diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 55c884246f..2d00545515 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -1,4 +1,5 @@ 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_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) \ No newline at end of file +add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) 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..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 @@ -69,18 +69,21 @@ 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 -///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| -///######| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| -///######| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| -///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S| -///###### 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>; + , 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>; // clang-format on int main(int argc, char* argv[]) @@ -229,7 +232,7 @@ int main(int argc, char* argv[]) "not support this GEMM problem"); } - float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50}); + float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 20, 50, true, 50}); std::size_t flop = std::size_t(2) * M * N * K; std::size_t num_btype = 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 new file mode 100644 index 0000000000..9a81ef5ea7 --- /dev/null +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp @@ -0,0 +1,396 @@ +// 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_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" + +#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 F16 = ck::half_t; +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 B0DataType = FP8; +using AccDataType = F32; +using CShuffleDataType = F32; +using D0DataType = F32; +using D1DataType = F32; +using DsDataType = ck::Tuple; +using EDataType = F16; + +using A0Layout = Row; +using B0Layout = Col; +using D0Layout = Row; +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; + + template <> + __host__ __device__ constexpr void operator()(F16& e, + const float& c, + const float& d0, + const float& d1) const + { + const float x0_f = c * d0 * d1; + + e = ck::type_convert(x0_f); + } + + template <> + __host__ __device__ constexpr void operator()(BF16& e, + const float& c, + const float& d0, + const float& d1) const + { + const float x0_f = c * d0 * d1; + + e = ck::type_convert(x0_f); + } + + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const int& c, const float& d0, const float& d1) const + { + const float x0_f = + ck::type_convert(c) * ck::type_convert(d0) * ck::type_convert(d1); + + e = ck::type_convert(x0_f); + } + + template <> + __host__ __device__ constexpr void operator()( + ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const + { + const float x0_f = + ck::type_convert(c) * ck::type_convert(d0) * ck::type_convert(d1); + + e = ck::type_convert(x0_f); + } +}; + +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 = MultiplyMultiply; + +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; + +using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle + // clang-format off + < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, + AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, + 256, 256, 128, + 16, 16, + 16, 16, + 8, 8, + S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, + 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, + 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; + + // GEMM shape + ck::index_t M = 3840; + ck::index_t N = 4096; + ck::index_t K = 4096; + + ck::index_t StrideA = K; + ck::index_t StrideB = K; + ck::index_t StrideD = 0; + ck::index_t StrideE = N; + + ck::index_t KBatch = 1; + + ck::index_t Warmup = 50; + ck::index_t Repeat = 50; + + 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 == 12) + { + 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]); + + StrideA = std::stoi(argv[7]); + StrideB = std::stoi(argv[8]); + StrideD = std::stoi(argv[9]); + StrideE = std::stoi(argv[10]); + + KBatch = std::stoi(argv[11]); + } + else if(argc == 14) + { + 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]); + + StrideA = std::stoi(argv[7]); + StrideB = std::stoi(argv[8]); + StrideD = std::stoi(argv[9]); + StrideE = std::stoi(argv[10]); + + KBatch = std::stoi(argv[11]); + + Warmup = std::stoi(argv[12]); + Repeat = std::stoi(argv[13]); + } + 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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, KBatch\n"); + printf("arg10 to 11: Warmup, Repeat\n"); + exit(0); + } + + 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 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 d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{})); + Tensor d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{})); + 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 << "b0_k_n: " << b0_k_n.mDesc << std::endl; + std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl; + std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl; + std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{0, 2}); + d0_m_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + d1_m_n.GenerateTensorValue(GeneratorTensor_2{-2, 2}); + break; + case 2: + a0_m_k.GenerateTensorValue(GeneratorTensor_1{}); + b0_k_n.GenerateTensorValue(GeneratorTensor_1{}); + d0_m_n.GenerateTensorValue(GeneratorTensor_1{}); + d1_m_n.GenerateTensorValue(GeneratorTensor_1{}); + break; + default: + a0_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + d0_m_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + d1_m_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + } + DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize()); + DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_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()); + d0_device_buf.ToDevice(d0_m_n.mData.data()); + d1_device_buf.ToDevice(d1_m_n.mData.data()); + e_device_buf.ToDevice(e_m_n_device_result.mData.data()); + + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto cde_element_op = CDEElementOp{}; + + constexpr ck::index_t NumDTensor = DsDataType::Size(); + + constexpr auto I0 = ck::Number<0>{}; + + // 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{d0_device_buf.GetDeviceBuffer(), + d1_device_buf.GetDeviceBuffer()}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{I0, I0}, + StrideE, + KBatch, + 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"); + } + + size_t total_size = + (M * K * sizeof(A0DataType) + N * K * sizeof(B0DataType) + M * sizeof(D0DataType) + + N * sizeof(D1DataType) + M * N * sizeof(EDataType)); + int rotate_buf_num = + ck::math::min(size_t(Repeat), ck::math::integer_divide_ceil(512 * 1024 * 1024, total_size)); + + float ave_time = invoker.Run( + argument, StreamConfig{nullptr, time_kernel, 0, Warmup, Repeat, true, rotate_buf_num}); + + 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: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" + << std::endl; + + if(do_verification) + { + invoker.Run(argument, StreamConfig{nullptr, false}); + + e_device_buf.FromDevice(e_m_n_device_result.mData.data()); + + Tensor c_m_n({M, N}); + + using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm; + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = ref_gemm.MakeArgument( + a0_m_k, b0_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) + { + cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n)); + } + } + + 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!", 1e-3, 5e-2) + ? 0 + : 1; + } + + return 0; +} diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp index fb1642bba5..cbbd37408e 100644 --- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp @@ -36,9 +36,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; using A0DataType = I8; using B0DataType = I8; using AccDataType = I32; -using CShuffleDataType = I32; -using D0DataType = F32; -using D1DataType = F32; +using CShuffleDataType = F16; +using D0DataType = F16; +using D1DataType = F16; using DsDataType = ck::Tuple; using EDataType = F16; @@ -74,6 +74,24 @@ struct MultiplyMultiply e = ck::type_convert(x0_f); } + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const int& c, const ck::half_t& d0, const ck::half_t& d1) const + { + const ck::half_t x0_f = ck::type_convert(c) * d0 * d1; + + e = x0_f; + } + + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const ck::half_t& c, const ck::half_t& d0, const ck::half_t& d1) const + { + const ck::half_t x0_f = c * d0 * d1; + + e = x0_f; + } + template <> __host__ __device__ constexpr void operator()( ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const @@ -91,7 +109,7 @@ 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 @@ -102,7 +120,17 @@ 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, I8>; ///###### 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, I8>; + < Row, Col, DsLayout, ELayout, + A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, + AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, + 64, 128, 256, + 16, 16, + 32, 32, + 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, I8>; // clang-format on int main(int argc, char* argv[]) @@ -196,6 +224,12 @@ int main(int argc, char* argv[]) d0_m_n.GenerateTensorValue(GeneratorTensor_2{0, 2}); d1_m_n.GenerateTensorValue(GeneratorTensor_2{0, 2}); break; + case 2: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-25, 25}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{0, 25}); + d0_m_n.GenerateTensorValue(GeneratorTensor_2{0, 200}); + d1_m_n.GenerateTensorValue(GeneratorTensor_2{0, 200}); + break; default: a0_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); @@ -251,7 +285,10 @@ int main(int argc, char* argv[]) "not support this GEMM problem"); } - float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50}); + hipStream_t stream; + hip_check_error(hipStreamCreate(&stream)); + + float ave_time = invoker.Run(argument, StreamConfig{stream, time_kernel, 0, 20, 50, true, 50}); std::size_t flop = std::size_t(2) * M * N * K; std::size_t num_btype = diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp index 43bc9a6cfe..574edf64d3 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp @@ -3,7 +3,7 @@ #include // different threshold for different dtype -template +template auto get_elimit() { double rtol = 1e-2; @@ -39,6 +39,7 @@ auto create_args(int argc, char* argv[]) .insert("v", "1", "cpu validation or not") .insert("kname", "1", "print kernel name or not") .insert("prec", "fp16", "precision") + .insert("quant", "int8", "precision") .insert("warmup", "5", "cold iter") .insert("repeat", "20", "hot iter"); @@ -46,7 +47,7 @@ auto create_args(int argc, char* argv[]) return std::make_tuple(result, arg_parser); } -template +template bool run(const ck_tile::ArgParser& arg_parser) { ck_tile::index_t m = arg_parser.get_int("m"); @@ -54,16 +55,17 @@ bool run(const ck_tile::ArgParser& arg_parser) ck_tile::index_t stride = arg_parser.get_int("stride"); if(stride < 0) stride = n; - float epsilon = arg_parser.get_float("e"); - std::string data_type = arg_parser.get_str("prec"); - int kname = arg_parser.get_int("kname"); - int do_validation = arg_parser.get_int("v"); - int warmup = arg_parser.get_int("warmup"); - int repeat = arg_parser.get_int("repeat"); + float epsilon = arg_parser.get_float("e"); + std::string input_data_type = arg_parser.get_str("prec"); + std::string quantized_data_type = arg_parser.get_str("quant"); + int kname = arg_parser.get_int("kname"); + int do_validation = arg_parser.get_int("v"); + int warmup = arg_parser.get_int("warmup"); + int repeat = arg_parser.get_int("repeat"); assert(stride >= n); - using TypeConfig = AddRmsnormRdquantTypeConfig; + using TypeConfig = AddRmsnormRdquantTypeConfig; using ADataType = typename TypeConfig::ADataType; using BDataType = typename TypeConfig::BDataType; @@ -102,10 +104,10 @@ bool run(const ck_tile::ArgParser& arg_parser) b_buf.ToDevice(b_host.data()); gamma_buf.ToDevice(gamma_host.data()); - std::cout << "[" << data_type << "]" + std::cout << "[" << input_data_type << ", " << quantized_data_type << "]" << " m:" << m << ", n:" << n << ", stride:" << stride << std::flush; - add_rmsnorm2d_rdquant_fwd_traits traits{data_type, SaveX}; + add_rmsnorm2d_rdquant_fwd_traits traits{input_data_type, quantized_data_type, SaveX}; add_rmsnorm2d_rdquant_fwd_args args{a_buf.GetDeviceBuffer(), b_buf.GetDeviceBuffer(), @@ -129,14 +131,14 @@ bool run(const ck_tile::ArgParser& arg_parser) num_byte += sizeof(XDataType) * m * n; float gb_per_sec = num_byte / 1.E6 / ave_time; - std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush; + std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::endl; bool pass = true; if(do_validation) { using YDataType = ComputeDataType; - using InvRmsDataType = DataType; + using InvRmsDataType = InputDataType; // Add { @@ -144,28 +146,36 @@ bool run(const ck_tile::ArgParser& arg_parser) ck_tile::reference_binary_elementwise( a_host, b_host, x_host_ref, op); - x_buf.FromDevice(x_host_dev.data()); + if constexpr(SaveX) + { + x_buf.FromDevice(x_host_dev.data()); - auto [rtol, atol] = get_elimit(); - if(stride == n) - { - pass = ck_tile::check_err( - x_host_dev, x_host_ref, std::string("x Error: Incorrect results!"), rtol, atol); - } - else - { - for(int i_r = 0; i_r < m; i_r++) + auto [rtol, atol] = get_elimit(); + if(stride == n) { - std::vector x_host_dev_row(x_host_dev.begin() + i_r * stride, - x_host_dev.begin() + i_r * stride + n); - std::vector x_host_ref_row(x_host_ref.begin() + i_r * stride, - x_host_ref.begin() + i_r * stride + n); - pass &= ck_tile::check_err(x_host_dev_row, - x_host_ref_row, - std::string("x[") + std::to_string(i_r) + - std::string("] Error: Incorrect results!"), - rtol, - atol); + pass = ck_tile::check_err(x_host_dev, + x_host_ref, + std::string("x Error: Incorrect results!"), + rtol, + atol); + } + else + { + for(int i_r = 0; i_r < m; i_r++) + { + std::vector x_host_dev_row(x_host_dev.begin() + i_r * stride, + x_host_dev.begin() + i_r * stride + + n); + std::vector x_host_ref_row(x_host_ref.begin() + i_r * stride, + x_host_ref.begin() + i_r * stride + + n); + pass &= ck_tile::check_err(x_host_dev_row, + x_host_ref_row, + std::string("x[") + std::to_string(i_r) + + std::string("] Error: Incorrect results!"), + rtol, + atol); + } } } } @@ -256,23 +266,40 @@ int main(int argc, char* argv[]) if(!result) return -1; - const std::string data_type = arg_parser.get_str("prec"); - int save_x = arg_parser.get_int("save_x"); - if(data_type == "fp16" && save_x) + const std::string input_data_type = arg_parser.get_str("prec"); + const std::string quantized_data_type = arg_parser.get_str("quant"); + int save_x = arg_parser.get_int("save_x"); + if(input_data_type == "fp16" && quantized_data_type == "int8" && save_x) { - return run(arg_parser) ? 