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

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

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck_tile/core/numeric/math.hpp"
 #include "ck_tile/core/utility/literals.hpp"
 #include "ck_tile/core/utility/type_traits.hpp"
 #include "ck_tile/host/stream_utils.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
 #include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
 #include "ck_tile/host.hpp"
  
 #include <hip/hip_runtime.h>
  
 namespace ck_tile {
  
 struct GroupedGemmHostArgs
 {
     CK_TILE_HOST GroupedGemmHostArgs(const void* a_ptr_,
                                      const void* b_ptr_,
                                      void* e_ptr_,
                                      index_t k_batch_,
                                      index_t M_,
                                      index_t N_,
                                      index_t K_,
                                      index_t stride_A_,
                                      index_t stride_B_,
                                      index_t stride_E_)
         : a_ptr(a_ptr_),
           b_ptr(b_ptr_),
           e_ptr(e_ptr_),
           M(M_),
           N(N_),
           K(K_),
           stride_A(stride_A_),
           stride_B(stride_B_),
           stride_E(stride_E_),
           k_batch(k_batch_)
     {
     }
  
     const void* a_ptr;
     const void* b_ptr;
     union
     {
         void* e_ptr;
         void* c_ptr;
     };
  
     index_t M;
     index_t N;
     index_t K;
     index_t stride_A;
     index_t stride_B;
  
     union
     {
         index_t stride_E;
         index_t stride_C;
     };
  
     index_t k_batch;
 };
  
 struct GemmTransKernelArg
 {
     UniversalGemmKernelArgs<> group_karg;
     ck_tile::index_t block_start;
     ck_tile::index_t block_end;
  
     GemmTransKernelArg() = delete;
     GemmTransKernelArg(UniversalGemmKernelArgs<>&& karg, index_t bl_start, index_t bl_end)
         : group_karg{karg}, block_start{bl_start}, block_end{bl_end}
     {
     }
  
     GemmTransKernelArg(UniversalGemmKernelArgs<>&& karg)
         : group_karg{karg}, block_start{0}, block_end{0}
     {
     }
 };
  
 template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
 struct GroupedGemmKernel
 {
     using Base = UniversalGemmKernel<TilePartitioner_, GemmPipeline_, EpiloguePipeline_>;
  
     using TilePartitioner  = remove_cvref_t<TilePartitioner_>;
     using GemmPipeline     = remove_cvref_t<GemmPipeline_>;
     using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
  
     using ALayout = remove_cvref_t<typename GemmPipeline::ALayout>;
     using BLayout = remove_cvref_t<typename GemmPipeline::BLayout>;
     using CLayout = remove_cvref_t<typename GemmPipeline::CLayout>;
  
     using ADataType = remove_cvref_t<typename GemmPipeline::ADataType>;
     using BDataType = remove_cvref_t<typename GemmPipeline::BDataType>;
     using CDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
  
     static_assert(
         !is_detected<is_tuple, ALayout>::value && !is_detected<is_tuple, ADataType>::value,
         "ALayout and ADataType must be scalars. Multiple parameters are not currently supported.");
  
     static_assert(
         !is_detected<is_tuple, BLayout>::value && !is_detected<is_tuple, BDataType>::value,
         "BLayout and BDataType must be scalars. Multiple parameters are not currently supported.");
  
     static_assert(!is_detected<is_tuple, CLayout>::value &&
                       !is_detected<is_tuple, CDataType>::value,
                   "C/ELayout and C/EDataType must be scalars.");
  
     using OffsetTile1DPartitioner = OffsettedTile1DPartitioner<TilePartitioner>;
     using Kernel = GroupedGemmKernel<TilePartitioner, GemmPipeline, EpiloguePipeline>;
  
     static constexpr index_t kBlockSize       = GemmPipeline::BlockSize;
     static constexpr bool UsePersistentKernel = GemmPipeline::UsePersistentKernel;
  
     [[nodiscard]] CK_TILE_HOST static const std::string GetName()
     {
         // clang-format off
         using P_ = GemmPipeline;
  
         return concat('_', "gemm_grouped", gemm_prec_str<ADataType, BDataType>(),
                       concat('x', P_::MPerBlock, P_::NPerBlock, P_::KPerBlock),
                       concat('x', P_::GetVectorSizeA(), P_::GetVectorSizeB(), P_::GetVectorSizeC()),
                       concat('x', P_::kPadM, P_::kPadN, P_::kPadK),
                       (UsePersistentKernel ? "Persistent" : "NonPersistent"));
         // clang-format on
     }
  
