( - a_lds_block_desc_permuted, - make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number{})), - make_pass_through_transform(Number{}), - make_pass_through_transform(AK1Number)), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), - make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); - - constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( - a_lds_block_desc_ak0_mldslayer_m_ak1, - make_tuple(make_pass_through_transform(AK0Number), - make_merge_transform_v3_division_mod( - make_tuple(Number{}, Number{})), - make_pass_through_transform(AK1Number)), - make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); - - return a_lds_block_desc_ak0_m_ak1; + return a_lds_block_desc_permuted; } else // ColumnMajor A { @@ -791,42 +770,19 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 } else if constexpr(is_same::value) { - // NLdsLayer * K0 as logical Bank - constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1 - ? 1 - : 32 * 4 / KPerBlock / sizeof(LDSTypeB); - ; - constexpr auto b_lds_block_desc = make_naive_tensor_descriptor( - make_tuple( - BK0Number * Number{}, Number{}, BK1Number), - make_tuple(BK1Number, Number{}, I1)); + constexpr auto b_lds_block_desc = + make_naive_tensor_descriptor(make_tuple(BK0Number, Number{}, BK1Number), + make_tuple(BK1Number, Number{}, I1)); constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor( b_lds_block_desc, - make_tuple(make_xor_with_modulo_transform(make_tuple( - Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), make_pass_through_transform(BK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor( - b_lds_block_desc_permuted, - make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number{})), - make_pass_through_transform(Number{}), - make_pass_through_transform(BK1Number)), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), - make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); - - constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor( - b_lds_block_desc_bk0_nldslayer_n_bk1, - make_tuple(make_pass_through_transform(BK0Number), - make_merge_transform_v3_division_mod( - make_tuple(Number{}, Number{})), - make_pass_through_transform(BK1Number)), - make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), - make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); - - return b_lds_block_desc_bk0_n_bk1; + return b_lds_block_desc_permuted; } else // RowMajor B { @@ -1363,15 +1319,19 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed( make_tuple(Number{}, Number{})); - + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl; - // auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 128) * MPerXdl; + // auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / + // 128) * MPerXdl; constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( - make_tuple(Number{}, Number{})); + make_tuple(Number{}, Number{})); + + constexpr auto c_scale_thread_desc = make_naive_tensor_descriptor_packed(make_tuple( + Number{}, Number{}, Number{})); auto a_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2( - a_scale_grid_desc_am_ak, + a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); auto b_scale_thread_copy = @@ -1407,9 +1367,9 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_multi_index(-MPerBlock, 1)); constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1); - const index_t num_k_block_per_scale = ScaleBlockK / KPerBlock; + constexpr auto NumKBlockPerScale = ScaleBlockK / KPerBlock; - blockwise_gemm_pipeline.template Run( + blockwise_gemm_pipeline.template Run( a_grid_desc_ak0_m_ak1, a_block_desc_ak0_m_ak1, a_blockwise_copy, @@ -1422,6 +1382,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 b_grid_buf, b_block_buf, b_block_slice_copy_step, + + c_scale_thread_desc, c_thread_buf, a_scale_grid_desc_am_ak, @@ -1436,8 +1398,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 b_scale_grid_buf, b_scale_thread_slice_copy_step, - num_k_block_main_loop, - num_k_block_per_scale); + num_k_block_main_loop); // shuffle C and write out { @@ -1447,7 +1408,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); - + // transposed XDL // // TODO: hacky, fix it! constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 = @@ -1467,9 +1428,6 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6); constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7); - - - constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp index a466bb23dd..569911e3de 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp @@ -29,7 +29,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_i add_device_operation_instances( instances, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); + GemmDefault>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp index a1722fade8..d1e5b6b535 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp @@ -29,7 +29,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_ add_device_operation_instances( instances, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); + GemmKPadding>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp index 0f91fdefcc..f51fa43bcb 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp @@ -29,7 +29,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpaddin add_device_operation_instances( instances, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); + GemmMNKPadding>{}); } } // namespace instance diff --git a/profiler/src/profile_gemm_ab_scale.cpp b/profiler/src/profile_gemm_ab_scale.cpp index 1b46b07864..c17ab74536 100644 --- a/profiler/src/profile_gemm_ab_scale.cpp +++ b/profiler/src/profile_gemm_ab_scale.cpp @@ -35,7 +35,6 @@ enum struct ScaleBlockTile Tile_1_128_128, // 1 }; - #define OP_NAME "gemm_ab_scale" #define OP_DESC "GEMM_AB_Scale" diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh index 6089fc7a7e..7d1d5e60e9 100755 --- a/script/cmake-ck-dev.sh +++ b/script/cmake-ck-dev.sh @@ -17,7 +17,7 @@ fi cmake \ -D CMAKE_PREFIX_PATH=/opt/rocm/ \ -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \ --D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \ +-D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \ -D CMAKE_BUILD_TYPE=Release \ -D BUILD_DEV=ON \ -D GPU_TARGETS=$GPU_TARGETS \ From b9a97f4d8df37f93f4fc2ec8820595bd4c6900b1 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Fri, 21 Feb 2025 05:33:29 +0000 Subject: [PATCH 11/28] fix a correctness bug --- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 18 ++++-- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 5 +- ...mm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp | 58 ++++++++++++++++--- ..._gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 14 +++-- profiler/src/CMakeLists.txt | 8 +-- 5 files changed, 81 insertions(+), 22 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index 76ebe93c36..b5ad01e0a5 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -182,13 +182,19 @@ int main(int argc, char* argv[]) b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); break; case 4: - a0_m_k.GenerateTensorValue(GeneratorTensor_1{}); - b0_k_n.GenerateTensorValue(GeneratorTensor_1{}); - // a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); // b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); break; + case 5: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + // b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; default: a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); @@ -241,12 +247,14 @@ int main(int argc, char* argv[]) "not support this GEMM problem"); } - float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50}); - std::size_t flop = std::size_t(2) * M * N * K; std::size_t num_btype = sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; + int rotating_buf = (512*1024*1024 + num_btype-1)/num_btype; + + float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf}); + float tflops = static_cast(flop) / 1.E9 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time; diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index dbbc27d706..f310683a48 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -193,7 +193,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale @@ -370,7 +370,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto m0) { @@ -414,6 +413,8 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale 1) - hipGetErrorString(hipMemsetAsync(arg.p_c_grid, - 0, - arg.M * arg.N * sizeof(CDataType), - stream_config.stream_id_)); + if(stream_config.flush_cache) + { + Argument arg_ = arg; - ave_time = launch_and_time_kernel( - stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg); + const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1( + arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0); + const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1( + arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0); + + auto size_a_buffer = + a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType); + auto size_b_buffer = + b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType); + + ck::utility::RotatingMemWrapper rotating_mem( + arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer); + rotating_mem.Print(); + + auto run_flush_cache = [&]() { + // flush icache + ck::utility::flush_icache(); + // rotating mem + rotating_mem.Next(); + // clear c mem + if(arg_.KBatch > 1) + hipGetErrorString(hipMemsetAsync(arg_.p_c_grid, + 0, + arg_.M * arg_.N * sizeof(CDataType), + stream_config.stream_id_)); + }; + + ave_time = ck::utility::launch_and_time_kernel_with_preprocess( + stream_config, + run_flush_cache, + kernel, + dim3(gdx, gdy, gdz), + dim3(BlockSize), + 0, + arg_); + } + else + { + if(arg.KBatch > 1) + hipGetErrorString(hipMemsetAsync(arg.p_c_grid, + 0, + arg.M * arg.N * sizeof(CDataType), + stream_config.stream_id_)); + + ave_time = launch_and_time_kernel( + stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg); + } }; constexpr index_t minimum_occupancy = diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp index 2710ab7a48..bcb2f8718c 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp @@ -225,7 +225,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{})); } - __device__ static auto MakeAGridDescriptor_AK0_M_AK1( + __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1( index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0) { const auto a_grid_desc_mraw_kraw = [&]() { @@ -307,7 +307,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 } } - __device__ static auto MakeBGridDescriptor_BK0_N_BK1( + __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1( index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0) { const auto b_grid_desc_nraw_kraw = [&]() { @@ -422,6 +422,13 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 } }(); + // pad M and N + return transform_tensor_descriptor(c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); +#if 0 using GemmSpecialization = tensor_operation::device::GemmSpecialization; if constexpr(GemmSpec == GemmSpecialization::MNPadding || @@ -459,6 +466,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 // not pad M or N return c_grid_desc_mraw_nraw; } +#endif } __host__ __device__ static auto MakeDsGridDescriptor_M_N( @@ -1324,8 +1332,6 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl; - // auto a_thread_offset = get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / - // 128) * MPerXdl; constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( make_tuple(Number{}, Number{})); diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index bdd41831c3..75d7dd9d9d 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -49,8 +49,8 @@ set(PROFILER_SOURCES # endif() # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) endif() # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) @@ -136,8 +136,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) endif() # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) From dd6d8797c9fe05623f2e2b754f29117c7f563883 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Tue, 25 Feb 2025 02:16:24 +0000 Subject: [PATCH 12/28] sanity checked --- ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 50 +- ...kwise_gemm_pipeline_xdlops_v1_ab_scale.hpp | 591 +++++++++++++++--- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 8 +- ...mm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp | 131 +--- ..._gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp | 30 +- .../gpu/gemm_ab_scale.hpp | 60 -- .../gpu/gemm_ab_scale/CMakeLists.txt | 7 +- ...le_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp | 65 +- ...n_128_128_128_comp_mnkpadding_instance.cpp | 37 -- ...mn_128_128_128_comp_mnpadding_instance.cpp | 37 -- ...128_128_128_mem_v1_mnkpadding_instance.cpp | 38 -- profiler/src/CMakeLists.txt | 8 +- profiler/src/profile_gemm_ab_scale.cpp | 3 +- 13 files changed, 618 insertions(+), 447 deletions(-) delete mode 100644 library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index b5ad01e0a5..b54ba5ddfb 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -65,14 +65,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, - 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>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; + 16, 128, + 256, 16, 16, + 16, 16, + 1, 2, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + 1, 2, S<1, 16, 1, 16>, S<8>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on int main(int argc, char* argv[]) @@ -80,6 +80,7 @@ int main(int argc, char* argv[]) bool do_verification = true; int init_method = 1; bool time_kernel = false; + bool flush_cache = true; // GEMM shape ck::index_t M = 128; @@ -100,7 +101,7 @@ int main(int argc, char* argv[]) init_method = std::stoi(argv[2]); time_kernel = std::stoi(argv[3]); } - else if(argc == 7) + else if(argc == 8) { do_verification = std::stoi(argv[1]); init_method = std::stoi(argv[2]); @@ -110,6 +111,8 @@ int main(int argc, char* argv[]) N = std::stoi(argv[5]); K = std::stoi(argv[6]); + flush_cache = std::stoi(argv[7]); + StrideA = K; StrideB = K; StrideE = N; @@ -119,7 +122,8 @@ int main(int argc, char* argv[]) printf("arg1: verification (0=no, 1=yes)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg3: time kernel (0=no, 1=yes)\n"); - printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"); + printf("arg4 to 6: M, N, K\n"); + printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n"); exit(0); } @@ -185,7 +189,6 @@ int main(int argc, char* argv[]) a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - // b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); break; case 5: @@ -193,7 +196,6 @@ int main(int argc, char* argv[]) b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); - // b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); break; default: a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); @@ -202,6 +204,16 @@ int main(int argc, char* argv[]) b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); } #endif +#if 0 + for(int im =0; im< (M + Scale_Block_M - 1) / Scale_Block_M; im++){ + float row_sum = .0; + for(int ik =0; ik< (K + Scale_Block_K - 1) / Scale_Block_K; ik++){ + printf("%lf ",a1_m_k(im, ik)); + row_sum += a1_m_k(im, ik); + } + printf("sum: %lf\n", row_sum * 128); + } +#endif DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize()); @@ -251,9 +263,19 @@ int main(int argc, char* argv[]) std::size_t num_btype = sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; - int rotating_buf = (512*1024*1024 + num_btype-1)/num_btype; + float ave_time = .0; - float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf}); + if(flush_cache) + { + int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype; + + ave_time = invoker.Run(argument, + StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf}); + } + else + { + ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100}); + } float tflops = static_cast(flop) / 1.E9 / ave_time; diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp index 02bc0a941c..fc65c202fa 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp @@ -7,10 +7,10 @@ namespace ck { -// Naive pipeline with lowest resource request per WGP -// GlobalPrefetchStages: 1 +// Compute optimized pipeline +// GlobalPrefetchStages: 2 // LocalPreFillStages: 1 -// LocalPreFetchStages: 0 +// LocalPreFetchStages: 1 // LocalSharedMemoryBuffer: 1 template ; + using Base::A_K1; + using Base::B_K1; using Base::I0; + using Base::I1; using Base::KRepeat; using Base::xdlops_gemm; + using typename Base::HotLoopInstList; using Base::CalculateCThreadOriginDataIndex; using Base::CalculateCThreadOriginDataIndex8D; @@ -133,19 +137,43 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale PrefetchStages; @@ -153,11 +181,116 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale + // sizeof(ComputeDataType) / sizeof(BDataType) + // ? sizeof(ComputeDataType) / sizeof(ADataType) + // : sizeof(ComputeDataType) / sizeof(BDataType); + constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma); + constexpr auto num_mfma_per_issue = + num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b); + constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a; + constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b; + + static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { + ignore = i; + static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) { + ignore = idswrite; + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + }); + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + __builtin_amdgcn_sched_group_barrier( + 0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA + }); + static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { + ignore = i; + static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) { + ignore = idswrite; + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + }); + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + __builtin_amdgcn_sched_group_barrier( + 0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA + }); + + // stage 2 + static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) { + if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >= + ds_read_a_mfma_rate) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read + } + else + { + __builtin_amdgcn_sched_group_barrier(0x100, + num_ds_read_inst_a - (num_dsread_a_mfma - 1) * + ds_read_a_mfma_rate, + 0); // DS read + } + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + }); + + static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) { + if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >= + ds_read_b_mfma_rate) + { + __builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read + } + else + { + __builtin_amdgcn_sched_group_barrier(0x100, + num_ds_read_inst_b - (num_dsread_b_mfma - 1) * + ds_read_b_mfma_rate, + 0); // DS read + } + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + }); } template ( a_thread_desc_.GetElementSpaceSize()); auto b_thread_buf = make_static_buffer( @@ -225,6 +359,8 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale( b_scale_thread_desc.GetElementSpaceSize()); + auto c_scale_thread_buf = make_static_buffer( + c_scale_thread_desc.GetElementSpaceSize()); // Global prefetch 1 a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); @@ -243,7 +379,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{})); }); - if(num_loop_per_scale == 1) + if constexpr(NumKBlockPerScale == 1) { a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); @@ -262,14 +398,99 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}); + constexpr auto num_scale_m_block = CScaleThreadDesc{}.GetLength(Number<1>{}); + constexpr auto num_scale_n_block = CScaleThreadDesc{}.GetLength(Number<2>{}); + + static_for<0, num_scale_m_block, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { + constexpr index_t c_offset = + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); + + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; + }); + }); + }); + // Local prefill 1 a_blockwise_copy.RunWrite(a_block_desc, a_block_buf); b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); + // Global prefetch 2 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf); + + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step); + // Initialize C c_thread_buf.Clear(); - auto c_thread_buf_per_scale = remove_cvref_t(); + StaticBufferTupleOfVector + c_thread_buf_per_scale; + + // Local prefetch 1 + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k, + make_tuple(m0, I0, I0, Number{}), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, k0, I0), + a_thread_buf); + }); + static_for<0, NRepeat, 1>{}([&](auto n0) { + b_thread_copy_.Run(b_block_desc_n0_n1_n2_k, + make_tuple(n0, I0, I0, Number{}), + b_block_buf, + b_thread_desc_, + make_tuple(n0, I0, k0, I0), + b_thread_buf); + }); + }); + + __builtin_amdgcn_sched_barrier(0); // main body if constexpr(HasMainLoop) @@ -277,13 +498,85 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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_buf[Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = + CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); + }); + }); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { + constexpr index_t c_offset = + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); + + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; + }); + }); + }); + block_sync_lds(); static_for<0, KRepeat, 1>{}([&](auto k) { static_for<0, MRepeat, 1>{}([&](auto m0) { @@ -304,42 +597,6 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto m0) { - static_for<0, NRepeat, 1>{}([&](auto n0) { - c_thread_buf_per_scale.Clear(); - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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_buf[Number{}]; - }); - - using mfma_input_type = - typename vector_type::type; - - xdlops_gemm.template Run<>( - a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf_per_scale.GetVectorTypeReference(I0)); - }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += - c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[m0]) * - type_convert(b_scale_thread_buf[I0]); - }); - }); - }); - static_for<0, MRepeat, 1>{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, a_scale_grid_buf, @@ -350,7 +607,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{})); }); - if(num_loop_per_scale == 1) + if constexpr(NumKBlockPerScale == 1) { a_scale_thread_copy.MoveSrcSliceWindow( a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); @@ -368,19 +625,87 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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_buf[Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); + }); + }); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { + constexpr index_t c_offset = + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); + + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; + }); + }); + }); + block_sync_lds(); static_for<0, KRepeat, 1>{}([&](auto k) { static_for<0, MRepeat, 1>{}([&](auto m0) { @@ -401,49 +726,143 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - c_thread_buf_per_scale.Clear(); - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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_buf[Number{}]; + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; - using mfma_input_type = - typename vector_type::type; + 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_buf[Number{}]; + }); - xdlops_gemm.template Run<>( - a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf_per_scale.GetVectorTypeReference(I0)); - }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += - c_thread_buf_per_scale[Number{}] * - type_convert(a_scale_thread_buf[m0]) * - type_convert(b_scale_thread_buf[I0]); + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); }); }); }); + __builtin_amdgcn_sched_barrier(0); + } + else if constexpr(TailNum == TailNumber::Odd) + { + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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_buf[Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); + }); + }); + }); + __builtin_amdgcn_sched_barrier(0); } } protected: - using Base::a_thread_copy_; using Base::a_thread_desc_; - using Base::b_thread_copy_; using Base::b_thread_desc_; using Base::c_thread_desc_; + using AThreadCopy = ThreadwiseTensorSliceTransfer_v4, + Sequence<0, 1, 2, 3>, + 3, + A_K1, + A_K1>; + + using BThreadCopy = ThreadwiseTensorSliceTransfer_v4, + Sequence<0, 1, 2, 3>, + 3, + B_K1, + B_K1>; + + AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()}; + BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()}; }; } // namespace ck diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index f310683a48..fc0075b196 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -193,7 +193,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale @@ -324,6 +324,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}) == 1, + "Pipeline v3 only support scaleblocksliceK=1"); + static_assert(CScaleThreadDesc{}.GetLength(Number<2>{}) == 1, + "Pipeline v3 only support scaleblocksliceN=1"); // assume kperblock = scaleblockk auto a_thread_buf = make_static_buffer( a_thread_desc_.GetElementSpaceSize()); @@ -413,7 +417,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale; - Run(kernel); - } - else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Full) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Three) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Four) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Five) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Seven) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - } - } } else { // Tail number always 1 if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Full) { const auto kernel = kernel_gemm_xdl_cshuffle_v3; Run(kernel); } + else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } } } return ave_time; diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp index bcb2f8718c..25be9bebb7 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp @@ -225,7 +225,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{})); } - __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1( + __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1( index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0) { const auto a_grid_desc_mraw_kraw = [&]() { @@ -307,7 +307,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 } } - __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1( + __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1( index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0) { const auto b_grid_desc_nraw_kraw = [&]() { @@ -956,7 +956,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || - GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) && + !(is_same::value)) { if(!(karg.M % MPerBlock == 0)) { @@ -973,7 +974,8 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || - GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) && + (is_same::value)) { if(!