0 : -2; + return run(arg_parser) ? 0 : -2; } - else if(data_type == "fp16" && !save_x) + else if(input_data_type == "fp16" && quantized_data_type == "int8" && !save_x) { - return run(arg_parser) ? 0 : -2; + return run(arg_parser) ? 0 : -2; } - else if(data_type == "bf16" && save_x) + else if(input_data_type == "bf16" && quantized_data_type == "int8" && save_x) { - return run(arg_parser) ? 0 : -2; + return run(arg_parser) ? 0 : -2; } - else if(data_type == "bf16" && !save_x) + else if(input_data_type == "bf16" && quantized_data_type == "int8" && !save_x) { - return run(arg_parser) ? 0 : -2; + return run(arg_parser) ? 0 : -2; + } + else if(input_data_type == "fp16" && quantized_data_type == "fp8" && save_x) + { + return run(arg_parser) ? 0 : -2; + } + else if(input_data_type == "fp16" && quantized_data_type == "fp8" && !save_x) + { + return run(arg_parser) ? 0 : -2; + } + else if(input_data_type == "bf16" && quantized_data_type == "fp8" && save_x) + { + return run(arg_parser) ? 0 : -2; + } + else if(input_data_type == "bf16" && quantized_data_type == "fp8" && !save_x) + { + return run(arg_parser) ? 0 : -2; } return -3; diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp index 443b9b1024..c91b387d62 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp @@ -8,11 +8,11 @@ #include "ck_tile/ops/add_rmsnorm2d_rdquant.hpp" #include -template +template struct AddRmsnormRdquantTypeConfig; template <> -struct AddRmsnormRdquantTypeConfig +struct AddRmsnormRdquantTypeConfig { using ADataType = ck_tile::half_t; using BDataType = ck_tile::half_t; @@ -24,7 +24,7 @@ struct AddRmsnormRdquantTypeConfig }; template <> -struct AddRmsnormRdquantTypeConfig +struct AddRmsnormRdquantTypeConfig { using ADataType = ck_tile::bf16_t; using BDataType = ck_tile::bf16_t; @@ -35,13 +35,38 @@ struct AddRmsnormRdquantTypeConfig using ComputeDataType = float; }; +template <> +struct AddRmsnormRdquantTypeConfig +{ + using ADataType = ck_tile::half_t; + using BDataType = ck_tile::half_t; + using GammaDataType = ck_tile::half_t; + using XDataType = ck_tile::half_t; + using YScaleDataType = float; + using QYDataType = ck_tile::fp8_t; + using ComputeDataType = float; +}; + +template <> +struct AddRmsnormRdquantTypeConfig +{ + using ADataType = ck_tile::bf16_t; + using BDataType = ck_tile::bf16_t; + using GammaDataType = ck_tile::bf16_t; + using XDataType = ck_tile::bf16_t; + using YScaleDataType = float; + using QYDataType = ck_tile::fp8_t; + using ComputeDataType = float; +}; + // runtime args struct add_rmsnorm2d_rdquant_fwd_args : public ck_tile::AddRmsnorm2dRdquantFwdHostArgs { }; // this is used to pattern-match internl kernel implementation, not to instantiate kernel -template struct add_rmsnorm2d_rdquant_fwd_traits_ { - using DataType = ck_tile::remove_cvref_t; + using InputDataType = ck_tile::remove_cvref_t; + using QuantizedDataType = ck_tile::remove_cvref_t; static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize; static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0); @@ -114,7 +140,8 @@ float add_rmsnorm2d_rdquant_fwd_(const ck_tile::stream_config& s, add_rmsnorm2d_ // This is the public API, will be generated by script struct add_rmsnorm2d_rdquant_fwd_traits { - std::string data_type; + std::string input_data_type; + std::string quantized_data_type; bool save_x; }; diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_api.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_api.cpp index 966c5bd02f..f695ea30b2 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_api.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_api.cpp @@ -4,7 +4,8 @@ #include #include "add_rmsnorm2d_rdquant_fwd.hpp" -template -using trait_ = add_rmsnorm2d_rdquant_fwd_traits_; -template -float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits /*t*/, +template +float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits t, add_rmsnorm2d_rdquant_fwd_args a, const ck_tile::stream_config& s) { @@ -32,99 +34,145 @@ float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits /*t*/, // clang-format off // rm rn tm tn vn pd x 3p if(a.n <= 64) { - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 128) { if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 256) { if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 512) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 768) { if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 1024) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 1536) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 2048) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 3072) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } else if(a.n <= 4096) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } - else if(a.n > 4096) { + else if(a.n <= 8192) { + if(a.n<8192){ + if(t.save_x){ + if (a.n % 8 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 4 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 2 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + } + else{ + if (a.n % 8 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 4 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 2 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + } + } + else{ + if(t.save_x){ + if (a.n % 8 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 4 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 2 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + } + else{ + if (a.n % 8 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 4 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else if (a.n % 2 == 0) + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + else + r = add_rmsnorm2d_rdquant_fwd_>(s, a); + } + } + } + else if(a.n > 8192) { if (a.n % 8 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 4 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else if (a.n % 2 == 0) - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); else - r = add_rmsnorm2d_rdquant_fwd_>(s, a); + r = add_rmsnorm2d_rdquant_fwd_>(s, a); } return r; // clang-format on @@ -134,16 +182,45 @@ float add_rmsnorm2d_rdquant_fwd(add_rmsnorm2d_rdquant_fwd_traits t, add_rmsnorm2d_rdquant_fwd_args a, const ck_tile::stream_config& s) { - - // Only support instance of save_x == true for now - assert(t.save_x); - if(t.data_type.compare("fp16") == 0) + if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("int8") == 0 && + t.save_x) { - return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); } - else if(t.data_type.compare("bf16") == 0) + else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("int8") == 0 && + !t.save_x) { - return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("int8") == 0 && + t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("int8") == 0 && + !t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("fp8") == 0 && + t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("fp8") == 0 && + !t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("fp8") == 0 && + t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); + } + else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("fp8") == 0 && + !t.save_x) + { + return add_rmsnorm2d_rdquant_fwd_b16_(t, a, s); } else throw std::runtime_error("Without supported instances!"); diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1024_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1024_instance.cpp index 5495e3c9ab..00df2f5082 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1024_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1024_instance.cpp @@ -15,8 +15,12 @@ template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); #endif -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp index 8bbfdc8589..2adb54c078 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp @@ -6,8 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp index 381a11fc80..39089843a2 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp @@ -6,9 +6,13 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp index 2fefac6934..ddb8e1b354 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp @@ -6,7 +6,10 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp index 263713bbc7..2a87614403 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp @@ -6,9 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp index c62c596fab..045a3b8880 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp @@ -6,9 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_tp_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_tp_instance.cpp deleted file mode 100644 index e4951f6ab9..0000000000 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_tp_instance.cpp +++ /dev/null @@ -1,14 +0,0 @@ - -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. - -#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" - -// clang-format off -// rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - -// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp index 4c7ee48e8e..1028973e74 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp @@ -6,8 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp index 8659dc82b3..b8439a0ce9 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp @@ -6,7 +6,10 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp index 5f15f11b47..b24b245757 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp @@ -6,7 +6,10 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_instance.cpp new file mode 100644 index 0000000000..14f0ec8525 --- /dev/null +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_instance.cpp @@ -0,0 +1,42 @@ + +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" + +// clang-format off +// rm rn tm tn vn pd x 3p +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); + +// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_tp_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_tp_instance.cpp new file mode 100644 index 0000000000..3e3a6d75b9 --- /dev/null +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_tp_instance.cpp @@ -0,0 +1,17 @@ + +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" + +// clang-format off +// rm rn tm tn vn pd x 3p +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp index 8ffdacbdcd..04d735c12c 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp @@ -15,8 +15,12 @@ template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); #endif -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp index 3551099651..5893d6c3ee 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp @@ -6,8 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp index d4d0474c27..ec9c417bf3 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp @@ -6,9 +6,13 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp index 2cb300eda6..5bc8245106 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp @@ -6,7 +6,10 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp index fb0ceb4c58..c022c62de6 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp @@ -6,9 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp index 3a241a3c93..19172b0793 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp @@ -6,9 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_tp_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_tp_instance.cpp deleted file mode 100644 index d3094679f9..0000000000 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_tp_instance.cpp +++ /dev/null @@ -1,14 +0,0 @@ - -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. - -#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" - -// clang-format off -// rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); - -// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp index 919bc177e8..f491d92787 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp @@ -6,8 +6,12 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp index 8a44f5e00f..065f0ea4cc 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp @@ -5,8 +5,11 @@ #include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" // clang-format off -// rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +// rm rn tm tn vn pd x 3p +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp index 5c4f05ec3c..be8c6c4de5 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp @@ -6,7 +6,10 @@ // clang-format off // rm rn tm tn vn pd x 3p -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); -template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); // clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_instance.cpp new file mode 100644 index 0000000000..ad2dfd931e --- /dev/null +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_instance.cpp @@ -0,0 +1,41 @@ + +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" + +// clang-format off +// rm rn tm tn vn pd x 3p +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_tp_instance.cpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_tp_instance.cpp new file mode 100644 index 0000000000..e3afa07fa4 --- /dev/null +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n8192_tp_instance.cpp @@ -0,0 +1,17 @@ + +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp" + +// clang-format off +// rm rn tm tn vn pd x 3p +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +template float add_rmsnorm2d_rdquant_fwd_>(const S&, A); +// clang-format on diff --git a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp index 6baaad471a..25b10e1dc4 100644 --- a/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp +++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp @@ -11,7 +11,8 @@ using S = ck_tile::stream_config; using A = add_rmsnorm2d_rdquant_fwd_args; -template -using trait_ = add_rmsnorm2d_rdquant_fwd_traits_ float add_rmsnorm2d_rdquant_fwd_(const S& s, A a) { - using DataType = typename Traits_::DataType; + using InputDataType = typename Traits_::InputDataType; + using QuantizedDataType = typename Traits_::QuantizedDataType; using PipelineProblem = ck_tile::AddRmsnorm2dRdquantFwdPipelineProblem< - typename