     CK_TILE_HOST static auto
     GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs) -> std::size_t
     {
         return gemm_descs.size() * sizeof(GemmTransKernelArg);
     }
  
     CK_TILE_HOST static auto GetWorkSpaceSize(index_t group_count) -> std::size_t
     {
         return group_count * sizeof(GemmTransKernelArg);
     }
  
     CK_TILE_HOST static auto BlockSize() -> dim3
     {
         if(is_wave32())
         {
             return dim3(kBlockSize / 2);
         }
         else
         {
             return dim3(kBlockSize);
         }
     }
  
     CK_TILE_HOST static auto MaxOccupancyGridSize(const stream_config& s) -> dim3
     {
         using ConstantPointer = const void CK_CONSTANT_ADDRESS_SPACE*;
         const auto kernel     = kentry<1, Kernel, ConstantPointer, index_t>;
         int occupancy;
         HIP_CHECK_ERROR(
             hipOccupancyMaxActiveBlocksPerMultiprocessor(&occupancy, kernel, kBlockSize, 0));
         const int grid_size = get_available_compute_units(s) * occupancy;
         return dim3(grid_size, 1, 1);
     }
  
     CK_TILE_HOST static auto GridSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
     {
         index_t grid_size = 0;
         for(const auto& it_desc : gemm_descs)
         {
             const auto local_grid_size = TilePartitioner::GridSize(it_desc.M, it_desc.N);
             grid_size += local_grid_size * it_desc.k_batch;
         }
         return dim3(grid_size, 1, 1);
     }
  
     CK_TILE_HOST static auto
     MakeKargs(const std::vector<GroupedGemmHostArgs>& gemm_descs) -> std::vector<GemmTransKernelArg>
     {
         std::vector<GemmTransKernelArg> gemm_kernel_args_;
         index_t group_count = ck_tile::type_convert<ck_tile::index_t>(gemm_descs.size());
         index_t grid_size   = 0;
         gemm_kernel_args_.reserve(group_count);
  
         for(std::size_t i = 0; i < gemm_descs.size(); ++i)
         {
             const index_t M = gemm_descs[i].M;
             const index_t N = gemm_descs[i].N;
             const index_t K = gemm_descs[i].K;
  
             if(M == 0 || N == 0 || K == 0)
             {
                 continue;
             }
  
             const index_t stride_a = gemm_descs[i].stride_A;
             const index_t stride_b = gemm_descs[i].stride_B;
             const index_t stride_e = gemm_descs[i].stride_E;
  
             const index_t grid_size_grp = TilePartitioner::GridSize(M, N) * gemm_descs[i].k_batch;
  
             const index_t block_start = grid_size;
             const index_t block_end   = grid_size + grid_size_grp;
  
             grid_size += grid_size_grp;
  
             auto karg =
                 UniversalGemmKernelArgs<>{{type_convert<const ADataType*>(gemm_descs[i].a_ptr)},
                                           {type_convert<const BDataType*>(gemm_descs[i].b_ptr)},
                                           {/*ds_ptr*/},
                                           type_convert<CDataType*>(gemm_descs[i].e_ptr),
                                           M,
                                           N,
                                           K,
                                           {stride_a},
                                           {stride_b},
                                           {/*stride_ds*/},
                                           stride_e,
                                           gemm_descs[i].k_batch};
  
             gemm_kernel_args_.emplace_back(std::move(karg), block_start, block_end);
         }
  
         return gemm_kernel_args_;
     }
  
     CK_TILE_HOST static bool IsSupportedArgument(const std::vector<GemmTransKernelArg>& kargs)
     {
         for(const auto& karg : kargs)
         {
             if(!Base::IsSupportedArgument(karg.group_karg))
             {
                 return false;
             }
         }
         return true;
     }
  
     CK_TILE_HOST_DEVICE static constexpr auto GetSmemSize() -> index_t
     {
         return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
     }
  
     CK_TILE_DEVICE void Run(const UniversalGemmKernelArgs<>& kargs,
                             const tuple<index_t, index_t>& block_idx_2d,
                             const index_t block_idx_z) const
     {
         const auto [iM, iN] = block_idx_2d;
  
         const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
         const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
  
         const typename Base::SplitKBatchOffset splitk_batch_offset(kargs, block_idx_z);
  
         const ADataType* a_ptr = static_cast<const ADataType*>(kargs.as_ptr[0]) +
                                  splitk_batch_offset.as_k_split_offset[0];
         const BDataType* b_ptr = static_cast<const BDataType*>(kargs.bs_ptr[0]) +
                                  splitk_batch_offset.bs_k_split_offset[0];
         CDataType* c_ptr = static_cast<CDataType*>(kargs.e_ptr);
  
