wp_pipeline_agmem_bgmem_creg_base_policy.hpp Source File

wp_pipeline_agmem_bgmem_creg_base_policy.hpp Source File#

Composable Kernel: wp_pipeline_agmem_bgmem_creg_base_policy.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
 
namespace ck_tile {
 
struct UniversalWeightPreshufflePipelineAgBgCrPolicy
    : public UniversalGemmBasePolicy<UniversalWeightPreshufflePipelineAgBgCrPolicy>
{
    using BasePolicy = UniversalGemmBasePolicy<UniversalWeightPreshufflePipelineAgBgCrPolicy>;
 
    // 3d + padding
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
    {
        using namespace ck_tile;
        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
        constexpr index_t kKPack     = GetSmemPackA<Problem>();
        using ADataType              = remove_cvref_t<typename Problem::ADataType>;
 
        constexpr auto DataTypeSize = sizeof(ADataType);
        constexpr auto MLdsLayer =
            (32 * 4 / kKPerBlock / DataTypeSize) < 1 ? 1 : (32 * 4 / kKPerBlock / DataTypeSize);
 
        constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
            make_tuple(number<kKPerBlock / kKPack * MLdsLayer>{},
                       number<kMPerBlock / MLdsLayer>{},
                       number<kKPack>{}),
            make_tuple(number<kKPack>{}, number<kKPerBlock * MLdsLayer>{}, number<1>{}),
            number<kKPack>{},
            number<1>{});
 
        constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
            a_lds_block_desc_0,
            make_tuple(make_xor_transform(make_tuple(number<kMPerBlock / MLdsLayer>{},
                                                     number<kKPerBlock / kKPack * MLdsLayer>{})),
                       make_pass_through_transform(number<kKPack>{})),
            make_tuple(sequence<1, 0>{}, sequence<2>{}),
            make_tuple(sequence<1, 0>{}, sequence<2>{}));
 
        constexpr auto a_lds_block_desc_xk0_mnldslayer_mn_xk1 = transform_tensor_descriptor(
            a_lds_block_desc_permuted,
            make_tuple(make_unmerge_transform(
                           make_tuple(number<MLdsLayer>{}, number<kKPerBlock / kKPack>{})),
                       make_pass_through_transform(number<kMPerBlock / MLdsLayer>{}),
                       make_pass_through_transform(number<kKPack>{})),
            make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
            make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
 
        constexpr auto a_lds_block_desc = transform_tensor_descriptor(
            a_lds_block_desc_xk0_mnldslayer_mn_xk1,
            make_tuple(
                make_merge_transform(
                    make_tuple(number<kMPerBlock / MLdsLayer>{}, number<MLdsLayer>{})),
                make_merge_transform(make_tuple(number<kKPerBlock / kKPack>{}, number<kKPack>{}))),
            make_tuple(sequence<1, 0>{}, sequence<2, 3>{}),
            make_tuple(sequence<0>{}, sequence<1>{}));
        return a_lds_block_desc;
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
    {
        constexpr index_t smem_size_a = sizeof(typename Problem::ADataType) *
                                        MakeALdsBlockDescriptor<Problem>().get_element_space_size();
        return smem_size_a;
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
    {
        constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
 
        return smem_size_a;
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
    {
        return Problem::VectorLoadSize / sizeof(typename Problem::ADataType);
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto GetKBPerLoad()
    {
        using TileShape = typename Problem::BlockGemmShape;
#if defined(__gfx11__)
        constexpr index_t scale = 4;
#else
        constexpr index_t scale = get_warp_size() == 32 ? 2 : 1;
#endif
        if constexpr(TileShape::WarpTile::at(I1) == 32)
        {
            return TileShape::WarpTile::at(I2) * scale / 2;
        }
        else
        {
            static_assert(TileShape::WarpTile::at(I1) == 16);
            return TileShape::WarpTile::at(I2) * scale / 4;
        }
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
    {
        using ADataType = remove_cvref_t<typename Problem::ADataType>;
        using ALayout   = remove_cvref_t<typename Problem::ALayout>;
 
        constexpr index_t BlockSize = Problem::kBlockSize;
 
        constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
 
        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
        {
            constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
            constexpr index_t M0           = MPerBlock / M1;
            constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
            static_assert(total_pixels % M1 == 0);
            constexpr index_t K3    = total_pixels / M1;
            constexpr index_t KPack = GetSmemPackA<Problem>();
            static_assert(KPack % K3 == 0);
            constexpr index_t K2 = KPack / K3;
            if constexpr(get_warp_size() >= (K2 * M0))
            {
                constexpr index_t K1 = get_warp_size() / (K2 * M0);
                constexpr index_t K0 = BlockSize / get_warp_size();
                static_assert(KPerBlock == K0 * K1 * K2 * K3);
                return make_static_tile_distribution(
                    tile_distribution_encoding<sequence<1>,
                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
                                               tuple<sequence<2>, sequence<2, 1, 2>>,
                                               tuple<sequence<0>, sequence<1, 0, 2>>,
                                               sequence<2, 1>,
                                               sequence<3, 1>>{});
            }
            else
            {
                constexpr index_t K1   = (K2 * M0) / get_warp_size();
                constexpr index_t K2_m = K2 / K1;
                constexpr index_t K0   = BlockSize / get_warp_size() / K1;
                static_assert(KPerBlock == K0 * K1 * K2_m * K3);
                return make_static_tile_distribution(
                    tile_distribution_encoding<sequence<1>,
                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
                                               tuple<sequence<2, 2>, sequence<1, 2>>,
                                               tuple<sequence<0, 1>, sequence<0, 2>>,
                                               sequence<2, 1>,
                                               sequence<3, 1>>{});
            }
        }
        else
        {
            constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType);
            constexpr index_t K0 = KPerBlock / K1;
            constexpr index_t M2 = get_warp_size() / K0;
            // coalesce reading for each blocks
            if constexpr(get_warp_size() % (M2 * K0) == 0)
            {
                constexpr index_t M1 = BlockSize / get_warp_size();
                static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
                static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
                constexpr index_t M0 = MPerBlock / (M2 * M1);
                static_assert(M0 * M1 * M2 == MPerBlock,
                              "Incorrect M0, M2, M1 configuration! "
                              "M0, M1, M2 must cover whole MPerBlock!");
 
                return make_static_tile_distribution(
                    tile_distribution_encoding<sequence<1>,
                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
                                               tuple<sequence<1>, sequence<1, 2>>,
                                               tuple<sequence<1>, sequence<2, 0>>,
                                               sequence<1, 2>,
                                               sequence<0, 1>>{});
            }
            else
            {
                constexpr index_t M0 = BlockSize / get_warp_size();
                constexpr index_t M1 = MPerBlock / (M2 * M0);
                static_assert(M0 * M1 * M2 == MPerBlock,
                              "Incorrect M0, M1, M2 configuration! "
                              "M0, M1, M2 must cover whole MPerBlock!");
                return make_static_tile_distribution(
                    tile_distribution_encoding<sequence<1>,
                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
                                               tuple<sequence<1>, sequence<1, 2>>,
                                               tuple<sequence<0>, sequence<2, 0>>,
                                               sequence<1, 2>,
                                               sequence<1, 1>>{});
            }
        }
    }
 
    template <typename Problem>
    CK_TILE_DEVICE static constexpr auto MakeBFlatDramTileDistribution()
    {
        using TileShape = typename Problem::BlockGemmShape;
 
        constexpr index_t BlockSize = Problem::kBlockSize;
        constexpr index_t WaveSize  = get_warp_size();
        constexpr index_t WaveNum   = BlockSize / WaveSize;
 
        constexpr index_t KBPerLoad = GetKBPerLoad<Problem>();
#if defined(__gfx11__)
        constexpr index_t KRepeatInWave = 2;
#else
        constexpr index_t KRepeatInWave = 1;
#endif
        constexpr index_t KThdPerWave = WaveSize / KRepeatInWave; // threads cnt in K dim
        constexpr index_t KWavePerBlk = 1;
        constexpr index_t KRepeat     = 1;
        static_assert(TileShape::flatKPerWarp == KThdPerWave * KBPerLoad, "wrong");
 
        constexpr index_t NBPerLoad   = 1;
        constexpr index_t NThdPerWave = 1;
        constexpr index_t NWavePerBlk = TileShape::BlockWarps::at(number<1>{}); // N_Warp
        constexpr index_t NRepeat     = 1;
 
        constexpr index_t WaveRepeat = WaveNum / TileShape::flatNPerWarp;
        return make_static_tile_distribution(
            tile_distribution_encoding<
                sequence<WaveRepeat, KRepeatInWave>,                           // ?
                tuple<sequence<NRepeat, NWavePerBlk, NThdPerWave, NBPerLoad>,  // second direction
                      sequence<KRepeat, KWavePerBlk, KThdPerWave, KBPerLoad>>, // first  direction
                // wave in blk,     // thd in wave
                // <M, K>           // <M, K>
                tuple<sequence<0, 1, 2>, sequence<0, 1, 2>>, // which direction
                tuple<sequence<0, 1, 1>, sequence<1, 2, 2>>, // which index
                // <repeat, vec_load>
                sequence<1, 1, 2, 2>,
                sequence<0, 3, 0, 3>>{});
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDistribution()
    {
        using ALayout   = remove_cvref_t<typename Problem::ALayout>;
        using ADataType = remove_cvref_t<typename Problem::ADataType>;
        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
        constexpr index_t kBlockSize = Problem::kBlockSize;
        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
 
        constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
        constexpr index_t M0           = kMPerBlock / M1;
        constexpr index_t total_pixels = kMPerBlock * kKPerBlock / kBlockSize;
        static_assert(total_pixels % M1 == 0);
        constexpr index_t K3     = total_pixels / M1;
        constexpr index_t kKPack = GetSmemPackA<Problem>();
        static_assert(kKPack % K3 == 0);
        constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
        constexpr index_t warp_size = get_warp_size();
        if constexpr(warp_size >= (K2 * M0))
        {
            constexpr index_t K1 = warp_size / (K2 * M0);
            constexpr index_t K0 = kBlockSize / warp_size;
 
            return make_static_tile_distribution(
                tile_distribution_encoding<sequence<1>,
                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
                                           tuple<sequence<2>, sequence<2, 1, 2>>,
                                           tuple<sequence<0>, sequence<1, 0, 2>>,
                                           sequence<1, 2>,
                                           sequence<1, 3>>{});
        }
        else
        {
            constexpr index_t K1   = (K2 * M0) / get_warp_size();
            constexpr index_t K2_m = K2 / K1;
            constexpr index_t K0   = kBlockSize / get_warp_size() / K1;
            static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
            return make_static_tile_distribution(
                tile_distribution_encoding<sequence<1>,
                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
                                           tuple<sequence<2, 2>, sequence<1, 2>>,
                                           tuple<sequence<0, 1>, sequence<0, 2>>,
                                           sequence<1, 2>,
                                           sequence<1, 3>>{});
        }
    }
 
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto GetBlockWeightPreshuffle()
    {
        using BlockWarps = typename Problem::BlockGemmShape::BlockWarps;
        using WarpTile   = typename Problem::BlockGemmShape::WarpTile;
        using BTypeToUse =
            std::conditional_t<std::is_same_v<typename Problem::BDataType, ck_tile::pk_int4_t>,
                               typename Problem::ADataType,
                               typename Problem::BDataType>;
        using WarpGemm = WarpGemmDispatcher<typename Problem::ADataType,
                                            BTypeToUse,
                                            typename Problem::CDataType,
                                            WarpTile::at(I0),
                                            WarpTile::at(I1),
                                            WarpTile::at(I2),
                                            Problem::TransposeC>;
 
        using BlockWeightPreshufflePolicy =
            BlockWeightPreshuffleASmemBSmemCRegV1CustomPolicy<typename Problem::ADataType,
                                                              typename Problem::BDataType,
                                                              typename Problem::CDataType,
                                                              BlockWarps,
                                                              WarpGemm>;
        return BlockWeightPreshuffleASmemBSmemCRegV1<Problem, BlockWeightPreshufflePolicy>{};
    }
    template <typename Problem>
    CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeC()
    {
        using BlockGemm = remove_cvref_t<decltype(GetBlockWeightPreshuffle<Problem>())>;
        using WG_       = typename BlockGemm::WG;
 
        constexpr bool TransposeC = Problem::TransposeC;
        using CLayout             = typename Problem::CLayout;
        using CWarpDstr           = typename WG_::CWarpDstr;
 
        // N is contiguous dimension
        if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
        {
            if constexpr(TransposeC)
            {
                // In this case each thread has multiple consecutive elements in
                // N dimension, however consecutive threads' elements have stride.
                constexpr index_t NDimY = CWarpDstr::NDimY;
                constexpr auto c_warp_y_lengths =
                    CWarpDstr{}.get_ys_to_d_descriptor().get_lengths();
                static_assert(WG_::WarpGemmAttribute::Impl::kCM1PerLane ==
                              c_warp_y_lengths.get(number<NDimY - 1>{}));
                return c_warp_y_lengths.get(number<NDimY - 1>{});
            }
            else
            {
                // In this case each thread has just a single item in Ndim
                return WG_::WarpGemmAttribute::Impl::kCNLane / WG_::kN;
            }
        }
        // M is contiguous dimension
        else if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::ColumnMajor>)
        {
            if constexpr(TransposeC)
            {
                // In this case each thread has just a single item in Mdim
                return WG_::WarpGemmAttribute::Impl::kCNLane / WG_::kN;
            }
            else
            {
                // In this case each thread has multiple consecutive elements in
                // M dimension, however consecutive threads' elements have stride.
                constexpr index_t NDimY = CWarpDstr::NDimY;
                constexpr auto c_warp_y_lengths =
                    CWarpDstr{}.get_ys_to_d_descriptor().get_lengths();
                static_assert(WG_::WarpGemmAttribute::Impl::kCM1PerLane ==
                              c_warp_y_lengths.get(number<NDimY - 1>{}));
                return c_warp_y_lengths.get(number<NDimY - 1>{});
            }
        }
        else
        {
            static_assert(false, "Unsupported CLayout!");
        }
    }
};
 
} // namespace ck_tile