From ede7126dc6f791ecbdb273a9335698f342f86053 Mon Sep 17 00:00:00 2001 From: Anton Gorenko Date: Mon, 28 Apr 2025 14:27:36 +0500 Subject: [PATCH] Clone BlockwiseGemmXdlops_pipeline_v1 for WMMA implementation --- .../blockwise_gemm_pipeline_wmmaops_v1.hpp | 731 ++++++++++++++++++ 1 file changed, 731 insertions(+) create mode 100644 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v1.hpp diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v1.hpp new file mode 100644 index 0000000000..e2f9b8432a --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v1.hpp @@ -0,0 +1,731 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp" + +namespace ck { + +// Naive pipeline with lowest resource request per WGP +// GlobalPrefetchStages: 1 +// LocalPreFillStages: 1 +// LocalPreFetchStages: 0 +// LocalSharedMemoryBuffer: 1 + +template +struct BlockwiseGemmXdlops_pipeline_v1 +{ +}; + +template +struct BlockwiseGemmXdlops_pipeline_v1 + : BlockwiseGemmXdlops_pipeline_base + +{ + using Base = BlockwiseGemmXdlops_pipeline_base; + using Base::I0; + using Base::KRepeat; + using Base::xdlops_gemm; + + 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::a_block_desc_m0_m1_m2_k; + using Base::b_block_desc_n0_n1_n2_k; + + using Base::AMmaKStride; + using Base::BMmaKStride; + + static constexpr index_t PrefetchStages = 1; + static constexpr index_t PrefillStages = 1; + static constexpr index_t GlobalBufferNum = 1; + + __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop) + { + return num_loop > PrefetchStages; + } + + __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop) + { + ignore = num_loop; + return TailNumber::Full; + } + + 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, + const BBlockDesc& b_block_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 + { + auto a_thread_buf = make_static_buffer( + a_thread_desc_.GetElementSpaceSize()); + auto b_thread_buf = make_static_buffer( + b_thread_desc_.GetElementSpaceSize()); + + // Global prefetch 1 + 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); + b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); + + // Initialize C + c_thread_buf.Clear(); + + // main body + if constexpr(HasMainLoop) + { + index_t i = 0; + do + { + // ------------------------------------------------------------------------------------------- + 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); + + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](auto k) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k, + make_tuple(m0, I0, I0, Number{}), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, k, I0), + a_thread_buf); + static_for<0, NRepeat, 1>{}([&](auto n0) { + b_thread_copy_.Run(b_block_desc_n0_n1_n2_k, + make_tuple(n0, I0, I0, Number{}), + b_block_buf, + b_thread_desc_, + make_tuple(n0, I0, k, I0), + b_thread_buf); + }); + }); + }); + + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + 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_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(); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf); + b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); + + i += 1; + } while(i < (num_loop - 1)); + } + + // tail + if constexpr(TailNum == TailNumber::Full) + { + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](auto k) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k, + make_tuple(m0, I0, I0, Number{}), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, k, I0), + a_thread_buf); + static_for<0, NRepeat, 1>{}([&](auto n0) { + b_thread_copy_.Run(b_block_desc_n0_n1_n2_k, + make_tuple(n0, I0, I0, Number{}), + b_block_buf, + b_thread_desc_, + make_tuple(n0, I0, k, I0), + b_thread_buf); + }); + }); + }); + + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + 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_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{})); + }); + }); + }); + } + } + + protected: + using Base::a_thread_copy_; + using Base::a_thread_desc_; + using Base::b_thread_copy_; + using Base::b_thread_desc_; + using Base::c_thread_desc_; +}; + +template +struct BlockwiseGemmXdlops_pipeline_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::KPerThread; + using Base::xdlops_gemm; + + 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::a_block_desc_m0_m1_m2_k; + using Base::b_block_desc_n0_n1_n2_k; + + static constexpr index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS; + static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack); + static constexpr index_t KRepeat = KPerThread / KPerInnerLoop; + static constexpr index_t PrefetchStages = 1; + static constexpr index_t PrefillStages = 1; + static constexpr index_t GlobalBufferNum = 1; + __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop) + { + return num_loop > PrefetchStages; + } + + __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop) + { + ignore = num_loop; + return TailNumber::Full; + } + + 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, + const BBlockDesc& b_block_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 + { + auto a_thread_buf = make_static_buffer( + a_thread_desc_.GetElementSpaceSize()); + auto b_thread_buf = make_static_buffer( + b_thread_desc_.GetElementSpaceSize()); + + // Global