include/ck_tile/ops/gemm_group_quant/kernel/gemm_aquant_kernel.hpp Source File

include/ck_tile/ops/gemm_group_quant/kernel/gemm_aquant_kernel.hpp Source File#

Composable Kernel: include/ck_tile/ops/gemm_group_quant/kernel/gemm_aquant_kernel.hpp Source File
 // SPDX-License-Identifier: MIT
 // Copyright (c) 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/host/concat.hpp"
  
 namespace ck_tile {
  
 struct AQuantGemmProblem
 {
     CK_TILE_HOST AQuantGemmProblem() = default;
     CK_TILE_HOST AQuantGemmProblem(index_t M_,
                                    index_t N_,
                                    index_t K_,
                                    index_t QK_,
                                    index_t stride_A_,
                                    index_t stride_B_,
                                    index_t stride_C_,
                                    index_t stride_AQ_)
         : M(M_),
           N(N_),
           K(K_),
           QK(QK_),
           stride_A(stride_A_),
           stride_B(stride_B_),
           stride_C(stride_C_),
           stride_AQ(stride_AQ_)
     {
     }
  
     index_t M;
     index_t N;
     index_t K;
     index_t QK;
     index_t stride_A;
     index_t stride_B;
     index_t stride_C;
     index_t stride_AQ;
 };
  
 struct AQuantGemmHostArgs : public AQuantGemmProblem
 {
     CK_TILE_HOST AQuantGemmHostArgs() = default;
     CK_TILE_HOST AQuantGemmHostArgs(const void* a_ptr_,
                                     const void* b_ptr_,
                                     void* c_ptr_,
                                     const void* aq_ptr_,
                                     index_t k_batch_,
                                     index_t M_,
                                     index_t N_,
                                     index_t K_,
                                     index_t QK_,
                                     index_t stride_A_,
                                     index_t stride_B_,
                                     index_t stride_C_,
                                     index_t stride_AQ_)
         : AQuantGemmProblem(M_, N_, K_, QK_, stride_A_, stride_B_, stride_C_, stride_AQ_),
           a_ptr(a_ptr_),
           b_ptr(b_ptr_),
           aq_ptr(aq_ptr_),
           c_ptr(c_ptr_),
           k_batch(k_batch_)
     {
     }
  
     const void* a_ptr;
     const void* b_ptr;
     const void* aq_ptr;
     void* c_ptr;
     index_t k_batch;
 };
  
 struct AQuantGemmKernelArgs
 {
     const void* a_ptr;
     const void* b_ptr;
     const void* aq_ptr;
     void* c_ptr;
     index_t M;
     index_t N;
     index_t K;
     index_t QK;
     index_t stride_A;
     index_t stride_B;
     index_t stride_C;
     index_t stride_AQ;
     index_t k_batch;
 };
  
 template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
 struct AQuantGemmKernel
 {
     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 AQLayout                           = remove_cvref_t<typename GemmPipeline::AQLayout>;
     using BLayout                            = remove_cvref_t<typename GemmPipeline::BLayout>;
     using CLayout                            = remove_cvref_t<typename GemmPipeline::CLayout>;
     static constexpr index_t KernelBlockSize = GemmPipeline::BlockSize;
  
     using ADataType  = remove_cvref_t<typename GemmPipeline::ADataType>;
     using AQDataType = remove_cvref_t<typename GemmPipeline::AQDataType>;
     using BDataType  = remove_cvref_t<typename GemmPipeline::BDataType>;
     using CDataType  = remove_cvref_t<typename EpiloguePipeline::ODataType>;
  
     static constexpr auto I0 = number<0>();
     static constexpr auto I1 = number<1>();
     static constexpr auto I2 = number<2>();
     static constexpr auto I3 = number<3>();
  
     [[nodiscard]] CK_TILE_HOST static const std::string GetName()
     {
         // clang-format off
         return concat('_', "gemm", gemm_prec_str<ADataType, BDataType>, GemmPipeline::GetName());
         // clang-format on
     }
  
     CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N, index_t KBatch)
     {
         return dim3(TilePartitioner::GridSize(M, N), 1, KBatch);
     }
  
     CK_TILE_HOST static constexpr auto BlockSize() { return dim3(KernelBlockSize); }
  
     CK_TILE_HOST static constexpr AQuantGemmKernelArgs
     MakeKernelArgs(const AQuantGemmHostArgs& hostArgs)
     {
         return AQuantGemmKernelArgs{hostArgs.a_ptr,
                                     hostArgs.b_ptr,
                                     hostArgs.aq_ptr,
                                     hostArgs.c_ptr,
                                     hostArgs.M,
                                     hostArgs.N,
                                     hostArgs.K,
                                     hostArgs.QK,
                                     hostArgs.stride_A,
                                     hostArgs.stride_B,
                                     hostArgs.stride_C,
                                     hostArgs.stride_AQ,
                                     hostArgs.k_batch};
     }
  