(karg.N % NPerBlock == 0)) { @@ -1321,10 +1323,12 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / KPerBlock); - const index_t ScaleSliceSizeM = MXdlPerWave; - const index_t ScaleSliceSizeN = 1; - const index_t ScaleSliceSizeK = 1; + constexpr index_t ScaleSliceSizeM = MXdlPerWave; + constexpr index_t ScaleSliceSizeN = math::integer_divide_ceil(NPerBlock, ScaleBlockN); + constexpr index_t ScaleSliceSizeK = math::integer_divide_ceil(KPerBlock, ScaleBlockK); + // ScaleSliceSizeK is last dimension in A/B scale for vector memory access + // ScaleSliceSizeK is first dimension in C scale for packed math constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed( make_tuple(Number{}, Number{})); @@ -1337,7 +1341,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 make_tuple(Number{}, Number{})); constexpr auto c_scale_thread_desc = make_naive_tensor_descriptor_packed(make_tuple( - Number{}, Number{}, Number{})); + Number{}, Number{}, Number{})); auto a_scale_thread_copy = ThreadwiseTensorSliceTransfer_v2, Sequence<0, 1>, 1, - 1, + ScaleSliceSizeK, 1, false>( a_scale_grid_desc_am_ak, @@ -1361,7 +1365,7 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 Sequence, Sequence<0, 1>, 1, - 1, + ScaleSliceSizeK, 1, false>( b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0)); @@ -1370,10 +1374,10 @@ struct GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 constexpr auto a_scale_thread_slice_copy_step = make_tuple(make_multi_index(MWaves * MPerXdl, 0), make_multi_index(-MPerBlock, 0), - make_multi_index(-MPerBlock, 1)); - constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1); + make_multi_index(-MPerBlock, ScaleSliceSizeK)); + constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, ScaleSliceSizeK); - constexpr auto NumKBlockPerScale = ScaleBlockK / KPerBlock; + constexpr auto NumKBlockPerScale = math::integer_divide_ceil(ScaleBlockK, KPerBlock); blockwise_gemm_pipeline.template Run( a_grid_desc_ak0_m_ak1, diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp index fb9d6fb078..3fa82ae53a 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp @@ -53,42 +53,6 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_in PassThrough, PassThrough>>>& instances); -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances); - -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances); - void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( std::vector>>& instances); - -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances); #endif template using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances = std::tuple< // clang-format off - //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| - //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| - //################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | 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| - //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + //################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | 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| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Compute friendly - // Spill in current compiler - // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, 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, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 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>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, 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>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 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>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> // clang-format on >; template using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances = std::tuple< // clang-format off - //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| - //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| - //################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | 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| - //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + //################################| ALayout| BLayout| DsLayout| ELayout|AData | BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - // Latency friendly - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, - // Memory friendly - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 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>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, 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>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, 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>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8> + // Memory friendly + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 256, 128, 8, 16, 16, 16, 1, 4, S<16, 16, 1>, 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, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 128, 8, 16, 16, 16, 1, 2, S<16, 16, 1>, 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, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 128, 8, 16, 16, 16, 1, 1, S<16, 16, 1>, 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, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 256, 128, 16, 16, 32, 32, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 128, 16, 16, 16, 16, 2, 1, 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, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 256, 16, 16, 16, 16, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 256, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 2, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8> // clang-format on >; } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp deleted file mode 100644 index 7b61cadea9..0000000000 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp +++ /dev/null @@ -1,37 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. - -#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances) -{ - add_device_operation_instances( - instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp deleted file mode 100644 index aa9b0c828a..0000000000 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp +++ /dev/null @@ -1,37 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. - -#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances) -{ - add_device_operation_instances( - instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp deleted file mode 100644 index f51fa43bcb..0000000000 --- a/library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp +++ /dev/null @@ -1,38 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. - -#include "device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances( - std::vector, - Row, - F8, - F32, - F8, - F32, - Tuple<>, - BF16, - 1, - 128, - 128, - PassThrough, - PassThrough, - PassThrough>>>& instances) -{ - add_device_operation_instances( - instances, - device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 75d7dd9d9d..bdd41831c3 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -49,8 +49,8 @@ set(PROFILER_SOURCES # endif() # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) - # list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) endif() # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) @@ -136,8 +136,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) - # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) endif() # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) diff --git a/profiler/src/profile_gemm_ab_scale.cpp b/profiler/src/profile_gemm_ab_scale.cpp index c17ab74536..c6904a2812 100644 --- a/profiler/src/profile_gemm_ab_scale.cpp +++ b/profiler/src/profile_gemm_ab_scale.cpp @@ -50,7 +50,8 @@ int profile_gemm_ab_scale(int argc, char* argv[]) printf(" 1: A[m, k] * B[n, k] = C[m, n];\n"); printf(" 2: A[k, m] * B[k, n] = C[m, n];\n"); printf(" 3: A[k, m] * B[n, k] = C[m, n])\n"); - printf("arg4: scale block tile (0: ScaleBlockM/N/K = [128, 128, 128];\n"); + printf("arg4: scale block tile (0: ScaleBlockM/N/K = [128, 128, 128]; 1: ScaleBlockM/N/K = " + "[1, 128, 128];\n"); printf("arg5: verification (0: no; 1: yes)\n"); printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n"); printf("arg7: print tensor value (0: no; 1: yes)\n"); From 00c5f0fc1afc482a9fc4213955acab5d0eebfc8d Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Tue, 25 Feb 2025 02:33:41 +0000 Subject: [PATCH 13/28] revert ckprofiler cmake changes --- profiler/src/CMakeLists.txt | 300 ++++++++++++++++++------------------ 1 file changed, 150 insertions(+), 150 deletions(-) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index bdd41831c3..35e91f8172 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -1,87 +1,87 @@ # ckProfiler set(PROFILER_SOURCES profiler.cpp - # profile_gemm.cpp - # profile_reduce.cpp - # profile_groupnorm_bwd_data.cpp - # profile_groupnorm_fwd.cpp - # profile_layernorm_bwd_data.cpp - # profile_layernorm_bwd_gamma_beta.cpp - # profile_groupnorm_bwd_gamma_beta.cpp - # profile_layernorm_fwd.cpp - # profile_max_pool2d_fwd.cpp - # profile_pool3d_fwd.cpp - # profile_avg_pool3d_bwd.cpp - # profile_max_pool3d_bwd.cpp - # profile_avg_pool2d_bwd.cpp - # profile_max_pool2d_bwd.cpp - # profile_softmax.cpp - # profile_batchnorm_fwd.cpp - # profile_batchnorm_bwd.cpp - # profile_batchnorm_infer.cpp - # profile_conv_tensor_rearrange.cpp - # profile_transpose.cpp - # profile_permute_scale.cpp + profile_gemm.cpp + profile_reduce.cpp + profile_groupnorm_bwd_data.cpp + profile_groupnorm_fwd.cpp + profile_layernorm_bwd_data.cpp + profile_layernorm_bwd_gamma_beta.cpp + profile_groupnorm_bwd_gamma_beta.cpp + profile_layernorm_fwd.cpp + profile_max_pool2d_fwd.cpp + profile_pool3d_fwd.cpp + profile_avg_pool3d_bwd.cpp + profile_max_pool3d_bwd.cpp + profile_avg_pool2d_bwd.cpp + profile_max_pool2d_bwd.cpp + profile_softmax.cpp + profile_batchnorm_fwd.cpp + profile_batchnorm_bwd.cpp + profile_batchnorm_infer.cpp + profile_conv_tensor_rearrange.cpp + profile_transpose.cpp + profile_permute_scale.cpp ) -# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") -# if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) -# list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) -# list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) -# endif() -# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) -# list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) -# list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) -# list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) -# endif() -# list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") + if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) + list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) + list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) + endif() + if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) + list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) + list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) + endif() + list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) endif() -# list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) -# list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) -# list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) -# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) -# list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) -# list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) -# list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) + list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) + list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) + list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) + list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) + list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) + list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) -# endif() +endif() -# if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") -# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) -# list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) -# endif() -# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -# endif() +if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") + if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) + endif() + list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +endif() -# if(DL_KERNELS) -# list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) -# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -# endif() +if(DL_KERNELS) + list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) + list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +endif() set(PROFILER_EXECUTABLE ckProfiler) @@ -94,88 +94,88 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) endif() target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) -# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") -# if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) -# endif() -# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) -# endif() -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") + if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) + endif() + if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) + endif() + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) endif() -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) -# endif() + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) +endif() -# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") -# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) -# endif() -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -# endif() +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") + if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) + endif() + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +endif() -# if(DL_KERNELS) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) -# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -# endif() +if(DL_KERNELS) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +endif() rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) From 2367a4f9b9286ff576c7cd1475e49a43a220467f Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Tue, 25 Feb 2025 02:35:20 +0000 Subject: [PATCH 14/28] clang format --- .../gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp | 2 +- .../impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp | 3 ++- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp index fc65c202fa..8375e81fa0 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp @@ -181,7 +181,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale Date: Tue, 25 Feb 2025 02:41:48 +0000 Subject: [PATCH 15/28] revert unnecessary changes. --- CMakeLists.txt | 7 -- .../gemm_multiply_multiply_xdl_fp8.cpp | 37 +++++----- .../gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp | 68 +++++++++++++++---- profiler/src/CMakeLists.txt | 4 ++ 4 files changed, 77 insertions(+), 39 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index cd03961477..3558666e5d 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -246,13 +246,6 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500500000) add_compile_options("SHELL: -mllvm --lsr-drop-solution=1") endif() endif() -# if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) -# check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED) -# if(HAS_ENABLE_POST_MISCHED) -# message("Adding the enable-post-misched=0 compiler flag") -# add_compile_options("SHELL: -mllvm -enable-post-misched=0") -# endif() -# endif() set(check-coerce) check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce) if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp index 34b9e0f6b1..c33fe357d8 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp @@ -24,10 +24,9 @@ template using S = ck::Sequence; -using F16 = ck::half_t; -using FP8 = ck::f8_t; -using F32 = float; -using BF16 = ck::bhalf_t; +using F16 = ck::half_t; +using FP8 = ck::f8_t; +using F32 = float; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -39,7 +38,7 @@ using CShuffleDataType = F32; using D0DataType = F32; using D1DataType = F32; using DsDataType = ck::Tuple; -using EDataType = BF16; +using EDataType = F16; using A0Layout = Row; using B0Layout = Col; @@ -48,23 +47,21 @@ using D1Layout = Col; using DsLayout = ck::Tuple; using ELayout = Row; -// struct MultiplyMultiply -// { -// template -// __host__ __device__ constexpr void -// operator()(E& e, const C& c, const D0& d0, const D1& d1) const; +struct MultiplyMultiply +{ + template + __host__ __device__ constexpr void + operator()(E& e, const C& c, const D0& d0, const D1& d1) const; -// template <> -// __host__ __device__ constexpr void operator()( -// ck::half_t& e, const float& c, const float& d0, const float& d1) const -// { -// const float x0_f = c * d0 * d1; + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const float& c, const float& d0, const float& d1) const + { + const float x0_f = c * d0 * d1; -// e = ck::type_convert(x0_f); -// } -// }; - -using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply; + e = ck::type_convert(x0_f); + } +}; using PassThrough = ck::tensor_operation::element_wise::PassThrough; diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp index 1c5080ea1b..a9e73bf461 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp @@ -688,19 +688,40 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 // in some cases. else if constexpr(is_same::value) { - constexpr auto a_lds_block_desc = - make_naive_tensor_descriptor(make_tuple(AK0Number, Number{}, AK1Number), - make_tuple(AK1Number, Number{}, I1)); + constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1 + ? 1 + : 32 * 4 / KPerBlock / sizeof(LDSTypeA); + constexpr auto a_lds_block_desc = make_naive_tensor_descriptor( + make_tuple( + AK0Number * Number{}, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( a_lds_block_desc, - make_tuple(make_xor_with_modulo_transform( - make_tuple(Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform(make_tuple( + Number{}, Number{})), make_pass_through_transform(AK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - return a_lds_block_desc_permuted; + constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_permuted, + make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); + + constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_ak0_mldslayer_m_ak1, + make_tuple(make_pass_through_transform(AK0Number), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return a_lds_block_desc_ak0_m_ak1; } else // ColumnMajor A { @@ -804,19 +825,42 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 } else if constexpr(is_same::value) { - constexpr auto b_lds_block_desc = - make_naive_tensor_descriptor(make_tuple(BK0Number, Number{}, BK1Number), - make_tuple(BK1Number, Number{}, I1)); + // NLdsLayer * K0 as logical Bank + constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1 + ? 1 + : 32 * 4 / KPerBlock / sizeof(LDSTypeB); + ; + constexpr auto b_lds_block_desc = make_naive_tensor_descriptor( + make_tuple( + BK0Number * Number{}, Number{}, BK1Number), + make_tuple(BK1Number, Number{}, I1)); constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor( b_lds_block_desc, - make_tuple(make_xor_with_modulo_transform( - make_tuple(Number{}, Number{})), + make_tuple(make_xor_with_modulo_transform(make_tuple( + Number{}, Number{})), make_pass_through_transform(BK1Number)), make_tuple(Sequence<1, 0>{}, Sequence<2>{}), make_tuple(Sequence<1, 0>{}, Sequence<2>{})); - return b_lds_block_desc_permuted; + constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor( + b_lds_block_desc_permuted, + make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); + + constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_lds_block_desc_bk0_nldslayer_n_bk1, + make_tuple(make_pass_through_transform(BK0Number), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{})), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return b_lds_block_desc_bk0_n_bk1; } else // RowMajor B { diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 58a6dc6f5f..2b059592e8 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -59,6 +59,8 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp) + list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) @@ -143,6 +145,8 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") endif() target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) From 41fab2d09d83a7e3a7696233749c66a693d0a969 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Tue, 25 Feb 2025 02:46:29 +0000 Subject: [PATCH 16/28] remove commented codes. --- profiler/src/CMakeLists.txt | 1 - profiler/src/profile_gemm_ab_scale.cpp | 17 ----------------- 2 files changed, 18 deletions(-) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 2b059592e8..5ed28b9826 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -183,5 +183,4 @@ if(DL_KERNELS) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) endif() - rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) diff --git a/profiler/src/profile_gemm_ab_scale.cpp b/profiler/src/profile_gemm_ab_scale.cpp index c6904a2812..3956038a30 100644 --- a/profiler/src/profile_gemm_ab_scale.cpp +++ b/profiler/src/profile_gemm_ab_scale.cpp @@ -156,23 +156,6 @@ int profile_gemm_ab_scale(int argc, char* argv[]) return pass ? 0 : 1; }; - // if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && - // scale_block_tile == ScaleBlockTile::Tile_128_128_128) - // { - // return profile(F8{}, - // F32{}, - // F8{}, - // F32{}, - // F8{}, - // F32{}, - // BF16{}, - // ck::Number<128>{}, - // ck::Number<128>{}, - // ck::Number<128>{}, - // Row{}, - // Col{}, - // Row{}); - // } if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && scale_block_tile == ScaleBlockTile::Tile_1_128_128) { From 4a332700874ae5cc69972cd721f7b1dcc24bb229 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Tue, 25 Feb 2025 10:02:18 +0000 Subject: [PATCH 17/28] v3 sanity checked --- .../65_gemm_multiply_multiply/CMakeLists.txt | 6 + ...emm_multiply_multiply_xdl_fp8_ab_scale.cpp | 16 +- ...ultiply_xdl_fp8_blockscale_bpreshuffle.cpp | 382 +++ ..._multiply_multiply_xdl_fp8_bpreshuffle.cpp | 1 - ...e_gemm_pipeline_xdlops_b_preshuffle_v3.hpp | 4 +- ...dlops_blockscale_b_preshuffle_selector.hpp | 116 + ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 506 ++++ ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 1051 ++++++++ .../device_gemm_multiple_d_ab_scale.hpp | 43 + ...xdl_cshuffle_v3_blockscale_bpreshuffle.hpp | 484 ++++ ...fle_v3_multi_d_blockscale_b_preshuffle.hpp | 2111 +++++++++++++++++ 11 files changed, 4709 insertions(+), 11 deletions(-) create mode 100644 example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp create mode 100644 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp create mode 100644 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp create mode 100644 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp create mode 100644 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp create mode 100644 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 2d00545515..466f2a9f3c 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -1,5 +1,11 @@ add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp) +add_example_executable(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp) add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) +set(EXAMPLE_COMPILE_OPTIONS) +list(APPEND EXAMPLE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) +list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) +target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) +target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp index b54ba5ddfb..ced39f4a0b 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp @@ -65,14 +65,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_ A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, - 16, 128, - 256, 16, 16, - 16, 16, - 1, 2, - S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, - S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, - 1, 2, S<1, 16, 1, 16>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; + 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>, S<8>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp new file mode 100644 index 0000000000..bc049a68c6 --- /dev/null +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp @@ -0,0 +1,382 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" +#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp" + +#include "ck/library/utility/device_memory.hpp" +#include "ck/library/utility/host_tensor.hpp" +#include "ck/library/utility/host_tensor_generator.hpp" +#include "ck/library/utility/literals.