AddRmsnormRdquantTypeConfig::ADataType, - typename AddRmsnormRdquantTypeConfig::BDataType, - typename AddRmsnormRdquantTypeConfig::GammaDataType, - typename AddRmsnormRdquantTypeConfig::ComputeDataType, - typename AddRmsnormRdquantTypeConfig::XDataType, - typename AddRmsnormRdquantTypeConfig::YScaleDataType, - typename AddRmsnormRdquantTypeConfig::QYDataType, + typename AddRmsnormRdquantTypeConfig::ADataType, + typename AddRmsnormRdquantTypeConfig::BDataType, + typename AddRmsnormRdquantTypeConfig::GammaDataType, + typename AddRmsnormRdquantTypeConfig::ComputeDataType, + typename AddRmsnormRdquantTypeConfig::XDataType, + typename AddRmsnormRdquantTypeConfig::YScaleDataType, + typename AddRmsnormRdquantTypeConfig::QYDataType, typename Traits_::Shape, Traits_::kPadN, Traits_::kSaveX, diff --git a/include/ck/host_utility/hip_check_error.hpp b/include/ck/host_utility/hip_check_error.hpp index c0894f1d70..0dfd275269 100644 --- a/include/ck/host_utility/hip_check_error.hpp +++ b/include/ck/host_utility/hip_check_error.hpp @@ -19,16 +19,15 @@ inline void hip_check_error(hipError_t x) } } -#define HIP_CHECK_ERROR(retval_or_funcall) \ - do \ - { \ - hipError_t _tmpVal = retval_or_funcall; \ - if(_tmpVal != hipSuccess) \ - { \ - std::ostringstream ostr; \ - ostr << "HIP Function Failed (" \ - << "hip_check_error.hpp" \ - << "," << __LINE__ << ") " << hipGetErrorString(_tmpVal); \ - throw std::runtime_error(ostr.str()); \ - } \ +#define HIP_CHECK_ERROR(retval_or_funcall) \ + do \ + { \ + hipError_t _tmpVal = retval_or_funcall; \ + if(_tmpVal != hipSuccess) \ + { \ + std::ostringstream ostr; \ + ostr << "HIP Function Failed (" << __FILE__ << "," << __LINE__ << ") " \ + << hipGetErrorString(_tmpVal); \ + throw std::runtime_error(ostr.str()); \ + } \ } while(0) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_ab_scale_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_ab_scale_selector.hpp index 1ab460fa8a..5997b47790 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_ab_scale_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_ab_scale_selector.hpp @@ -9,13 +9,6 @@ namespace ck { -enum struct BlockGemmPipelineVersion -{ - v1, // Naive - v2, // Mem - v3, // Comp -}; - template +constexpr auto BlockGemmBPreshufflePipeline_Selector() +{ + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + return BlockwiseGemmXdlops_pipeline_bpreshuffle_v1{}; + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) + { + return BlockwiseGemmXdlops_pipeline_bpreshuffle_v2{}; + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3"); + return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3{}; + } + 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_b_preshuffle_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp new file mode 100644 index 0000000000..8ed25895b5 --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_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_b_preshuffle_v2.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp new file mode 100644 index 0000000000..4c019a41a4 --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp @@ -0,0 +1,558 @@ +// 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: 3 +// LocalPreFillStages: 2 +// LocalPreFetchStages: 2 +// LocalSharedMemoryBuffer: 2 + +template +struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2 + : 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 = 3; + static constexpr index_t PrefillStages = 2; + 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 + A local + static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) { + ignore = i; + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read A + }); + + static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) { + ignore = i; + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read A + }); + + // 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{}> a_thread_bufs; + 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); + + // Local prefill A1 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0); + + // Global prefetch A2 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1); + 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.At(I0), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, I0), + a_thread_bufs(I0)); + }); + }); + + // Local prefill A2 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1); + + // // Global prefetch A3 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + // 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) { + block_sync_lds(); + + 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); + + 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.At(local_read_buf), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, I0), + a_thread_bufs(local_read_buf)); + }); + }); + + a_blockwise_copy.RunWrite( + a_block_desc, a_block_buf.At(mfma_reg_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_bufs[mfma_reg_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{})); + }); + }); + }); + + HotLoopScheduler(); + __builtin_amdgcn_sched_barrier(0); + }; + + LoopFunc(I0, I1); + LoopFunc(I1, I0); + + i += 2; + } while(i < (num_loop - 3)); + } + // tail + + auto ReadWriteCompFunc = [&](auto mfma_reg, auto local_read_reg) { + block_sync_lds(); + + 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_reg)); + b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); + + 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.At(local_read_reg), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, I0), + a_thread_bufs(local_read_reg)); + }); + }); + + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(mfma_reg), mfma_reg); + + 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_bufs[mfma_reg][Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[mfma_reg][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{})); + }); + }); + }); + + HotLoopScheduler(); + __builtin_amdgcn_sched_barrier(0); + }; + + auto ReadCompFunc = [&](auto mfma_reg, auto local_read_reg) { + block_sync_lds(); + + 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_reg)); + + 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.At(local_read_reg), + a_thread_desc_, + make_tuple(m0, I0, I0, k0, I0, I0), + a_thread_bufs(local_read_reg)); + }); + }); + + 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_bufs[mfma_reg][Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[mfma_reg][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{})); + }); + }); + }); + + HotLoopScheduler(); + __builtin_amdgcn_sched_barrier(0); + }; + + auto CompFunc = [&](auto mfma_reg) { + 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_bufs[mfma_reg][Number{}]; + b_thread_vec.template AsType()(ik) = + b_thread_bufs[mfma_reg][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{})); + }); + }); + }); + }; + + if constexpr(TailNum == TailNumber::Even) + { + ReadCompFunc(I0, I1); + CompFunc(I1); + } + else if constexpr(TailNum == TailNumber::Odd) + { + ReadWriteCompFunc(I0, I1); + ReadCompFunc(I1, I0); + CompFunc(I0); + } + } + + 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_b_preshuffle_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp new file mode 100644 index 0000000000..49af782132 --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp @@ -0,0 +1,860 @@ +// 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_v3 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_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(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 + 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); + + // // Local prefill A1 + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); + + // // Global prefetch A2 + a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf); + a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); + + // Local prefetch A1 + block_sync_lds(); + static_for<0, 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); + }); + + // 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) { + 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, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + 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{})); + }); + }); + + 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); + }); + }; + + 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, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + 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{})); + }); + }); + + 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) { + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + 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{})); + }); + }); + + 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, KRepeat, 1>{}([&](auto k0) { + static_for<0, NRepeat, 1>{}([&](auto n0) { + 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{})); + }); + }); + + 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/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp index ea0c511da3..48fe5131d2 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp @@ -11,15 +11,6 @@ namespace ck { -enum struct BlockGemmPipelineVersion -{ - v1, // Naive - v2, // Mem - v3, // Comp - v4, // Comp, double lds buffer - v5, // Comp, double global prefetch register buffer -}; - template - __device__ static constexpr void HotLoopScheduler(ScheduleGroup schedule_group) + __device__ static constexpr void HotLoopScheduler() { // TODO: Take data type into consideration as pipe ver 3 // A-B splited schedule @@ -195,42 +194,42 @@ struct BlockwiseGemmXdlops_pipeline_v4{}([&](auto idsread) { ignore = idsread; - __builtin_amdgcn_sched_group_barrier(0x100, 1, schedule_group); // DS read - __builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA }); static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) { ignore = idswrite; - __builtin_amdgcn_sched_group_barrier(0x200, 1, schedule_group); // DS write - __builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // 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, schedule_group); // VMEM read + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_issue - num_dsread_per_issue_a - num_dswrite_per_issue_a, - schedule_group); // MFMA + 0); // MFMA }); static_for<0, num_issue_b, 1>{}([&](auto i) { ignore = i; static_for<0, num_dsread_per_issue_b, 1>{}([&](auto idsread) { ignore = idsread; - __builtin_amdgcn_sched_group_barrier(0x100, 1, schedule_group); // DS read - __builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // MFMA + __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read + __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA }); static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) { ignore = idswrite; - __builtin_amdgcn_sched_group_barrier(0x200, 1, schedule_group); // DS write - __builtin_amdgcn_sched_group_barrier(0x008, 1, schedule_group); // 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, schedule_group); // VMEM read + __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_issue - num_dsread_per_issue_a - num_dswrite_per_issue_b, - schedule_group); // MFMA + 0); // MFMA }); __builtin_amdgcn_sched_barrier(0); } @@ -274,26 +273,15 @@ struct BlockwiseGemmXdlops_pipeline_v4{}> b_thread_bufs; // Global prefetch 1 - a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0); - b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0); - - a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step); - b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step); - - // Global prefetch 2 - a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1); - b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1); + 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); // Local prefill 1 - a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0); - b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0), I0); - - // Local prefill 2 - a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1); - b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1), I1); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); + b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0)); // Local prefetch 1 block_sync_lds(); @@ -316,16 +304,20 @@ struct BlockwiseGemmXdlops_pipeline_v4{}([&](auto k) { @@ -367,13 +357,11 @@ struct BlockwiseGemmXdlops_pipeline_v4{}([&](auto k) { @@ -447,8 +433,8 @@ struct BlockwiseGemmXdlops_pipeline_v4{}([&](auto k0) { static_for<0, MRepeat, 1>{}([&](auto m0) { @@ -478,13 +464,10 @@ struct BlockwiseGemmXdlops_pipeline_v4{}([&](auto k) { @@ -534,7 +517,7 @@ struct BlockwiseGemmXdlops_pipeline_v4 + __device__ constexpr auto GetSrcThreadScratchIdx() + { + return threadwise_transfer_.template GetSrcThreadScratchIdx(); + } + template __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp index 48fca67f56..403a1cb085 100644 --- a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp +++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp @@ -96,6 +96,51 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator virtual std::unique_ptr MakeInvokerPointer() = 0; }; +// GEMM: +// input : A[M, K], B[K, N], +// input : D0[M, N], D1[M, N], ... +// output : E[M, N] +// C = a_op(A) * b_op(B) +// E = cde_op(C, D0, D1, ...) +// Assume: +// D0, D1, ... and E have the same layout +template +struct DeviceGemmMultipleDSplitKBPreShuffle : 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, + ck::index_t M, + ck::index_t N, + ck::index_t K, + ck::index_t StrideA, + ck::index_t StrideB, + std::array StrideDs, + ck::index_t StrideE, + ck::index_t KBatch, + 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.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp index 7661114ef8..0796614bd4 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp @@ -227,8 +227,20 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK +#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.