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
         return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
     }
  
     struct SplitKBatchOffset
     {
         __device__ SplitKBatchOffset(const AQuantGemmKernelArgs& kargs,
                                      const std::size_t k_id = blockIdx.z)
         {
             constexpr auto K1   = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
             const index_t K_t   = __builtin_amdgcn_readfirstlane(kargs.k_batch * K1);
             const index_t KRead = __builtin_amdgcn_readfirstlane((kargs.K + K_t - 1) / K_t * K1);
  
             if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
             {
                 a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
             }
             else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
             {
                 a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_A);
             }
  
             if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
             {
                 b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_B);
             }
             else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
             {
                 b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
             }
  
             if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
             {
                 splitted_k = __builtin_amdgcn_readfirstlane(KRead);
             }
             else
             {
                 splitted_k = __builtin_amdgcn_readfirstlane(kargs.K - KRead * (kargs.k_batch - 1));
             }
         }
  
         index_t a_k_split_offset;
         index_t b_k_split_offset;
         index_t splitted_k;
     };
  
     CK_TILE_HOST static bool IsSupportedArgument(const AQuantGemmKernelArgs& kargs)
     {
         if(kargs.k_batch != 1)
         {
             if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
             {
                 CK_TILE_ERROR("Conditions not met for Kbatch >1 !");
             }
             return false;
         }
  
         static_assert(std::is_same_v<AQLayout, tensor_layout::gemm::RowMajor>);
         if(kargs.QK % GemmPipeline::GetVectorSizeAQ() != 0)
         {
             if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
             {
                 CK_TILE_ERROR("K is not a multiple of vector load size for A tensor!");
             }
             return false;
         }
  
         if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
         {
             if(kargs.K % (TilePartitioner::KPerBlock * kargs.k_batch) != 0 &&
                GemmPipeline::kPadK == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("Can't support K that is not a multiple of k_batch * KPerBlock "
                                   "without padding!");
                 }
                 return false;
             }
             if(kargs.K % GemmPipeline::GetVectorSizeA() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("K is not a multiple of vector load size for A tensor!");
                 }
                 return false;
             }
         }
         else
         {
             if(kargs.M % TilePartitioner::MPerBlock != 0 && GemmPipeline::kPadM == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR(
                         "Can't support M that is not a multiple of MPerBlock without padding!");
                 }
                 return false;
             }
             if(kargs.M % GemmPipeline::GetVectorSizeA() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("M is not a multiple of vector load size for A tensor!");
                 }
                 return false;
             }
         }
  
         if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
         {
             if(kargs.N % TilePartitioner::NPerBlock != 0 && GemmPipeline::kPadN == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR(
                         "Can't support N that is not a multiple of NPerBlock without padding!");
                 }
                 return false;
             }
             if(kargs.N % GemmPipeline::GetVectorSizeB() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("N is not a multiple of vector load size for B tensor!");
                 }
                 return false;
             }
         }
         else
         {
             if(kargs.K % (TilePartitioner::KPerBlock * kargs.k_batch) != 0 &&
                GemmPipeline::kPadK == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("Can't support K that is not a multiple of k_batch * KPerBlock "
                                   "without padding!");
                 }
                 return false;
             }
             if(kargs.K % GemmPipeline::GetVectorSizeB() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("K is not a multiple of vector load size for B tensor!");
                 }
                 return false;
             }
         }
  
         if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
         {
             if(kargs.N % TilePartitioner::NPerBlock != 0 && GemmPipeline::kPadN == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR(
                         "Can't support N that is not a multiple of NPerBlock without padding!");
                 }
                 return false;
             }
             if(kargs.N % EpiloguePipeline::GetVectorSizeC() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("N is not a multiple of vector load size for C tensor!");
                 }
                 return false;
             }
         }
         else
         {
             if(kargs.M % TilePartitioner::MPerBlock != 0 && GemmPipeline::kPadM == false)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR(
                         "Can't support M that is not a multiple of MPerBlock without padding!");
                 }
                 return false;
             }
             if(kargs.M % EpiloguePipeline::GetVectorSizeC() != 0)
             {
                 if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                 {
                     CK_TILE_ERROR("M is not a multiple of vector load size for C tensor!");
                 }
                 return false;
             }
         }
         return true;
     }
  