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" +#include "ck/library/utility/check_err.hpp" + +#include "ck/utility/blkgemmpipe_scheduler.hpp" + +template +using S = ck::Sequence; + +using BF16 = ck::bhalf_t; +using FP8 = ck::f8_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using A0DataType = FP8; +using A1DataType = F32; +using B0DataType = FP8; +using B1DataType = F32; +using AccDataType = F32; +using CShuffleDataType = F32; +using DsDataType = ck::Tuple<>; +using EDataType = BF16; + +using A0Layout = Row; +using B0Layout = Col; +using D0Layout = Row; +using D1Layout = Col; +using DsLayout = ck::Tuple<>; +using ELayout = Row; + +void preShuffleBuffer(const FP8* src, FP8* dst, int N, int K, int NXdl) +{ + int KPack = 16; + int NLane = NXdl; + int KLane = 64 / NLane; + + int K0 = K / (KLane * KPack); + // K -> K0 KLane KPack + // N -> N0 NLane + // N, K -> N0 K0 KLane NLane KPack + int tempk; + for(int n = 0; n < N; ++n) + { + for(int k = 0; k < K; ++k) + { + int n0 = n / NLane; + int n1 = n % NLane; + + int k0 = k / (KLane * KPack); + tempk = k % (KLane * KPack); + int k1 = tempk / KPack; + int k2 = tempk % KPack; + + int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane + + k1 * KPack * NLane + n1 * KPack + k2; + + dst[outputIndex] = src[n * K + k]; + } + } +} +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = PassThrough; + +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; + +static constexpr ck::index_t Scale_Block_M = 1; +static constexpr ck::index_t Scale_Block_N = 128; +static constexpr ck::index_t Scale_Block_K = 128; + +using DeviceOpInstance = + ck::tensor_operation::device::DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle + // clang-format off + , 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>, S<8>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; +// clang-format on + +int main(int argc, char* argv[]) +{ + bool do_verification = true; + int init_method = 1; + bool time_kernel = false; + bool flush_cache = true; + + // GEMM shape + ck::index_t M = 128; + ck::index_t N = 1024; + ck::index_t K = 1024; + + ck::index_t StrideA = K; + ck::index_t StrideB = K; + ck::index_t StrideE = N; + + if(argc == 1) + { + // use default case + } + else if(argc == 4) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + } + else if(argc == 8) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + + M = std::stoi(argv[4]); + N = std::stoi(argv[5]); + K = std::stoi(argv[6]); + + flush_cache = std::stoi(argv[7]); + + StrideA = K; + StrideB = K; + StrideE = N; + } + else + { + printf("arg1: verification (0=no, 1=yes)\n"); + printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); + printf("arg3: time kernel (0=no, 1=yes)\n"); + printf("arg4 to 6: M, N, K\n"); + printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n"); + exit(0); + } + + ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K; + ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K; + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + using namespace ck::literals; + + if(std::is_same::value) + { + return HostTensorDescriptor({row, col}, {stride, 1_uz}); + } + else + { + return HostTensorDescriptor({row, col}, {1_uz, stride}); + } + }; + + Tensor a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{})); + Tensor a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M, + (K + Scale_Block_K - 1) / Scale_Block_K, + Scale_Stride_AM, + A0Layout{})); + Tensor b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); + Tensor b0_preshuffled( + f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size + Tensor b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K, + (N + Scale_Block_N - 1) / Scale_Block_N, + Scale_Stride_BN, + B0Layout{})); + Tensor e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + Tensor e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + + std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl; + std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl; + std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl; + std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl; + std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl; + +#if 1 + switch(init_method) + { + case 0: break; + case 1: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + case 2: + a0_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b0_k_n.GenerateTensorValue(GeneratorTensor_1{}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 3: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 4: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 5: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + default: + a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + } +#endif +#if 0 + for(int im =0; im< (M + Scale_Block_M - 1) / Scale_Block_M; im++){ + float row_sum = .0; + for(int ik =0; ik< (K + Scale_Block_K - 1) / Scale_Block_K; ik++){ + printf("%lf ",a1_m_k(im, ik)); + row_sum += a1_m_k(im, ik); + } + printf("sum: %lf\n", row_sum * 128); + } +#endif + + DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); + DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize()); + DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); + + a0_device_buf.ToDevice(a0_m_k.mData.data()); + a1_device_buf.ToDevice(a1_m_k.mData.data()); + b1_device_buf.ToDevice(b1_k_n.mData.data()); + + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto cde_element_op = CDEElementOp{}; + + constexpr ck::index_t NumDTensor = DsDataType::Size(); + + // do GEMM + auto device_op = DeviceOpInstance{}; + int NPerXdl = device_op.GetPreShuffleParameters(); + + preShuffleBuffer(b0_k_n.mData.data(), b0_preshuffled.mData.data(), N, K, NPerXdl); + + b0_device_buf.ToDevice(b0_preshuffled.mData.data()); + auto invoker = device_op.MakeInvoker(); + auto argument = device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(), + b0_device_buf.GetDeviceBuffer(), + std::array{}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{}, + StrideE, + a1_device_buf.GetDeviceBuffer(), + b1_device_buf.GetDeviceBuffer(), + a_element_op, + b_element_op, + cde_element_op); + + if(!device_op.IsSupportedArgument(argument)) + { + throw std::runtime_error( + "wrong! device_gemm with the specified compilation parameters does " + "not support this GEMM problem"); + } + + std::size_t flop = std::size_t(2) * M * N * K; + std::size_t num_btype = + sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; + + float ave_time = .0; + + if(flush_cache) + { + int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype; + + ave_time = invoker.Run(argument, + StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf}); + } + else + { + ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100}); + } + + float tflops = static_cast(flop) / 1.E9 / ave_time; + + float gb_per_sec = num_btype / 1.E6 / ave_time; + + std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" + << std::endl; + + if(do_verification) + { + Tensor c_m_n({M, N}); + Tensor a_m_k({M, K}); + Tensor b_k_n({K, N}); + + for(int m = 0; m < M; m++) + { + for(int k = 0; k < K; k++) + { + a_m_k(m, k) = ck::type_convert(a0_m_k(m, k)) * + a1_m_k(m / Scale_Block_M, k / Scale_Block_K); + } + } + + for(int n = 0; n < N; n++) + { + for(int k = 0; k < K; k++) + { + b_k_n(k, n) = ck::type_convert(b0_k_n(k, n)) * + b1_k_n(k / Scale_Block_K, n / Scale_Block_N); + } + } + + using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm; + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = + ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{}); + + ref_invoker.Run(ref_argument); + +#if 1 + for(int m = 0; m < M; ++m) + { + for(int n = 0; n < N; ++n) + { + e_m_n_host_result(m, n) = ck::type_convert(c_m_n(m, n)); + } + } +#endif + + e_device_buf.FromDevice(e_m_n_device_result.mData.data()); + + return ck::utils::check_err( + e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2) + ? 0 + : 1; + } + + return 0; +} diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp index 9a81ef5ea7..7319d345c9 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp @@ -9,7 +9,6 @@ #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp" -#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp" 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 49af782132..e13fa2e5c6 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 @@ -528,10 +528,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3 +constexpr auto BlockGemmBlockScaleBPreshufflePipeline_Selector() +{ +#if 0 + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1< + BlkGemmPipeSche, + BlockSize, + ADataType, + BDataType, + ComputeDataType, + AccDataType, + ATileDesc, + BTileDesc, + AMmaTileDesc, + BMmaTileDesc, + ABlockTransferSrcScalarPerVector, + BBlockTransferSrcScalarPerVector, + MPerBlock, + NPerBlock, + KPerBlock, + MPerXDL, + NPerXDL, + MRepeat, + NRepeat, + KPack>{}; + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) + { + return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v2< + BlkGemmPipeSche, + BlockSize, + ADataType, + BDataType, + ComputeDataType, + AccDataType, + ATileDesc, + BTileDesc, + AMmaTileDesc, + BMmaTileDesc, + ABlockTransferSrcScalarPerVector, + BBlockTransferSrcScalarPerVector, + MPerBlock, + NPerBlock, + KPerBlock, + MPerXDL, + NPerXDL, + MRepeat, + NRepeat, + KPack>{}; + } +#endif + // else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3"); + return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3< + BlkGemmPipeSche, + BlockSize, + ADataType, + BDataType, + ComputeDataType, + AccDataType, + ATileDesc, + BTileDesc, + AMmaTileDesc, + BMmaTileDesc, + ABlockTransferSrcScalarPerVector, + BBlockTransferSrcScalarPerVector, + MPerBlock, + NPerBlock, + KPerBlock, + MPerXDL, + NPerXDL, + MRepeat, + NRepeat, + KPack>{}; + } + else + { + std::cerr << "BlockGemmPipeline configuration is not available" << std::endl; + } +} + +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp new file mode 100644 index 0000000000..8ed25895b5 --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -0,0 +1,506 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, 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_v1 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 + : BlockwiseGemmXdlops_pipeline_base + +{ + using Base = BlockwiseGemmXdlops_pipeline_base; + using Base::A_K1; + using Base::B_K1; + using Base::I0; + using Base::I1; + 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; + + static constexpr index_t PrefetchStages = 2; + static constexpr index_t PrefillStages = 1; + static constexpr index_t GlobalBufferNum = 2; + + 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; + constexpr index_t K1 = 64 / NPerXDL; + constexpr index_t K0 = KRepeat; + + 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() + { + constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_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; + + // B global + static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { + ignore = i; + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + }); + + // A global + static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { + ignore = i; + __builtin_amdgcn_sched_group_barrier(0x008, 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(0x020, 1, 0); // VMEM read + }); + + // A local + static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) { + ignore = i; + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read + }); + } + + 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, + const BGridBuffer& b_grid_buf, + BBlockBuffer& b_block_buf, + const BBlockTransferStep& b_block_copy_step, + CThreadBuffer& c_thread_buf, + 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; + constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0); + + // Global prefetch A1 B1 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); + 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)); + + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + __builtin_amdgcn_sched_barrier(0); + + // // Local prefill A1 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0); + + // // Global prefetch A2 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + // Local prefetch A1 + 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); + }); + }); + + // Initialize C + c_thread_buf.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); + + block_sync_lds(); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf); + + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + 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{})); + }); + }); + }); + + 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); + }); + }); + + HotLoopScheduler(); + __builtin_amdgcn_sched_barrier(0); + }; + + 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)); + + block_sync_lds(); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + 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{})); + }); + }); + }); + + 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); + }); + }); + + __builtin_amdgcn_sched_barrier(0); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + 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{})); + }); + }); + }); + // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle + // latency + // __builtin_amdgcn_sched_barrier(0); + } + else + { + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + 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{})); + }); + }); + }); + } + } + + protected: + // MRepeat MWave MLane KRepeat KLane KPack + // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack + static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, 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_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp new file mode 100644 index 0000000000..83759a3192 --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -0,0 +1,1051 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, 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_blockscale_bpreshuffle_v3 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_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::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; + constexpr index_t K1 = 64 / NPerXDL; + constexpr index_t K0 = KRepeat; + + 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; + } + + template + __device__ static constexpr auto HotLoopScheduler(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_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; + + 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; + + 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( + // ABlockCopy + 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, + // BBlockCopy + const BGridDesc& b_grid_desc, + const BBlockDesc& b_block_desc, + BBlockTransfer& b_blockwise_copy, + const BGridBuffer& b_grid_buf, + BBlockBuffer& b_block_buf, + const BBlockTransferStep& b_block_copy_step, + // CThread + const CScaleThreadDesc& c_scale_thread_desc, + CThreadBuffer& c_thread_buf, + // AScaleThreadCopy + const AScaleGridDesc& a_scale_grid_desc, + const AScaleThreadDesc& a_scale_thread_desc, + AScaleThreadTransfer& a_scale_thread_copy, + const AScaleGridBuffer& a_scale_grid_buf, + const AScaleThreadTransferStep& a_scale_thread_copy_step, + // BScaleThreadCopy + const BScaleGridDesc& b_scale_grid_desc, + const BScaleThreadDesc& b_scale_thread_desc, + BScaleThreadTransfer& b_scale_thread_copy, + const BScaleGridBuffer& b_scale_grid_buf, + const BScaleThreadTransferStep& b_scale_thread_copy_step, + // num_loop + index_t num_loop) const + { + ignore = b_block_desc; + ignore = b_block_buf; + __builtin_amdgcn_sched_barrier(0); + static_assert(CScaleThreadDesc{}.GetLength(Number<0>{}) == 1, + "Pipeline v3 only support scaleblocksliceK=1"); + static_assert(CScaleThreadDesc{}.GetLength(Number<2>{}) == 1, + "Pipeline v3 only support scaleblocksliceN=1"); + // assume kperblock = scaleblockk + 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; + constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0); + auto a_scale_thread_buf = make_static_buffer( + a_scale_thread_desc.GetElementSpaceSize()); + auto b_scale_thread_buf = make_static_buffer( + b_scale_thread_desc.GetElementSpaceSize()); + auto c_scale_thread_buf = make_static_buffer( + c_scale_thread_desc.GetElementSpaceSize()); + + // Global prefetch A1 B1, AScale1 BScale1 + 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.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); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0]; + }); + + // Local prefill A1 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); + + // Global prefetch A2, AScale2 BScale2 + 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) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step); + + // Initialize C + c_thread_buf.Clear(); + + StaticBufferTupleOfVector + c_thread_buf_per_scale; + + // Local prefetch A1 + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](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); + }); + + __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) { + static_for<0, MRepeat, 1>{}([&](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); + } + + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); + }); + }); + + if constexpr(m0.value == MRepeat - 1) + { + block_sync_lds(); + + 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(local_read_buf), + a_thread_desc_, + make_tuple( + Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) % + 2>{}, + I0, + I0, + k0, + I0, + I0), + 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); + }); + } + + HotLoopScheduler(m0); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0]; + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, + b_scale_thread_copy_step); + }; + + LoopFunc(I0, I1); + LoopFunc(I1, I0); + + i += 2; + } while(i < (num_loop - 2)); + } + + // tail + if constexpr(TailNum == TailNumber::Even) + { + static_for<0, MRepeat, 1>{}([&](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)); + } + + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); + }); + }); + + if constexpr(m0.value == MRepeat - 1) + { + block_sync_lds(); + + 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(I1), + a_thread_desc_, + make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0), + 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); + }); + } + + EpilogueScheduler_1(m0); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + c_scale_thread_buf(m0) = a_scale_thread_buf[m0] * b_scale_thread_buf[I0]; + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); + }); + }); + + if constexpr(m0.value != (MRepeat - 1)) + { + static_for<0, KRepeat, 1>{}([&](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); + }); + + EpilogueScheduler_2(); + } + }); + // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle + // latency + // __builtin_amdgcn_sched_barrier(0); + } + else + { + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; + }); + static_for<0, KRepeat, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; + + 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{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert(c_scale_thread_buf[m0]); + }); + }); + + if constexpr(m0.value != (MRepeat - 1)) + { + static_for<0, KRepeat, 1>{}([&](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); + }); + + EpilogueScheduler_2(); + } + }); + } + } + + 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/device/device_gemm_multiple_d_ab_scale.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_ab_scale.hpp index 7171715250..785eab0fda 100644 --- a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_ab_scale.hpp @@ -60,6 +60,49 @@ struct DeviceGemmMultipleD_ABScale : public BaseOperator virtual std::unique_ptr MakeInvokerPointer() = 0; }; +template +struct DeviceGemmMultipleD_BlockScale_BPreshuffle : public BaseOperator +{ + static constexpr index_t NumDTensor = DsDataType::Size(); + + virtual std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_e, + const ck::index_t M, + const ck::index_t N, + const ck::index_t K, + const ck::index_t StrideA, + const ck::index_t StrideB, + const std::array StrideDs, + const ck::index_t StrideE, + const void* p_a_scale, + const void* p_b_scale, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) = 0; + + virtual std::unique_ptr MakeInvokerPointer() = 0; + + virtual int GetPreShuffleParameters() = 0; +}; + } // namespace device } // namespace tensor_operation } // namespace ck diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp new file mode 100644 index 0000000000..6a067163c7 --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp @@ -0,0 +1,484 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_ab_scale.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp" +#include "ck/host_utility/device_prop.hpp" +#include "ck/host_utility/kernel_launch.hpp" +#include "ck/host_utility/flush_cache.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +template +struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle + : public DeviceGemmMultipleD_BlockScale_BPreshuffle +{ + static constexpr index_t NumDTensor = DsDataType::Size(); + + // GridwiseGemm + using GridwiseGemm = GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle< + ALayout, + BLayout, + DsLayout, + CLayout, + ADataType, + BDataType, + GemmAccDataType, + CShuffleDataType, + DsDataType, + CDataType, + AElementwiseOperation, + BElementwiseOperation, + CElementwiseOperation, + GemmSpec, + BlockSize, + ScaleBlockM, + ScaleBlockN, + ScaleBlockK, + MPerBlock, + NPerBlock, + KPerBlock, + AK1, + BK1, + MPerXDL, + NPerXDL, + MXdlPerWave, + NXdlPerWave, + ABlockTransferThreadClusterLengths_AK0_M_AK1, + ABlockTransferThreadClusterArrangeOrder, + ABlockTransferSrcAccessOrder, + ABlockTransferSrcVectorDim, + ABlockTransferSrcScalarPerVector, + ABlockTransferDstScalarPerVector_AK1, + false, + ABlockLdsExtraM, + BBlockTransferThreadClusterLengths_BK0_N_BK1, + BBlockTransferThreadClusterArrangeOrder, + BBlockTransferSrcAccessOrder, + BBlockTransferSrcVectorDim, + BBlockTransferSrcScalarPerVector, + BBlockTransferDstScalarPerVector_BK1, + false, + BBlockLdsExtraN, + CShuffleMXdlPerWavePerShuffle, + CShuffleNXdlPerWavePerShuffle, + CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, + CDEShuffleBlockTransferScalarPerVectors, + BlkGemmPipeSched, + BlkGemmPipelineVer, + ComputeTypeA, + ComputeTypeB, + LDSTypeA, + LDSTypeB>; + + using Argument = typename GridwiseGemm::Argument; + + int GetPreShuffleParameters() override { return NPerXDL; } + + // Invoker + struct Invoker : public BaseInvoker + { + float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) + { + if(stream_config.log_level_ > 0) + { + arg.Print(); + } + + if(!GridwiseGemm::CheckValidity(arg)) + { + throw std::runtime_error("wrong! GridwiseGemm has invalid setting"); + } + + index_t gdx, gdy, gdz; + std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch); + + float ave_time = 0; + + index_t k_grain = arg.KBatch * KPerBlock; + index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock; + + const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split); + + const auto Run = [&](const auto& kernel) { + if(stream_config.flush_cache) + { + Argument arg_ = arg; + + const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1( + arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0); + const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1( + arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0); + + auto size_a_buffer = + a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType); + auto size_b_buffer = + b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType); + + ck::utility::RotatingMemWrapper rotating_mem( + arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer); + rotating_mem.Print(); + + auto run_flush_cache = [&]() { + // flush icache + ck::utility::flush_icache(); + // rotating mem + rotating_mem.Next(); + // clear c mem + if(arg_.KBatch > 1) + hipGetErrorString(hipMemsetAsync(arg_.p_c_grid, + 0, + arg_.M * arg_.N * sizeof(CDataType), + stream_config.stream_id_)); + }; + + ave_time = ck::utility::launch_and_time_kernel_with_preprocess( + stream_config, + run_flush_cache, + kernel, + dim3(gdx, gdy, gdz), + dim3(BlockSize), + 0, + arg_); + } + else + { + if(arg.KBatch > 1) + hipGetErrorString(hipMemsetAsync(arg.p_c_grid, + 0, + arg.M * arg.N * sizeof(CDataType), + stream_config.stream_id_)); + + ave_time = launch_and_time_kernel( + stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg); + } + }; + + constexpr index_t minimum_occupancy = + (BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave && + MPerBlock * NPerBlock / BlockSize > 64) + ? 1 + : 2; + + if(has_main_k_block_loop) + { + // Tail number always full + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy>; + Run(kernel); + } + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + } + else + { + // Tail number always 1 + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy>; + Run(kernel); + } + } + } + return ave_time; + } + + // polymorphic + float Run(const BaseArgument* p_arg, + const StreamConfig& stream_config = StreamConfig{}) override + { + return Run(*dynamic_cast(p_arg), stream_config); + } + }; + + static constexpr bool IsValidCompilationParameter() + { + // TODO: properly implement this check + return true; + } + + static bool IsSupportedArgument(const Argument& arg) + { + if(!ck::is_xdl_supported()) + { + return false; + } + + // if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK != + // KPerBlock) + // { + // return false; + // } + if(!is_bf16_atomic_supported() && std::is_same_v && arg.KBatch > 1) + { + return false; + } + + if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding || + GemmSpec == GemmSpecialization::NKPadding || + GemmSpec == GemmSpecialization::MNKPadding || + GemmSpec == GemmSpecialization::KPadding)) + { + return false; + } + + // Padding to release this restriction + if(arg.N % NPerBlock != 0 || arg.K % KPerBlock != 0) + { + return false; + } + + return GridwiseGemm::CheckValidity(arg); + } + + // polymorphic + bool IsSupportedArgument(const BaseArgument* p_arg) override + { + return IsSupportedArgument(*dynamic_cast(p_arg)); + } + + static auto MakeArgument(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_c, + const index_t M, + const index_t N, + const index_t K, + const index_t StrideA, + const index_t StrideB, + const std::array StrideDs, + const index_t StrideC, + const void* p_a_scale, + const void* p_b_scale, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + return Argument{static_cast(p_a), + static_cast(p_b), + p_ds, + static_cast(p_c), + M, + N, + K, + StrideA, + StrideB, + StrideDs, + StrideC, + static_cast(p_a_scale), + static_cast(p_b_scale), + 1, + a_element_op, + b_element_op, + c_element_op}; + } + + static auto MakeInvoker() { return Invoker{}; } + + // polymorphic + std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_c, + const index_t M, + const index_t N, + const index_t K, + const index_t StrideA, + const index_t StrideB, + const std::array StrideDs, + const index_t StrideC, + const void* p_a_scale, + const void* p_b_scale, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) override + { + return std::make_unique(static_cast(p_a), + static_cast(p_b), + p_ds, + static_cast(p_c), + M, + N, + K, + StrideA, + StrideB, + StrideDs, + StrideC, + static_cast(p_a_scale), + static_cast(p_b_scale), + 1, + a_element_op, + b_element_op, + c_element_op); + } + + // polymorphic + std::unique_ptr MakeInvokerPointer() override + { + return std::make_unique(Invoker{}); + } + + // polymorphic + std::string GetTypeString() const override + { + auto str = std::stringstream(); + + std::map BlkGemmPipelineSchedulerToString{ + {BlockGemmPipelineScheduler::Intrawave, "Intrawave"}, + {BlockGemmPipelineScheduler::Interwave, "Interwave"}}; + + std::map BlkGemmPipelineVersionToString{ + {BlockGemmPipelineVersion::v1, "v1"}, + {BlockGemmPipelineVersion::v2, "v2"}, + {BlockGemmPipelineVersion::v3, "v3"}}; + + // clang-format off + str << "DeviceGemmXdlUniversal" + << "<" + << getGemmSpecializationString(GemmSpec) << ", " + << std::string(ALayout::name)[0] + << std::string(BLayout::name)[0] + << std::string(CLayout::name)[0] + << ">" + << " BlkSize: " + << BlockSize << ", " + << "BlkTile: " + << MPerBlock<<"x"< +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy) +#endif + // __attribute__((amdgpu_waves_per_eu(1, 1))) + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle( + typename GridwiseGemm::Argument karg) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__)) + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + GridwiseGemm::template Run( + karg.p_a_grid, + karg.p_b_grid, + karg.p_ds_grid, + karg.p_c_grid, + karg.p_a_scale_grid, + karg.p_b_scale_grid, + p_shared, + karg, + karg.a_element_op, + karg.b_element_op, + karg.c_element_op); +#else + ignore = karg; +#endif // end of if (defined(__gfx908__) || defined(__gfx90a__)) +} + +template +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy) +#endif + // __attribute__((amdgpu_waves_per_eu(1, 1))) + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds( + typename GridwiseGemm::Argument karg) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__)) + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + __shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + GridwiseGemm::template Run_2Lds( + karg.p_a_grid, + karg.p_b_grid, + karg.p_ds_grid, + karg.p_c_grid, + karg.p_a_scale_grid, + karg.p_b_scale_grid, + p_shared, + p_shared1, + karg, + karg.a_element_op, + karg.b_element_op, + karg.c_element_op); +#else + ignore = karg; +#endif // end of if (defined(__gfx908__) || defined(__gfx90a__)) +} + +template +struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle +{ + using AScaleType = float; + using BScaleType = float; + + static constexpr auto I0 = Number<0>{}; + static constexpr auto I1 = Number<1>{}; + static constexpr auto I2 = Number<2>{}; + static constexpr auto I3 = Number<3>{}; + static constexpr auto I4 = Number<4>{}; + static constexpr auto I5 = Number<5>{}; + static constexpr auto I6 = Number<6>{}; + static constexpr auto I7 = Number<7>{}; + + static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock = + CDEShuffleBlockTransferScalarPerVectors{}[I0]; + + // K1 should be Number<...