         // allocate LDS
         __shared__ char smem_ptr_0[GetSmemSize()];
  
         if constexpr(GemmPipeline::DoubleSmemBuffer == true)
         {
             __shared__ char smem_ptr_1[GetSmemSize()];
             if constexpr(UsePersistentKernel)
             {
                 RunGemmWithPipelineSelection2LDS(a_ptr,
                                                  b_ptr,
                                                  c_ptr,
                                                  smem_ptr_0,
                                                  smem_ptr_1,
                                                  kargs,
                                                  splitk_batch_offset,
                                                  i_m,
                                                  i_n);
             }
             else
             {
                 Base::RunGemm2LDS({a_ptr},
                                   {b_ptr},
                                   {/*ds_ptr*/},
                                   c_ptr,
                                   smem_ptr_0,
                                   smem_ptr_1,
                                   kargs,
                                   splitk_batch_offset,
                                   i_m,
                                   i_n);
             }
         }
         else
         {
             if constexpr(UsePersistentKernel)
             {
                 RunGemmWithPipelineSelection(
                     a_ptr, b_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
             }
             else
             {
                 Base::RunGemm({a_ptr},
                               {b_ptr},
                               {/*ds_ptr*/},
                               c_ptr,
                               smem_ptr_0,
                               kargs,
                               splitk_batch_offset,
                               i_m,
                               i_n);
             }
         }
     }
  
     CK_TILE_DEVICE static void
     RunGemmWithPipelineSelection(const ADataType* a_ptr,
                                  const BDataType* b_ptr,
                                  CDataType* c_ptr,
                                  void* smem_ptr_0,
                                  const UniversalGemmKernelArgs<>& kargs,
                                  const typename Base::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 =
             Base::template MakeGemmTensorViews<EpiloguePipeline::MemoryOperation>(
                 {a_ptr}, {b_ptr}, {/*ds_ptr*/}, c_ptr, kargs, splitk_batch_offset);
  
         const auto& gemm_pad_views = Base::MakeGemmPadViews(gemm_tensor_views_tuple);
         auto gemm_tile_windows =
             Base::MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
         const auto& a_block_window = gemm_tile_windows.at(Base::I0);
         const auto& b_block_window = gemm_tile_windows.at(Base::I1);
         const auto& d_block_window = gemm_tile_windows.at(Base::I2);
  
         // Get hot-loop and tail configuration
         const index_t num_loop = __builtin_amdgcn_readfirstlane(
             TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
         const bool has_hot_loop   = GemmPipeline::BlockHasHotloop(num_loop);
         const TailNumber tail_num = GemmPipeline::GetBlockLoopTailNum(num_loop);
  
         // Run GEMM pipeline
         const auto& c_block_tile = GemmPipeline{}.template operator()(a_block_window[Base::I0],
                                                                       b_block_window[Base::I0],
                                                                       num_loop,
                                                                       has_hot_loop,
                                                                       tail_num,
                                                                       smem_ptr_0);
         // Run Epilogue Pipeline
         auto& c_block_window = gemm_tile_windows.at(Base::I3);
         EpiloguePipeline{}.template
         operator()<decltype(c_block_window), decltype(c_block_tile), decltype(d_block_window)>(
             c_block_window, c_block_tile, d_block_window, smem_ptr_0);
     }
  
     CK_TILE_DEVICE static void
     RunGemmWithPipelineSelection2LDS(const ADataType* a_ptr,
                                      const BDataType* b_ptr,
                                      CDataType* c_ptr,
                                      void* __restrict__ smem_ptr_0,
                                      void* __restrict__ smem_ptr_1,
                                      const UniversalGemmKernelArgs<>& kargs,
                                      const typename Base::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 =
             Base::template MakeGemmTensorViews<EpiloguePipeline::MemoryOperation>(
                 {a_ptr}, {b_ptr}, {/*ds_ptr*/}, c_ptr, kargs, splitk_batch_offset);
  
         const auto& gemm_pad_views = Base::MakeGemmPadViews(gemm_tensor_views_tuple);
         auto gemm_tile_windows =
             Base::MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
         const auto& a_block_window = gemm_tile_windows.at(Base::I0);
         const auto& b_block_window = gemm_tile_windows.at(Base::I1);
         const auto& d_block_window = gemm_tile_windows.at(Base::I2);
  
