/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/kernel/universal_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/universal_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/universal_gemm_kernel.hpp Source File
Go to the documentation of this file.
 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
  
 #include <iostream>
 #include <string>
  
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/common.hpp"
 #include "ck_tile/host/concat.hpp"
 #include "ck_tile/host/kernel_launch.hpp"
 #include "ck_tile/host/stream_utils.hpp"
 #include "ck_tile/core/utility/env.hpp"
 #include "ck_tile/core/utility/type_traits.hpp"
  
 namespace ck_tile {
  
 template <index_t NumATensor = 1, index_t NumBTensor = 1, index_t NumDTensor = 0>
 struct UniversalGemmHostArgs
 {
     CK_TILE_HOST UniversalGemmHostArgs(const std::array<const void*, NumATensor>& as_ptr_,
                                        const std::array<const void*, NumBTensor>& bs_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_,
                                        const std::array<index_t, NumATensor>& stride_As_,
                                        const std::array<index_t, NumBTensor>& stride_Bs_,
                                        const std::array<index_t, NumDTensor>& stride_Ds_,
                                        index_t stride_E_)
         : as_ptr(as_ptr_),
           bs_ptr(bs_ptr_),
           ds_ptr(ds_ptr_),
           e_ptr(e_ptr_),
           M(M_),
           N(N_),
           K(K_),
           stride_As(stride_As_),
           stride_Bs(stride_Bs_),
           stride_Ds(stride_Ds_),
           stride_E(stride_E_),
           k_batch(k_batch_)
     {
     }
  
     const std::array<const void*, NumATensor> as_ptr;
     const std::array<const void*, NumBTensor> bs_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;
     const std::array<index_t, NumATensor> stride_As;
     const std::array<index_t, NumBTensor> stride_Bs;
     const std::array<index_t, NumDTensor> stride_Ds;
     union
     {
         index_t stride_E;
         index_t stride_C;
     };
  
     index_t k_batch;
 };
  
 template <index_t NumATensor = 1, index_t NumBTensor = 1, index_t NumDTensor = 0>
 struct UniversalGemmKernelArgs
 {
     const std::array<const void*, NumATensor> as_ptr;
     const std::array<const void*, NumBTensor> bs_ptr;
     const std::array<const void*, NumDTensor> ds_ptr;
     void* e_ptr;
     index_t M;
     index_t N;
     index_t K;
     std::array<index_t, NumATensor> stride_As;
     std::array<index_t, NumBTensor> stride_Bs;
     std::array<index_t, NumDTensor> stride_Ds;
     index_t stride_E;
     index_t k_batch;
 };
  
 template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
 struct UniversalGemmKernel
 {
     using TilePartitioner  = remove_cvref_t<TilePartitioner_>;
     using GemmPipeline     = remove_cvref_t<GemmPipeline_>;
     using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
  
     static constexpr bool ADataTypeIsTuple =
         is_detected<is_tuple, typename GemmPipeline::AsDataType>::value;
     static constexpr bool BDataTypeIsTuple =
         is_detected<is_tuple, typename GemmPipeline::BsDataType>::value;
     static constexpr bool DDataTypeIsTuple =
         is_detected<is_tuple, typename EpiloguePipeline::DsDataType>::value;
     static constexpr bool ALayoutIsTuple =
         is_detected<is_tuple, typename GemmPipeline::AsLayout>::value;
     static constexpr bool BLayoutIsTuple =
         is_detected<is_tuple, typename GemmPipeline::BsLayout>::value;
     static constexpr bool DLayoutIsTuple =
         is_detected<is_tuple, typename EpiloguePipeline::DsLayout>::value;
  
     using AsLayout = std::conditional_t<ALayoutIsTuple,
                                         remove_cvref_t<typename GemmPipeline::AsLayout>,
                                         remove_cvref_t<tuple<typename GemmPipeline::ALayout>>>;
     using BsLayout = std::conditional_t<BLayoutIsTuple,
                                         remove_cvref_t<typename GemmPipeline::BsLayout>,
                                         remove_cvref_t<tuple<typename GemmPipeline::BLayout>>>;
  