> + static constexpr auto AK0Number = Number{}; + static constexpr auto BK0Number = Number{}; + static constexpr auto AK1Number = Number{}; + static constexpr auto BK1Number = Number{}; + static constexpr auto BlockSizeNumber = Number{}; + + static constexpr index_t NumDTensor = DsDataType::Size(); + using mfma_selector = MfmaSelector; + static constexpr index_t KPack = + 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 NLane = NPerXdl; + static constexpr index_t NWave = NPerBlock / NPerXdl / NXdlPerWave; + + static constexpr auto MakeDsGridPointer() + { + return generate_tuple( + [&](auto i) { + using DDataType = remove_cvref_t>; + + return static_cast(nullptr); + }, + Number{}); + } + + using DsGridPointer = decltype(MakeDsGridPointer()); + + using ThisThreadBlock = ThisThreadBlock; + + __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch) + { + return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch); + } + + __host__ __device__ static auto CalculateMPadded(index_t M) + { + return math::integer_least_multiple(M, MPerBlock); + } + + __host__ __device__ static auto CalculateNPadded(index_t N) + { + return math::integer_least_multiple(N, NPerBlock); + } + + __host__ __device__ static auto CalculateBN0Shuffled(index_t N) + { + return math::integer_divide_ceil(N, NLane); + } + __host__ __device__ static auto CalculateBK0Shuffled(index_t K) + { + return math::integer_divide_ceil(K, KLane * KPack); + } + + __host__ __device__ static auto CalculateKPadded(index_t K) + { + return math::integer_divide_ceil(K, KPerBlock) * KPerBlock; + } + + __host__ __device__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * (KPerBlock / AK1Value); + } + + __host__ __device__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * (KPerBlock / BK1Value); + } + + __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * KPerBlock; + } + + __host__ __device__ static auto CalculateKRead(index_t K, index_t K_Batch = 1) + { + constexpr auto KReadVec = math::lcm(AK1Number, BK1Number); + auto K_t = K_Batch * KReadVec; + return (K + K_t - 1) / K_t * KReadVec; + } + + __host__ __device__ static auto CalculateMBlock(index_t M) + { + return math::integer_divide_ceil(M, MPerBlock); + } + + __host__ __device__ static auto CalculateNBlock(index_t N) + { + return math::integer_divide_ceil(N, NPerBlock); + } + + template + __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&) + { + constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{}); + constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{}); + + return transform_tensor_descriptor( + TileDesc_K0_MN_K1{}, + make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number{}, Number{})), + make_unmerge_transform(make_tuple( + Number{}, Number{}, Number{}))), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}), + make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{})); + } + + __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1( + index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0) + { + const auto a_grid_desc_mraw_kraw = [&]() { + if constexpr(is_same_v) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1)); + } + else if constexpr(is_same_v) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA)); + } + }(); + + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::MKPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad both M and K + const auto a_grid_desc_m_k = + transform_tensor_descriptor(a_grid_desc_mraw_kraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(MPad)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else if constexpr(GemmSpec == GemmSpecialization::MPadding || + GemmSpec == GemmSpecialization::MNPadding) + { + // pad M, but not K + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_right_pad_transform(M, MPad - M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else if constexpr(GemmSpec == GemmSpecialization::KPadding || + GemmSpec == GemmSpecialization::NKPadding) + { + // pad K, but not M + const auto a_grid_desc_m_k = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else + { + // not pad M or K + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + } + + __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0) + { + 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)); + } + + __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1( + index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0) + { + const auto b_grid_desc_nraw_kraw = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1)); + } + }(); + + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::NKPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad both N and K + const auto b_grid_desc_n_k = + transform_tensor_descriptor(b_grid_desc_nraw_kraw, + make_tuple(make_right_pad_transform(N, NPad - N), + make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_n_k, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(NPad)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else if constexpr(GemmSpec == GemmSpecialization::NPadding || + GemmSpec == GemmSpecialization::MNPadding) + { + // pad N, but not K + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else if constexpr(GemmSpec == GemmSpecialization::KPadding || + GemmSpec == GemmSpecialization::MKPadding) + { + // pad K, but not N + const auto b_grid_desc_n_k = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_n_k, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else + { + // not pad N or K + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + } + + template + __host__ __device__ static constexpr auto + MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&) + { + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + + return MakeGemmMmaTileDescriptor(ABlockDesc_AK0_M_AK1{}); + } + + template + __host__ __device__ static constexpr auto + MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&) + { + constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); + + return MakeGemmMmaTileDescriptor(BBlockDesc_BK0_N_BK1{}); + } + + template + __host__ __device__ static auto + MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC) + { + const auto c_grid_desc_mraw_nraw = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC)); + } + }(); + + // pad M and N + return transform_tensor_descriptor(c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); +#if 0 + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::MNPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad M and N + return transform_tensor_descriptor(c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else if constexpr(GemmSpec == GemmSpecialization::MPadding || + GemmSpec == GemmSpecialization::MKPadding) + { + // pad M, but not N + return transform_tensor_descriptor( + c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else if constexpr(GemmSpec == GemmSpecialization::NPadding || + GemmSpec == GemmSpecialization::NKPadding) + { + // pad N, but not M + return transform_tensor_descriptor( + c_grid_desc_mraw_nraw, + make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else + { + // not pad M or N + return c_grid_desc_mraw_nraw; + } +#endif + } + + __host__ __device__ static auto MakeDsGridDescriptor_M_N( + index_t M, index_t MPad, index_t N, index_t NPad, std::array StrideDs) + { + return generate_tuple( + [&](auto i) { + using DLayout = remove_cvref_t>; + return MakeCGridDescriptor_M_N(M, MPad, N, NPad, StrideDs[i]); + }, + Number{}); + } + + template + __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock) + { + return generate_tuple( + [&](auto i) { + return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + ds_grid_desc_m_n[i], MBlock, NBlock); + }, + Number{}); + } + + using DsGridDesc_M_N = remove_cvref_t; + + struct Problem + { + __host__ __device__ Problem(index_t M_, + index_t N_, + index_t K_, + index_t StrideA_, + index_t StrideB_, + std::array StrideDs_, + index_t StrideC_, + index_t KBatch_) + : M{M_}, + N{N_}, + K{K_}, + StrideA{StrideA_}, + StrideB{StrideB_}, + StrideDs{StrideDs_}, + StrideC{StrideC_}, + KBatch{KBatch_}, + MPadded{CalculateMPadded(M_)}, + NPadded{CalculateNPadded(N_)}, + KRead{CalculateKRead(K_, KBatch_)}, + KPadded{CalculateKPadded(K_, KBatch_)}, + AK0{CalculateAK0Padded(K_, KBatch_)}, + BK0{CalculateBK0Padded(K_, KBatch_)}, + MBlock{CalculateMBlock(M_)}, + NBlock{CalculateNBlock(N_)}, + BN0Shuffled{CalculateBN0Shuffled(N_)}, + BK0Shuffled{CalculateBK0Shuffled(K_)} + { + } + + __host__ void Print() const + { + std::cout << "problem {" + << "M:" << M << ", " + << "N:" << N << ", " + << "K:" << K << ", " + << "SA:" << StrideA << ", " + << "SB:" << StrideB << ", " + << "SC:" << StrideC << ", " + << "MP:" << MPadded << ", " + << "NP:" << NPadded << ", " + << "KRead:" << KRead << ", " + << "KP:" << KPadded << ", " + << "AK0:" << AK0 << ", " + << "BK0:" << BK0 << ", " + << "MBlock: " << MBlock << ", " + << "NBlock: " << NBlock << "}" << std::endl; + } + + index_t M; + index_t N; + index_t K; + index_t StrideA; + index_t StrideB; + std::array StrideDs; + index_t StrideC; + + index_t KBatch; + index_t MPadded; + index_t NPadded; + index_t KRead; + index_t KPadded; + index_t AK0; + index_t BK0; + index_t MBlock; + index_t NBlock; + // FOR PRESHUFFLE ONLY + index_t BN0Shuffled; + index_t BK0Shuffled; + }; + + // Argument + struct Argument : public tensor_operation::device::BaseArgument, public Problem + { + __host__ Argument(const ADataType* p_a_grid_, + const BDataType* p_b_grid_, + std::array p_ds_grid_, + CDataType* p_c_grid_, + index_t M_, + index_t N_, + index_t K_, + index_t StrideA_, + index_t StrideB_, + std::array StrideDs_, + index_t StrideC_, + const AScaleType* p_a_scale_grid_, + const BScaleType* p_b_scale_grid_, + index_t k_batch_, + AElementwiseOperation a_element_op_, + BElementwiseOperation b_element_op_, + CElementwiseOperation c_element_op_) + : Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_}, + p_a_grid{p_a_grid_}, + p_b_grid{p_b_grid_}, + p_ds_grid{}, + p_c_grid{p_c_grid_}, + p_a_scale_grid{p_a_scale_grid_}, + p_b_scale_grid{p_b_scale_grid_}, + a_element_op{a_element_op_}, + b_element_op{b_element_op_}, + c_element_op{c_element_op_} + { + + // populate pointer, desc for Ds + static_for<0, NumDTensor, 1>{}([&](auto i) { + using DDataType_ = remove_cvref_t>; + + // D pointer + p_ds_grid(i) = static_cast(p_ds_grid_[i]); + }); + } + + const ADataType* p_a_grid; + const BDataType* p_b_grid; + DsGridPointer p_ds_grid; + CDataType* p_c_grid; + + const AScaleType* p_a_scale_grid; + const BScaleType* p_b_scale_grid; + + const AElementwiseOperation a_element_op; + const BElementwiseOperation b_element_op; + const CElementwiseOperation c_element_op; + }; + + struct SplitKBatchOffset + { + __device__ SplitKBatchOffset(Argument& karg) + { + if constexpr(is_same_v) + { + a_k_split_offset = blockIdx.z * karg.KRead; + } + else if constexpr(is_same_v) + { + a_k_split_offset = blockIdx.z * karg.KRead * karg.M; + } + + if constexpr(is_same_v) + { + b_k_split_offset = blockIdx.z * karg.KRead * karg.N; + } + else if constexpr(is_same_v) + { + b_k_split_offset = blockIdx.z * karg.KRead; + } + + if(blockIdx.z < static_cast(karg.KBatch - 1)) + { + karg.K = karg.KRead; + } + else + { + karg.K = karg.K - karg.KRead * (karg.KBatch - 1); + } + } + + index_t a_k_split_offset; + index_t b_k_split_offset; + }; + + __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1() + { + // A matrix in LDS memory, dst of blockwise copy + if constexpr(ABlockLdsExtraM) + { + return make_naive_tensor_descriptor( + make_tuple(AK0Number, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); + } + // xor tensor transformation request more unnecessary vgpr usage, would cause register spill + // in some cases. + else if constexpr(is_same::value) + { + constexpr auto a_lds_block_desc = + make_naive_tensor_descriptor(make_tuple(AK0Number, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); + + constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( + a_lds_block_desc, + make_tuple(make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<1, 0>{}, Sequence<2>{}), + make_tuple(Sequence<1, 0>{}, Sequence<2>{})); + + return a_lds_block_desc_permuted; + } + else // ColumnMajor A + { + // kfold and mpair dimension is not always required. + // more dimension in merge_transform increase the difficulty of generating immarg offset + // for compiler. + constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1); + constexpr auto M1 = MPerBlock / M0; + + constexpr auto KThreadWrite = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0); + constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite; + constexpr auto KThreadRead = 64 / MPerXdl; + constexpr auto K0PerThreadRead = AK0Number / KThreadRead; + + constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128) + ? 1 + : 128 / (AK1Number * M0 * sizeof(LDSTypeA)); + constexpr auto KThreadReadPerm = + (kfold * K0PerThreadWrite / K0PerThreadRead) > 1 + ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead) + : KThreadRead; + + // 1<=mpair<=n0 + constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128) + ? 1 + : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0 + ? M0 + : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))); + + constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, + Number{}, + Number{}, + Number{}, + Number{}, + AK1Number)); + + constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( + a_lds_block_desc, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_xor_with_modulo_transform( + make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{})); + + constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor( + a_lds_block_desc_permuted, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, + Sequence<1>{}, + Sequence<2>{}, + Sequence<3>{}, + Sequence<4>{}, + Sequence<5>{}), + make_tuple(Sequence<1>{}, + Sequence<2>{}, + Sequence<0, 3>{}, + Sequence<4, 5>{}, + Sequence<6>{}, + Sequence<7>{})); + + constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_unmerged, + make_tuple(make_merge_transform_v3_division_mod( + make_tuple(Number{}, + Number{}, + Number{}, + Number{})), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{}, Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return a_lds_block_desc_ak0_m_ak1; + } + } + + __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1() + { + // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack + return make_naive_tensor_descriptor_packed( + make_tuple(Number{}, I1, Number{}, Number{})); + } + + __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock() + { + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + make_naive_tensor_descriptor_packed( + make_tuple(I1, + Number{}, + I1, + Number{})); + + return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock; + } + + using BlockwiseGemmPipe = + remove_cvref_t())>; + + __device__ static constexpr index_t GetSharedMemoryNumberOfByte() + { + // LDS allocation for A and B: be careful of alignment + constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); + // lds max alignment + constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); + + constexpr auto a_block_space_size_aligned = math::integer_least_multiple( + a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align); + + // LDS allocation for C shuffle in LDS + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); + + constexpr auto c_block_size = + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(); + + return math::max(a_block_space_size_aligned * sizeof(LDSTypeA), + c_block_size * sizeof(CShuffleDataType)); + } + + // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01} + __host__ static constexpr bool CheckValidity(const Argument& karg) + { + static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) && + (NPerBlock % (NXdlPerWave * NPerXdl)) == 0, + "Invalid tuning param!"); + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) && + !(is_same::value)) + { + if(!(karg.M % MPerBlock == 0)) + { +#if DEBUG_LOG + std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " " + << __FILE__ << ":" << __LINE__ << ", in function: " << __func__ + << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) && + (is_same::value)) + { + if(!(karg.N % NPerBlock == 0)) + { +#if DEBUG_LOG + std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " " + << __FILE__ << ":" << __LINE__ << ", in function: " << __func__ + << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + { + + auto K_t = karg.KBatch * KPerBlock; + if(!(karg.K % K_t == 0)) + { +#if DEBUG_LOG + std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: " + << karg.K << " " << __FILE__ << ":" << __LINE__ + << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + constexpr auto KReadVec = math::lcm(AK1Number, BK1Number); + auto K_t = karg.KBatch * KReadVec; + auto KReadPadSplited = math::integer_divide_ceil(karg.K, K_t) * KReadVec; + if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K) + { + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.K % ABlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg K (" << karg.K + << ") value is not a multiple of ABlockTransferSrcScalarPerVector (" + << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.M % ABlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg M (" << karg.M + << ") value is not a multiple of ABlockTransferSrcScalarPerVector (" + << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.N % BBlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg N (" << karg.N + << ") value is not a multiple of BBlockTransferSrcScalarPerVector (" + << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.K % BBlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg K (" << karg.K + << ") value is not a multiple of BBlockTransferSrcScalarPerVector (" + << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0) + { +#if DEBUG_LOG + std::cout << "Arg N (" << karg.N + << ") value is not a multiple of " + "CShuffleBlockTransferScalarPerVector_NPerBlock (" + << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__ + << ":" << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0) + { +#if DEBUG_LOG + std::cout << "Arg M (" << karg.M + << ") value is not a multiple of " + "CShuffleBlockTransferScalarPerVector_NPerBlock (" + << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__ + << ":" << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + // check gridwise gemm pipeline + const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value); + + if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1) + { + if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages) + { + return false; + } + } + + // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc) + return true; + } + + __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K) + { + const index_t num_loop = K / KPerBlock; + + return BlockwiseGemmPipe::BlockHasHotloop(num_loop); + } + + __host__ __device__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K) + { + const index_t num_loop = K / KPerBlock; + + return BlockwiseGemmPipe::BlockLoopTailNum(num_loop); + } + + template + __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock) + { + const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor( + c_grid_desc_m_n, + make_tuple(make_unmerge_transform(make_tuple(MBlock, Number{})), + make_unmerge_transform(make_tuple(NBlock, Number{}))), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{})); + + return c_grid_desc_mblock_mperblock_nblock_nperblock; + } + + // return block_id to C matrix tile idx (m0, n0) mapping + // if arch = gfx942 + using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>; + + template + __device__ static void Run(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + const AScaleType* p_a_scale_grid, + const BScaleType* p_b_scale_grid, + void* p_shared, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + ignore = b_element_op; + const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( + problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); + const auto b_grid_desc_bpreshuffled = + MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled); + const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC); + + const auto a_scale_grid_desc_am_ak = make_naive_tensor_descriptor( + make_tuple(math::integer_divide_ceil(problem.M, ScaleBlockM), + math::integer_divide_ceil(problem.K, ScaleBlockK)), + make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1)); + const auto b_scale_grid_desc_bn_ak = make_naive_tensor_descriptor( + make_tuple(math::integer_divide_ceil(problem.N, ScaleBlockN), + math::integer_divide_ceil(problem.K, ScaleBlockK)), + make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1)); + + const auto c_grid_desc_mblock_mperblock_nblock_nperblock = + MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + c_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto a_grid_buf = make_dynamic_buffer( + p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize()); + const auto b_grid_buf = make_dynamic_buffer( + p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize()); + auto c_grid_buf = make_dynamic_buffer( + p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + const auto a_scale_grid_buf = make_dynamic_buffer( + p_a_scale_grid, a_scale_grid_desc_am_ak.GetElementSpaceSize()); + + const auto b_scale_grid_buf = make_dynamic_buffer( + p_b_scale_grid, b_scale_grid_desc_bn_ak.GetElementSpaceSize()); + + // divide block work by [M, N] + const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4}; + + const auto block_work_idx = + block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id())); + + if(!block_2_ctile_map.ValidCTileIndex( + block_work_idx, + make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0), + c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2)))) + { + return; + } + + const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]); + const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]); + + // HACK: this force m/n_block_data_idx_on_grid into SGPR + const index_t m_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock); + + const index_t n_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave); + + // lds max alignment + constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); + + // A matrix in LDS memory, dst of blockwise copy + constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); + + // B matrix in LDS memory, dst of blockwise copy + constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1(); + + // A matrix blockwise copy + auto a_blockwise_copy = + ThreadGroupTensorSliceTransfer_v4r1, + ABlockTransferThreadClusterLengths_AK0_M_AK1, + ABlockTransferThreadClusterArrangeOrder, + ADataType, + LDSTypeA, + decltype(a_grid_desc_ak0_m_ak1), + decltype(a_block_desc_ak0_m_ak1), + ABlockTransferSrcAccessOrder, + Sequence<0, 1, 2>, + ABlockTransferSrcVectorDim, + 2, + ABlockTransferSrcScalarPerVector, + ABlockTransferDstScalarPerVector_AK1, + 1, + 1, + AThreadTransferSrcResetCoordinateAfterRun, + true, + BlockwiseGemmPipe::GlobalBufferNum>( + a_grid_desc_ak0_m_ak1, + make_multi_index(0, m_block_data_idx_on_grid, 0), + a_element_op, + a_block_desc_ak0_m_ak1, + make_multi_index(0, 0, 0), + ck::tensor_operation::element_wise::PassThrough{}); + + // Thread-wise copy + // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack + auto b_block_buf = make_static_buffer( + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + + auto b_blockwise_copy = 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 * (get_thread_local_1d_id() % warpSize))); + + // LDS allocation for A and B: be careful of alignment + // Cast after lds + auto a_block_buf = make_dynamic_buffer( + static_cast(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + + constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0); + constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0); + + // Blockwise GEMM pipeline + static_assert(std::is_default_constructible_v); + auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; + auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); + + 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); + + constexpr index_t ScaleSliceSizeM = MXdlPerWave; + constexpr index_t ScaleSliceSizeN = math::integer_divide_ceil(NPerBlock, ScaleBlockN); + constexpr index_t ScaleSliceSizeK = math::integer_divide_ceil(KPerBlock, ScaleBlockK); + + // ScaleSliceSizeK is last dimension in A/B scale for vector memory access + // ScaleSliceSizeK is first dimension in C scale for packed math + constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, Number{})); + + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); + auto a_thread_offset = + get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl; + + constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, Number{})); + + constexpr auto c_scale_thread_desc = make_naive_tensor_descriptor_packed(make_tuple( + Number{}, Number{}, Number{})); + + auto a_scale_thread_copy = + ThreadwiseTensorSliceTransfer_v2, + Sequence<0, 1>, + 1, + ScaleSliceSizeK, + 1, + false>( + a_scale_grid_desc_am_ak, + make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); + + auto b_scale_thread_copy = + ThreadwiseTensorSliceTransfer_v2, + Sequence<0, 1>, + 1, + ScaleSliceSizeK, + 1, + false>( + b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0)); + + // constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1); + constexpr auto a_scale_thread_slice_copy_step = + make_tuple(make_multi_index(MWaves * MPerXdl, 0), + make_multi_index(-MPerBlock, 0), + make_multi_index(-MPerBlock, ScaleSliceSizeK)); + constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, ScaleSliceSizeK); + + constexpr auto NumKBlockPerScale = math::integer_divide_ceil(ScaleBlockK, 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_buf, + a_block_slice_copy_step, + b_grid_desc_bpreshuffled, + b_block_desc_bk0_n_bk1, + b_blockwise_copy, + b_grid_buf, + b_block_buf, + b_block_slice_copy_step, + + c_scale_thread_desc, + c_thread_buf, + + a_scale_grid_desc_am_ak, + a_scale_thread_desc, + a_scale_thread_copy, + a_scale_grid_buf, + a_scale_thread_slice_copy_step, + + b_scale_grid_desc_bn_ak, + b_scale_thread_desc, + b_scale_thread_copy, + b_scale_grid_buf, + b_scale_thread_slice_copy_step, + + num_k_block_main_loop); + + // shuffle C and write out + { + static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && + NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, + "wrong!"); + + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + + // transposed XDL + // // TODO: hacky, fix it! + constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 = + blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + // // TODO: hacky, fix it! + // only used to get lengths + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp = + blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0); + constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1); + constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2); + constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3); + constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4); + constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5); + constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6); + constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7); + + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); + + auto c_shuffle_block_buf = make_dynamic_buffer( + static_cast(p_shared), + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor( + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, + make_tuple( + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // M0 (MXdlPerWave) per shuffle + M1, // M1 = MWave + M2)), // M2 = MPerXdl + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // N0 (NXdlPerWave) per shuffle + N1, // N1 = NWave + N2, // N2 * N3 * N4 = NPerXdl + N3, + N4))), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple( + Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{})); + + // calculate