     template <memory_operation_enum DstInMemOp = memory_operation_enum::set>
     CK_TILE_DEVICE static auto MakeGemmTensorViews(const ADataType* a_ptr,
                                                    const BDataType* b_ptr,
                                                    const AQDataType* aq_ptr,
                                                    CDataType* c_ptr,
                                                    const AQuantGemmKernelArgs& kargs,
                                                    const SplitKBatchOffset& splitk_batch_offset)
     {
         static_assert(!TilePartitioner::BlockGemmShape::PermuteA, "Not implemented!");
         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<GemmPipeline::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<GemmPipeline::GetVectorSizeA()>{},
                     number<1>{});
             }
         }();
  
         const auto& aq_tensor_view = [&]() {
             static_assert(std::is_same_v<AQLayout, tensor_layout::gemm::RowMajor>);
             return make_naive_tensor_view<address_space_enum::global>(
                 aq_ptr,
                 make_tuple(kargs.M, kargs.QK),
                 make_tuple(kargs.stride_AQ, 1),
                 number<GemmPipeline::GetVectorSizeAQ()>{},
                 number<1>{});
         }();
  
         const auto& b_tensor_view = [&]() {
             if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
             {
                 if constexpr(TilePartitioner::BlockGemmShape::PermuteB)
                 {
                     constexpr index_t K1          = GemmPipeline::GetSmemPackB();
                     const index_t K0              = splitk_batch_offset.splitted_k / K1;
                     constexpr index_t VectorSizeB = std::min(K1, GemmPipeline::GetVectorSizeB());
                     const auto b_k0_n_k1_desc =
                         make_naive_tensor_descriptor(make_tuple(K0, kargs.N, K1),
                                                      make_tuple(kargs.N * K1, K1, I1),
                                                      number<VectorSizeB>{},
                                                      number<1>{});
                     const auto b_n_k_desc = transform_tensor_descriptor(
                         b_k0_n_k1_desc,
                         make_tuple(make_merge_transform(make_tuple(K0, K1)),
                                    make_pass_through_transform(kargs.N)),
                         make_tuple(sequence<0, 2>{}, sequence<1>{}),
                         make_tuple(sequence<0>{}, sequence<1>{}));
                     return make_tensor_view<address_space_enum::global>(b_ptr, b_n_k_desc);
                 }
                 else
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         b_ptr,
                         make_tuple(splitk_batch_offset.splitted_k, kargs.N),
                         make_tuple(kargs.stride_B, 1),
                         number<GemmPipeline::GetVectorSizeB()>{},
                         number<1>{});
                 }
             }
             else
             {
                 if constexpr(TilePartitioner::BlockGemmShape::PermuteB)
                 {
                     constexpr index_t K1          = GemmPipeline::GetSmemPackB();
                     const index_t K0              = splitk_batch_offset.splitted_k / K1;
                     constexpr index_t VectorSizeB = std::min(K1, GemmPipeline::GetVectorSizeB());
                     const auto b_k0_n_k1_desc =
                         make_naive_tensor_descriptor(make_tuple(K0, kargs.N, K1),
                                                      make_tuple(kargs.N * K1, K1, I1),
                                                      number<VectorSizeB>{},
                                                      number<1>{});
                     const auto b_n_k_desc = transform_tensor_descriptor(
                         b_k0_n_k1_desc,
                         make_tuple(make_merge_transform(make_tuple(K0, K1)),
                                    make_pass_through_transform(kargs.N)),
                         make_tuple(sequence<0, 2>{}, sequence<1>{}),
                         make_tuple(sequence<1>{}, sequence<0>{}));
                     return make_tensor_view<address_space_enum::global>(b_ptr, b_n_k_desc);
                 }
                 else
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         b_ptr,
                         make_tuple(kargs.N, splitk_batch_offset.splitted_k),
                         make_tuple(kargs.stride_B, 1),
                         number<GemmPipeline::GetVectorSizeB()>{},
                         number<1>{});
                 }
             }
         }();
  
         // TODO: enable vector write for C in ColMajor
         const auto& c_tensor_view = [&]() {
             if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     c_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(kargs.stride_C, 1),
                     number<EpiloguePipeline::GetVectorSizeC()>{},
                     number<1>{});
             }
             else
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     c_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(1, kargs.stride_C),
                     number<1>{},
                     number<1>{});
             }
         }();
  
         return make_tuple(a_tensor_view, aq_tensor_view, b_tensor_view, c_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, GemmPipeline::kPadK>{});
             }
             else
             {
                 return pad_tensor_view(a_tensor_view,
                                        make_tuple(number<TilePartitioner::KPerBlock>{},
                                                   number<TilePartitioner::MPerBlock>{}),
                                        sequence<false, GemmPipeline::kPadM>{});
             }
         }();
  