     using DsLayout = std::conditional_t<DLayoutIsTuple,
                                         remove_cvref_t<typename EpiloguePipeline::DsLayout>,
                                         remove_cvref_t<tuple<typename EpiloguePipeline::DsLayout>>>;
  
     using AsDataType = std::conditional_t<ADataTypeIsTuple,
                                           remove_cvref_t<typename GemmPipeline::AsDataType>,
                                           remove_cvref_t<tuple<typename GemmPipeline::ADataType>>>;
  
     using BsDataType = std::conditional_t<BDataTypeIsTuple,
                                           remove_cvref_t<typename GemmPipeline::BsDataType>,
                                           remove_cvref_t<tuple<typename GemmPipeline::BDataType>>>;
  
     using DsDataType =
         std::conditional_t<DDataTypeIsTuple,
                            remove_cvref_t<typename EpiloguePipeline::DsDataType>,
                            remove_cvref_t<tuple<typename EpiloguePipeline::DsDataType>>>;
  
     using CLayout   = remove_cvref_t<typename GemmPipeline::CLayout>;
     using EDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
  
     using AElementWise = remove_cvref_t<typename GemmPipeline::AElementWise>;
     using BElementWise = remove_cvref_t<typename GemmPipeline::BElementWise>;
  
     static constexpr index_t kBlockSize = GemmPipeline::BlockSize;
  
     // Get the persistent kernel if the pipeline has it available
     struct has_persistent_kernel
     {
         template <typename T>
         using has_persistent_type = decltype(T::UsePersistentKernel);
  
         static constexpr bool value = []() {
             if constexpr(is_detected<has_persistent_type, GemmPipeline>{})
                 return GemmPipeline::UsePersistentKernel;
             else
                 return false;
         }();
     };
     static constexpr bool PersistentKernel = has_persistent_kernel::value;
  
     // Check if TilePartitioner has GetOutputOffset method with kargs and k_id
     struct has_tile_partitioner_output_offset_impl
     {
         template <typename T, typename KernelArgs>
         using has_get_output_offset_t =
             decltype(T::GetOutputOffset(std::declval<KernelArgs>(), std::declval<index_t>()));
  
         static constexpr bool value = []() {
             if constexpr(is_detected<has_get_output_offset_t, TilePartitioner>{})
                 return true;
             else
                 return false;
         }();
     };
     static constexpr bool has_tile_partitioner_output_offset =
         has_tile_partitioner_output_offset_impl::value;
  
     static constexpr auto I0 = number<0>();
     static constexpr auto I1 = number<1>();
     static constexpr auto I2 = number<2>();
     static constexpr auto I3 = number<3>{};
  
     static constexpr index_t NumATensor = AsDataType::size();
     static constexpr index_t NumBTensor = BsDataType::size();
     static constexpr index_t NumDTensor = DsDataType::size();
  
     using ADataType = remove_cvref_t<std::tuple_element_t<I0, AsDataType>>;
     using BDataType = remove_cvref_t<std::tuple_element_t<I0, BsDataType>>;
  
     static_assert(AsLayout::size() == AsDataType::size(),
                   "The size of AsLayout and AsDataType should be the same");
  
     static_assert(BsLayout::size() == BsDataType::size(),
                   "The size of BsLayout and BsDataType should be the same");
  
     static_assert(DsLayout::size() == DsDataType::size(),
                   "The size of DsLayout and DsDataType should be the same");
  
     static_assert(!GemmPipeline::BlockGemmShape::PermuteA, "Not implemented!");
  
     using KernelArgs =
         UniversalGemmKernelArgs<AsLayout::size(), BsLayout::size(), DsLayout::size()>;
  
     [[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 auto MaxOccupancyGridSize(const stream_config& s) -> dim3
     {
         using Kernel      = UniversalGemmKernel<TilePartitioner, GemmPipeline, EpiloguePipeline>;
         const auto kernel = kentry<1, Kernel, KernelArgs>;
         int occupancy;
         ck_tile::hip_check_error(
             hipOccupancyMaxActiveBlocksPerMultiprocessor(&occupancy, kernel, BlockSize().x, 0));
  
         const int grid_size = get_available_compute_units(s) * occupancy;
         return dim3(grid_size, 1, 1);
     }
  