origin of thread output tensor on global memory + // blockwise GEMM c matrix starting index + const auto c_thread_mtx_on_block = + blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0); + + const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0]; + const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1]; + + const auto m_thread_data_on_block_to_m0_m1_m2_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(M0, M1, M2))), + make_tuple(Sequence<0, 1, 2>{}), + make_tuple(Sequence<0>{})); + + const auto m_thread_data_on_block_idx = + m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex( + make_multi_index(m_thread_data_on_block)); + + const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))), + make_tuple(Sequence<0, 1, 2, 3, 4>{}), + make_tuple(Sequence<0>{})); + + const auto n_thread_data_on_block_idx = + n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex( + make_multi_index(n_thread_data_on_block)); + + // shuffle: threadwise copy C from VGPR to LDS + auto c_thread_copy_vgpr_to_lds = + ThreadwiseTensorSliceTransfer_v1r3, + Sequence<0, 1, 2, 3, 4, 5, 6, 7>, + 7, + 1, + InMemoryDataOperationEnum::Set, + 1, + true>{ + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, + make_multi_index(0, + 0, + m_thread_data_on_block_idx[I1], + n_thread_data_on_block_idx[I1], + m_thread_data_on_block_idx[I2], + n_thread_data_on_block_idx[I2], + n_thread_data_on_block_idx[I3], + n_thread_data_on_block_idx[I4]), + tensor_operation::element_wise::PassThrough{}}; + + using EDataType = CDataType; + + const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs); + + const auto ds_grid_desc_mblock_mperblock_nblock_nperblock = + MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + ds_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto ds_grid_buf = generate_tuple( + [&](auto i) { + return make_dynamic_buffer( + p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize()); + }, + Number{}); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_desc_refs = concat_tuple_of_reference( + tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; }, + Number{})); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_buf_refs = concat_tuple_of_reference( + tie(c_shuffle_block_buf), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_buf[i]; }, + Number{})); + + // tuple of starting index of C/Ds blockwise copy + const auto idx_c_ds_block_begin = container_concat( + make_tuple(make_multi_index(0, 0, 0, 0)), + generate_tuple( + [&](auto) { + return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0); + }, + Number{})); + + const auto e_grid_desc_mblock_mperblock_nblock_nperblock = + c_grid_desc_mblock_mperblock_nblock_nperblock; + + using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock = + CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock; + const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation; + + auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3< + ThisThreadBlock, + decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})), + Tuple, + decltype(c_ds_desc_refs), + decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)), + CElementwiseOperation, + Sequence(EGlobalMemoryDataOperation)>, // FIXME: make Sequence + // support arbitray type + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths, + CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, + Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder, + Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder, + Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder, + 3, // index_t SrcVectorDim, + 3, // index_t DstVectorDim, + CDEShuffleBlockTransferScalarPerVectors, + CShuffleBlockTransferScalarPerVector_NPerBlock, + sequence_merge_t< + Sequence, + uniform_sequence_gen_t>, // ThreadTransferSrcResetCoordinateAfterRunFlags + Sequence> // ThreadTransferDstResetCoordinateAfterRunFlags + {c_ds_desc_refs, + idx_c_ds_block_begin, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)), + c_element_op}; + + constexpr auto sfc_c_vgpr = + SpaceFillingCurve, + Sequence<0, 1, 2, 3, 4, 5, 6, 7>, + Sequence>{}; + + constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess(); + + // space filling curve for shuffled blockwise C/D/E + constexpr auto sfc_cde_block = + SpaceFillingCurve, + Sequence<0, 2, 1, 3>, + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{}; + + static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!"); + + static_for<0, num_access, 1>{}([&](auto access_id) { + // make sure it's safe to write to LDS + block_sync_lds(); + + // each thread write its data from VGPR to LDS + c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4, + sfc_c_vgpr.GetIndexTupleOfNumber(access_id), + c_thread_buf, + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, + c_shuffle_block_buf); + + // make sure it's safe to read from LDS + block_sync_lds(); + + // each block copy its data from LDS to global + cde_block_copy_lds_and_global.Run( + c_ds_desc_refs, + c_ds_buf_refs, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + tie(c_grid_buf)); + + if constexpr(access_id < num_access - 1) + { + constexpr auto cde_lds_and_global_step = + sfc_cde_block.GetForwardStep(access_id); + + // move on Ds + static_for<0, NumDTensor, 1>{}([&](auto i) { + cde_block_copy_lds_and_global.MoveSrcSliceWindow( + c_ds_desc_refs, i + I1, cde_lds_and_global_step); + }); + + // move on E + cde_block_copy_lds_and_global.MoveDstSliceWindow( + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + I0, + cde_lds_and_global_step); + } + }); + } + } + + template + __device__ static void Run_2Lds(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + const AScaleType* p_a_scale_grid, + const BScaleType* p_b_scale_grid, + void* p_shared, + void* p_shared1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + ignore = b_element_op; + const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( + problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); + const auto b_grid_desc_bpreshuffled = + MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled); + const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC); + + const auto a_scale_grid_desc_am_ak = make_naive_tensor_descriptor( + make_tuple(math::integer_divide_ceil(problem.M, ScaleBlockM), + math::integer_divide_ceil(problem.K, ScaleBlockK)), + make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1)); + const auto b_scale_grid_desc_bn_ak = make_naive_tensor_descriptor( + make_tuple(math::integer_divide_ceil(problem.N, ScaleBlockN), + math::integer_divide_ceil(problem.K, ScaleBlockK)), + make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1)); + + const auto c_grid_desc_mblock_mperblock_nblock_nperblock = + MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + c_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto a_grid_buf = make_dynamic_buffer( + p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize()); + const auto b_grid_buf = make_dynamic_buffer( + p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize()); + auto c_grid_buf = make_dynamic_buffer( + p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + const auto a_scale_grid_buf = make_dynamic_buffer( + p_a_scale_grid, a_scale_grid_desc_am_ak.GetElementSpaceSize()); + + const auto b_scale_grid_buf = make_dynamic_buffer( + p_b_scale_grid, b_scale_grid_desc_bn_ak.GetElementSpaceSize()); + + // divide block work by [M, N] + const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4}; + + const auto block_work_idx = + block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id())); + + if(!block_2_ctile_map.ValidCTileIndex( + block_work_idx, + make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0), + c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2)))) + { + return; + } + + const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]); + const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]); + + // HACK: this force m/n_block_data_idx_on_grid into SGPR + const index_t m_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock); + + const index_t n_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave); + + // A matrix in LDS memory, dst of blockwise copy + constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); + + // B matrix in LDS memory, dst of blockwise copy + constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1(); + + // A matrix blockwise copy + auto a_blockwise_copy = + ThreadGroupTensorSliceTransfer_v4r1, + ABlockTransferThreadClusterLengths_AK0_M_AK1, + ABlockTransferThreadClusterArrangeOrder, + ADataType, + LDSTypeA, + decltype(a_grid_desc_ak0_m_ak1), + decltype(a_block_desc_ak0_m_ak1), + ABlockTransferSrcAccessOrder, + Sequence<0, 1, 2>, + ABlockTransferSrcVectorDim, + 2, + ABlockTransferSrcScalarPerVector, + ABlockTransferDstScalarPerVector_AK1, + 1, + 1, + AThreadTransferSrcResetCoordinateAfterRun, + true, + BlockwiseGemmPipe::GlobalBufferNum>( + a_grid_desc_ak0_m_ak1, + make_multi_index(0, m_block_data_idx_on_grid, 0), + a_element_op, + a_block_desc_ak0_m_ak1, + make_multi_index(0, 0, 0), + ck::tensor_operation::element_wise::PassThrough{}); + + // Thread-wise copy + // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack + auto b_block_buf_ping = make_static_buffer( + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + auto b_block_buf_pong = make_static_buffer( + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong); + + auto b_blockwise_copy = 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 * (get_thread_local_1d_id() % warpSize))); + + // LDS allocation for A and B: be careful of alignment + // Cast after lds + auto a_block_buf_ping = make_dynamic_buffer( + static_cast(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + auto a_block_buf_pong = make_dynamic_buffer( + static_cast(p_shared1), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong); + + constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0); + constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0); + + // Blockwise GEMM pipeline + static_assert(std::is_default_constructible_v); + auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; + auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); + + 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); + + constexpr index_t ScaleSliceSizeM = MXdlPerWave; + constexpr index_t ScaleSliceSizeN = math::integer_divide_ceil(NPerBlock, ScaleBlockN); + constexpr index_t ScaleSliceSizeK = math::integer_divide_ceil(KPerBlock, ScaleBlockK); + + // ScaleSliceSizeK is last dimension in A/B scale for vector memory access + // ScaleSliceSizeK is first dimension in C scale for packed math + constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, Number{})); + + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); + auto a_thread_offset = + get_thread_local_1d_id() % MPerXdl + (get_thread_local_1d_id() / 64) / NWaves * MPerXdl; + + constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, Number{})); + + constexpr auto c_scale_thread_desc = make_naive_tensor_descriptor_packed(make_tuple( + Number{}, Number{}, Number{})); + + auto a_scale_thread_copy = + ThreadwiseTensorSliceTransfer_v2, + Sequence<0, 1>, + 1, + ScaleSliceSizeK, + 1, + false>( + a_scale_grid_desc_am_ak, + make_multi_index(block_m_id * MPerBlock / ScaleBlockM + a_thread_offset, 0)); + + auto b_scale_thread_copy = + ThreadwiseTensorSliceTransfer_v2, + Sequence<0, 1>, + 1, + ScaleSliceSizeK, + 1, + false>( + b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0)); + + // constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1); + constexpr auto a_scale_thread_slice_copy_step = + make_tuple(make_multi_index(MWaves * MPerXdl, 0), + make_multi_index(-MPerBlock, 0), + make_multi_index(-MPerBlock, ScaleSliceSizeK)); + constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, ScaleSliceSizeK); + + constexpr auto NumKBlockPerScale = math::integer_divide_ceil(ScaleBlockK, 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_block_desc_bk0_n_bk1, + b_blockwise_copy, + b_grid_buf, + b_block_bufs, + b_block_slice_copy_step, + + c_scale_thread_desc, + c_thread_buf, + + a_scale_grid_desc_am_ak, + a_scale_thread_desc, + a_scale_thread_copy, + a_scale_grid_buf, + a_scale_thread_slice_copy_step, + + b_scale_grid_desc_bn_ak, + b_scale_thread_desc, + b_scale_thread_copy, + b_scale_grid_buf, + b_scale_thread_slice_copy_step, + + num_k_block_main_loop); + + // shuffle C and write out + { + static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && + NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, + "wrong!"); + + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + // constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); + + // transposed XDL + // // TODO: hacky, fix it! + constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 = + blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + // // TODO: hacky, fix it! + // only used to get lengths + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp = + blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(); + + constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0); + constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1); + constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2); + constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3); + constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4); + constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5); + constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6); + constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7); + + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); + + auto c_shuffle_block_buf = make_dynamic_buffer( + static_cast(p_shared), + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor( + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, + make_tuple( + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // M0 (MXdlPerWave) per shuffle + M1, // M1 = MWave + M2)), // M2 = MPerXdl + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // N0 (NXdlPerWave) per shuffle + N1, // N1 = NWave + N2, // N2 * N3 * N4 = NPerXdl + N3, + N4))), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple( + Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{})); + + // calculate origin of thread output tensor on global memory + // blockwise GEMM c matrix starting index + const auto c_thread_mtx_on_block = + blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0); + + const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0]; + const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1]; + + const auto m_thread_data_on_block_to_m0_m1_m2_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(M0, M1, M2))), + make_tuple(Sequence<0, 1, 2>{}), + make_tuple(Sequence<0>{})); + + const auto m_thread_data_on_block_idx = + m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex( + make_multi_index(m_thread_data_on_block)); + + const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))), + make_tuple(Sequence<0, 1, 2, 3, 4>{}), + make_tuple(Sequence<0>{})); + + const auto n_thread_data_on_block_idx = + n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex( + make_multi_index(n_thread_data_on_block)); + + // shuffle: threadwise copy C from VGPR to LDS + auto c_thread_copy_vgpr_to_lds = + ThreadwiseTensorSliceTransfer_v1r3, + Sequence<0, 1, 2, 3, 4, 5, 6, 7>, + 7, + 1, + InMemoryDataOperationEnum::Set, + 1, + true>{ + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, + make_multi_index(0, + 0, + m_thread_data_on_block_idx[I1], + n_thread_data_on_block_idx[I1], + m_thread_data_on_block_idx[I2], + n_thread_data_on_block_idx[I2], + n_thread_data_on_block_idx[I3], + n_thread_data_on_block_idx[I4]), + tensor_operation::element_wise::PassThrough{}}; + + using EDataType = CDataType; + + const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs); + + const auto ds_grid_desc_mblock_mperblock_nblock_nperblock = + MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + ds_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto ds_grid_buf = generate_tuple( + [&](auto i) { + return make_dynamic_buffer( + p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize()); + }, + Number{}); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_desc_refs = concat_tuple_of_reference( + tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; }, + Number{})); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_buf_refs = concat_tuple_of_reference( + tie(c_shuffle_block_buf), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_buf[i]; }, + Number{})); + + // tuple of starting index of C/Ds blockwise copy + const auto idx_c_ds_block_begin = container_concat( + make_tuple(make_multi_index(0, 0, 0, 0)), + generate_tuple( + [&](auto) { + return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0); + }, + Number{})); + + const auto e_grid_desc_mblock_mperblock_nblock_nperblock = + c_grid_desc_mblock_mperblock_nblock_nperblock; + + using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock = + CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock; + const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation; + + auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3< + ThisThreadBlock, + decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})), + Tuple, + decltype(c_ds_desc_refs), + decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)), + CElementwiseOperation, + Sequence(EGlobalMemoryDataOperation)>, // FIXME: make Sequence + // support arbitray type + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths, + CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, + Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder, + Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder, + Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder, + 3, // index_t SrcVectorDim, + 3, // index_t DstVectorDim, + CDEShuffleBlockTransferScalarPerVectors, + CShuffleBlockTransferScalarPerVector_NPerBlock, + sequence_merge_t< + Sequence, + uniform_sequence_gen_t>, // ThreadTransferSrcResetCoordinateAfterRunFlags + Sequence> // ThreadTransferDstResetCoordinateAfterRunFlags + {c_ds_desc_refs, + idx_c_ds_block_begin, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)), + c_element_op}; + + constexpr auto sfc_c_vgpr = + SpaceFillingCurve, + Sequence<0, 1, 2, 3, 4, 5, 6, 7>, + Sequence>{}; + + constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess(); + + // space filling curve for shuffled blockwise C/D/E + constexpr auto sfc_cde_block = + SpaceFillingCurve, + Sequence<0, 2, 1, 3>, + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{}; + + static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!"); + + static_for<0, num_access, 1>{}([&](auto access_id) { + // make sure it's safe to write to LDS + block_sync_lds(); + + // each thread write its data from VGPR to LDS + c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4, + sfc_c_vgpr.GetIndexTupleOfNumber(access_id), + c_thread_buf, + c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4, + c_shuffle_block_buf); + + // make sure it's safe to read from LDS + block_sync_lds(); + + // each block copy its data from LDS to global + cde_block_copy_lds_and_global.Run( + c_ds_desc_refs, + c_ds_buf_refs, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + tie(c_grid_buf)); + + if constexpr(access_id < num_access - 1) + { + constexpr auto cde_lds_and_global_step = + sfc_cde_block.GetForwardStep(access_id); + + // move on Ds + static_for<0, NumDTensor, 1>{}([&](auto i) { + cde_block_copy_lds_and_global.MoveSrcSliceWindow( + c_ds_desc_refs, i + I1, cde_lds_and_global_step); + }); + + // move on E + cde_block_copy_lds_and_global.MoveDstSliceWindow( + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + I0, + cde_lds_and_global_step); + } + }); + } + } +}; + +} // namespace ck From 818c671876ad8640ddb69e5a97d15b9bf755ec9e Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Wed, 26 Feb 2025 10:52:42 +0000 Subject: [PATCH 18/28] Add control route for packed valu math --- ...dlops_blockscale_b_preshuffle_selector.hpp | 3 +- ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 220 ++++++++++-- ...kwise_gemm_pipeline_xdlops_v1_ab_scale.hpp | 63 +++- ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 69 +++- ...xdl_cshuffle_v3_blockscale_bpreshuffle.hpp | 30 +- ...fle_v3_multi_d_blockscale_b_preshuffle.hpp | 4 +- profiler/src/CMakeLists.txt | 312 +++++++++--------- 7 files changed, 483 insertions(+), 218 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp index da96a89f4f..be1610a7a3 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp @@ -4,7 +4,8 @@ #pragma once #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp" -// #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v2.hpp" +// #include +// "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v2.hpp" #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp" namespace ck { diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp index 83759a3192..d67a339183 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -201,34 +201,73 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](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 + static_for<0, staged_num_mfma_per_buffer_load_b, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * + staged_num_buffer_load_b_per_ds_read_a + + ibuf_inst * staged_num_mfma_per_buffer_load_b + imfma + 1) % + num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * + staged_num_buffer_load_b_per_ds_read_a + + (staged_num_buffer_load_b_per_ds_read_a - 1) * + staged_num_mfma_per_buffer_load_b + + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __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 + + static_for<0, staged_num_mfma_per_buffer_load_b - 1, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * + staged_num_buffer_load_b_per_ds_read_a + + (staged_num_buffer_load_b_per_ds_read_a - 1) * + staged_num_mfma_per_buffer_load_b + + imfma + 2) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); __builtin_amdgcn_sched_barrier(0); } else if constexpr(stage.value == 1) { + // A LDS write access. constexpr auto staged_num_mfma_per_ds_write_a = math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a); @@ -241,16 +280,44 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_ds_write_a + i_mfma + 1) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr(((i_inst+1) * staged_num_mfma_per_ds_write_a) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + __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 + static_for<0, staged_num_mfma_per_ds_write_a, 1>{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_ds_write_a + i_mfma + 1) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write } } @@ -258,16 +325,47 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + + (i_inst-stage_more_mfma) * (staged_num_mfma_per_ds_write_a -1) + + i_mfma + 1) % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + + (i_inst-stage_more_mfma + 1) * (staged_num_mfma_per_ds_write_a -1)) + % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __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 + static_for<0, staged_num_mfma_per_ds_write_a - 1, 1>{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + + (i_inst-stage_more_mfma) * (staged_num_mfma_per_ds_write_a -1) + + i_mfma + 1) % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write } } @@ -277,6 +375,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_buffer_load_a + i_mfma + 1) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr(((i_inst+1) * staged_num_mfma_per_buffer_load_a) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __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 + static_for<0, staged_num_mfma_per_buffer_load_a, 1>{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_buffer_load_a + i_mfma + 1) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read } } @@ -306,16 +431,46 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + + (i_inst-stage_more_mfma) * (staged_num_mfma_per_buffer_load_a -1) + + i_mfma + 1) % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + + (i_inst-stage_more_mfma + 1) * (staged_num_mfma_per_buffer_load_a -1)) + % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __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 + static_for<0, staged_num_mfma_per_buffer_load_a - 1, 1>{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + + (i_inst-stage_more_mfma) * (staged_num_mfma_per_buffer_load_a -1) + + i_mfma + 1) % num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read } } @@ -327,9 +482,17 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_inst) { - ignore = i_inst; - __builtin_amdgcn_sched_group_barrier( - 0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA + static_for<0, staged_num_mfma_per_ds_read_a, 1>{}([&](auto i_mfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i_inst * staged_num_mfma_per_ds_read_a + i_mfma + 1) % + num_mfma_per_kscaleblock ==0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read }); @@ -816,6 +979,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto m0) { diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp index 8375e81fa0..8147745a6e 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp @@ -220,6 +220,9 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale @@ -233,26 +236,59 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto i) { - ignore = i; static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) { - ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((i * num_mfma_per_issue + idswrite + 1) % num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier( - 0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA + if constexpr(num_mfma_per_issue - num_dswrite_per_issue_a >= 1) + { + static_for<0, num_mfma_per_issue - num_dswrite_per_issue_a, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((i * num_mfma_per_issue + num_dswrite_per_issue_a + imfma + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + } }); + static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { - ignore = i; static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) { - ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + idswrite + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier( - 0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA + if constexpr(num_mfma_per_issue - num_dswrite_per_issue_b >= 1) + { + static_for<0, num_mfma_per_issue - num_dswrite_per_issue_b, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + + num_dswrite_per_issue_b + imfma + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + } }); // stage 2 @@ -270,6 +306,11 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale{}([&](auto i) { @@ -286,6 +327,12 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale @@ -207,26 +210,61 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto i) { - ignore = i; static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) { - ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i * num_mfma_per_issue + idswrite + 1) % num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier( - 0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA + static_for<0, num_mfma_per_issue - num_dswrite_per_issue_a, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr((i * num_mfma_per_issue + num_dswrite_per_issue_a + imfma + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); }); + static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { - ignore = i; static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) { - ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + idswrite + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier( - 0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA + static_for<0, num_mfma_per_issue - num_dswrite_per_issue_b, 1>{}([&](auto imfma) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + + /* Judging issue v_pk_fma */ + if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + + num_dswrite_per_issue_b + imfma + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); }); // stage 2 @@ -244,6 +282,13 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto i) { @@ -260,6 +305,14 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultipleD_BlockScale_BPreshuffle + BLayout, + DsLayout, + CLayout, + ADataType, + AScaleDataType, + BDataType, + BScaleDataType, + DsDataType, + CDataType, + ScaleBlockM, + ScaleBlockN, + ScaleBlockK, + AElementwiseOperation, + BElementwiseOperation, + CElementwiseOperation> { static constexpr index_t NumDTensor = DsDataType::Size(); diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp index 82e413ac5f..5529c358cf 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp @@ -1121,7 +1121,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle BElementwiseOperation b_element_op, CElementwiseOperation c_element_op) { - ignore = b_element_op; + ignore = b_element_op; const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); const auto b_grid_desc_bpreshuffled = @@ -1620,7 +1620,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle BElementwiseOperation b_element_op, CElementwiseOperation c_element_op) { - ignore = b_element_op; + ignore = b_element_op; const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); const auto b_grid_desc_bpreshuffled = diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 5ed28b9826..9fbac4bc24 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -1,90 +1,90 @@ # ckProfiler set(PROFILER_SOURCES profiler.cpp - profile_gemm.cpp - profile_reduce.cpp - profile_groupnorm_bwd_data.cpp - profile_groupnorm_fwd.cpp - profile_layernorm_bwd_data.cpp - profile_layernorm_bwd_gamma_beta.cpp - profile_groupnorm_bwd_gamma_beta.cpp - profile_layernorm_fwd.cpp - profile_max_pool2d_fwd.cpp - profile_pool3d_fwd.cpp - profile_avg_pool3d_bwd.cpp - profile_max_pool3d_bwd.cpp - profile_avg_pool2d_bwd.cpp - profile_max_pool2d_bwd.cpp - profile_softmax.cpp - profile_batchnorm_fwd.cpp - profile_batchnorm_bwd.cpp - profile_batchnorm_infer.cpp - profile_conv_tensor_rearrange.cpp - profile_transpose.cpp - profile_permute_scale.cpp + # profile_gemm.cpp + # profile_reduce.cpp + # profile_groupnorm_bwd_data.cpp + # profile_groupnorm_fwd.cpp + # profile_layernorm_bwd_data.cpp + # profile_layernorm_bwd_gamma_beta.cpp + # profile_groupnorm_bwd_gamma_beta.cpp + # profile_layernorm_fwd.cpp + # profile_max_pool2d_fwd.cpp + # profile_pool3d_fwd.cpp + # profile_avg_pool3d_bwd.cpp + # profile_max_pool3d_bwd.cpp + # profile_avg_pool2d_bwd.cpp + # profile_max_pool2d_bwd.cpp + # profile_softmax.cpp + # profile_batchnorm_fwd.cpp + # profile_batchnorm_bwd.cpp + # profile_batchnorm_infer.cpp + # profile_conv_tensor_rearrange.cpp + # profile_transpose.cpp + # profile_permute_scale.cpp ) if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) - list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) - endif() - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) - endif() - list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) - if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_weight_preshuffle.cpp) + # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) + # list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) + # list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) + # endif() + # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + # list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) + # list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) + # list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) + # endif() + # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) + # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") + # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_weight_preshuffle.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) - endif() - list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) - list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) + # endif() + # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) + # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp) + # list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) + # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) + # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) + # list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) + # list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) + # list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) - endif() - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -endif() +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) +# endif() +# list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +# endif() -if(DL_KERNELS) - list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) -endif() +# if(DL_KERNELS) +# list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) +# list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) +# endif() set(PROFILER_EXECUTABLE ckProfiler) @@ -97,90 +97,90 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) endif() target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") - if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) - endif() - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) - endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) - if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_weight_preshuffle_instance) + # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) + # endif() + # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) + # endif() + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) + # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_weight_preshuffle_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) - endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) + # endif() + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") - if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) - endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -endif() +# if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") +# if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) +# endif() +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +# endif() -if(DL_KERNELS) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) -endif() +# if(DL_KERNELS) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) +# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) +# endif() rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) From a9bcf3dda869eb3d3a5851c86b8222f1519957ac Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Thu, 27 Feb 2025 08:59:15 +0000 Subject: [PATCH 19/28] correctness checked. performance issue: unnecessary a lot v_perm generated --- ...ultiply_xdl_fp8_blockscale_bpreshuffle.cpp | 12 +- ...dlops_blockscale_b_preshuffle_selector.hpp | 5 +- ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 556 ++++++++++++++---- ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 108 ++-- ...xdl_cshuffle_v3_blockscale_bpreshuffle.hpp | 30 +- ...fle_v3_multi_d_blockscale_b_preshuffle.hpp | 12 +- 6 files changed, 542 insertions(+), 181 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp index bc049a68c6..c7917adada 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp @@ -96,14 +96,14 @@ using DeviceOpInstance = A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, - 128, 128, - 128, 16, 16, + 32, 128, + 256, 16, 16, 32, 32, - 4, 1, - 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<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp index be1610a7a3..5a4fa047b4 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp @@ -32,7 +32,6 @@ template constexpr auto BlockGemmBlockScaleBPreshufflePipeline_Selector() { -#if 0 if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) { return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1< @@ -57,6 +56,7 @@ constexpr auto BlockGemmBlockScaleBPreshufflePipeline_Selector() NRepeat, KPack>{}; } +#if 0 else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) { return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v2< @@ -82,8 +82,7 @@ constexpr auto BlockGemmBlockScaleBPreshufflePipeline_Selector() KPack>{}; } #endif - // else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) - if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) { static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3"); return BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3< diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index 8ed25895b5..2ebb76ab0c 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -33,7 +33,7 @@ template -struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 +struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1 { }; @@ -58,26 +58,26 @@ template -struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 +struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1 : BlockwiseGemmXdlops_pipeline_base + KPack, + true> { using Base = BlockwiseGemmXdlops_pipeline_base; + KPack, + true>; using Base::A_K1; using Base::B_K1; using Base::I0; @@ -125,8 +127,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&) { @@ -211,6 +239,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 - __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, - const BGridBuffer& b_grid_buf, - BBlockBuffer& b_block_buf, - const BBlockTransferStep& b_block_copy_step, - CThreadBuffer& c_thread_buf, - index_t num_loop) const + typename CScaleThreadDesc, + typename CThreadBuffer, + typename AScaleGridBuffer, + typename AScaleGridDesc, + typename AScaleThreadDesc, + typename AScaleThreadTransfer, + typename AScaleThreadTransferStep, + typename BScaleGridBuffer, + typename BScaleGridDesc, + typename BScaleThreadDesc, + typename BScaleThreadTransfer, + typename BScaleThreadTransferStep> + __device__ void Run( + // ABlockCopy + 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, + // BBlockCopy + const BGridDesc& b_grid_desc, + const BBlockDesc& b_block_desc, + BBlockTransfer& b_blockwise_copy, + const BGridBuffer& b_grid_buf, + BBlockBuffer& b_block_buf, + const BBlockTransferStep& b_block_copy_step, + // CThread + const CScaleThreadDesc& c_scale_thread_desc, + CThreadBuffer& c_thread_buf, + // AScaleThreadCopy + const AScaleGridDesc& a_scale_grid_desc, + const AScaleThreadDesc& a_scale_thread_desc, + AScaleThreadTransfer& a_scale_thread_copy, + const AScaleGridBuffer& a_scale_grid_buf, + const AScaleThreadTransferStep& a_scale_thread_copy_step, + // BScaleThreadCopy + const BScaleGridDesc& b_scale_grid_desc, + const BScaleThreadDesc& b_scale_thread_desc, + BScaleThreadTransfer& b_scale_thread_copy, + const BScaleGridBuffer& b_scale_grid_buf, + const BScaleThreadTransferStep& b_scale_thread_copy_step, + // num_loop + index_t num_loop) const { + ignore = b_block_desc; ignore = b_block_buf; __builtin_amdgcn_sched_barrier(0); auto a_thread_buf = make_static_buffer( @@ -248,6 +309,13 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}> b_thread_bufs; constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0); + auto a_scale_thread_buf = make_static_buffer( + a_scale_thread_desc.GetElementSpaceSize()); + auto b_scale_thread_buf = make_static_buffer( + b_scale_thread_desc.GetElementSpaceSize()); + auto c_scale_thread_buf = make_static_buffer( + c_scale_thread_desc.GetElementSpaceSize()); + // Global prefetch A1 B1 a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); b_blockwise_copy.Run(b_grid_desc, @@ -260,13 +328,99 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step); + + constexpr auto num_scale_k_block = CScaleThreadDesc{}.GetLength(Number<0>{}); + constexpr auto num_scale_m_block = CScaleThreadDesc{}.GetLength(Number<1>{}); + constexpr auto num_scale_n_block = CScaleThreadDesc{}.GetLength(Number<2>{}); + + static_for<0, num_scale_m_block, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { + constexpr index_t c_offset = + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); + + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; + }); + }); + }); + + // Local prefill A1 a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0); - // // Global prefetch A2 + // Global prefetch A2 a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, + a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step); + + StaticBufferTupleOfVector + c_thread_buf_per_scale; + // Local prefetch A1 block_sync_lds(); static_for<0, MRepeat, 1>{}([&](auto m0) { @@ -307,31 +461,73 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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{}]; + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; - using mfma_input_type = - typename vector_type::type; + 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_desc_.CalculateOffset(make_tuple( + n0, + I0, + kscale0 * KRepeat / num_scale_k_block + k0, + ik))>{}]; + }); + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = + CScaleThreadDesc{}.CalculateOffset(make_tuple( + kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); + }); + }); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); - xdlops_gemm.Run( - a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf.GetVectorTypeReference(Number{})); + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; }); }); }); @@ -349,6 +545,35 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}([&](auto m0) { + a_scale_thread_copy.Run(a_scale_grid_desc, + a_scale_grid_buf, + a_scale_thread_desc, + make_tuple(m0, I0), + a_scale_thread_buf); + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<0>{})); + }); + + if constexpr(NumKBlockPerScale == 1) + { + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<2>{})); + } + else + { + a_scale_thread_copy.MoveSrcSliceWindow( + a_scale_grid_desc, a_scale_thread_copy_step.At(Number<1>{})); + } + + b_scale_thread_copy.Run(b_scale_grid_desc, + b_scale_grid_buf, + b_scale_thread_desc, + make_tuple(I0, I0), + b_scale_thread_buf); + + b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, + b_scale_thread_copy_step); HotLoopScheduler(); __builtin_amdgcn_sched_barrier(0); }; @@ -359,6 +584,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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{}]; + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; - using mfma_input_type = - typename vector_type::type; + 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{}]; + }); + using mfma_input_type = + typename vector_type::type; + + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); + }); + }); + }); + + static_for<0, MRepeat, 1>{}([&](auto m0) { + static_for<0, num_scale_n_block, 1>{}([&](auto n0) { + static_for<0, num_scale_k_block, 1>{}([&](auto k0) { constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + CScaleThreadDesc{}.CalculateOffset(make_tuple(k0, m0, n0)); + constexpr index_t a_offset = + AScaleThreadDesc{}.CalculateOffset(make_tuple(m0, k0)); + constexpr index_t b_offset = + BScaleThreadDesc{}.CalculateOffset(make_tuple(n0, k0)); - xdlops_gemm.Run(a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf.GetVectorTypeReference(Number{})); + c_scale_thread_buf(Number{}) = + a_scale_thread_buf[Number{}] * + b_scale_thread_buf[Number{}]; }); }); }); @@ -416,61 +683,108 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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{}]; + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; - using mfma_input_type = - typename vector_type::type; + 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{}]; + }); - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + using mfma_input_type = + typename vector_type::type; - xdlops_gemm.Run(a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf.GetVectorTypeReference(Number{})); + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); }); }); }); - // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle - // latency - // __builtin_amdgcn_sched_barrier(0); } - else + else if constexpr(TailNum == TailNumber::Odd) { static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { - static_for<0, KRepeat, 1>{}([&](auto k0) { - vector_type a_thread_vec; - vector_type b_thread_vec; - - 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{}]; + static_for<0, num_scale_k_block, 1>{}([&](auto kscale0) { + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()(Number{}) = 0; }); + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { + vector_type a_thread_vec; + vector_type b_thread_vec; - using mfma_input_type = - typename vector_type::type; + 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{}]; + }); - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + using mfma_input_type = + typename vector_type::type; - xdlops_gemm.Run(a_thread_vec.template AsType(), - b_thread_vec.template AsType(), - c_thread_buf.GetVectorTypeReference(Number{})); + xdlops_gemm.template Run<>( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); + }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_thread_buf(Number{}) += + c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) + .template AsType()[Number{}] * + type_convert( + c_scale_thread_buf[Number{}]); + }); }); }); }); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp index d67a339183..bc3747dc03 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -221,7 +221,8 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - + /* Judging issue v_pk_fma */ if constexpr((i_inst * staged_num_mfma_per_ds_write_a + i_mfma + 1) % - num_mfma_per_kscaleblock ==0) + num_mfma_per_kscaleblock == + 0) { __builtin_amdgcn_sched_group_barrier( 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA @@ -295,8 +299,9 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - + /* Judging issue v_pk_fma */ if constexpr((i_inst * staged_num_mfma_per_buffer_load_a + i_mfma + 1) % - num_mfma_per_kscaleblock ==0) + num_mfma_per_kscaleblock == + 0) { __builtin_amdgcn_sched_group_barrier( 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA } }); - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA /* Judging issue v_pk_fma */ - if constexpr(((i_inst+1) * staged_num_mfma_per_buffer_load_a) % - num_mfma_per_kscaleblock ==0) + if constexpr(((i_inst + 1) * staged_num_mfma_per_buffer_load_a) % + num_mfma_per_kscaleblock == + 0) { __builtin_amdgcn_sched_group_barrier( 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA } - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read } else { @@ -418,7 +434,8 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_inst) { static_for<0, staged_num_mfma_per_ds_read_a, 1>{}([&](auto i_mfma) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA /* Judging issue v_pk_fma */ if constexpr((i_inst * staged_num_mfma_per_ds_read_a + i_mfma + 1) % - num_mfma_per_kscaleblock ==0) + num_mfma_per_kscaleblock == + 0) { __builtin_amdgcn_sched_group_barrier( 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp index e6ddb53246..9bfcc19fd4 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp @@ -244,13 +244,26 @@ struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle // Tail number always full if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) { const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< GridwiseGemm, true, InMemoryDataOperationEnum::Set, - minimum_occupancy>; + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; Run(kernel); } } @@ -285,13 +298,26 @@ struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle // Tail number always 1 if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) { const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< GridwiseGemm, false, InMemoryDataOperationEnum::Set, - minimum_occupancy>; + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; Run(kernel); } } diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp index 5529c358cf..37bc05ae3b 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp @@ -30,7 +30,7 @@ template + TailNumber TailNum = TailNumber::Even> __global__ void #if CK_USE_LAUNCH_BOUNDS __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy) @@ -163,7 +163,6 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock = CDEShuffleBlockTransferScalarPerVectors{}[I0]; - // K1 should be Number<...> static constexpr auto AK0Number = Number{}; static constexpr auto BK0Number = Number{}; @@ -172,6 +171,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle static constexpr auto BlockSizeNumber = Number{}; static constexpr index_t NumDTensor = DsDataType::Size(); + using mfma_selector = MfmaSelector; static constexpr index_t KPack = math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk); @@ -460,9 +460,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle __host__ __device__ static constexpr auto MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&) { - constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); - - return MakeGemmMmaTileDescriptor(BBlockDesc_BK0_N_BK1{}); + return MakeGemmMmaTileDescriptor(BBlockDesc_BK0_N_BK1{}); } template @@ -1124,6 +1122,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle ignore = b_element_op; const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); + const auto b_grid_desc_bpreshuffled = MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled); const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( @@ -1179,9 +1178,6 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave); - // lds max alignment - constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); - // A matrix in LDS memory, dst of blockwise copy constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); From 7bb92cdc67c81ac217b3054319f4a1e807f7efaa Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Fri, 28 Feb 2025 02:31:41 +0000 Subject: [PATCH 20/28] tempsave --- example/65_gemm_multiply_multiply/CMakeLists.txt | 1 + .../gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp | 14 +++++++------- ...ckwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp | 4 ++-- ...e_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp | 2 -- 4 files changed, 10 insertions(+), 11 deletions(-) diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 466f2a9f3c..5eae71a021 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -9,3 +9,4 @@ list(APPEND EXAMPLE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) +target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp index 7319d345c9..49eccc5e03 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp @@ -139,14 +139,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu // clang-format off < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, - 256, 256, 128, + 32, 128, 256, 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>, S<8, 8, 1>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; + 32, 32, + 1, 1, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp index 8ed25895b5..5f41fc1f98 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp @@ -260,10 +260,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1 Date: Tue, 4 Mar 2025 12:17:05 +0000 Subject: [PATCH 21/28] add reproducable problem cases --- .../65_gemm_multiply_multiply/CMakeLists.txt | 2 + ...ultiply_xdl_fp8_blockscale_bpreshuffle.cpp | 12 +- ...iply_xdl_fp8_blockscale_bpreshuffle_v1.cpp | 382 ++++++++++++++++++ ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 67 ++- ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 50 +-- 5 files changed, 432 insertions(+), 81 deletions(-) create mode 100644 example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 5eae71a021..2b0eeefe21 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -1,6 +1,7 @@ add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp) +add_example_executable(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1 gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp) add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) @@ -9,4 +10,5 @@ list(APPEND EXAMPLE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) +target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1 PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp index c7917adada..bc049a68c6 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp @@ -96,14 +96,14 @@ using DeviceOpInstance = A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, - 32, 128, - 256, 16, 16, + 128, 128, + 128, 16, 16, 32, 32, - 1, 1, - S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, - S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + 4, 1, + 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>, S<8>, - ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>; // clang-format on int main(int argc, char* argv[]) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp new file mode 100644 index 0000000000..035749d20b --- /dev/null +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp @@ -0,0 +1,382 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" +#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp" + +#include "ck/library/utility/device_memory.hpp" +#include "ck/library/utility/host_tensor.hpp" +#include "ck/library/utility/host_tensor_generator.hpp" +#include "ck/library/utility/literals.