/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include <iostream>
 #include <string>
  
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/common.hpp"
  
 #include "ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp"
  
 namespace ck_tile {
  
 template <typename TilePartitioner_, typename MXFlatmmPipeline_, typename EpiloguePipeline_>
 struct MXFlatmmKernel : FlatmmKernel<TilePartitioner_, MXFlatmmPipeline_, EpiloguePipeline_>
 {
     using Underlying = FlatmmKernel<TilePartitioner_, MXFlatmmPipeline_, EpiloguePipeline_>;
  
     using TilePartitioner = remove_cvref_t<TilePartitioner_>;
     using FlatmmPipeline  = remove_cvref_t<MXFlatmmPipeline_>;
     using BlockGemmShape =
         remove_cvref_t<typename MXFlatmmPipeline_::BlockGemmShape>; // TileFlatmmShape
     using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
     using ALayout          = remove_cvref_t<typename FlatmmPipeline::ALayout>;
     using BLayout          = remove_cvref_t<typename FlatmmPipeline::BLayout>;
     using ELayout          = remove_cvref_t<typename FlatmmPipeline::CLayout>;
     using DsLayout         = remove_cvref_t<typename EpiloguePipeline::DsLayout>;
     using DsDataType       = remove_cvref_t<typename EpiloguePipeline::DsDataType>;
     static constexpr index_t KernelBlockSize  = FlatmmPipeline::BlockSize;
     static constexpr bool UsePersistentKernel = FlatmmPipeline::UsePersistentKernel;
  
     using ADataType = remove_cvref_t<typename FlatmmPipeline::ADataType>;
     using BDataType = remove_cvref_t<typename FlatmmPipeline::BDataType>;
     // Below type is actually accumulation data type - the output of block GEMM.
     using EDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
  
     static constexpr int MThreadPerXdl = BlockGemmShape::WarpTile::at(number<0>{});
     static constexpr int NThreadPerXdl = BlockGemmShape::WarpTile::at(number<1>{});
     static constexpr int KThreadPerXdl = 64 / MThreadPerXdl;
  
     static constexpr int APackedSize = numeric_traits<ADataType>::PackedSize;
     static constexpr int BPackedSize = numeric_traits<BDataType>::PackedSize;
  
     static constexpr int MXdlPack = FlatmmPipeline::MXdlPack;
     static constexpr int NXdlPack = FlatmmPipeline::NXdlPack;
     static constexpr int KXdlPack = FlatmmPipeline::KXdlPack;
  
     static constexpr index_t NumDTensor = DsDataType::size();
  
     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_assert(DsLayout::size() == DsDataType::size(),
                   "The size of DsLayout and DsDataType should be the same");
     // using KernelArgs = FlatmmKernelArgs<DsLayout::size()>;
  
     [[nodiscard]] CK_TILE_HOST static const std::string GetName()
     {
         // clang-format off
         return concat('_', "mx_flatmm_gemm", gemm_prec_str<ADataType, BDataType>, FlatmmPipeline::GetName());
         // clang-format on
     }
  
     template <class ScaleM, class ScaleN>
     CK_TILE_HOST static constexpr auto
     GridSize(const FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()>& kargs)
     {
         if constexpr(UsePersistentKernel)
         {
             hipDeviceProp_t prop;
             int deviceId = 0; // default device
  
             constexpr int block_size = MXFlatmmKernel::BlockSize().x;
             int dync_smem_size       = 0;
             int maxActiveBlocksPerCU = 0;
  
             if(hipGetDeviceProperties(&prop, deviceId) != hipSuccess)
                 throw std::runtime_error(std::string("hipGetDeviceProperties failed: ") +
                                          hipGetErrorName(hipGetLastError()));
  
             if(hipOccupancyMaxActiveBlocksPerMultiprocessor(
                    &maxActiveBlocksPerCU,
                    reinterpret_cast<void*>(
                        kentry<1, MXFlatmmKernel, remove_cvref_t<decltype(kargs)>>),
                    block_size,
                    dync_smem_size) != hipSuccess)
                 throw std::runtime_error(
                     std::string("hipOccupancyMaxActiveBlocksPerMultiprocessor failed: ") +
                     hipGetErrorName(hipGetLastError()));
  
             const int persistent_block_size = prop.multiProcessorCount * maxActiveBlocksPerCU;
             const int total_work_tile_cnt   = TilePartitioner::GridSize(kargs.M, kargs.N);
  