     CK_TILE_HOST static auto BlockSize()
     {
         if(ck_tile::is_wave32())
         {
             return dim3(kBlockSize / 2);
         }
         else
         {
             return dim3(kBlockSize);
         }
     }
  
     CK_TILE_HOST static constexpr KernelArgs
     MakeKernelArgs(const UniversalGemmHostArgs<NumATensor, NumBTensor, NumDTensor>& hostArgs)
     {
         return KernelArgs{hostArgs.as_ptr,
                           hostArgs.bs_ptr,
                           hostArgs.ds_ptr,
                           hostArgs.e_ptr,
                           hostArgs.M,
                           hostArgs.N,
                           hostArgs.K,
                           hostArgs.stride_As,
                           hostArgs.stride_Bs,
                           hostArgs.stride_Ds,
                           hostArgs.stride_E,
                           hostArgs.k_batch};
     }
  
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
         return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
     }
  
     struct SplitKBatchOffset
     {
         // This structure distributes work evenly among splitkk workgroups
         // It's based on a principle that if there is enough work to fill all workgroups,
         // then we can distribute the (K / K1) parts among k_batch workgroups in such a way
         // that each workgroup will be doing ceil((K / K1) / splitk) or ceil((K / K1) / splitk) - 1
         // and leave the potential tail for last(splitk - 1) indexed workgroup.
         __device__ SplitKBatchOffset(const KernelArgs& kargs, const index_t k_id = blockIdx.z)
         {
             constexpr auto K1     = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
             const index_t num_all = amd_wave_read_first_lane(
                 kargs.K / K1); // num of all loops not including potential tail
             index_t num_full = amd_wave_read_first_lane(num_all % kargs.k_batch);
             num_full         = num_full == 0 ? kargs.k_batch : num_full;
  
             const index_t num_full_iters =
                 amd_wave_read_first_lane(std::max(integer_divide_ceil(num_all, kargs.k_batch), 1));
             const index_t full_k_read    = num_full_iters * K1;
             const index_t partial_k_read = (num_full_iters - 1) * K1;
  
             static_for<0, NumATensor, 1>{}([&](auto index) {
                 using AiLayout = remove_cvref_t<std::tuple_element_t<index.value, AsLayout>>;
                 if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, AiLayout>)
                 {
                     as_k_split_offset[index] =
                         amd_wave_read_first_lane(std::min(k_id, num_full) * full_k_read +
                                                  std::max(k_id - num_full, 0) * partial_k_read);
                 }
                 else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, AiLayout>)
                 {
                     as_k_split_offset[index] =
                         amd_wave_read_first_lane((std::min(k_id, num_full) * full_k_read +
                                                   std::max(k_id - num_full, 0) * partial_k_read) *
                                                  kargs.stride_As[index]);
                 }
             });
  
             static_for<0, NumBTensor, 1>{}([&](auto index) {
                 using BiLayout = remove_cvref_t<std::tuple_element_t<index.value, BsLayout>>;
                 if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BiLayout>)
                 {
                     bs_k_split_offset[index] =
                         amd_wave_read_first_lane((std::min(k_id, num_full) * full_k_read +
                                                   std::max(k_id - num_full, 0) * partial_k_read) *
                                                  kargs.stride_Bs[index]);
                 }
                 else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BiLayout>)
                 {
                     bs_k_split_offset[index] =
                         amd_wave_read_first_lane(std::min(k_id, num_full) * full_k_read +
                                                  std::max(k_id - num_full, 0) * partial_k_read);
                 }
             });
  
             if(k_id == kargs.k_batch - 1)
             {
                 splitted_k = kargs.K - std::min(k_id, num_full) * full_k_read -
                              std::max(k_id - num_full, 0) * partial_k_read;
             }
             else if(k_id < num_full)
             {
                 splitted_k = full_k_read;
             }
             else
             {
                 splitted_k = partial_k_read;
             }
         }
  
         std::array<index_t, NumATensor> as_k_split_offset;
         std::array<index_t, NumBTensor> bs_k_split_offset;
         index_t splitted_k;
     };
  