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_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_Xdl_CShuffle_V3_BPreshuffle + : public DeviceGemmMultipleDSplitKBPreShuffle +{ + static constexpr index_t NumDTensor = DsDataType::Size(); + + // GridwiseGemm + using GridwiseGemm = GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle< + ALayout, + BLayout, + DsLayout, + CLayout, + ADataType, + BDataType, + GemmAccDataType, + CShuffleDataType, + DsDataType, + CDataType, + AElementwiseOperation, + BElementwiseOperation, + CElementwiseOperation, + GemmSpec, + BlockSize, + 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) + { + + std::array DsSize; + + 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); + + const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N( + arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs); + + static_for<0, NumDTensor, 1>{}([&](auto i) { + using DDataType = remove_cvref_t>; + DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType); + }); + ck::utility::RotatingMemWrapperMultiD rotating_mem( + arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer, DsSize); + 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 auto estimated_reg_a = MPerBlock * KPerBlock * sizeof(ADataType) / BlockSize / + 4 * (1 + GridwiseGemm::NWave); + constexpr auto estimated_reg_b = + NPerBlock * KPerBlock * sizeof(BDataType) / BlockSize / 4 * (2); + constexpr auto estimated_reg_c = + MPerBlock * NPerBlock * sizeof(GemmAccDataType) / BlockSize / 4; + constexpr auto estimated_reg_total = + estimated_reg_a + estimated_reg_b + estimated_reg_c; + + constexpr index_t minimum_occupancy = (estimated_reg_total >= 256) ? 1 : 2; + + // static_assert(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 && + // has_main_k_block_loop, "only impl BlockGemmPipelineVersion::v3 and has mainloop right + // now"); + if(has_main_k_block_loop) + { + // Tail number always full + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || + BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + } + else + { + throw std::runtime_error("todo: only v1 v2 and v3 support now"); + } + } +#if 0 + else + { + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { +#if 0 + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } +#endif + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + false, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + false, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + } + else + { + throw std::runtime_error("todo: only v3 support now"); + } + } +#endif + + 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(!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; + } + + 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, + index_t M, + index_t N, + index_t K, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideC, + index_t KBatch, + 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, + KBatch, + 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, + index_t M, + index_t N, + index_t K, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideC, + index_t KBatch, + 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, + KBatch, + 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"<(x0_f); } + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const int& c, const float& d0, const float& d1) const + { + const float x0_f = + ck::type_convert(c) * ck::type_convert(d0) * ck::type_convert(d1); + + e = ck::type_convert(x0_f); + } + template <> __host__ __device__ constexpr void operator()( ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const diff --git a/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp index f1055d1eff..be4e68bffa 100644 --- a/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp +++ b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp @@ -284,6 +284,12 @@ struct PassThrough y = type_convert(x); } + template <> + __host__ __device__ void operator()(half_t& y, const int32_t& x) const + { + y = type_convert(x); + } + template <> __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const { diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp index 55360fc0d0..d4c915aa5e 100755 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp @@ -674,11 +674,13 @@ struct GridwiseGemm_xdl_cshuffle_v3 __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1() { // A matrix in LDS memory, dst of blockwise copy - if constexpr(ABlockLdsExtraM) + if constexpr(ABlockLdsExtraM || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) { + // bank conflict when writting the data into LDS, but don't worry, we have whole entire + // loop to hide it in v4. it may give you some benefit from less valu in compute address return make_naive_tensor_descriptor( make_tuple(AK0Number, Number{}, AK1Number), - make_tuple(AK1Number, Number{}, I1)); + make_tuple(Number{} * AK1Number, AK1Number, I1)); } // xor tensor transformation request more unnecessary vgpr usage, would cause register spill // in some cases. @@ -810,11 +812,13 @@ struct GridwiseGemm_xdl_cshuffle_v3 __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1() { // B matrix in LDS memory, dst of blockwise copy - if constexpr(BBlockLdsExtraN) + if constexpr(BBlockLdsExtraN || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) { + // bank conflict when writting the data into LDS, but don't worry, we have whole entire + // loop to hide it in v4. it may give you some benefit from less valu in compute address return make_naive_tensor_descriptor( make_tuple(BK0Number, Number{}, BK1Number), - make_tuple(BK1Number, Number{}, I1)); + make_tuple(Number{} * BK1Number, BK1Number, I1)); } else if constexpr(is_same::value) { 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..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 @@ -676,11 +676,13 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1() { // A matrix in LDS memory, dst of blockwise copy - if constexpr(ABlockLdsExtraM) + if constexpr(ABlockLdsExtraM || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) { + // bank conflict when writting the data into LDS, but don't worry, we have whole entire + // loop to hide it in v4. it may give you some benefit from less valu in compute address return make_naive_tensor_descriptor( make_tuple(AK0Number, Number{}, AK1Number), - make_tuple(AK1Number, Number{}, I1)); + make_tuple(Number{} * AK1Number, AK1Number, I1)); } // xor tensor transformation request more unnecessary vgpr usage, would cause register spill // in some cases. @@ -813,11 +815,13 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1() { // B matrix in LDS memory, dst of blockwise copy - if constexpr(BBlockLdsExtraN) + if constexpr(BBlockLdsExtraN || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) { + // bank conflict when writting the data into LDS, but don't worry, we have whole entire + // loop to hide it in v4. it may give you some benefit from less valu in compute address return make_naive_tensor_descriptor( make_tuple(BK0Number, Number{}, BK1Number), - make_tuple(BK1Number, Number{}, I1)); + make_tuple(Number{} * BK1Number, BK1Number, I1)); } else if constexpr(is_same::value) { diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp new file mode 100644 index 0000000000..238ab14606 --- /dev/null +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp @@ -0,0 +1,1968 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/multi_index_transform_helper.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp" +#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp" +#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp" +#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp" +#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp" + +#define DEBUG_LOG 0 + +namespace ck { + +// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same +// kernel function Blockers: +// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on +// two lds chunks. +// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds +// buffer when we declare __shared__ inside blkgemmpipe +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_b_preshuffle(typename GridwiseGemm::Argument karg) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__)) + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z); + + GridwiseGemm::template Run( + karg.p_a_grid + splitk_batch_offset.a_k_split_offset, + karg.p_b_grid + splitk_batch_offset.b_k_split_offset, + karg.p_ds_grid, + karg.p_c_grid, + p_shared, + karg, + karg.a_element_op, + karg.b_element_op, + karg.c_element_op); +#else + ignore = karg; +#endif // end of if (defined(__gfx9__)) +} + +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_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()]; + + auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z); + + GridwiseGemm::template Run_2Lds( + karg.p_a_grid + splitk_batch_offset.a_k_split_offset, + karg.p_b_grid + splitk_batch_offset.b_k_split_offset, + karg.p_ds_grid, + karg.p_c_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(__gfx9__)) +} + +template +struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle +{ + 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(Block2CTileMapDefault::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&) + { + 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_, + 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_}, + 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 AElementwiseOperation a_element_op; + const BElementwiseOperation b_element_op; + const CElementwiseOperation c_element_op; + }; + + struct SplitKBatchOffset + { + __device__ SplitKBatchOffset(Argument& karg, index_t k_id) + { + if constexpr(is_same_v) + { + a_k_split_offset = k_id * karg.KRead; + } + else if constexpr(is_same_v) + { + a_k_split_offset = k_id * karg.KRead * karg.StrideA; + } + + if constexpr(is_same_v) + { + b_k_split_offset = k_id * karg.KRead * karg.StrideB; + } + else if constexpr(is_same_v) + { + // KPack * NLane * KLane * K0 * N0 + b_k_split_offset = k_id * karg.KRead * NLane; + } + + if(k_id < 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 +#if 1 + const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value); + + if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages) + { + return false; + } +#endif + // 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 Block2CTileMapDefault = 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, + void* p_shared, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4}; + Run( + p_a_grid, + p_b_grid, + p_ds_grid, + p_c_grid, + p_shared, + problem, + a_element_op, + b_element_op, + c_element_op, + block_2_ctile_map); + } + + template + __device__ static void Run(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + void* p_shared, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op, + const Block2CTileMap& block_2_ctile_map) + { + 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 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 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); + + // N0, K0, Blocksize*KPack + 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 + // dummy + 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); + + 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_blockwise_copy, + b_grid_buf, + b_block_buf, + b_block_slice_copy_step, + c_thread_buf, + 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); + + // 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(); + + 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_m3_m4_n2 = 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)), + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // N0 (NXdlPerWave) per shuffle + N1, // N1 = NWave + N2))), // N2 = NPerXdl + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple( + Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 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_m3_m4_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(Sequence<0, 1, 2>{}), + make_tuple(Sequence<0>{})); + + const auto n_thread_data_on_block_idx = + n_thread_data_on_block_to_n0_n1_n2_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_m3_m4_n2, + 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{}}; + + 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}; + + // space filling curve for threadwise C in VGPR + 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_m3_m4_n2, + sfc_c_vgpr.GetIndexTupleOfNumber(access_id), + c_thread_buf, + c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, + 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, + void* p_shared, + void* p_shared1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4}; + Run_2Lds( + p_a_grid, + p_b_grid, + p_ds_grid, + p_c_grid, + p_shared, + p_shared1, + problem, + a_element_op, + b_element_op, + c_element_op, + block_2_ctile_map); + } + + template + __device__ static void Run_2Lds(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + void* p_shared, + void* p_shared1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op, + const Block2CTileMap& block_2_ctile_map) + { + 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 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 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); + + // N0, K0, Blocksize*KPack + 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 + // dummy + 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, + 2>( + 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); + + blockwise_gemm_pipeline.template Run(a_grid_desc_ak0_m_ak1, + a_block_desc_ak0_m_ak1, + a_blockwise_copy, + a_grid_buf, + a_block_bufs, + a_block_slice_copy_step, + b_grid_desc_bpreshuffled, + b_blockwise_copy, + b_grid_buf, + b_block_bufs, + b_block_slice_copy_step, + c_thread_buf, + num_k_block_main_loop); + + // shuffle C and write out + { + static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && + NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, + "wrong!"); + + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + + // 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(); + + 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_m3_m4_n2 = 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)), + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // N0 (NXdlPerWave) per shuffle + N1, // N1 = NWave + N2))), // N2 = NPerXdl + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple( + Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 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_m3_m4_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(Sequence<0, 1, 2>{}), + make_tuple(Sequence<0>{})); + + const auto n_thread_data_on_block_idx = + n_thread_data_on_block_to_n0_n1_n2_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_m3_m4_n2, + 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{}}; + + 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}; + + // space filling curve for threadwise C in VGPR + 