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" +#include "ck/library/utility/check_err.hpp" + +#include "ck/utility/blkgemmpipe_scheduler.hpp" + +template +using S = ck::Sequence; + +using BF16 = ck::bhalf_t; +using FP8 = ck::f8_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using A0DataType = FP8; +using A1DataType = F32; +using B0DataType = FP8; +using B1DataType = F32; +using AccDataType = F32; +using CShuffleDataType = F32; +using DsDataType = ck::Tuple<>; +using EDataType = BF16; + +using A0Layout = Row; +using B0Layout = Col; +using D0Layout = Row; +using D1Layout = Col; +using DsLayout = ck::Tuple<>; +using ELayout = Row; + +void preShuffleBuffer(const FP8* src, FP8* dst, int N, int K, int NXdl) +{ + int KPack = 16; + int NLane = NXdl; + int KLane = 64 / NLane; + + int K0 = K / (KLane * KPack); + // K -> K0 KLane KPack + // N -> N0 NLane + // N, K -> N0 K0 KLane NLane KPack + int tempk; + for(int n = 0; n < N; ++n) + { + for(int k = 0; k < K; ++k) + { + int n0 = n / NLane; + int n1 = n % NLane; + + int k0 = k / (KLane * KPack); + tempk = k % (KLane * KPack); + int k1 = tempk / KPack; + int k2 = tempk % KPack; + + int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane + + k1 * KPack * NLane + n1 * KPack + k2; + + dst[outputIndex] = src[n * K + k]; + } + } +} +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = PassThrough; + +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; + +static constexpr ck::index_t Scale_Block_M = 1; +static constexpr ck::index_t Scale_Block_N = 128; +static constexpr ck::index_t Scale_Block_K = 128; + +using DeviceOpInstance = + ck::tensor_operation::device::DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle + // clang-format off + , 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, 16, 1, 16>, S<8>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; +// clang-format on + +int main(int argc, char* argv[]) +{ + bool do_verification = true; + int init_method = 1; + bool time_kernel = false; + bool flush_cache = true; + + // GEMM shape + ck::index_t M = 128; + ck::index_t N = 1024; + ck::index_t K = 1024; + + ck::index_t StrideA = K; + ck::index_t StrideB = K; + ck::index_t StrideE = N; + + if(argc == 1) + { + // use default case + } + else if(argc == 4) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + } + else if(argc == 8) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + + M = std::stoi(argv[4]); + N = std::stoi(argv[5]); + K = std::stoi(argv[6]); + + flush_cache = std::stoi(argv[7]); + + StrideA = K; + StrideB = K; + StrideE = N; + } + else + { + printf("arg1: verification (0=no, 1=yes)\n"); + printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); + printf("arg3: time kernel (0=no, 1=yes)\n"); + printf("arg4 to 6: M, N, K\n"); + printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n"); + exit(0); + } + + ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K; + ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K; + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + using namespace ck::literals; + + if(std::is_same::value) + { + return HostTensorDescriptor({row, col}, {stride, 1_uz}); + } + else + { + return HostTensorDescriptor({row, col}, {1_uz, stride}); + } + }; + + Tensor a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{})); + Tensor a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M, + (K + Scale_Block_K - 1) / Scale_Block_K, + Scale_Stride_AM, + A0Layout{})); + Tensor b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); + Tensor b0_preshuffled( + f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size + Tensor b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K, + (N + Scale_Block_N - 1) / Scale_Block_N, + Scale_Stride_BN, + B0Layout{})); + Tensor e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + Tensor e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + + std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl; + std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl; + std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl; + std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl; + std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl; + +#if 1 + switch(init_method) + { + case 0: break; + case 1: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + case 2: + a0_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b0_k_n.GenerateTensorValue(GeneratorTensor_1{}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 3: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 4: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + case 5: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + default: + a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + } +#endif +#if 0 + for(int im =0; im< (M + Scale_Block_M - 1) / Scale_Block_M; im++){ + float row_sum = .0; + for(int ik =0; ik< (K + Scale_Block_K - 1) / Scale_Block_K; ik++){ + printf("%lf ",a1_m_k(im, ik)); + row_sum += a1_m_k(im, ik); + } + printf("sum: %lf\n", row_sum * 128); + } +#endif + + DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); + DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize()); + DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); + + a0_device_buf.ToDevice(a0_m_k.mData.data()); + a1_device_buf.ToDevice(a1_m_k.mData.data()); + b1_device_buf.ToDevice(b1_k_n.mData.data()); + + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto cde_element_op = CDEElementOp{}; + + constexpr ck::index_t NumDTensor = DsDataType::Size(); + + // do GEMM + auto device_op = DeviceOpInstance{}; + int NPerXdl = device_op.GetPreShuffleParameters(); + + preShuffleBuffer(b0_k_n.mData.data(), b0_preshuffled.mData.data(), N, K, NPerXdl); + + b0_device_buf.ToDevice(b0_preshuffled.mData.data()); + auto invoker = device_op.MakeInvoker(); + auto argument = device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(), + b0_device_buf.GetDeviceBuffer(), + std::array{}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{}, + StrideE, + a1_device_buf.GetDeviceBuffer(), + b1_device_buf.GetDeviceBuffer(), + a_element_op, + b_element_op, + cde_element_op); + + if(!device_op.IsSupportedArgument(argument)) + { + throw std::runtime_error( + "wrong! device_gemm with the specified compilation parameters does " + "not support this GEMM problem"); + } + + std::size_t flop = std::size_t(2) * M * N * K; + std::size_t num_btype = + sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; + + float ave_time = .0; + + if(flush_cache) + { + int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype; + + ave_time = invoker.Run(argument, + StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf}); + } + else + { + ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100}); + } + + float tflops = static_cast(flop) / 1.E9 / ave_time; + + float gb_per_sec = num_btype / 1.E6 / ave_time; + + std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" + << std::endl; + + if(do_verification) + { + Tensor c_m_n({M, N}); + Tensor a_m_k({M, K}); + Tensor b_k_n({K, N}); + + for(int m = 0; m < M; m++) + { + for(int k = 0; k < K; k++) + { + a_m_k(m, k) = ck::type_convert(a0_m_k(m, k)) * + a1_m_k(m / Scale_Block_M, k / Scale_Block_K); + } + } + + for(int n = 0; n < N; n++) + { + for(int k = 0; k < K; k++) + { + b_k_n(k, n) = ck::type_convert(b0_k_n(k, n)) * + b1_k_n(k / Scale_Block_K, n / Scale_Block_N); + } + } + + using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm; + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = + ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{}); + + ref_invoker.Run(ref_argument); + +#if 1 + for(int m = 0; m < M; ++m) + { + for(int n = 0; n < N; ++n) + { + e_m_n_host_result(m, n) = ck::type_convert(c_m_n(m, n)); + } + } +#endif + + e_device_buf.FromDevice(e_m_n_device_result.mData.data()); + + return ck::utils::check_err( + e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2) + ? 0 + : 1; + } + + return 0; +} diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index 2ebb76ab0c..edb33e6694 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -127,8 +127,8 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1 __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&) { @@ -212,28 +186,45 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i) { - ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + /* Judging issue v_pk_fma */ + if constexpr((i + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); // A global static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { - ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((num_buffer_load_inst_b + 2*i + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((num_buffer_load_inst_b + 2*i + 2) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); // A local static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) { - ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + if constexpr((num_buffer_load_inst_b + 2*num_buffer_load_inst_a + i + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read }); } @@ -300,7 +291,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1( a_thread_desc_.GetElementSpaceSize()); auto b_thread_buf = make_static_buffer( @@ -326,7 +317,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -437,7 +428,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto m0) { static_for<0, NRepeat, 1>{}([&](auto n0) { diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp index bc3747dc03..b0c058b091 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -141,10 +141,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](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); @@ -631,8 +607,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}) == 1, "Pipeline v3 only support scaleblocksliceK=1"); static_assert(CScaleThreadDesc{}.GetLength(Number<2>{}) == 1, @@ -760,7 +735,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -852,7 +827,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3 Date: Wed, 5 Mar 2025 08:38:19 +0000 Subject: [PATCH 22/28] replace pk_fma with llvm builtin --- ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 186 ++++++++++++------ ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 104 +++++++--- 2 files changed, 199 insertions(+), 91 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index edb33e6694..4d74f1ef69 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -59,25 +59,25 @@ template struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1 + BlockSize, + ADataType, + BDataType, + ComputeDataType, + AccDataType, + ATileDesc, + BTileDesc, + AMmaTileDesc, + BMmaTileDesc, + ABlockTransferSrcScalarPerVector, + BBlockTransferSrcScalarPerVector, + MPerBlock, + NPerBlock, + KPerBlock, + MPerXDL, + NPerXDL, + MRepeat, + NRepeat, + KPack> : BlockwiseGemmXdlops_pipeline_base{}([&](auto i) { __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2*i + 1) % num_mfma_per_kscaleblock == 0) + if constexpr((num_buffer_load_inst_b + 2 * i + 1) % num_mfma_per_kscaleblock == 0) { - __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA } __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2*i + 2) % num_mfma_per_kscaleblock == 0) + if constexpr((num_buffer_load_inst_b + 2 * i + 2) % num_mfma_per_kscaleblock == 0) { - __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA } __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); @@ -221,9 +224,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i) { __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2*num_buffer_load_inst_a + i + 1) % num_mfma_per_kscaleblock == 0) + if constexpr((num_buffer_load_inst_b + 2 * num_buffer_load_inst_a + i + 1) % + num_mfma_per_kscaleblock == + 0) { - __builtin_amdgcn_sched_group_barrier(0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA } __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read }); @@ -457,6 +463,16 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}) .template AsType()(Number{}) = 0; }); + vector_type c_scale_thread_vec; + constexpr index_t cscale_offset = + CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[Number{}]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[Number{}]; + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -489,19 +505,26 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - constexpr index_t cscale_offset = - CScaleThreadDesc{}.CalculateOffset(make_tuple( - kscale0, m0, n0 * num_scale_n_block / NRepeat)); + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); - c_thread_buf(Number{}) += - c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert( - c_scale_thread_buf[Number{}]); - }); + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}( + [&](auto t) { + using pk_fma_type = + typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = + __builtin_elementwise_fma( + c_thread_buf_per_scale + .GetVectorTypeReference(Number<0>{}) + .template AsType()[t], + c_scale_thread_vec + .template AsType()[Number<0>{}], + c_thread_buf + .GetVectorTypeReference(Number{}) + .template AsType()[t]); + }); }); }); }); @@ -594,6 +617,15 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}) .template AsType()(Number{}) = 0; }); + vector_type c_scale_thread_vec; + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[Number{}]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[Number{}]; + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -624,17 +656,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( - make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); - c_thread_buf(Number{}) += + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert( - c_scale_thread_buf[Number{}]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); }); @@ -679,6 +713,15 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}) .template AsType()(Number{}) = 0; }); + vector_type c_scale_thread_vec; + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[Number{}]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[Number{}]; + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -709,17 +752,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( - make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); - c_thread_buf(Number{}) += + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert( - c_scale_thread_buf[Number{}]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); }); @@ -734,6 +779,15 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}) .template AsType()(Number{}) = 0; }); + vector_type c_scale_thread_vec; + constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( + make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[Number{}]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[Number{}]; + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -764,17 +818,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset( - make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat)); + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); - c_thread_buf(Number{}) += + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert( - c_scale_thread_buf[Number{}]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); }); diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp index b0c058b091..73e8cb9960 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -735,7 +735,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -856,6 +856,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3 c_scale_thread_vec; + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[m0]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[m0]; + static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) @@ -890,13 +896,20 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert(c_scale_thread_buf[m0]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); @@ -977,11 +990,13 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3 c_scale_thread_vec; + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[m0]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[m0]; + static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) @@ -1031,13 +1052,20 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert(c_scale_thread_buf[m0]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); @@ -1076,6 +1104,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto m0) { + vector_type c_scale_thread_vec; + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[m0]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[m0]; + static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) @@ -1102,13 +1136,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert(c_scale_thread_buf[m0]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); @@ -1135,6 +1175,12 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto m0) { + vector_type c_scale_thread_vec; + c_scale_thread_vec.template AsType()(Number<0>{}) = + c_scale_thread_buf[m0]; + c_scale_thread_vec.template AsType()(Number<1>{}) = + c_scale_thread_buf[m0]; + static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) @@ -1161,13 +1207,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3(), c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})); }); - static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) { - constexpr index_t c_offset = - c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t)); - c_thread_buf(Number{}) += + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) { + using pk_fma_type = typename vector_type::type; + + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()(t) = __builtin_elementwise_fma( c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{}) - .template AsType()[Number{}] * - type_convert(c_scale_thread_buf[m0]); + .template AsType()[t], + c_scale_thread_vec.template AsType()[Number<0>{}], + c_thread_buf.GetVectorTypeReference(Number{}) + .template AsType()[t]); }); }); From 616ab3dd1397b3e484f7cb3a6abb839f98d5906a Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Thu, 13 Mar 2025 08:52:12 +0000 Subject: [PATCH 23/28] optimized decode pipeline for blockscale gemm --- ...multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp | 6 +++--- ...ne_xdlops_blockscale_b_preshuffle_selector.hpp | 9 +++++++++ ...pipeline_xdlops_blockscale_b_preshuffle_v1.hpp | 15 ++++++++++++--- ...pipeline_xdlops_blockscale_b_preshuffle_v3.hpp | 9 +++++++++ ...shuffle_v3_multi_d_blockscale_b_preshuffle.hpp | 3 +++ 5 files changed, 36 insertions(+), 6 deletions(-) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp index 035749d20b..16f7a79367 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1.cpp @@ -96,12 +96,12 @@ using DeviceOpInstance = A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, - 32, 128, 128, + 32, 128, 256, 16, 16, 32, 32, 1, 1, - 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, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<8>, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>; // clang-format on diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp index 5a4fa047b4..6ac939d748 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp @@ -25,6 +25,9 @@ template {}([&](auto i) { @@ -323,7 +332,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -589,7 +598,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1 Date: Mon, 17 Mar 2025 07:19:54 +0000 Subject: [PATCH 24/28] add ckProfiler. performance debugging for blockscale_wp prefill --- ...dlops_blockscale_b_preshuffle_selector.hpp | 8 +- ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 11 +- ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 23 +- ...xdl_cshuffle_v3_blockscale_bpreshuffle.hpp | 10 +- ...fle_v3_multi_d_blockscale_b_preshuffle.hpp | 2 +- .../gpu/gemm_blockscale_wp.hpp | 172 ++++++++ .../gpu/gemm_blockscale_wp/CMakeLists.txt | 16 + ...wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp | 86 ++++ ...k_mn_128_128_128_comp_default_instance.cpp | 38 ++ ..._mn_128_128_128_comp_kpadding_instance.cpp | 38 ++ ...mn_128_128_128_mem_v1_default_instance.cpp | 39 ++ ...n_128_128_128_mem_v1_kpadding_instance.cpp | 39 ++ .../profile_gemm_blockscale_wp_impl.hpp | 408 ++++++++++++++++++ profiler/src/CMakeLists.txt | 6 +- profiler/src/profile_gemm_blockscale_wp.cpp | 184 ++++++++ 15 files changed, 1051 insertions(+), 29 deletions(-) create mode 100644 library/include/ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/CMakeLists.txt create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp create mode 100644 profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp create mode 100644 profiler/src/profile_gemm_blockscale_wp.cpp diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp index 6ac939d748..b7a5cc8919 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_selector.hpp @@ -54,8 +54,8 @@ constexpr auto BlockGemmBlockScaleBPreshufflePipeline_Selector() NPerBlock, KPerBlock, MScaleBlock, -NScaleBlock, -KScaleBlock, + NScaleBlock, + KScaleBlock, MPerXDL, NPerXDL, MRepeat, @@ -108,8 +108,8 @@ KScaleBlock, NPerBlock, KPerBlock, MScaleBlock, -NScaleBlock, -KScaleBlock, + NScaleBlock, + KScaleBlock, MPerXDL, NPerXDL, MRepeat, diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index 45bcb91425..c2167a3db4 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -55,8 +55,8 @@ template {}([&](auto i) { @@ -481,7 +482,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}]; c_scale_thread_vec.template AsType()(Number<1>{}) = c_scale_thread_buf[Number{}]; - + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; @@ -730,7 +731,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}]; c_scale_thread_vec.template AsType()(Number<1>{}) = c_scale_thread_buf[Number{}]; - + static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) { vector_type a_thread_vec; vector_type b_thread_vec; diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp index 7e537d9a1f..a9e362464d 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v3.hpp @@ -271,7 +271,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}) == 1, "Pipeline v3 only support scaleblocksliceK=1"); static_assert(CScaleThreadDesc{}.GetLength(Number<2>{}) == 1, @@ -836,7 +836,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3 64) - ? 1 - : 2; + constexpr index_t minimum_occupancy = 2; + // (BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave && + // MPerBlock * NPerBlock / BlockSize > 64) + // ? 1 + // : 2; if(has_main_k_block_loop) { diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp index db97450e80..b20287f824 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp @@ -1125,7 +1125,7 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle ignore = b_element_op; const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); - + const auto b_grid_desc_bpreshuffled = MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled); const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp new file mode 100644 index 0000000000..1a75db60e4 --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp @@ -0,0 +1,172 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8)) +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances); + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances); + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances); + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances); +#endif + +template +struct DeviceOperationInstanceFactory< + ck::tensor_operation::device::DeviceGemmMultipleD_BlockScale_BPreshuffle< + ALayout, + BLayout, + Tuple<>, + CLayout, + A0DataType, + A1DataType, + B0DataType, + B1DataType, + Tuple<>, + CDataType, + 1, + 128, + 128, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough>> +{ + using DeviceOp = + DeviceGemmMultipleD_BlockScale_BPreshuffle, + CLayout, + A0DataType, + A1DataType, + B0DataType, + B1DataType, + Tuple<>, + CDataType, + 1, + 128, + 128, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough>; + + static auto GetInstances() + { + std::vector> op_ptrs; + +#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8)) + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( + op_ptrs); + add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( + op_ptrs); + + add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( + op_ptrs); + add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( + op_ptrs); + } + } +#endif + return op_ptrs; + } +}; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/CMakeLists.txt new file mode 100644 index 0000000000..f13ab883a1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/CMakeLists.txt @@ -0,0 +1,16 @@ +# ONLY XDL_KERNELS +set(GEMM_BLOCKSCALE_WP_INSTANCES) + +list(APPEND GEMM_BLOCKSCALE_WP_INSTANCES + device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp + device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp + device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp + device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp + ) + +set_source_files_properties(device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + +add_instance_library(device_gemm_blockscale_wp_instance ${GEMM_BLOCKSCALE_WP_INSTANCES}) diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp new file mode 100644 index 0000000000..a0c95cf2ab --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp @@ -0,0 +1,86 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using BF16 = bhalf_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +template +using device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances = std::tuple< + // clang-format off + //################################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | 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| + //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + + // Compute friendly + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 128, 128, 16, 16, 32, 32, 4, 1, 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, 2, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 16, 16, 4, 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, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 16, 16, 4, 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, 2, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 16, 16, 4, 1, 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, 2, 1, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> + // clang-format on + >; + +template +using device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances = std::tuple< + // clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData | BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | M| N| K| | | | | | | | 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| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + + // Memory friendly + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 256, 128, 8, 16, 16, 16, 1, 4, S<16, 16, 1>, 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, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 128, 8, 16, 16, 16, 1, 2, S<16, 16, 1>, 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, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 128, 8, 16, 16, 16, 1, 1, S<16, 16, 1>, 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, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 128, 256, 16, 16, 16, 16, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 16, 64, 256, 16, 16, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 256, 128, 16, 16, 32, 32, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 128, 16, 16, 16, 16, 2, 1, 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, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 32, 64, 256, 16, 16, 16, 16, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 256, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 2, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, 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>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 256, 16, 16, 32, 32, 2, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, + DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp new file mode 100644 index 0000000000..747210d2e2 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp @@ -0,0 +1,38 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances) +{ + add_device_operation_instances( + instances, + device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp new file mode 100644 index 0000000000..47b19e8afe --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp @@ -0,0 +1,38 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances) +{ + add_device_operation_instances( + instances, + device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp new file mode 100644 index 0000000000..27d592f4c0 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp @@ -0,0 +1,39 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances) +{ + add_device_operation_instances( + instances, + device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp new file mode 100644 index 0000000000..dd9b249420 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_blockscale_wp/device_gemm_blockscale_wp_xdl_f8_f8_bf16/device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp @@ -0,0 +1,39 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances( + std::vector, + Row, + F8, + F32, + F8, + F32, + Tuple<>, + BF16, + 1, + 128, + 128, + PassThrough, + PassThrough, + PassThrough>>>& + instances) +{ + add_device_operation_instances( + instances, + device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp b/profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp new file mode 100644 index 0000000000..e3844b1ef7 --- /dev/null +++ b/profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp @@ -0,0 +1,408 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, 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/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp" + +#include "ck/library/utility/check_err.hpp" +#include "ck/library/utility/device_memory.hpp" +#include "ck/library/utility/host_tensor.hpp" +#include "ck/library/utility/host_tensor_generator.hpp" +#include "ck/library/utility/literals.