     CK_TILE_HOST static bool IsSupportedArgument(const KernelArgs& kargs)
     {
         if constexpr(EpiloguePipeline::GetVectorSizeC() % 2 != 0 &&
                      is_any_of<EDataType, fp16_t, bf16_t>::value)
         {
             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;
             }
         }
  
         if(kargs.K < GemmPipeline::BlockGemmShape::WarpTile::at(number<2>{}) * kargs.k_batch)
         {
             if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
             {
                 CK_TILE_ERROR("KBatch is too large, part of GPU wouldn't be utilized!");
             }
             return false;
         }
  
         const auto vectorSizeA = is_wave32() ? GemmPipeline::template GetVectorSizeA<true>()
                                              : GemmPipeline::template GetVectorSizeA<false>();
         bool AsTensorIsValid   = {true};
         static_for<0, NumATensor, 1>{}([&](auto index) {
             using AiLayout = remove_cvref_t<std::tuple_element_t<index.value, AsLayout>>;
             if constexpr(std::is_same_v<AiLayout, 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!");
                     }
                     AsTensorIsValid = false;
                 }
                 if(kargs.K % vectorSizeA != 0)
                 {
                     const auto remainder = kargs.K % vectorSizeA;
                     constexpr ck_tile::index_t APackedSize =
                         ck_tile::numeric_traits<ADataType>::PackedSize;
                     const auto remainder_in_bytes = remainder * sizeof(ADataType) / APackedSize;
                     // oob can support to dword level
                     if(remainder_in_bytes % 4 == 0)
                     {
                         AsTensorIsValid = true;
                     }
                     else
                     {
                         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!");
                         }
                         AsTensorIsValid = 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!");
                     }
                     AsTensorIsValid = false;
                 }
                 if(kargs.M % vectorSizeA != 0)
                 {
                     const auto remainder = kargs.M % vectorSizeA;
                     constexpr ck_tile::index_t APackedSize =
                         ck_tile::numeric_traits<ADataType>::PackedSize;
                     const auto remainder_in_bytes = remainder * sizeof(ADataType) / APackedSize;
                     // oob can support to dword level
                     if(remainder_in_bytes % 4 == 0)
                     {
  
                         AsTensorIsValid = true;
                     }
                     else
                     {
                         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!");
                         }
                         AsTensorIsValid = false;
                     }
                 }
             }
         });
  
         bool BsTensorIsValid   = {true};
         const auto vectorSizeB = is_wave32() ? GemmPipeline::template GetVectorSizeB<true>()
                                              : GemmPipeline::template GetVectorSizeB<false>();
         static_for<0, NumBTensor, 1>{}([&](auto index) {
             using BiLayout = remove_cvref_t<std::tuple_element_t<index.value, BsLayout>>;
             if constexpr(std::is_same_v<BiLayout, 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!");
                     }
                     BsTensorIsValid = false;
                 }
                 if(kargs.N % vectorSizeB != 0)
                 {
                     const auto remainder = kargs.N % vectorSizeB;
                     constexpr ck_tile::index_t BPackedSize =
                         ck_tile::numeric_traits<BDataType>::PackedSize;
                     const auto remainder_in_bytes = remainder * sizeof(BDataType) / BPackedSize;
                     // oob can support to dword level
                     if(remainder_in_bytes % 4 == 0)
                     {
                         BsTensorIsValid = true;
                     }
                     else
                     {
                         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!");
                         }
                         BsTensorIsValid = 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!");
                         }
                         BsTensorIsValid = false;
                     }
                     if(kargs.K % vectorSizeB != 0)
                     {
                         const auto remainder = kargs.K % vectorSizeB;
                         constexpr ck_tile::index_t BPackedSize =
                             ck_tile::numeric_traits<BDataType>::PackedSize;
                         const auto remainder_in_bytes = remainder * sizeof(BDataType) / BPackedSize;
                         // oob can support to dword level
                         if(remainder_in_bytes % 4 == 0)
                         {
                             BsTensorIsValid = true;
                         }
                         else
                         {
                             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!");
                             }
                             BsTensorIsValid = false;
                         }
                     }
                 }
             }
         });
  