             // std::cout << "maxActiveBlocksPerCU: " << maxActiveBlocksPerCU
             //           << ", persistent_block_size: " << persistent_block_size
             //           << ", total_work_tile_cnt: " << total_work_tile_cnt << std::endl;
  
             if(kargs.k_batch != 1)
                 throw std::runtime_error("Wrong! k_batch != 1 not supported in persistent kernel");
             return dim3(min(persistent_block_size, total_work_tile_cnt), 1, kargs.k_batch);
         }
         else
         {
             return dim3(TilePartitioner::GridSize(kargs.M, kargs.N), 1, kargs.k_batch);
         }
     }
  
     using SplitKBatchOffset = typename Underlying::SplitKBatchOffset;
  
     template <memory_operation_enum DstInMemOp = memory_operation_enum::set, class KernelArgs>
     CK_TILE_DEVICE static auto
     MakeGemmTensorViews(const ADataType* a_ptr,
                         const BDataType* b_flat_ptr,
                         const std::array<const void*, NumDTensor>& ds_ptr,
                         EDataType* e_ptr,
                         const KernelArgs& kargs,
                         const SplitKBatchOffset& splitk_batch_offset)
     {
         const auto& a_tensor_view = [&]() {
             if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global>(
                     a_ptr,
                     make_tuple(kargs.M, splitk_batch_offset.splitted_k),
                     make_tuple(kargs.stride_A, 1),
                     number<FlatmmPipeline::GetVectorSizeA()>{},
                     number<1>{});
             }
             else
             {
                 return make_naive_tensor_view<address_space_enum::global>(
                     a_ptr,
                     make_tuple(splitk_batch_offset.splitted_k, kargs.M),
                     make_tuple(kargs.stride_A, 1),
                     number<FlatmmPipeline::GetVectorSizeA()>{},
                     number<1>{});
             }
         }();
  
         index_t kFlatK = kargs.K * BlockGemmShape::WarpTile::at(I1);
         index_t kFlatN = kargs.N * kargs.K / kFlatK;
  
         const auto& b_flat_tensor_view = [&]() {
             return make_naive_tensor_view<address_space_enum::global>(
                 b_flat_ptr,
                 make_tuple(kFlatN, kFlatK),
                 make_tuple(kFlatK, 1),
                 number<FlatmmPipeline::GetVectorSizeB()>{},
                 number<1>{});
         }();
  
         const auto& ds_tensor_view = generate_tuple(
             [&](auto i) {
                 using DiLayout   = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 using DDataType_ = remove_cvref_t<std::tuple_element_t<i.value, DsDataType>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const DDataType_*>(ds_ptr[i]),
                         make_tuple(kargs.M, kargs.N),
                         make_tuple(kargs.stride_Ds[i], 1),
                         number<EpiloguePipeline::GetVectorSizeD(i)>{},
                         number<1>{});
                 }
                 else
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const DDataType_*>(ds_ptr[i]),
                         make_tuple(kargs.N, kargs.M),
                         make_tuple(kargs.stride_Ds[i], 1),
                         number<EpiloguePipeline::GetVectorSizeD(i)>{},
                         number<1>{});
                 }
             },
             number<NumDTensor>{});
  
         // TODO: enable vector write for C in ColMajor
         const auto& e_tensor_view = [&]() {
             if constexpr(std::is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     e_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(kargs.stride_E, 1),
                     number<EpiloguePipeline::GetVectorSizeC()>{},
                     number<1>{});
             }
             else
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     e_ptr,
                     make_tuple(kargs.N, kargs.M),
                     make_tuple(kargs.stride_E, 1),
                     number<1>{},
                     number<1>{});
             }
         }();
  
         auto scale_a = kargs.scale_m_ptr;
         auto scale_b = kargs.scale_n_ptr;
  
         static constexpr int BlockScaleSize = 32; // decltype(scale_n)::GranularityK;
         const auto&& scale_packs_m = integer_divide_ceil(kargs.M, (MXdlPack * MThreadPerXdl));
         const auto&& scale_packs_n = integer_divide_ceil(kargs.N, (NXdlPack * NThreadPerXdl));
         const auto&& scale_packs_k = kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl);
  