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" + +namespace ck { +namespace profiler { + +template +void preShuffleBuffer(const InOutDataType* src, InOutDataType* dst, int N, int K, int NXdl) +{ + int KPack = 16; + int NLane = NXdl; + int KLane = 64 / NLane; + + int K0 = K / (KLane * KPack); + // K -> K0 KLane KPack + // N -> N0 NLane + // N, K -> N0 K0 KLane NLane KPack + int tempk; + for(int n = 0; n < N; ++n) + { + for(int k = 0; k < K; ++k) + { + int n0 = n / NLane; + int n1 = n % NLane; + + int k0 = k / (KLane * KPack); + tempk = k % (KLane * KPack); + int k1 = tempk / KPack; + int k2 = tempk % KPack; + + int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane + + k1 * KPack * NLane + n1 * KPack + k2; + + dst[outputIndex] = src[n * K + k]; + } + } +} + +template +bool profile_gemm_blockscale_weighpreshuffle_impl(int do_verification, + int init_method, + bool do_log, + bool time_kernel, + int M, + int N, + int K, + int StrideA, + int StrideB, + int StrideE, + int n_warmup, + int n_iter, + uint64_t rotating = 0) +{ + bool pass = true; + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + using namespace ck::literals; + + if(is_same::value) + { + return HostTensorDescriptor({row, col}, {stride, 1_uz}); + } + else + { + return HostTensorDescriptor({row, col}, {1_uz, stride}); + } + }; + + ck::index_t Scale_Stride_AM = ck::is_same_v + ? ((K + ScaleBlockK - 1) / ScaleBlockK) + : ((M + ScaleBlockM - 1) / ScaleBlockM); + ck::index_t Scale_Stride_BN = ck::is_same_v + ? ((K + ScaleBlockK - 1) / ScaleBlockK) + : ((N + ScaleBlockN - 1) / ScaleBlockN); + + Tensor a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); + Tensor a1_m_k(f_host_tensor_descriptor((M + ScaleBlockM - 1) / ScaleBlockM, + (K + ScaleBlockK - 1) / ScaleBlockK, + Scale_Stride_AM, + ALayout{})); + Tensor b0_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); + Tensor b_preshuffled_mfma16( + f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // use layout only for size + Tensor b_preshuffled_mfma32( + f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // use layout only for size + Tensor b1_k_n(f_host_tensor_descriptor((K + ScaleBlockK - 1) / ScaleBlockK, + (N + ScaleBlockN - 1) / ScaleBlockN, + Scale_Stride_BN, + BLayout{})); + Tensor e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + Tensor e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + + int total_gemm_needed = + a0_m_k.GetElementSpaceSizeInBytes() + b0_k_n.GetElementSpaceSizeInBytes() + + a1_m_k.GetElementSpaceSizeInBytes() + b1_k_n.GetElementSpaceSizeInBytes(); + int rotating_count = std::max( + 1, + std::min(n_iter, + static_cast(std::ceil(static_cast(rotating) / total_gemm_needed)))); + + std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl; + std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl; + std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl; + std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl; + std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl; + std::cout << "rotating count: " << rotating_count << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + break; + default: + a0_m_k.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + a1_m_k.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + b1_k_n.GenerateTensorValue(GeneratorTensor_3{0, 1.0}); + } + + preShuffleBuffer(b0_k_n.mData.data(), b_preshuffled_mfma16.mData.data(), N, K, 16); + preShuffleBuffer(b0_k_n.mData.data(), b_preshuffled_mfma32.mData.data(), N, K, 32); + + using PassThrough = ck::tensor_operation::element_wise::PassThrough; + + using AElementOp = PassThrough; + using BElementOp = PassThrough; + using CElementOp = PassThrough; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto c_element_op = CElementOp{}; + + DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b_device_buf_mfma16(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem b_device_buf_mfma32(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize()); + DeviceMem c_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); + + a0_device_buf.ToDevice(a0_m_k.mData.data()); + b_device_buf_mfma16.ToDevice(b_preshuffled_mfma16.mData.data()); + b_device_buf_mfma32.ToDevice(b_preshuffled_mfma32.mData.data()); + a1_device_buf.ToDevice(a1_m_k.mData.data()); + b1_device_buf.ToDevice(b1_k_n.mData.data()); + + using DeviceOp = + ck::tensor_operation::device::DeviceGemmMultipleD_BlockScale_BPreshuffle, + ELayout, + A0DataType, + A1DataType, + B0DataType, + B1DataType, + ck::Tuple<>, + EDataType, + ScaleBlockM, + ScaleBlockN, + ScaleBlockK, + AElementOp, + BElementOp, + CElementOp>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + // Run reference GEMM + if(do_verification) + { + Tensor c_m_n({M, N}); + Tensor a_m_k({M, K}); + Tensor b_k_n({K, N}); + + for(int m = 0; m < M; m++) + { + for(int k = 0; k < K; k++) + { + a_m_k(m, k) = ck::type_convert(a0_m_k(m, k)) * + a1_m_k(m / ScaleBlockM, k / ScaleBlockK); + } + } + + for(int n = 0; n < N; n++) + { + for(int k = 0; k < K; k++) + { + b_k_n(k, n) = ck::type_convert(b0_k_n(k, n)) * + b1_k_n(k / ScaleBlockK, n / ScaleBlockN); + } + } + + using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm; + + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = + ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{}); + + ref_invoker.Run(ref_argument); + + for(int m = 0; m < M; ++m) + { + for(int n = 0; n < N; ++n) + { + e_m_n_host_result(m, n) = ck::type_convert(c_m_n(m, n)); + } + } + } + + std::string best_op_name; + float best_ave_time = 0; + float best_tflops = 0; + float best_gb_per_sec = 0; + + // profile device GEMM instances + for(auto& op_ptr : op_ptrs) + { + int NPerXdl = op_ptr->GetPreShuffleParameters(); + + auto argument_ptr = op_ptr->MakeArgumentPointer( + static_cast(a0_device_buf.GetDeviceBuffer()), + static_cast(NPerXdl == 16 ? b_device_buf_mfma16.GetDeviceBuffer() + : b_device_buf_mfma32.GetDeviceBuffer()), + std::array{}, + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + std::array{}, + StrideE, + a1_device_buf.GetDeviceBuffer(), + b1_device_buf.GetDeviceBuffer(), + a_element_op, + b_element_op, + c_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + + // re-init C to zero before profiling next kernel + c_device_buf.SetZero(); + + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter}); + + if(do_verification) + { + c_device_buf.FromDevice(e_m_n_device_result.mData.data()); + +#if defined CK_ENABLE_FP8 + // set softer tolerances for fp8 + if constexpr(is_same_v || is_same_v || + is_same_v) + { + std::string msg = "Error: Incorrect results!"; + double rtol = 5e-2; + double atol = 5e-2; + pass = pass & ck::utils::check_err( + e_m_n_device_result, e_m_n_host_result, msg, rtol, atol); + } + else + { +#endif + pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result); +#if defined CK_ENABLE_FP8 + } +#endif + + if(do_log) + { + LogRangeAsType(std::cout << "a : ", a0_m_k.mData, ",") << std::endl; + LogRangeAsType(std::cout << "b: ", b0_k_n.mData, ",") << std::endl; + LogRangeAsType(std::cout << "c_host : ", e_m_n_host_result.mData, ",") + << std::endl; + LogRangeAsType(std::cout << "c_device: ", e_m_n_device_result.mData, ",") + << std::endl; + } + } + + std::string op_name = op_ptr->GetTypeString(); + + float ave_time = invoker_ptr->Run( + argument_ptr.get(), + StreamConfig{ + nullptr, time_kernel, 0, n_warmup, n_iter, rotating_count > 1, rotating_count}); + + std::size_t flop = std::size_t(2) * M * N * K; + + std::size_t num_btype = + sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; + + float tflops = static_cast(flop) / 1.E9 / ave_time; + + float gb_per_sec = num_btype / 1.E6 / ave_time; + + std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, " + << gb_per_sec << " GB/s, " << op_name << std::endl; + + if(tflops > best_tflops) + { + best_op_name = op_name; + best_tflops = tflops; + best_ave_time = ave_time; + best_gb_per_sec = gb_per_sec; + } + } + else + { + std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl; + } + } + + if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f32"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = bf16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = int8"; + } + + if constexpr(is_same::value) + { + std::cout << " ALayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " ALayout = ColumnMajor"; + } + + if constexpr(is_same::value) + { + std::cout << " BLayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " BLayout = ColumnMajor"; + } + + std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA + << " StrideB = " << StrideB << " StrideE = " << StrideE << " : " << best_ave_time + << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, " + << best_op_name << std::endl; + + return pass; +} + +} // namespace profiler +} // namespace ck diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 9fbac4bc24..02d0e0a192 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -51,7 +51,8 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_weight_preshuffle.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + # list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_blockscale_wp.cpp) # endif() # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) @@ -141,7 +142,8 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_weight_preshuffle_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_blockscale_wp_instance) # endif() # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) diff --git a/profiler/src/profile_gemm_blockscale_wp.cpp b/profiler/src/profile_gemm_blockscale_wp.cpp new file mode 100644 index 0000000000..01df933f7d --- /dev/null +++ b/profiler/src/profile_gemm_blockscale_wp.cpp @@ -0,0 +1,184 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "profiler/profile_gemm_blockscale_wp_impl.hpp" +#include "profiler_operation_registry.hpp" + +enum struct GemmMatrixLayout +{ + MK_KN_MN, // 0 + MK_NK_MN, // 1 + KM_KN_MN, // 2 + KM_NK_MN, // 3 +}; + +enum struct GemmDataType +{ + F32_F32_F32, // 0 + F16_F16_F16, // 1 + BF16_BF16_BF16, // 2 + INT8_INT8_INT8, // 3 + F8_F16_F16, // 4 + F16_F8_F16, // 5 + F16_F16_F16_F8, // 6 + F8_F8_BF16, // 7 +}; + +enum struct ScaleBlockTile +{ + Tile_128_128_128, // 0 + Tile_1_128_128, // 1 +}; + +#define OP_NAME "gemm_blockscale_weighpreshuffle" +#define OP_DESC "GEMM_BlockScale_WeightPreshuffle" + +int profile_gemm_blockscale_weighpreshuffle(int argc, char* argv[]) +{ + if(argc != 15 && argc != 18) + { + printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); + printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: " + "f16->f8; 7: f8->bf16, " + "comp f8)\n"); + printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"); + printf(" 1: A[m, k] * B[n, k] = C[m, n];\n"); + printf(" 2: A[k, m] * B[k, n] = C[m, n];\n"); + printf(" 3: A[k, m] * B[n, k] = C[m, n])\n"); + printf("arg4: scale block tile (0: ScaleBlockM/N/K = [128, 128, 128]; 1: ScaleBlockM/N/K = " + "[1, 128, 128];\n"); + printf("arg5: verification (0: no; 1: yes)\n"); + printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n"); + printf("arg7: print tensor value (0: no; 1: yes)\n"); + printf("arg8: time kernel (0=no, 1=yes)\n"); + printf("arg9 to 14: M, N, K, StrideA, StrideB, StrideE\n"); + printf("optional:\n"); + printf("arg15: number of warm-up cycles (default 1)\n"); + printf("arg16: number of iterations (default 10)\n"); + printf("arg17: memory for rotating buffer (default 0, size in MB)\n"); + exit(1); + } + + const auto data_type = static_cast(std::stoi(argv[2])); + const auto layout = static_cast(std::stoi(argv[3])); + const auto scale_block_tile = static_cast(std::stoi(argv[4])); + const bool do_verification = std::stoi(argv[5]); + const int init_method = std::stoi(argv[6]); + const bool do_log = std::stoi(argv[7]); + const bool time_kernel = std::stoi(argv[8]); + + const int M = std::stoi(argv[9]); + const int N = std::stoi(argv[10]); + const int K = std::stoi(argv[11]); + + const int StrideA = std::stoi(argv[12]); + const int StrideB = std::stoi(argv[13]); + const int StrideE = std::stoi(argv[14]); + + int n_warmup = 1; + int n_iter = 10; + uint64_t rotating = 0; + if(argc == 18) + { + n_warmup = std::stoi(argv[15]); + n_iter = std::stoi(argv[16]); + rotating = std::stoull(argv[17]) * 1024 * 1024; + } + + using F32 = float; + using BF16 = ck::bhalf_t; + using F8 = ck::f8_t; + + using Row = ck::tensor_layout::gemm::RowMajor; + using Col = ck::tensor_layout::gemm::ColumnMajor; + + auto profile = [&](auto a0_type, + auto a1_type, + auto b0_type, + auto b1_type, + auto comp_type, + auto acc_type, + auto c_type, + auto scale_block_m, + auto scale_block_n, + auto scale_block_k, + auto a_layout, + auto b_layout, + auto e_layout) { + using A0DataType = decltype(a0_type); + using A1DataType = decltype(a1_type); + using B0DataType = decltype(b0_type); + using B1DataType = decltype(b1_type); + using ComputeDataType = decltype(comp_type); + using AccDataType = decltype(acc_type); + using EDataType = decltype(c_type); + + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using ELayout = decltype(e_layout); + + const int DefaultStrideA = ck::is_same_v ? K : M; + const int DefaultStrideB = ck::is_same_v ? N : K; + const int DefaultStrideE = ck::is_same_v ? N : M; + + bool pass = ck::profiler::profile_gemm_blockscale_weighpreshuffle_impl( + do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + (StrideA < 0) ? DefaultStrideA : StrideA, + (StrideB < 0) ? DefaultStrideB : StrideB, + (StrideE < 0) ? DefaultStrideE : StrideE, + n_warmup, + n_iter, + rotating); + + return pass ? 0 : 1; + }; + + if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && + scale_block_tile == ScaleBlockTile::Tile_1_128_128) + { + return profile(F8{}, + F32{}, + F8{}, + F32{}, + F8{}, + F32{}, + BF16{}, + ck::Number<1>{}, + ck::Number<128>{}, + ck::Number<128>{}, + Row{}, + Col{}, + Row{}); + } + else + { + std::cout << "this data_type & layout is not implemented" << std::endl; + + return 1; + } +} + +REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_blockscale_weighpreshuffle); From d60d23ea8e97ee202f175dc920e41f663950e496 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Wed, 19 Mar 2025 12:57:16 +0000 Subject: [PATCH 25/28] v1 performance debugging --- ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 49 ++++++++++++++++++- ...xdl_cshuffle_v3_blockscale_bpreshuffle.hpp | 5 +- 2 files changed, 48 insertions(+), 6 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index c2167a3db4..dde2f8d09d 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -200,7 +200,7 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i) { __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA @@ -243,6 +243,50 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + /* Judging issue v_pk_fma */ + if constexpr((i + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + + // A global + static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write + if constexpr((num_buffer_load_inst_b + 2 * i + 1) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read + if constexpr((num_buffer_load_inst_b + 2 * i + 2) % num_mfma_per_kscaleblock == 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); + + // A local + static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) { + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read + if constexpr((num_buffer_load_inst_b + 2 * num_buffer_load_inst_a + i + 1) % + num_mfma_per_kscaleblock == + 0) + { + __builtin_amdgcn_sched_group_barrier( + 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA + } + }); +#endif } template {}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, @@ -364,6 +407,8 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}); constexpr auto num_scale_m_block = CScaleThreadDesc{}.GetLength(Number<1>{}); constexpr auto num_scale_n_block = CScaleThreadDesc{}.GetLength(Number<2>{}); diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp index a412b12756..dd6d6ba316 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp @@ -233,11 +233,8 @@ struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle } }; + // unconditional 2 to remove agpr usage constexpr index_t minimum_occupancy = 2; - // (BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave && - // MPerBlock * NPerBlock / BlockSize > 64) - // ? 1 - // : 2; if(has_main_k_block_loop) { From 7613cd60d64e9a97408b1a9ae4220f67a3ea62a8 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Fri, 18 Apr 2025 06:32:25 +0000 Subject: [PATCH 26/28] temp save --- .../65_gemm_multiply_multiply/CMakeLists.txt | 3 +- ...line_xdlops_blockscale_b_preshuffle_v1.hpp | 105 ++-------- ...line_xdlops_blockscale_b_preshuffle_v3.hpp | 192 ++---------------- 3 files changed, 37 insertions(+), 263 deletions(-) diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 2b0eeefe21..5f44014784 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -7,7 +7,8 @@ add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) set(EXAMPLE_COMPILE_OPTIONS) list(APPEND EXAMPLE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) -list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) +# Open it when SGBPack branch landed on mainline +# list(APPEND EXAMPLE_COMPILE_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --schedmodel=0 -mllvm -misched=gcn-iterative-max-occupancy-experimental") target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle_v1 PRIVATE ${EXAMPLE_COMPILE_OPTIONS}) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp index dde2f8d09d..07496e6db4 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp @@ -197,96 +197,28 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i) { + ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - /* Judging issue v_pk_fma */ - if constexpr((i + 1) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); // A global static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { + ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2 * i + 1) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2 * i + 2) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); // A local static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - if constexpr((num_buffer_load_inst_b + 2 * num_buffer_load_inst_a + i + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read - }); -#elif 1 // v_mul occured too early causing vmcnt stall - // B global - static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - /* Judging issue v_pk_fma */ - if constexpr((i + 1) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - }); - - // A global - static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write - if constexpr((num_buffer_load_inst_b + 2 * i + 1) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - if constexpr((num_buffer_load_inst_b + 2 * i + 2) % num_mfma_per_kscaleblock == 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - }); - - // A local - static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) { + ignore = i; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read - if constexpr((num_buffer_load_inst_b + 2 * num_buffer_load_inst_a + i + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); -#endif } template {}([&](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); + }); + }); + + HotLoopScheduler(); + __builtin_amdgcn_sched_barrier(0); + static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, num_scale_n_block, 1>{}([&](auto n0) { static_for<0, num_scale_k_block, 1>{}([&](auto k0) { @@ -601,19 +549,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](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, MRepeat, 1>{}([&](auto m0) { a_scale_thread_copy.Run(a_scale_grid_desc, a_scale_grid_buf, @@ -643,8 +578,6 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1{}([&](auto i_inst) { + ignore = i_inst; static_for<0, staged_num_buffer_load_b_per_ds_read_a - 1, 1>{}([&](auto ibuf_inst) { - static_for<0, staged_num_mfma_per_buffer_load_b, 1>{}([&](auto imfma) { + ignore = ibuf_inst; + static_for<0, staged_num_mfma_per_buffer_load_b, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * - staged_num_buffer_load_b_per_ds_read_a + - ibuf_inst * staged_num_mfma_per_buffer_load_b + imfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * - staged_num_buffer_load_b_per_ds_read_a + - (staged_num_buffer_load_b_per_ds_read_a - 1) * - staged_num_mfma_per_buffer_load_b + - 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read - static_for<0, staged_num_mfma_per_buffer_load_b - 1, 1>{}([&](auto imfma) { + static_for<0, staged_num_mfma_per_buffer_load_b - 1, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_buffer_load_b * - staged_num_buffer_load_b_per_ds_read_a + - (staged_num_buffer_load_b_per_ds_read_a - 1) * - staged_num_mfma_per_buffer_load_b + - imfma + 2) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read }); @@ -289,44 +253,18 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_ds_write_a + i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr(((i_inst + 1) * staged_num_mfma_per_ds_write_a) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read } else { static_for<0, staged_num_mfma_per_ds_write_a, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_ds_write_a + i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write @@ -337,52 +275,19 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + - (i_inst - stage_more_mfma) * - (staged_num_mfma_per_ds_write_a - 1) + - i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + - (i_inst - stage_more_mfma + 1) * - (staged_num_mfma_per_ds_write_a - 1)) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read } else { static_for<0, staged_num_mfma_per_ds_write_a - 1, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_ds_write_a + - (i_inst - stage_more_mfma) * - (staged_num_mfma_per_ds_write_a - 1) + - i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write @@ -408,43 +313,18 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_buffer_load_a + i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr(((i_inst + 1) * staged_num_mfma_per_buffer_load_a) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read } else { static_for<0, staged_num_mfma_per_buffer_load_a, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_buffer_load_a + i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read } @@ -454,51 +334,18 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + - (i_inst - stage_more_mfma) * - (staged_num_mfma_per_buffer_load_a - 1) + - i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + - (i_inst - stage_more_mfma + 1) * - (staged_num_mfma_per_buffer_load_a - 1)) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read } else { static_for<0, staged_num_mfma_per_buffer_load_a - 1, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((stage_more_mfma * staged_num_mfma_per_buffer_load_a + - (i_inst - stage_more_mfma) * - (staged_num_mfma_per_buffer_load_a - 1) + - i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read @@ -512,17 +359,10 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v3{}([&](auto i_inst) { + ignore = i_inst; static_for<0, staged_num_mfma_per_ds_read_a, 1>{}([&](auto i_mfma) { + ignore = i_mfma; __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i_inst * staged_num_mfma_per_ds_read_a + i_mfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read }); From 6716e5119de57ef5aca6d608fdc6dd715bb14a33 Mon Sep 17 00:00:00 2001 From: aska-0096 Date: Fri, 18 Apr 2025 06:42:58 +0000 Subject: [PATCH 27/28] remove unnecessary changes --- CMakeLists.txt | 7 ++ ...kwise_gemm_pipeline_xdlops_v3_ab_scale.hpp | 71 +++---------------- 2 files changed, 16 insertions(+), 62 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index bdb6eab511..ba57ead09a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -247,6 +247,13 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500500000) add_compile_options("SHELL: -mllvm --lsr-drop-solution=1") endif() endif() +if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) + check_cxx_compiler_flag("-mllvm -enable-post-misched=0" HAS_ENABLE_POST_MISCHED) + if(HAS_ENABLE_POST_MISCHED) + message("Adding the enable-post-misched=0 compiler flag") + add_compile_options("SHELL: -mllvm -enable-post-misched=0") + endif() +endif() set(check-coerce) check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce) if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp index f28ecf6c6f..77f93aa6e3 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp @@ -194,9 +194,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale @@ -210,61 +207,26 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto i) { + ignore = i; static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) { + ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i * num_mfma_per_issue + idswrite + 1) % num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - static_for<0, num_mfma_per_issue - num_dswrite_per_issue_a, 1>{}([&](auto imfma) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr((i * num_mfma_per_issue + num_dswrite_per_issue_a + imfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - }); + __builtin_amdgcn_sched_group_barrier( + 0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA }); - static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) { + ignore = i; static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) { + ignore = idswrite; __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + idswrite + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } }); __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read - static_for<0, num_mfma_per_issue - num_dswrite_per_issue_b, 1>{}([&](auto imfma) { - __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA - - /* Judging issue v_pk_fma */ - if constexpr(((num_buffer_load_inst_a + i) * num_mfma_per_issue + - num_dswrite_per_issue_b + imfma + 1) % - num_mfma_per_kscaleblock == - 0) - { - __builtin_amdgcn_sched_group_barrier( - 0x800, num_pk_fma_per_kscaleblock, 0); // PK_FMA - } - }); + __builtin_amdgcn_sched_group_barrier( + 0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA }); // stage 2 @@ -282,13 +244,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale{}([&](auto i) { @@ -305,14 +260,6 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale Date: Fri, 18 Apr 2025 09:09:37 +0000 Subject: [PATCH 28/28] enable operator in ckprofiler --- profiler/src/CMakeLists.txt | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 9cb70e4670..f3fd3b8d2e 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -52,6 +52,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_wp.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + list(APPEND PROFILER_SOURCES profile_gemm_blockscale_wp.cpp) endif() list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) @@ -69,7 +70,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) - endif() if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") @@ -142,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_wp_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_blockscale_wp_instance) endif() target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)