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_m3_m4_n2, + sfc_c_vgpr.GetIndexTupleOfNumber(access_id), + c_thread_buf, + c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, + 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 diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp index baf14b2575..7ccea96dda 100644 --- a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp +++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp @@ -352,6 +352,14 @@ struct ThreadwiseTensorSliceTransfer_v3r1 } } + template + __device__ constexpr auto + GetSrcThreadScratchIdx(Number thread_scratch_id = Number{}) + { + using vector_t = typename vector_type_maker::type::type; + return src_thread_scratch_tuple_(thread_scratch_id).template GetAsType(SeqIdx{}); + } + template __device__ void TransferDataFromSrcThreadScratchToDstThreadScratch(Number thread_scratch_id) diff --git a/include/ck/utility/blkgemmpipe_scheduler.hpp b/include/ck/utility/blkgemmpipe_scheduler.hpp index 86dcb6c157..574be5c4ad 100644 --- a/include/ck/utility/blkgemmpipe_scheduler.hpp +++ b/include/ck/utility/blkgemmpipe_scheduler.hpp @@ -8,6 +8,19 @@ namespace ck { +enum struct BlockGemmPipelineVersion +{ + // For GEMM + v1, // Naive + v2, // Mem + v3, // Comp + v4, // Comp, double lds buffer + v5, // Comp, double global prefetch register buffer + + // For GEMM with preshuffled weight + // v1, single lds buffer + // v2, double lds buffer +}; enum struct BlockGemmPipelineScheduler { Intrawave, diff --git a/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_one_pass.hpp b/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_one_pass.hpp index 24f35d3636..64e5224780 100644 --- a/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_one_pass.hpp +++ b/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_one_pass.hpp @@ -157,7 +157,7 @@ struct AddRmsnorm2dRdquantFwdPipelineOnePass sweep_tile(qy, [&, yscale_ = yscale](auto idx) { constexpr auto i_idx = make_tuple(idx[number<0>{}]); auto qy_ = y[idx] / yscale_[i_idx]; - qy(idx) = saturates{}(qy_); + qy(idx) = type_convert(saturates{}(qy_)); }); store_tile(qy_window, qy); } diff --git a/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp b/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp index aec7368e27..ecd4e81b22 100644 --- a/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp +++ b/include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp @@ -260,7 +260,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass const auto x_ = type_convert(x[idx]); auto y_ = x_ * inv_rms[i_idx] * gamma_; auto qy_ = y_ / yscale[i_idx]; - qy(idx) = saturates{}(qy_); + qy(idx) = type_convert(saturates{}(qy_)); }); store_tile(qy_window, qy); diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp index 2815c7a8c4..58f8ed3c3c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp @@ -18,7 +18,7 @@ namespace device { namespace instance { #ifdef CK_ENABLE_FP8 #ifdef CK_ENABLE_BF16 -void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part1( std::vector, @@ -31,7 +31,85 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_inst PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( std::vector, @@ -177,81 +255,239 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_kpadding_in #endif #endif -#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8)) -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances( +#ifdef CK_ENABLE_FP16 +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part1( std::vector, Row, - I8, - I8, + F8, + F8, Tuple, - BF16, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part1( std::vector, Row, - I8, - I8, + F8, + F8, Tuple, - BF16, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part2( std::vector, Row, - I8, - I8, + F8, + F8, Tuple, - BF16, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part2( std::vector, Row, - I8, - I8, + F8, + F8, Tuple, - BF16, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part1( std::vector, Row, - I8, - I8, + F8, + F8, Tuple, - BF16, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); -void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_kpadding_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_kpadding_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); +#endif + +#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_INT8)) +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instances( std::vector, Row, I8, I8, - Tuple, - BF16, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances); + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, PassThrough, PassThrough, MultiplyMultiply>>>& instances); @@ -261,6 +497,7 @@ void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_i template @@ -271,7 +508,7 @@ struct DeviceOperationInstanceFactory, + DsDataType, CDataType, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, @@ -284,7 +521,7 @@ struct DeviceOperationInstanceFactory, + DsDataType, CDataType, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, @@ -302,9 +539,22 @@ struct DeviceOperationInstanceFactory && is_same_v && is_same_v) { - add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances( + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part1( op_ptrs); - add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances( + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part2( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part2( + op_ptrs); + + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( op_ptrs); add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances( @@ -326,9 +576,22 @@ struct DeviceOperationInstanceFactory && is_same_v && is_same_v) { - add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances( + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part1( op_ptrs); - add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances( + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part2( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part2( + op_ptrs); + + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + op_ptrs); + add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( op_ptrs); add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instances( @@ -344,26 +607,26 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v) + is_same_v) { if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instances( op_ptrs); - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instances( op_ptrs); - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instances( op_ptrs); - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instances( op_ptrs); - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instances( op_ptrs); - add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances( + add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instances( op_ptrs); } } diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.hpp new file mode 100644 index 0000000000..90da3ad0fe --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.hpp @@ -0,0 +1,317 @@ +// 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_b_preshuffle.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_F16) || defined(CK_ENABLE_FP8)) +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); +#endif + +#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8)) +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& + instances); +#endif + +template +struct DeviceOperationInstanceFactory< + ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle< + ALayout, + BLayout, + Tuple, + CLayout, + ADataType, + BDataType, + Tuple, + CDataType, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::MultiplyMultiply>> +{ + using DeviceOp = + DeviceGemmMultipleDSplitKBPreShuffle, + CLayout, + ADataType, + BDataType, + Tuple, + CDataType, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::PassThrough, + ck::tensor_operation::element_wise::MultiplyMultiply>; + + static auto GetInstances() + { + std::vector> op_ptrs; +// TODO: Add MFMA layout into tensor layout +#if(defined(CK_ENABLE_F16) || 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_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances( + op_ptrs); + + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2( + op_ptrs); + + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances( + op_ptrs); + } + } +#endif + +#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_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances( + op_ptrs); + + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2( + op_ptrs); + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2( + op_ptrs); + + add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_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_multiply_multiply/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt index 3b63c2c160..57bae7a2ac 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt @@ -2,35 +2,59 @@ set(GEMM_MULTIPLY_MULTIPLY_INSTANCES) list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES - device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp - device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp - device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance.cpp - device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp + device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instance.cpp - device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instance.cpp + device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp ) -set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") -set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") -set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") -set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") -set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") -set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") add_instance_library(device_gemm_multiply_multiply_instance ${GEMM_MULTIPLY_MULTIPLY_INSTANCES}) diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp index 3d0e0a0634..17da9ce5b7 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp @@ -35,7 +35,7 @@ static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; template -using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple< +using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1 = std::tuple< // clang-format off //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| @@ -44,9 +44,6 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std #if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) // Compute friendly DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 224, 128, 16, 16, 32, 32, 2, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 32, 32, 4, 3, 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>, @@ -62,6 +59,20 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 32, 32, 1, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part2 = std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 3, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, @@ -88,28 +99,85 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 192, 256, 16, 16, 16, 16, 1, 3, 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, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, 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, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; +template +using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part1 = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + // 256x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 16, 16, 8, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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> + // 224x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 192x[64, 256, 32]x128, 192x[64]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 192, 128, 16, 16, 16, 16, 6, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 256, 16, 16, 16, 16, 6, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> #endif - // clang-format on - >; + // clang-format on + >; + +template +using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part2 = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + + // 160x[64, 256, 32]x128, 160x[64, 96, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 192, 128, 16, 16, 16, 16, 5, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 256, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 256, 16, 16, 16, 16, 5, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 128x[64, 256, 32]x128, 128x[64, 128, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; template using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple< @@ -144,16 +212,18 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std: DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 32, 16, 512, 16, 16, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 2, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 16, 32, 512, 16, 16, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 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, 16>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 16, 1, 16>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + + // v3 more DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<8, 16, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> + +// clang-format on #endif - // clang-format on >; } // namespace instance } // namespace device diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp similarity index 95% rename from library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp rename to library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp index 6527d93473..f0e0d3d689 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part1( std::vector, @@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_inst { add_device_operation_instances( instances, - device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances{}); + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part2.cpp similarity index 95% rename from library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp rename to library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part2.cpp index 7f16a7a2c5..9fec958841 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part2.cpp @@ -8,7 +8,7 @@ namespace tensor_operation { namespace device { namespace instance { -void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances( +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances_part2( std::vector, @@ -23,7 +23,7 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_ins { add_device_operation_instances( instances, - device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances{}); + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part2{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp new file mode 100644 index 0000000000..686dcb67e1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part2.cpp new file mode 100644 index 0000000000..e9d1183199 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part2.