         // Get hot-loop and tail configuration
         const index_t num_loop = __builtin_amdgcn_readfirstlane(
             TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
         const bool has_hot_loop   = GemmPipeline::BlockHasHotloop(num_loop);
         const TailNumber tail_num = GemmPipeline::GetBlockLoopTailNum(num_loop);
  
         // Run GEMM pipeline
         const auto& c_block_tile = GemmPipeline{}.template operator()(a_block_window[Base::I0],
                                                                       b_block_window[Base::I0],
                                                                       num_loop,
                                                                       has_hot_loop,
                                                                       tail_num,
                                                                       smem_ptr_0,
                                                                       smem_ptr_1);
         // Run Epilogue Pipeline
         auto& c_block_window = gemm_tile_windows.at(Base::I3);
         EpiloguePipeline{}.template
         operator()<decltype(c_block_window), decltype(c_block_tile), decltype(d_block_window)>(
             c_block_window, c_block_tile, d_block_window, smem_ptr_0);
     }
  
     CK_TILE_DEVICE index_t FindGroupId(const GemmTransKernelArg* gemm_desc_ptr,
                                        index_t block_id,
                                        index_t group_count) const
     {
         index_t left     = 0;
         index_t right    = group_count;
         index_t group_id = index_t((left + right) >> 1);
  
         while((!(block_id >= gemm_desc_ptr[group_id].block_start &&
                  block_id < gemm_desc_ptr[group_id].block_end)) &&
               left <= right)
         {
             if(block_id < gemm_desc_ptr[group_id].block_start)
             {
                 right = group_id;
             }
             else
             {
                 left = group_id;
             }
             group_id = index_t((left + right) >> 1);
         }
  
         return group_id;
     }
  
     // For non-persistent kernels
     template <bool U = UsePersistentKernel, typename = std::enable_if_t<!U>>
     CK_TILE_DEVICE void operator()(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
                                    index_t group_count) const
     {
         const index_t block_id   = ck_tile::get_block_1d_id();
         const auto gemm_desc_ptr = reinterpret_cast<const GemmTransKernelArg*>(
             cast_pointer_to_generic_address_space(gemm_descs_const));
  
         const index_t group_id  = FindGroupId(gemm_desc_ptr, block_id, group_count);
         const auto& kargs       = gemm_desc_ptr[group_id];
         const auto grid_size_2d = TilePartitioner::GridSize(kargs.group_karg.M, kargs.group_karg.N);
         const auto block_idx_2d = OffsetTile1DPartitioner::GetOffsetedTileIndex(
             0,
             kargs.group_karg.M,
             kargs.group_karg.N,
             (block_id - kargs.block_start) % grid_size_2d);
         Run(kargs.group_karg, block_idx_2d, (block_id - kargs.block_start) / grid_size_2d);
     }
  
     // For persistent kernels
     template <bool U   = UsePersistentKernel,
               typename = std::enable_if_t<U>,
               typename = void> // extra template parameter to avoid redefinition
     CK_TILE_DEVICE void operator()(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
                                    const index_t group_count) const
     {
         const index_t grid_size  = ck_tile::get_grid_size();
         const auto gemm_desc_ptr = reinterpret_cast<const GemmTransKernelArg*>(
             cast_pointer_to_generic_address_space(gemm_descs_const));
         index_t block_id      = ck_tile::get_block_1d_id(); // initial block_id
         index_t cum_grid_size = 0;
         for(index_t group_id = 0; group_id < group_count; ++group_id)
         {
             const auto& kargs      = gemm_desc_ptr[group_id].group_karg;
             const auto& k_batch    = kargs.k_batch;
             const auto block_start = cum_grid_size;
             cum_grid_size += TilePartitioner::GridSize(kargs.M, kargs.N) * k_batch;
             while(block_id < cum_grid_size)
             {
                 const auto grid_size_2d = TilePartitioner::GridSize(kargs.M, kargs.N);
                 const auto block_idx_2d = OffsetTile1DPartitioner::GetOffsetedTileIndex(
                     0, kargs.M, kargs.N, (block_id - block_start) % grid_size_2d);
                 Run(kargs, block_idx_2d, (block_id - block_start) / grid_size_2d);
                 block_id = block_id + grid_size; // advance to next block
                 // NOTE: this check is redundant but helps the compiler avoid spilling some VGPR
                 if(block_id >= cum_grid_size)
                 {
                     break; // exit the loop if all blocks are processed
                 }
             }
         }
     }
 };
  
 } // namespace ck_tile