         bool DTensorIsValid = {true};
         static_for<0, NumDTensor, 1>{}([&](auto index) {
             using DiLayout = remove_cvref_t<std::tuple_element_t<index.value, DsLayout>>;
             if(std::is_same_v<DiLayout, CLayout> == false)
             {
                 DTensorIsValid = false;
             }
             if constexpr(std::is_same_v<DiLayout, 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 for tensor D that is not a multiple of "
                                       "NPerBlock without padding!");
                     }
                     DTensorIsValid = false;
                 }
                 if(kargs.N % EpiloguePipeline::GetVectorSizeD(index) != 0)
                 {
                     if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                     {
                         CK_TILE_ERROR("N is not a multiple of vector load size for D tensor!");
                     }
                     DTensorIsValid = 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 for tensor D that is not a multiple of "
                                       "MPerBlock without padding!");
                     }
                     DTensorIsValid = false;
                 }
                 if(kargs.M % EpiloguePipeline::GetVectorSizeD(index) != 0)
                 {
                     if(ck_tile::EnvIsEnabled(CK_TILE_ENV(CK_TILE_LOGGING)))
                     {
                         CK_TILE_ERROR("M is not a multiple of vector load size for D tensor!");
                     }
                     DTensorIsValid = 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 AsTensorIsValid && BsTensorIsValid && DTensorIsValid;
     }
  
     CK_TILE_DEVICE static auto
     MakeABlockWindows(const std::array<const ADataType*, NumATensor>& as_ptr,
                       const KernelArgs& kargs,
                       const index_t k_size,
                       const index_t i_m)
     {
         // Step 1: Create tensor views for A tensors (from MakeGemmTensorViews)
         const auto& as_tensor_view = generate_tuple(
             [&](auto i) {
                 using AiLayout   = remove_cvref_t<std::tuple_element_t<i.value, AsLayout>>;
                 using AiDataType = remove_cvref_t<std::tuple_element_t<i.value, AsDataType>>;
                 if constexpr(std::is_same_v<AiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const AiDataType*>(as_ptr[i]),
                         make_tuple(kargs.M, k_size),
                         make_tuple(kargs.stride_As[i], 1),
                         number<GemmPipeline::GetVectorSizeA()>{},
                         number<1>{});
                 }
                 else
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const AiDataType*>(as_ptr[i]),
                         make_tuple(k_size, kargs.M),
                         make_tuple(kargs.stride_As[i], 1),
                         number<GemmPipeline::GetVectorSizeA()>{},
                         number<1>{});
                 }
             },
             number<NumATensor>{});
  
         // Step 2: Create padded views (from MakeGemmPadViews)
         const auto& as_pad_view = generate_tuple(
             [&](auto i) {
                 using AiLayout = remove_cvref_t<std::tuple_element_t<i.value, AsLayout>>;
                 if constexpr(std::is_same_v<AiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return pad_tensor_view(as_tensor_view[i],
                                            make_tuple(number<TilePartitioner::MPerBlock>{},
                                                       number<TilePartitioner::KPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadK>{});
                 }
                 else
                 {
                     return pad_tensor_view(as_tensor_view[i],
                                            make_tuple(number<TilePartitioner::KPerBlock>{},
                                                       number<TilePartitioner::MPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadM>{});
                 }
             },
             number<NumATensor>{});
  