         const auto& aq_pad_view = [&]() {
             const auto& aq_tensor_view = views.at(I1);
             static_assert(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>);
             return pad_tensor_view(
                 aq_tensor_view,
                 make_tuple(number<TilePartitioner::MPerBlock>{},
                            number<TilePartitioner::KPerBlock / GemmPipeline::QuantGroupSize>{}),
                 // TODO: Add support for padding.
                 sequence<false, false>{});
         }();
  
         const auto& b_pad_view = [&]() {
             const auto& b_tensor_view = views.at(I2);
             if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
             {
                 return pad_tensor_view(b_tensor_view,
                                        make_tuple(number<TilePartitioner::NPerBlock>{},
                                                   number<TilePartitioner::KPerBlock>{}),
                                        sequence<false, GemmPipeline::kPadK>{});
             }
             else
             {
                 return pad_tensor_view(b_tensor_view,
                                        make_tuple(number<TilePartitioner::KPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<false, GemmPipeline::kPadN>{});
             }
         }();
  
         // TODO vector write in for C in ColMajor
         const auto& c_pad_view = [&]() {
             const auto& c_tensor_view = views.at(I3);
             if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
             {
                 return pad_tensor_view(c_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<false, GemmPipeline::kPadN>{});
             }
             else
             {
                 return pad_tensor_view(c_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<GemmPipeline::kPadM, false>{});
             }
         }();
  
         return make_tuple(a_pad_view, aq_pad_view, b_pad_view, c_pad_view);
     }
  
     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& aq_pad_view = views.at(I1);
         const auto& b_pad_view  = views.at(I2);
         const auto& c_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& aq_block_window = [&]() {
             static_assert(std::is_same_v<AQLayout, tensor_layout::gemm::RowMajor>);
             return make_tile_window(
                 aq_pad_view,
                 make_tuple(number<TilePartitioner::MPerBlock>{},
                            number<TilePartitioner::KPerBlock / GemmPipeline::QuantGroupSize>{}),
                 {i_m, 0});
         }();
  
         const auto& b_block_window = [&]() {
             if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
             {
                 return make_tile_window(b_pad_view,
                                         make_tuple(number<TilePartitioner::NPerBlock>{},
                                                    number<TilePartitioner::KPerBlock>{}),
                                         {i_n, 0});
             }
             else
             {
                 return make_tile_window(b_pad_view,
                                         make_tuple(number<TilePartitioner::KPerBlock>{},
                                                    number<TilePartitioner::NPerBlock>{}),
                                         {0, i_n});
             }
         }();
  
         auto c_block_window = make_tile_window(
             c_pad_view,
             make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
             {i_m, i_n});
  
         return make_tuple(a_block_window, aq_block_window, b_block_window, c_block_window);
     }
  
     template <memory_operation_enum DstInMemOp = memory_operation_enum::set>
     CK_TILE_DEVICE static void RunGemm(const ADataType* a_ptr,
                                        const BDataType* b_ptr,
                                        const AQDataType* aq_ptr,
                                        CDataType* c_ptr,
                                        void* smem_ptr_0,
                                        const AQuantGemmKernelArgs& 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<DstInMemOp>(
             a_ptr, b_ptr, aq_ptr, c_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 = __builtin_amdgcn_readfirstlane(
             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& aq_block_window = gemm_tile_windows.at(I1);
         const auto& b_block_window  = gemm_tile_windows.at(I2);
  
         const auto& c_block_tile = GemmPipeline{}.template operator()(
             a_block_window, b_block_window, aq_block_window, num_loop, smem_ptr_0);
  
         // Run Epilogue Pipeline
         auto& c_block_window = gemm_tile_windows.at(I3);
  
         EpiloguePipeline{}.template
         operator()<decltype(c_block_window), decltype(c_block_tile), decltype(c_block_window)>(
             c_block_window, c_block_tile, c_block_window, smem_ptr_0);
     }
  
     CK_TILE_DEVICE void operator()(AQuantGemmKernelArgs kargs) const
     {
         const auto blockId  = __builtin_amdgcn_readfirstlane(blockIdx.x);
         const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
         const index_t i_m   = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
         const index_t i_n   = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
  
         const SplitKBatchOffset splitk_batch_offset(kargs);
         // options
         const ADataType* a_ptr   = static_cast<const ADataType*>(kargs.a_ptr);
         const BDataType* b_ptr   = static_cast<const BDataType*>(kargs.b_ptr);
         const AQDataType* aq_ptr = static_cast<const AQDataType*>(kargs.aq_ptr);
         CDataType* c_ptr         = static_cast<CDataType*>(kargs.c_ptr);
  
         // allocate LDS
         __shared__ char smem_ptr_0[GetSmemSize()];
  
         assert(kargs.k_batch == 1);
         RunGemm(a_ptr, b_ptr, aq_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
     }
 };
  
 } // namespace ck_tile