prefetch 1 + 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); + b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); + + // Initialize C + c_thread_buf.Clear(); + + // main body + if constexpr(HasMainLoop) + { + index_t i = 0; + do + { + // ------------------------------------------------------------------------------------------- + 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); + + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k, + make_tuple(m0, I0, I0, Number{}), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, k0, I0), + a_thread_buf); + static_for<0, NRepeat, 1>{}([&](auto n0) { + b_thread_copy_.Run(b_block_desc_n0_n1_n2_k, + make_tuple(n0, I0, I0, Number{}), + b_block_buf, + b_thread_desc_, + make_tuple(n0, I0, k0, I0), + b_thread_buf); + }); + }); + __builtin_amdgcn_sched_barrier(0); + // NOTE: Synchronize threads in a workgroup at the start of each MAC cluster, + // but except the first, as we can shorten non-MAC cluster a bit and there's no + // observable negative impact. The desired effect is waves in a workgroup + // executing MAC in sync. This avoids some out-of-sync waves hijacking MAC + // resource from other workgroups and reducing the chance of latency hiding by + // waiting for the rest of the workgroup at the eventual sync point. + if constexpr(k0.value != 0 || KRepeat == 1) + { + __builtin_amdgcn_s_barrier(); + __builtin_amdgcn_sched_barrier(0); + } + static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + 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_buf[Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + // The block_sync_lds() here performs double duty: + // A) safeguard against data hazard because barrier from + // blockwise_gemm is moved here B) reduce VMEM FIFO congestion by + // applying small delays to different wavefronts It is performed + // near the end of MAC cluster to minimize lgkmcnt penalty + if constexpr(k0.value == KRepeat - 1 && + k_.value == KPerInnerLoop - KPack && + m0.value == MRepeat - 1 && n0.value == NRepeat - 1) + { + __builtin_amdgcn_sched_barrier(0); + block_sync_lds(); + __builtin_amdgcn_sched_barrier(0); + } + xdlops_gemm.Run( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0) + { + __builtin_amdgcn_sched_barrier(0); + __builtin_amdgcn_s_setprio(1); + __builtin_amdgcn_sched_barrier(0); + } + }); + }); + }); + __builtin_amdgcn_sched_barrier(0); + __builtin_amdgcn_s_setprio(0); + __builtin_amdgcn_sched_barrier(0); + }); + + // block_sync_lds(); + a_blockwise_copy.RunWrite(a_block_desc, a_block_buf); + b_blockwise_copy.RunWrite(b_block_desc, b_block_buf); + + i += 1; + } while(i < (num_loop - 1)); + } + + // tail + if constexpr(TailNum == TailNumber::Full) + { + block_sync_lds(); + static_for<0, KRepeat, 1>{}([&](auto k0) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + a_thread_copy_.Run(a_block_desc_m0_m1_m2_k, + make_tuple(m0, I0, I0, Number{}), + a_block_buf, + a_thread_desc_, + make_tuple(m0, I0, k0, I0), + a_thread_buf); + static_for<0, NRepeat, 1>{}([&](auto n0) { + b_thread_copy_.Run(b_block_desc_n0_n1_n2_k, + make_tuple(n0, I0, I0, Number{}), + b_block_buf, + b_thread_desc_, + make_tuple(n0, I0, k0, I0), + b_thread_buf); + }); + }); + + __builtin_amdgcn_sched_barrier(0); + if constexpr(k0.value != 0 || KRepeat == 1) + { + __builtin_amdgcn_s_barrier(); + __builtin_amdgcn_sched_barrier(0); + } + static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) { + static_for<0, MRepeat, 1>{}([&](auto m0) { + 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_buf[Number{}]; + }); + + using mfma_input_type = + typename vector_type::type; + + constexpr index_t c_offset = + c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); + + if constexpr(k0.value == KRepeat - 1 && + k_.value == KPerInnerLoop - KPack && + m0.value == MRepeat - 1 && n0.value == NRepeat - 1) + { + __builtin_amdgcn_sched_barrier(0); + block_sync_lds(); + __builtin_amdgcn_sched_barrier(0); + } + xdlops_gemm.Run( + a_thread_vec.template AsType(), + b_thread_vec.template AsType(), + c_thread_buf.GetVectorTypeReference(Number{})); + if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0) + { + __builtin_amdgcn_sched_barrier(0); + __builtin_amdgcn_s_setprio(1); + __builtin_amdgcn_sched_barrier(0); + } + }); + }); + }); + __builtin_amdgcn_sched_barrier(0); + __builtin_amdgcn_s_setprio(0); + __builtin_amdgcn_sched_barrier(0); + }); + } + } + + protected: + // K->M loopover + static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor( + make_tuple(Number{}, I1, Number{}, Number{}), + make_tuple(Number{}, + Number{}, + Number{}, + I1)); + + static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor( + make_tuple(Number{}, I1, Number{}, Number{}), + make_tuple(Number{}, + Number{}, + Number{}, + I1)); + + using AThreadCopy = ThreadwiseTensorSliceTransfer_v4, + Sequence<0, 1, 2, 3>, + 3, + A_K1, + A_K1>; + + using BThreadCopy = ThreadwiseTensorSliceTransfer_v4, + Sequence<0, 1, 2, 3>, + 3, + B_K1, + B_K1>; + + AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()}; + BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()}; + using Base::c_thread_desc_; +}; + +} // namespace ck