From ad04db9d00ee9f595285e38f173fc8956bd0d7d5 Mon Sep 17 00:00:00 2001 From: felix Date: Wed, 16 Apr 2025 10:53:21 +0800 Subject: [PATCH] add preshuffle gemm fp16 (#2036) * add preshuffle gemm fp16 * clang format and test ok * Update gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp remove useless comments in example * Update gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp remove 2 --------- Co-authored-by: coderfeli [ROCm/composable_kernel commit: c5975529bb016318ae135431d61761b885f0f5b9] --- .../65_gemm_multiply_multiply/CMakeLists.txt | 1 + ...multiply_multiply_xdl_fp16_bpreshuffle.cpp | 371 ++++++++++++++++++ 2 files changed, 372 insertions(+) create mode 100644 example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index 95fd8bace8..deca85ae64 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -1,6 +1,7 @@ add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp) +add_example_executable(example_gemm_multiply_multiply_xdl_fp16_bpreshuffle gemm_multiply_multiply_xdl_fp16_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) add_example_executable(example_moe_gemm1_xdl_fp8 moe_gemm1_xdl_fp8.cpp) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp new file mode 100644 index 0000000000..69803c7eeb --- /dev/null +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp @@ -0,0 +1,371 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, 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 F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using A0DataType = F16; +using B0DataType = F16; +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 F16* src, F16* dst, int N, int K, int NXdl) +{ + int KPack = 16 / sizeof(F16); + 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 +using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle + // 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| +///###### RCR + // kernel 1: 256->32x128x128 + < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, + AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, + 32, 128, 128, + 8, 8, + 32, 32, + 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>, S<8, 8, 1>, + ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, F16>; +// 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; + + 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 + { + 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"); + 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"); + } + + float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 50, false, 1}); + + 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; +}