         // A scale tensor view
         const auto& scale_a_tensor_view = [&]() {
             // Pack 2x2 e8m0 over M/K dimension into 1 int32_t to trigger dword width load
             const auto scale_a_naive_desc = make_naive_tensor_descriptor_packed(
                 make_tuple(scale_packs_m, scale_packs_k, KThreadPerXdl, MThreadPerXdl));
             const auto scale_a_desc = transform_tensor_descriptor(
                 scale_a_naive_desc,
                 make_tuple(make_merge_transform(make_tuple(scale_packs_m, MThreadPerXdl)),
                            make_merge_transform(make_tuple(scale_packs_k, KThreadPerXdl))),
                 make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
                 make_tuple(sequence<0>{}, sequence<1>{}));
  
             return make_tensor_view<address_space_enum::global>(
                 reinterpret_cast<const int32_t*>(scale_a.ptr), scale_a_desc);
         }();
  
         // B scale tensor view
         const auto& scale_b_tensor_view = [&]() {
             const auto scale_b_navie_desc = make_naive_tensor_descriptor_packed(
                 make_tuple(scale_packs_n, scale_packs_k, KThreadPerXdl, NThreadPerXdl));
             const auto scale_b_desc = transform_tensor_descriptor(
                 scale_b_navie_desc,
                 make_tuple(make_merge_transform(make_tuple(scale_packs_n, NThreadPerXdl)),
                            make_merge_transform(make_tuple(scale_packs_k, KThreadPerXdl))),
                 make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
                 make_tuple(sequence<0>{}, sequence<1>{}));
  
             return make_tensor_view<address_space_enum::global>(
                 reinterpret_cast<const int32_t*>(scale_b.ptr), scale_b_desc);
         }();
  
         return make_tuple(a_tensor_view,
                           b_flat_tensor_view,
                           ds_tensor_view,
                           e_tensor_view,
                           scale_a_tensor_view,
                           scale_b_tensor_view);
     }
  
     template <typename TensorView>
     CK_TILE_DEVICE static auto MakeGemmPadViews(const TensorView& views)
     {
         const auto& a_pad_view = [&]() {
             const auto& a_tensor_view = views.at(I0);
             if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
             {
                 return pad_tensor_view(a_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::KPerBlock>{}),
                                        sequence<false, FlatmmPipeline::kPadK>{});
             }
             else
             {
                 return pad_tensor_view(a_tensor_view,
                                        make_tuple(number<TilePartitioner::KPerBlock>{},
                                                   number<TilePartitioner::MPerBlock>{}),
                                        sequence<false, FlatmmPipeline::kPadM>{});
             }
         }();
  
         const auto& b_flat_tensor_view = views.at(I1);
  
         const auto& ds_pad_view = generate_tuple(
             [&](auto i) {
                 const auto& d_tensor_view = views.at(I2);
                 using DiLayout            = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return pad_tensor_view(d_tensor_view[i],
                                            make_tuple(number<TilePartitioner::MPerBlock>{},
                                                       number<TilePartitioner::NPerBlock>{}),
                                            sequence<false, FlatmmPipeline::kPadN>{});
                 }
                 else
                 {
                     return pad_tensor_view(d_tensor_view[i],
                                            make_tuple(number<TilePartitioner::NPerBlock>{},
                                                       number<TilePartitioner::MPerBlock>{}),
                                            sequence<false, FlatmmPipeline::kPadM>{});
                 }
             },
             number<NumDTensor>{});
  