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part2{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp new file mode 100644 index 0000000000..687b95f07a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part1< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp new file mode 100644 index 0000000000..3f2ffbf231 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part2< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp new file mode 100644 index 0000000000..9b6ec0fdba --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part1< + GemmKPadding>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp new file mode 100644 index 0000000000..e103ecfce1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mfma16x16_instances_part2< + GemmKPadding>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp index 1730eba3cc..5c854ee5d9 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp @@ -35,7 +35,7 @@ static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; template -using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances = std::tuple< +using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part1 = std::tuple< // clang-format off //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| @@ -44,9 +44,6 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances = std: #if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) // Compute friendly DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 224, 128, 16, 16, 32, 32, 2, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 32, 32, 4, 3, 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>, @@ -61,6 +58,20 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances = std: DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 32, 32, 1, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part2 = std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 128, 16, 16, 32, 32, 1, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 224, 128, 16, 16, 16, 16, 2, 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>, @@ -80,28 +91,85 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances = std: DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 192, 256, 16, 16, 16, 16, 1, 3, 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, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, 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, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; +template +using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part1 = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + // 256x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 16, 16, 8, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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> + // 224x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 192x[64, 256, 32]x128, 192x[64]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 192, 128, 16, 16, 16, 16, 6, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 256, 16, 16, 16, 16, 6, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> #endif - // clang-format on - >; + // clang-format on + >; + +template +using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part2 = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + + // 160x[64, 256, 32]x128, 160x[64, 96, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 192, 128, 16, 16, 16, 16, 5, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 256, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 256, 16, 16, 16, 16, 5, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 128x[64, 256, 32]x128, 128x[64, 128, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; template using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_instances = std::tuple< @@ -136,16 +204,18 @@ using device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_mem_instances = std:: DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 32, 16, 512, 16, 16, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 2, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 128, 16, 32, 512, 16, 16, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 4, 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_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 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, 16>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, - DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 16, 1, 16>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, + + // v3 more DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<8, 16, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> + +// clang-format on #endif - // clang-format on >; } // namespace instance } // namespace device diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp new file mode 100644 index 0000000000..463bbd1570 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part1.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part1{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part2.cpp new file mode 100644 index 0000000000..a78ba4692b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instance_part2.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part2{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp new file mode 100644 index 0000000000..f937a607d6 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part1.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part1{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part2.cpp new file mode 100644 index 0000000000..fa2b5ba6e7 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instance_part2.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_instances_part2{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp new file mode 100644 index 0000000000..38b19178dc --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part1.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part1< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp new file mode 100644 index 0000000000..95fda4fbfb --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instance_part2.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_default_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part2< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp new file mode 100644 index 0000000000..4a86e2e8d5 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part1.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part1( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part1< + GemmKPadding>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp new file mode 100644 index 0000000000..19c53e6287 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_f16/device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instance_part2.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_kpadding_instances_part2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_f8_f8_f16_mk_nk_mn_comp_mfma16x16_instances_part2< + GemmKPadding>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp new file mode 100644 index 0000000000..04d44ad6b4 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp @@ -0,0 +1,120 @@ +// 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.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using I8 = int8_t; +using I32 = int; +using F16 = half_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 MultiplyMultiply = element_wise::MultiplyMultiply; + +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; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_instances = std::tuple< + // clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + + // Compute friendly + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 224, 128, 16, 16, 32, 32, 2, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 32, 32, 4, 3, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 32, 32, 2, 5, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 32, 32, 2, 3, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 32, 32, 1, 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, 1, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 32, 32, 2, 3, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 32, 32, 1, 5, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 32, 32, 1, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 128, 16, 16, 32, 32, 1, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 224, 128, 16, 16, 16, 16, 2, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 192, 256, 16, 16, 32, 32, 1, 3, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 192, 128, 16, 16, 32, 32, 1, 3, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 160, 256, 16, 16, 16, 16, 2, 5, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 256, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 96, 256, 16, 16, 16, 16, 2, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 64, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 224, 256, 16, 16, 16, 16, 1, 7, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 192, 256, 16, 16, 16, 16, 1, 6, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 160, 256, 16, 16, 16, 16, 1, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 96, 256, 16, 16, 16, 16, 1, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 64, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 192, 256, 16, 16, 16, 16, 1, 3, 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, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, I8> + + // clang-format oI + >; + +template +using device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_instances = std::tuple< + // clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + + // Latency friendly + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + // Memory friendly + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 32, 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<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 16, 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, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 64, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 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, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 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, 16>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, I8>, + DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple, Row, I8, I8, Tuple, F16, I32, I32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 16, 1, 16>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, I8> + // clang-format oI + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..1162b7a110 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instance.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_default_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp new file mode 100644 index 0000000000..ac2994e868 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instance.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instance.cpp new file mode 100644 index 0000000000..5fb2d06543 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_default_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp new file mode 100644 index 0000000000..e948057284 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v1_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instance.cpp new file mode 100644 index 0000000000..82cc25f4eb --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_default_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp new file mode 100644 index 0000000000..8c3c2d2c58 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_f16/device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_v2_kpadding_instances( + std::vector, + Row, + I8, + I8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_xdl_i8_i8_f16_mk_nk_mn_mem_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/CMakeLists.txt new file mode 100644 index 0000000000..943b2bf4c7 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/CMakeLists.txt @@ -0,0 +1,42 @@ +# ONLY XDL_KERNELS +set(GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES) + +list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp + + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp + ) + +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + +add_instance_library(device_gemm_multiply_multiply_weight_preshuffle_instance ${GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES}) diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp new file mode 100644 index 0000000000..71383f5dc1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp new file mode 100644 index 0000000000..a138452295 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp @@ -0,0 +1,195 @@ +// 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_b_preshuffle.