composable_kernel/example/ck_tile/17_grouped_gemm/grouped_gemm.cpp

// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.

#include <hip/hip_runtime.h>

#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include <memory>

#include "ck_tile/core.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include "grouped_gemm.hpp"

template <typename GemmConfig,
          typename ALayout,
          typename BLayout,
          typename CLayout,
          typename ADataType,
          typename BDataType,
          typename AccDataType,
          typename CDataType>
float grouped_gemm_tileloop(const ck_tile::stream_config& s,
                            const ck_tile::index_t num_groups,
                            void* kargs_ptr,
                            bool splitk)
{
    constexpr bool kPadM = false;
    constexpr bool kPadN = false;
    constexpr bool kPadK = false;

    constexpr ck_tile::index_t TileParitionerGroupNum = 8;
    constexpr ck_tile::index_t TileParitionerM01      = 4;

    using GemmShape = ck_tile::TileGemmShape<
        ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
        ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
        ck_tile::
            sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>>;
    using TilePartitioner = ck_tile::
        GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;

    using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
    using GemmUniversalTraits =
        ck_tile::PersistentTileGemmUniversalTraits<GemmConfig::kPadM,
                                                   GemmConfig::kPadN,
                                                   GemmConfig::kPadK,
                                                   GemmConfig::DoubleSmemBuffer,
                                                   ALayout,
                                                   BLayout,
                                                   CLayout>;
    using GemmPipelineProblem =
        ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;

    float ave_time{0};

    const auto Run = [&](const auto memory_operation_) {
        constexpr auto scheduler        = GemmConfig::Scheduler;
        constexpr auto memory_operation = memory_operation_.value;

        // We create the GEMM pipeline without specifying hotloop or tailnumber.
        // These are automatically run inside the kernel based on the given input data.
        using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
                                                                           BDataType,
                                                                           AccDataType,
                                                                           GemmShape,
                                                                           GemmUniversalTraits,
                                                                           scheduler>;

        using GemmPipeline = typename PipelineTypeTraits<
            GemmConfig::Pipeline>::template GemmPipeline<UniversalGemmProblem>;
        using GemmEpilogue = ck_tile::CShuffleEpilogue<
            ck_tile::CShuffleEpilogueProblem<ADataType,
                                             BDataType,
                                             ck_tile::tuple<>,
                                             AccDataType,
                                             CDataType,
                                             ck_tile::tuple<>,
                                             CLayout,
                                             ck_tile::element_wise::PassThrough,
                                             GemmPipelineProblem::kBlockSize,
                                             TilePartitioner::MPerBlock,
                                             TilePartitioner::NPerBlock,
                                             GemmConfig::M_Warp,
                                             GemmConfig::N_Warp,
                                             GemmConfig::M_Warp_Tile,
                                             GemmConfig::N_Warp_Tile,
                                             GemmConfig::K_Warp_Tile,
                                             UniversalGemmProblem::TransposeC,
                                             memory_operation>>;
        using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
        constexpr dim3 blocks = Kernel::BlockSize();
        const dim3 grids      = Kernel::MaxOccupancyGridSize(s);

        if(s.log_level_ > 0)
        {
            std::cout << "Launching kernel: " << Kernel::GetName() << " with args:" << " grid: {"
                      << grids.x << ", " << grids.y << ", " << grids.z << "}" << ", blocks: {"
                      << blocks.x << ", " << blocks.y << ", " << blocks.z << "}" << std::endl;
        }

        ave_time =
            ck_tile::launch_kernel(s,
                                   ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
                                       Kernel{},
                                       grids,
                                       blocks,
                                       0,
                                       ck_tile::cast_pointer_to_constant_address_space(kargs_ptr),
                                       num_groups));

        return ave_time;
    };

    if(!splitk)
    {
        Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
                                       ck_tile::memory_operation_enum::set>{});
    }
    else
    {
        Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
                                       ck_tile::memory_operation_enum::atomic_add>{});
    }

    return ave_time;
}

#include "run_grouped_gemm_example.inc"

constexpr bool Persistent = true;
int main(int argc, char* argv[])
{
    return !run_grouped_gemm_example<Persistent, GemmConfigComputeV4>(argc, argv);
}