         // TODO vector write in for C in ColMajor
         const auto& e_pad_view = [&]() {
             const auto& e_tensor_view = views.at(I3);
             if constexpr(std::is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
             {
                 return pad_tensor_view(e_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<false, FlatmmPipeline::kPadN>{});
             }
             else
             {
                 return pad_tensor_view(e_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<FlatmmPipeline::kPadM, false>{});
             }
         }();
  
         return make_tuple(
             a_pad_view, b_flat_tensor_view, ds_pad_view, e_pad_view, views.at(I4), views.at(I5));
     }
  
     template <typename PadView>
     CK_TILE_DEVICE static auto
     MakeGemmTileWindows(const PadView& views, const index_t i_m, const index_t i_n)
     {
         const auto& a_pad_view      = views.at(I0);
         const auto& b_flat_pad_view = views.at(I1);
         const auto& ds_pad_view     = views.at(I2);
         const auto& e_pad_view      = views.at(I3);
  
         const auto& a_block_window = [&]() {
             if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_tile_window(a_pad_view,
                                         make_tuple(number<TilePartitioner::MPerBlock>{},
                                                    number<TilePartitioner::KPerBlock>{}),
                                         {i_m, 0});
             }
             else
             {
                 return make_tile_window(a_pad_view,
                                         make_tuple(number<TilePartitioner::KPerBlock>{},
                                                    number<TilePartitioner::MPerBlock>{}),
                                         {0, i_m});
             }
         }();
  
         const auto& b_flat_block_window =
             make_tile_window(b_flat_pad_view,
                              make_tuple(number<FlatmmPipeline::flatNPerWarp>{},
                                         number<FlatmmPipeline::flatKPerWarp>{}),
                              {static_cast<int>(i_n / BlockGemmShape::WarpTile::at(I1)), 0});
  
         const auto ds_block_window = generate_tuple(
             [&](auto i) {
                 using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_tile_window(ds_pad_view[i],
                                             make_tuple(number<TilePartitioner::MPerBlock>{},
                                                        number<TilePartitioner::NPerBlock>{}),
                                             {i_m, i_n});
                 }
                 else
                 {
                     return make_tile_window(ds_pad_view[i],
                                             make_tuple(number<TilePartitioner::NPerBlock>{},
                                                        number<TilePartitioner::MPerBlock>{}),
                                             {i_n, i_m});
                 }
             },
             number<NumDTensor>{});
  
         auto e_block_window = make_tile_window(
             e_pad_view,
             make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
             {i_m, i_n});
  
         static constexpr int BlockScaleSize = 32;
  
         auto scale_a_block_window = make_tile_window(
             views.at(I4),
             make_tuple(number<TilePartitioner::MPerBlock / MXdlPack>{},
                        number<TilePartitioner::KPerBlock / (BlockScaleSize * KXdlPack)>{}),
             {i_m / MXdlPack, 0});
  
         auto scale_b_block_window = make_tile_window(
             views.at(I5),
             make_tuple(number<TilePartitioner::NPerBlock / NXdlPack>{},
                        number<TilePartitioner::KPerBlock / (BlockScaleSize * KXdlPack)>{}),
             {i_n / NXdlPack, 0});
  
         return make_tuple(a_block_window,
                           b_flat_block_window,
                           ds_block_window,
                           e_block_window,
                           scale_a_block_window,
                           scale_b_block_window);
     }
  
     template <class ScaleM, class ScaleN, bool UseDefaultScheduler = true>
     CK_TILE_DEVICE static void
     RunFlatmm(const ADataType* a_ptr,
               const BDataType* b_flat_ptr,
               const std::array<const void*, NumDTensor>& ds_ptr,
               EDataType* e_ptr,
               void* smem_ptr_ping,
               void* smem_ptr_pong,
               const FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()>& kargs,
               const SplitKBatchOffset& splitk_batch_offset,
               const index_t block_idx_m,
               const index_t block_idx_n)
     {
         // Create Gemm tensor views, pad views and tile windows
         const auto& gemm_tensor_views_tuple =
             MakeGemmTensorViews<EpiloguePipeline::MemoryOperation>(
                 a_ptr, b_flat_ptr, ds_ptr, e_ptr, kargs, splitk_batch_offset);
         const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
         auto gemm_tile_windows     = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
  
         const index_t num_loop = TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k);
  