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 MultiplyMultiply = element_wise::MultiplyMultiply; + +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; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +static constexpr auto v1 = BlockGemmPipelineVersion::v1; +static constexpr auto v2 = BlockGemmPipelineVersion::v2; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + // N 256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + // N 512 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 128, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 128, 16, 16, 32, 32, 1, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 512, 16, 16, 32, 32, 2, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + // N 256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 256, 16, 16, 32, 32, 4, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + // N 512 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances = + std::tuple< +// clang-format off + //############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 32, 32, 7, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 32, 32, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 32, 32, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 32, 32, 7, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 32, 32, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_instances = + std::tuple< +// clang-format off + //############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + // 256x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 16, 16, 8, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 224x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 192x[64, 256, 32]x128, 192x[64]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 192, 128, 16, 16, 16, 16, 6, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 160x[64, 256, 32]x128, 160x[64, 96, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 192, 128, 16, 16, 16, 16, 5, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 128x[64, 256, 32]x128, 128x[64, 128, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp new file mode 100644 index 0000000000..e3ff079d99 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_compute_instances< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp new file mode 100644 index 0000000000..6e9b3ea172 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp new file mode 100644 index 0000000000..cc543b19c1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp new file mode 100644 index 0000000000..8557d0c80e --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp new file mode 100644 index 0000000000..9fcce478e7 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp new file mode 100644 index 0000000000..84c2c70e35 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp new file mode 100644 index 0000000000..0933b1fe18 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + BF16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp new file mode 100644 index 0000000000..05529f9cdd --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp new file mode 100644 index 0000000000..c4f53e834a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp @@ -0,0 +1,195 @@ +// 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_b_preshuffle.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 F16 = half_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 MultiplyMultiply = element_wise::MultiplyMultiply; + +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; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +static constexpr auto v1 = BlockGemmPipelineVersion::v1; +static constexpr auto v2 = BlockGemmPipelineVersion::v2; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + // N 256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + // N 512 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 128, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 128, 16, 16, 32, 32, 1, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 32, 32, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 128, 512, 16, 16, 32, 32, 2, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 128, 512, 16, 16, 32, 32, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + // N 256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 256, 16, 16, 32, 32, 4, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 256, 16, 16, 32, 32, 2, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 256, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 32, 32, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 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, BlkGemmPipeVer, F8>, + // N 512 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 32, 32, 2, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 32, 32, 1, 4, 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>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances = + std::tuple< +// clang-format off + //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 256, 16, 16, 16, 16, 1, 4, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 512, 256, 16, 16, 16, 16, 1, 8, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 64, 512, 16, 16, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 16>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 128, 512, 16, 16, 16, 16, 1, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 16, 256, 512, 16, 16, 16, 16, 1, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 16, 1, 16>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances = + std::tuple< +// clang-format off + //############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 32, 32, 7, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 32, 32, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 32, 32, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 32, 32, 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, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 32, 32, 7, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 32, 32, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 32, 32, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +template +using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_instances = + std::tuple< +// clang-format off + //############################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| + //############################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| + //############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| + //############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | +#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) + // Compute friendly + // 256x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 192, 128, 16, 16, 16, 16, 8, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 160, 128, 16, 16, 16, 16, 8, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 96, 128, 16, 16, 16, 16, 8, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 64, 128, 16, 16, 16, 16, 8, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 224x[64, 256, 32]x128 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 224, 128, 16, 16, 16, 16, 7, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 192, 128, 16, 16, 16, 16, 7, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 160, 128, 16, 16, 16, 16, 7, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 128, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 96, 128, 16, 16, 16, 16, 7, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 64, 128, 16, 16, 16, 16, 7, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 192x[64, 256, 32]x128, 192x[64]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 256, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 224, 128, 16, 16, 16, 16, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 192, 128, 16, 16, 16, 16, 6, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 160, 128, 16, 16, 16, 16, 6, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 128, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 96, 128, 16, 16, 16, 16, 6, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 192, 64, 128, 16, 16, 16, 16, 6, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 160x[64, 256, 32]x128, 160x[64, 96, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 256, 128, 16, 16, 16, 16, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 224, 128, 16, 16, 16, 16, 5, 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, 1, 1, S<1, 32, 1, 8>, S<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 192, 128, 16, 16, 16, 16, 5, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 160, 128, 16, 16, 16, 16, 5, 5, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 128, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 96, 128, 16, 16, 16, 16, 5, 3, 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<4, 4, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 160, 64, 128, 16, 16, 16, 16, 5, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + // 128x[64, 256, 32]x128, 128x[64, 128, 32]x256 + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 224, 128, 16, 16, 16, 16, 4, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 192, 128, 16, 16, 16, 16, 4, 6, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 160, 128, 16, 16, 16, 16, 4, 5, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 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>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 128, 16, 16, 16, 16, 4, 3, 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_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 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, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 256, 16, 16, 16, 16, 4, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 96, 256, 16, 16, 16, 16, 4, 3, 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, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple, Row, F8, F8, Tuple, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 64, 256, 16, 16, 16, 16, 4, 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, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> +#endif + // clang-format on + >; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp new file mode 100644 index 0000000000..c123a0fdd8 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_compute_instances< + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp new file mode 100644 index 0000000000..cb15688e6f --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp new file mode 100644 index 0000000000..c5a8448b59 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp new file mode 100644 index 0000000000..c9ab9c1071 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp new file mode 100644 index 0000000000..bb83bacb35 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp new file mode 100644 index 0000000000..fb43347ceb --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances< + v1, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp new file mode 100644 index 0000000000..c8ff03d6ef --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp @@ -0,0 +1,34 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances_v2( + std::vector, + Row, + F8, + F8, + Tuple, + F16, + PassThrough, + PassThrough, + MultiplyMultiply>>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances< + v2, + GemmDefault>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp index b57888f133..0619a98cf0 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp @@ -70,11 +70,10 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_instances = std::tup #if defined(CK_USE_AMD_MFMA_GFX950) #endif // defined(CK_USE_AMD_MFMA_GFX950) // Latency friendly - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 64, 128, 8, 4, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp index 3848187540..3edbd28cd8 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp @@ -57,6 +57,7 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_instances = std::tu // AGPR Spill when use permuted lds layout. so, use padding for these two. DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 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, 8, 8, 0, 1, 2, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 64, 8, 8, 16, 16, 8, 7, S<8, 32, 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, 8, 8, 0, 2, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 192, 64, 8, 8, 32, 32, 4, 3, S<8, 32, 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, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, #endif // !defined(CK_USE_AMD_MFMA_GFX950) DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 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, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 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, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, @@ -74,11 +75,12 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances = std::tup #if defined(CK_USE_AMD_MFMA_GFX950) #endif // defined(CK_USE_AMD_MFMA_GFX950) // Latency friendly - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 8, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 64, 128, 8, 8, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 8, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 4, 4, 16, 16, 4, 1, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp index b844a5c804..e3f3afff46 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp @@ -53,7 +53,6 @@ using device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 32, 8, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 32, 8, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 32, 32, 2, 2, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> @@ -70,15 +69,10 @@ using device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple< #if defined(CK_USE_AMD_MFMA_GFX950) #endif // defined(CK_USE_AMD_MFMA_GFX950) // Latency friendly - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 4, 4, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<32, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 2, 2, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 64, 128, 8, 4, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 64, 8, 2, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 64, 2, 2, 32, 32, 2, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn.hpp index ebb5e76b8a..e39c9a63b9 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn.hpp @@ -82,14 +82,11 @@ using device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_instances = std::tuple< #if defined(CK_USE_AMD_MFMA_GFX950) #endif // defined(CK_USE_AMD_MFMA_GFX950) // Latency friendly - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 64, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 8, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 64, 128, 8, 8, 16, 16, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, // Memory friendly DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 64, 8, 8, 32, 32, 2, 1, S<8, 32, 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, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 64, 4, 4, 32, 32, 2, 1, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp index 1a9756cea5..6388c13444 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp @@ -51,6 +51,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple< DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, 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>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, #endif // !defined(CK_USE_AMD_MFMA_GFX950) DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, 