grouped_gemm_kernel.hpp Source File

grouped_gemm_kernel.hpp Source File#

Composable Kernel: grouped_gemm_kernel.hpp Source 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 {
 
template <index_t NumDTensor = 0>
struct GroupedGemmHostArgs
{
    CK_TILE_HOST GroupedGemmHostArgs(const void* a_ptr_,
                                     const void* b_ptr_,
                                     const std::array<const void*, NumDTensor>& ds_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_,
                                     const std::array<index_t, NumDTensor>& stride_Ds_,
                                     index_t stride_E_)
        : a_ptr(a_ptr_),
          b_ptr(b_ptr_),
          ds_ptr(ds_ptr_),
          e_ptr(e_ptr_),
          M(M_),
          N(N_),
          K(K_),
          stride_A(stride_A_),
          stride_B(stride_B_),
          stride_Ds(stride_Ds_),
          stride_E(stride_E_),
          k_batch(k_batch_)
    {
    }
 
    const void* a_ptr;
    const void* b_ptr;
    const std::array<const void*, NumDTensor> ds_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;
    const std::array<index_t, NumDTensor> stride_Ds;
    union
    {
        index_t stride_E;
        index_t stride_C;
    };
 
    index_t k_batch;
};
 
template <index_t NumDTensor = 0>
struct GemmTransKernelArg
{
    UniversalGemmKernelArgs<1, 1, NumDTensor> group_karg;
    ck_tile::index_t block_start;
    ck_tile::index_t block_end;
 
    GemmTransKernelArg() = delete;
    GemmTransKernelArg(UniversalGemmKernelArgs<1, 1, NumDTensor>&& karg,
                       index_t bl_start,
                       index_t bl_end)
        : group_karg{std::move(karg)}, block_start{bl_start}, block_end{bl_end}
    {
    }
 
    GemmTransKernelArg(UniversalGemmKernelArgs<1, 1, NumDTensor>&& karg)
        : group_karg{std::move(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>;
    using DsDataType = remove_cvref_t<typename EpiloguePipeline::DsDataType>;
 
    static constexpr index_t NumDTensor_ = DsDataType::size();
    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/CLayout 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"),
                      (NumDTensor_ == 2 ? "MultiD" : "NoMultiD"),
                      (GemmPipeline::DoubleSmemBuffer ? "DoubleSmemBuffer" : "SingleSmemBuffer"));
        // clang-format on
    }
 
    CK_TILE_HOST static auto
    GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs<>>& gemm_descs) -> std::size_t
    {
        return gemm_descs.size() * sizeof(GemmTransKernelArg<NumDTensor_>);
    }
 
    CK_TILE_HOST static auto GetWorkSpaceSize(index_t group_count) -> std::size_t
    {
        return group_count * sizeof(GemmTransKernelArg<NumDTensor_>);
    }
 
    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<NumDTensor_>>& 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<NumDTensor_>>& gemm_descs)
        -> std::vector<GemmTransKernelArg<NumDTensor_>>
    {
        std::vector<GemmTransKernelArg<NumDTensor_>> 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;
            auto stride_ds         = gemm_descs[i].stride_Ds;
 
            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<1, 1, NumDTensor_>{
                {type_convert<const ADataType*>(gemm_descs[i].a_ptr)},
                {type_convert<const BDataType*>(gemm_descs[i].b_ptr)},
                {gemm_descs[i].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<NumDTensor_>>& 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<1, 1, NumDTensor_>& kargs,
                            const tuple<index_t, index_t>& block_idx_2d,
                            const index_t block_idx_z) const
    {
 
        static_assert(GemmPipeline::DoubleSmemBuffer || !GemmPipeline::Preshuffle,
                      "SingleSmemBuffer and Preshuffle cannot both be enabled simultaneously!");
 
        const auto [iM, iN] = block_idx_2d;
 
        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 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()];
 
        // TO DO:
        // Can we simplify this branching logic?
        if constexpr(GemmPipeline::DoubleSmemBuffer == true)
        {
 
            __shared__ char smem_ptr_1[GetSmemSize()];
            RunGemmWithPipelineSelection2LDS(a_ptr,
                                             b_ptr,
                                             c_ptr,
                                             kargs.ds_ptr,
                                             smem_ptr_0,
                                             smem_ptr_1,
                                             kargs,
                                             splitk_batch_offset,
                                             i_m,
                                             i_n);
        }
        else // SingleSmemBuffer
        {
 
            if constexpr(UsePersistentKernel)
            {
                RunGemmWithPipelineSelection(a_ptr,
                                             b_ptr,
                                             kargs.ds_ptr,
                                             c_ptr,
                                             smem_ptr_0,
                                             kargs,
                                             splitk_batch_offset,
                                             i_m,
                                             i_n);
            }
            else // Non-persistent kernel
            {
                Base::RunGemm({a_ptr},
                              {b_ptr},
                              kargs.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,
                                 const std::array<const void*, NumDTensor_>& ds_ptr,
                                 CDataType* c_ptr,
                                 void* smem_ptr_0,
                                 const UniversalGemmKernelArgs<1, 1, NumDTensor_>& 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.splitted_k);
 
        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 =
            amd_wave_read_first_lane(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, b_block_window, 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,
                                     const std::array<const void*, NumDTensor_>& ds_ptr,
                                     void* __restrict__ smem_ptr_0,
                                     void* __restrict__ smem_ptr_1,
                                     const UniversalGemmKernelArgs<1, 1, NumDTensor_>& 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.splitted_k);
 
        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 =
            amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
        const TailNumber tail_num = GemmPipeline::GetBlockLoopTailNum(num_loop);
 
        // Run GEMM pipeline with compile-time branching
        const auto& c_block_tile = [&]() {
            if constexpr(GemmPipeline::Preshuffle)
            {
                // Preshuffle version - without has_hot_loop parameter
                return GemmPipeline{}.template operator()(a_block_window[Base::I0],
                                                          b_block_window[Base::I0],
                                                          num_loop,
                                                          tail_num,
                                                          smem_ptr_0,
                                                          smem_ptr_1);
            }
            else
            {
                // Regular version - with has_hot_loop parameter
                const bool has_hot_loop = GemmPipeline::BlockHasHotloop(num_loop);
                return 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<NumDTensor_>* 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<NumDTensor_>*>(
            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<NumDTensor_>*>(
            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