         // Step 3: Create tile windows (from MakeGemmTileWindows)
         const auto& as_block_window = generate_tuple(
             [&](auto i) {
                 using AiLayout = remove_cvref_t<std::tuple_element_t<i.value, AsLayout>>;
                 if constexpr(std::is_same_v<AiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_tile_window(as_pad_view[i],
                                             make_tuple(number<TilePartitioner::MPerBlock>{},
                                                        number<TilePartitioner::KPerBlock>{}),
                                             {i_m, 0});
                 }
                 else
                 {
                     return make_tile_window(as_pad_view[i],
                                             make_tuple(number<TilePartitioner::KPerBlock>{},
                                                        number<TilePartitioner::MPerBlock>{}),
                                             {0, i_m});
                 }
             },
             number<NumATensor>{});
  
         return as_block_window;
     }
  
     CK_TILE_DEVICE static auto
     MakeBBlockWindows(const std::array<const BDataType*, NumBTensor>& bs_ptr,
                       const KernelArgs& kargs,
                       const index_t k_size,
                       const index_t i_n)
     {
         // Step 1: Create tensor views for B tensors (from MakeGemmTensorViews)
         const auto& bs_tensor_view = generate_tuple(
             [&](auto i) {
                 using BiLayout   = remove_cvref_t<std::tuple_element_t<i.value, BsLayout>>;
                 using BiDataType = remove_cvref_t<std::tuple_element_t<i.value, BsDataType>>;
                 if constexpr(std::is_same_v<BiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     if constexpr(GemmPipeline::BlockGemmShape::PermuteB)
                     {
                         constexpr index_t K1 = GemmPipeline::GetSmemPackB();
                         const index_t K0     = k_size / 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>(
                             static_cast<const BiDataType*>(bs_ptr[i]), b_n_k_desc);
                     }
                     else
                     {
                         return make_naive_tensor_view<address_space_enum::global>(
                             bs_ptr[i],
                             make_tuple(k_size, kargs.N),
                             make_tuple(kargs.stride_Bs[i], 1),
                             number<GemmPipeline::GetVectorSizeB()>{},
                             number<1>{});
                     }
                 }
                 else
                 {
                     if constexpr(GemmPipeline::BlockGemmShape::PermuteB)
                     {
                         constexpr index_t K1 = GemmPipeline::GetSmemPackB();
                         const index_t K0     = k_size / 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>(
                             static_cast<const BiDataType*>(bs_ptr[i]), b_n_k_desc);
                     }
                     else
                     {
                         if constexpr(GemmPipeline::Preshuffle)
                         {
                             index_t kFlatK =
                                 GemmPipeline::BlockGemmShape::flatKPerWarp *
                                 (k_size / GemmPipeline::BlockGemmShape::WarpTile::at(number<2>{}));
                             index_t kFlatN = kargs.N * kargs.K / kFlatK;
  
                             return make_naive_tensor_view<address_space_enum::global>(
                                 bs_ptr[i],
                                 make_tuple(kFlatN, kFlatK),
                                 make_tuple(kFlatK, 1),
                                 number<GemmPipeline::GetVectorSizeB()>{},
                                 number<1>{});
                         }
                         else
                         {
                             return make_naive_tensor_view<address_space_enum::global>(
                                 bs_ptr[i],
                                 make_tuple(kargs.N, k_size),
                                 make_tuple(kargs.stride_Bs[i], 1),
                                 number<GemmPipeline::GetVectorSizeB()>{},
                                 number<1>{});
                         }
                     }
                 }
             },
             number<NumBTensor>{});
  
         // Step 2: Create padded views (from MakeGemmPadViews)
         const auto& bs_pad_view = generate_tuple(
             [&](auto i) {
                 using BiLayout = remove_cvref_t<std::tuple_element_t<i.value, BsLayout>>;
                 if constexpr(std::is_same_v<BiLayout, tensor_layout::gemm::ColumnMajor>)
                 {
                     return pad_tensor_view(bs_tensor_view[i],
                                            make_tuple(number<TilePartitioner::NPerBlock>{},
                                                       number<TilePartitioner::KPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadK>{});
                 }
                 else
                 {
                     return pad_tensor_view(bs_tensor_view[i],
                                            make_tuple(number<TilePartitioner::KPerBlock>{},
                                                       number<TilePartitioner::NPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadN>{});
                 }
             },
             number<NumBTensor>{});
  