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>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, + DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 192, 128, 16, 16, 32, 32, 4, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, 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>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, 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>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1, F8>, diff --git a/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp index 29a645e9d7..dbfddeb8a4 100644 --- a/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp +++ b/profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp @@ -188,7 +188,8 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, // profile device GEMM instances for(auto& op_ptr : op_ptrs) { - std::vector kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38}; + // Seems like when performance measurement has bug when spiltK is large + std::vector kbatch_list = {1, 2, 4, 8, 16}; if(KBatch > 0) { @@ -232,7 +233,26 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, { c_device_buf.FromDevice(e_m_n_device_result.mData.data()); - pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result); +#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8 + // set softer tolerances for fp8 + if constexpr((is_same_v || is_same_v || + is_same_v) || + (is_same_v || is_same_v || + is_same_v)) + { + std::string msg = "Error: Incorrect results!"; + double rtol = 1e-1; + double atol = 1e-1; + 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 || defined CK_ENABLE_INT8 + } +#endif if(do_log) { @@ -249,6 +269,10 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, std::string op_name = op_ptr->GetTypeString(); + // timer of develop branch should only apply to empty hipstream + // hipStream_t stream; + // hip_check_error(hipStreamCreate(&stream)); + float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel, @@ -271,27 +295,6 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch " << kbatch_curr << std::endl; -#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8 - // set softer tolerances for fp8 - if constexpr((is_same_v || is_same_v || - is_same_v) || - (is_same_v || is_same_v || - is_same_v)) - { - std::string msg = "Error: Incorrect results!"; - double rtol = 1e-1; - double atol = 1e-1; - 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 || defined CK_ENABLE_INT8 - } -#endif - if(tflops > best_tflops && ave_time > 1e-10) { best_op_name = op_name; diff --git a/profiler/include/profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp new file mode 100644 index 0000000000..177e652cc3 --- /dev/null +++ b/profiler/include/profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp @@ -0,0 +1,396 @@ +// 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_b_preshuffle.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.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_multiply_multiply_weight_preshuffle_impl(int do_verification, + int init_method, + bool do_log, + bool time_kernel, + int M, + int N, + int K, + int StrideA, + int StrideB, + int StrideD0, + int StrideD1, + int StrideE, + int KBatch, + 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}); + } + }; + + Tensor a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); + Tensor b_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 d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{})); + Tensor d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{})); + 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 = + a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes() + + d0_m_n.GetElementSpaceSizeInBytes() + d1_m_n.GetElementSpaceSizeInBytes(); + int rotating_count = std::max( + 1, + std::min(n_iter, + static_cast(std::ceil(static_cast(rotating) / total_gemm_needed)))); + + std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; + std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; + std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl; + std::cout << "d1_m_n: " << d1_m_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: + a_m_k.GenerateTensorValue(GeneratorTensor_2{-1, 2}); + b_k_n.GenerateTensorValue(GeneratorTensor_2{-1, 2}); + d0_m_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + d1_m_n.GenerateTensorValue(GeneratorTensor_2{-1, 1}); + break; + default: + a_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + b_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + d0_m_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + d1_m_n.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + } + + preShuffleBuffer(b_k_n.mData.data(), b_preshuffled_mfma16.mData.data(), N, K, 16); + preShuffleBuffer(b_k_n.mData.data(), b_preshuffled_mfma32.mData.data(), N, K, 32); + + using PassThrough = ck::tensor_operation::element_wise::PassThrough; + using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply; + + using AElementOp = PassThrough; + using BElementOp = PassThrough; + using CElementOp = MultiplyMultiply; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto c_element_op = CElementOp{}; + + DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b_device_buf_mfma16(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); + DeviceMem b_device_buf_mfma32(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); + DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize()); + DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize()); + DeviceMem c_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); + + a_device_buf.ToDevice(a_m_k.mData.data()); + b_device_buf_mfma16.ToDevice(b_preshuffled_mfma16.mData.data()); + b_device_buf_mfma32.ToDevice(b_preshuffled_mfma32.mData.data()); + + d0_device_buf.ToDevice(d0_m_n.mData.data()); + d1_device_buf.ToDevice(d1_m_n.mData.data()); + + using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle< + ALayout, + BLayout, + ck::Tuple, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + 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}); + + 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) + { + c_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n)); + } + } + } + + std::string best_op_name; + float best_ave_time = 0; + float best_tflops = 0; + float best_gb_per_sec = 0; + float best_kbatch = 0; + + // profile device GEMM instances + for(auto& op_ptr : op_ptrs) + { + int NPerXdl = op_ptr->GetPreShuffleParameters(); + + std::vector kbatch_list = {1, 2, 4, 8}; + + if(KBatch > 0) + { + kbatch_list = {KBatch}; + } + + for(std::size_t i = 0; i < kbatch_list.size(); i++) + { + auto kbatch_curr = kbatch_list[i]; + + auto argument_ptr = op_ptr->MakeArgumentPointer( + static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(NPerXdl == 16 ? b_device_buf_mfma16.GetDeviceBuffer() + : b_device_buf_mfma32.GetDeviceBuffer()), + std::array{d0_device_buf.GetDeviceBuffer(), + d1_device_buf.GetDeviceBuffer()}, + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + std::array{StrideD0, StrideD1}, + StrideE, + kbatch_curr, + a_element_op, + b_element_op, + c_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + c_device_buf.SetZero(); + + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false}); + + if(do_verification) + { + c_device_buf.FromDevice(e_m_n_device_result.mData.data()); + +#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8 + // set softer tolerances for fp8 + if constexpr((is_same_v || is_same_v || + is_same_v) || + (is_same_v || is_same_v || + is_same_v)) + { + std::string msg = "Error: Incorrect results!"; + double rtol = 1e-3; + 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 || defined CK_ENABLE_INT8 + } +#endif + + if(do_log) + { + LogRangeAsType(std::cout << "a : ", a_m_k.mData, ",") << std::endl; + LogRangeAsType(std::cout << "b: ", b_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(ADataType) * M * K + sizeof(BDataType) * 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 << ", KBatch " + << kbatch_curr << std::endl; + + if(tflops > best_tflops && ave_time > 1e-10) + { + best_op_name = op_name; + best_tflops = tflops; + best_ave_time = ave_time; + best_gb_per_sec = gb_per_sec; + best_kbatch = kbatch_curr; + } + } + else + { + std::cout << op_ptr->GetTypeString() << " does not support this problem" + << std::endl; + } + } + } + + if constexpr(is_same::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 << " KBatch = " << best_kbatch + << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec + << " GB/s, " << best_op_name << std::endl; + + return pass; +} + +} // namespace profiler +} // namespace ck diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index d2d7753fb4..5ed28b9826 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -50,6 +50,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") 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) @@ -139,6 +140,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") 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) diff --git a/profiler/src/profile_gemm_multiply_multiply.cpp b/profiler/src/profile_gemm_multiply_multiply.cpp index 24c8630594..ad2bb77544 100644 --- a/profiler/src/profile_gemm_multiply_multiply.cpp +++ b/profiler/src/profile_gemm_multiply_multiply.cpp @@ -28,7 +28,8 @@ enum struct GemmDataType F16_F16_F16_F8, // 6 F8_F8_BF16, // 7 INT8_INT8_BF16, // 8 - F8_F8_F16, // 9 + INT8_INT8_F16, // 9 + F8_F8_F16, // 10 }; #define OP_NAME "gemm_multiply_multiply" @@ -177,6 +178,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[]) return profile( I8{}, I8{}, I8{}, I32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{}); } + else if(data_type == GemmDataType::INT8_INT8_F16 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile( + I8{}, I8{}, I8{}, I32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Col{}, Row{}); + } else { std::cout << "this data_type & layout is not implemented" << std::endl; diff --git a/profiler/src/profile_gemm_multiply_multiply_weight_preshuffle.cpp b/profiler/src/profile_gemm_multiply_multiply_weight_preshuffle.cpp new file mode 100644 index 0000000000..ee3be398e5 --- /dev/null +++ b/profiler/src/profile_gemm_multiply_multiply_weight_preshuffle.cpp @@ -0,0 +1,167 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp" +#include "profiler_operation_registry.hpp" + +enum struct GemmMatrixLayout +{ + MK_MFMA_MN, // 0 +}; + +enum struct GemmDataType +{ + F8_F8_F16, // 0 + F8_F8_BF16, // 1 +}; + +#define OP_NAME "gemm_multiply_multiply_weight_preshuffle" +#define OP_DESC "GEMM_Multiply_Multiply_Weight_PreShuffle" + +int profile_gemm_multiply_multiply_weight_preshuffle(int argc, char* argv[]) +{ + if(argc != 16 && argc != 20) + { + printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); + printf("arg2: data type (0: f8->f16; 1: f8->bf16;\n"); + printf("arg3: matrix layout (0: A[m, k] * B[MFMA] = C[m, n];\n"); + printf("arg4: verification (0: no; 1: yes)\n"); + printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"); + printf("arg6: print tensor value (0: no; 1: yes)\n"); + printf("arg7: time kernel (0=no, 1=yes)\n"); + printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n"); + printf("optional:\n"); + printf("arg16: number of kbatch (default 1)\n"); + printf("arg17: number of warm-up cycles (default 1)\n"); + printf("arg18: number of iterations (default 10)\n"); + printf("arg19: memory for rotating buffer (default 0, size in MB)\n"); + exit(1); + } + + const auto data_type = static_cast(std::stoi(argv[2])); + const auto layout = static_cast(std::stoi(argv[3])); + const bool do_verification = std::stoi(argv[4]); + const int init_method = std::stoi(argv[5]); + const bool do_log = std::stoi(argv[6]); + const bool time_kernel = std::stoi(argv[7]); + + const int M = std::stoi(argv[8]); + const int N = std::stoi(argv[9]); + const int K = std::stoi(argv[10]); + + const int StrideA = std::stoi(argv[11]); + const int StrideB = std::stoi(argv[12]); + const int StrideD0 = std::stoi(argv[13]); + const int StrideD1 = std::stoi(argv[14]); + const int StrideE = std::stoi(argv[15]); + + int n_warmup = 1; + int n_iter = 10; + uint64_t rotating = 0; + int KBatch = 1; + if(argc == 20) + { + KBatch = std::stoi(argv[16]); + n_warmup = std::stoi(argv[17]); + n_iter = std::stoi(argv[18]); + rotating = std::stoull(argv[19]) * 1024 * 1024; + } + + using F32 = float; + using BF16 = ck::bhalf_t; + using F16 = ck::half_t; + using F8 = ck::f8_t; + // using I8 = int8_t; + // using I32 = int; + + using Row = ck::tensor_layout::gemm::RowMajor; + using Col = ck::tensor_layout::gemm::ColumnMajor; + + auto profile = [&](auto a_type, + auto b_type, + auto comp_type, + auto acc_type, + auto d0_type, + auto d1_type, + auto c_type, + auto a_layout, + auto b_layout, + auto d0_layout, + auto d1_layout, + auto e_layout) { + using ADataType = decltype(a_type); + using BDataType = decltype(b_type); + using ComputeDataType = decltype(comp_type); + using D0DataType = decltype(d0_type); + using D1DataType = decltype(d1_type); + using AccDataType = decltype(acc_type); + using EDataType = decltype(c_type); + + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using D0Layout = decltype(d0_layout); + using D1Layout = decltype(d1_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 DefaultStrideD0 = ck::is_same_v ? N : M; + const int DefaultStrideD1 = ck::is_same_v ? N : M; + const int DefaultStrideE = ck::is_same_v ? N : M; + + bool pass = + ck::profiler::profile_gemm_multiply_multiply_weight_preshuffle_impl( + do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + (StrideA < 0) ? DefaultStrideA : StrideA, + (StrideB < 0) ? DefaultStrideB : StrideB, + (StrideD0 < 0) ? DefaultStrideD0 : StrideD0, + (StrideD1 < 0) ? DefaultStrideD1 : StrideD1, + (StrideE < 0) ? DefaultStrideE : StrideE, + KBatch, + n_warmup, + n_iter, + rotating); + + return pass ? 0 : 1; + }; + if(data_type == GemmDataType::F8_F8_F16 && layout == GemmMatrixLayout::MK_MFMA_MN) + { + return profile( + F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Col{}, Row{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_MFMA_MN) + { + return profile( + F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, 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_multiply_multiply_weight_preshuffle); diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh index 6089fc7a7e..cdf407d6cd 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 \