         // Run GEMM cooperatively by whole workgroup.
         const auto& a_block_window       = gemm_tile_windows.at(I0);
         const auto& b_flat_block_window  = gemm_tile_windows.at(I1);
         const auto& d_block_window       = gemm_tile_windows.at(I2);
         const auto& scale_a_block_window = gemm_tile_windows.at(I4);
         const auto& scale_b_block_window = gemm_tile_windows.at(I5);
  
         static_assert(ScaleM::GranularityK == ScaleN::GranularityK // have the same granK
                           || ScaleM::GranularityMN == -1           // or ScaleA is disable
                           || ScaleN::GranularityMN == -1,          // or ScaleB is disable
                       "ScaleM and ScaleN should have the same GranularityK");
         constexpr bool DoEpiScale =
             (ScaleM::GranularityMN != -1 && ScaleM::GranularityK == 0) || // per token
             (ScaleN::GranularityMN != -1 && ScaleN::GranularityK == 0);   // per channel
  
         auto a_block_window_with_distr =
             ck_tile::make_tile_window(a_block_window.get_bottom_tensor_view(),
                                       a_block_window.get_window_lengths(),
                                       a_block_window.get_window_origin(),
                                       FlatmmPipeline::GetADramTileDistribution());
         const auto& c_block_tile = FlatmmPipeline{}(a_block_window_with_distr,
                                                     b_flat_block_window,
                                                     scale_a_block_window,
                                                     scale_b_block_window,
                                                     num_loop,
                                                     smem_ptr_ping,
                                                     smem_ptr_pong);
  
         // Run Epilogue Pipeline
         if constexpr(DoEpiScale)
         {
             auto& c_block_window = gemm_tile_windows.at(I3);
             EpiloguePipeline{}(c_block_window,
                                c_block_tile,
                                d_block_window,
                                smem_ptr_ping,
                                kargs.scale_m_ptr + block_idx_m,
                                kargs.scale_n_ptr + block_idx_n);
         }
         else if(UseDefaultScheduler || (get_warp_id() == 0))
         {
             // Run Epilogue Pipeline
             auto& c_block_window = gemm_tile_windows.at(I3);
             EpiloguePipeline{}(c_block_window, c_block_tile, d_block_window, smem_ptr_ping);
         }
     }
  
     template <class ScaleM, class ScaleN>
     CK_TILE_DEVICE void operator()(FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()> kargs,
                                    int partition_idx = blockIdx.x) const
     {
         int total_work_tile_cnt = TilePartitioner::GridSize(kargs.M, kargs.N);
  
         do
         {
             const auto [iM, iN] =
                 TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(partition_idx);
             const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
             const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
  
             const SplitKBatchOffset splitk_batch_offset(kargs);
             // options
             const ADataType* a_ptr = static_cast<const ADataType*>(kargs.a_ptr) +
                                      splitk_batch_offset.a_k_split_offset / APackedSize;
             const BDataType* b_flat_ptr = static_cast<const BDataType*>(kargs.b_ptr) +
                                           splitk_batch_offset.b_k_split_offset / BPackedSize;
             EDataType* e_ptr = static_cast<EDataType*>(kargs.e_ptr);
  
             // allocate LDS
             __shared__ char smem_ptr_ping[Underlying::GetSmemPingSize()];
             __shared__ char smem_ptr_pong[Underlying::GetSmemPongSize()];
  
             if constexpr(!(EpiloguePipeline::MemoryOperation == memory_operation_enum::atomic_add &&
                            EpiloguePipeline::GetVectorSizeC() % 2 != 0 &&
                            is_any_of<EDataType, fp16_t, bf16_t>::value))
             {
                 constexpr auto scheduler_type = (FlatmmPipeline::NumWaveGroups == 1);
                 RunFlatmm<ScaleM, ScaleN, scheduler_type>(a_ptr,
                                                           b_flat_ptr,
                                                           kargs.ds_ptr,
                                                           e_ptr,
                                                           smem_ptr_ping,
                                                           smem_ptr_pong,
                                                           kargs,
                                                           splitk_batch_offset,
                                                           i_m,
                                                           i_n);
             }
             else
             {
                 static_assert(false,
                               "Unimplemented: atomic_add with odd vector size for fp16/bf16");
             }
             partition_idx += gridDim.x;
         } while(UsePersistentKernel && partition_idx < total_work_tile_cnt);
     }
 };
  
 } // namespace ck_tile