         // Step 3: Create tile windows (from MakeGemmTileWindows)
         const auto& bs_block_window = generate_tuple(
             [&](auto i) {
                 using BiLayout = remove_cvref_t<std::tuple_element_t<i.value, BsLayout>>;
                 if constexpr(GemmPipeline::Preshuffle)
                 {
                     return make_tile_window(
                         bs_pad_view[i],
                         make_tuple(number<GemmPipeline::BlockGemmShape::flatNPerWarp>{},
                                    number<GemmPipeline::BlockGemmShape::flatKPerWarp>{}),
                         {static_cast<int>(i_n / GemmPipeline::BlockGemmShape::WarpTile::at(I1)),
                          0});
                 }
                 else
                 {
                     if constexpr(std::is_same_v<BiLayout, tensor_layout::gemm::ColumnMajor>)
                     {
                         return make_tile_window(bs_pad_view[i],
                                                 make_tuple(number<TilePartitioner::NPerBlock>{},
                                                            number<TilePartitioner::KPerBlock>{}),
                                                 {i_n, 0});
                     }
                     else
                     {
                         return make_tile_window(bs_pad_view[i],
                                                 make_tuple(number<TilePartitioner::KPerBlock>{},
                                                            number<TilePartitioner::NPerBlock>{}),
                                                 {0, i_n});
                     }
                 }
             },
             number<NumBTensor>{});
  
         return bs_block_window;
     }
  
     CK_TILE_DEVICE static auto MakeDBlockWindows(const std::array<const void*, NumDTensor>& ds_ptr,
                                                  const KernelArgs& kargs,
                                                  const index_t i_m,
                                                  const index_t i_n)
     {
         // Step 1: Create tensor views for D tensors (from MakeGemmTensorViews)
         const auto& ds_tensor_view = generate_tuple(
             [&](auto i) {
                 using DiLayout   = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 using DDataType_ = remove_cvref_t<std::tuple_element_t<i.value, DsDataType>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const DDataType_*>(ds_ptr[i]),
                         make_tuple(kargs.M, kargs.N),
                         make_tuple(kargs.stride_Ds[i], 1),
                         number<EpiloguePipeline::GetVectorSizeD(i)>{},
                         number<1>{});
                 }
                 else
                 {
                     return make_naive_tensor_view<address_space_enum::global>(
                         static_cast<const DDataType_*>(ds_ptr[i]),
                         make_tuple(kargs.N, kargs.M),
                         make_tuple(kargs.stride_Ds[i], 1),
                         number<EpiloguePipeline::GetVectorSizeD(i)>{},
                         number<1>{});
                 }
             },
             number<NumDTensor>{});
  
         // Step 2: Create padded views (from MakeGemmPadViews)
         const auto& ds_pad_view = generate_tuple(
             [&](auto i) {
                 using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return pad_tensor_view(ds_tensor_view[i],
                                            make_tuple(number<TilePartitioner::MPerBlock>{},
                                                       number<TilePartitioner::NPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadN>{});
                 }
                 else
                 {
                     return pad_tensor_view(ds_tensor_view[i],
                                            make_tuple(number<TilePartitioner::NPerBlock>{},
                                                       number<TilePartitioner::MPerBlock>{}),
                                            sequence<false, GemmPipeline::kPadM>{});
                 }
             },
             number<NumDTensor>{});
  
         // Step 3: Create tile windows (from MakeGemmTileWindows)
         const auto& ds_block_window = generate_tuple(
             [&](auto i) {
                 using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
                 if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
                 {
                     return make_tile_window(ds_pad_view[i],
                                             make_tuple(number<TilePartitioner::MPerBlock>{},
                                                        number<TilePartitioner::NPerBlock>{}),
                                             {i_m, i_n});
                 }
                 else
                 {
                     return make_tile_window(ds_pad_view[i],
                                             make_tuple(number<TilePartitioner::NPerBlock>{},
                                                        number<TilePartitioner::MPerBlock>{}),
                                             {i_n, i_m});
                 }
             },
             number<NumDTensor>{});
  
         return ds_block_window;
     }
  
     template <memory_operation_enum DstInMemOp = memory_operation_enum::set>
     CK_TILE_DEVICE static auto MakeCBlockWindows(EDataType* e_ptr,
                                                  const KernelArgs& kargs,
                                                  const index_t i_m,
                                                  const index_t i_n)
     {
         // Step 1: Create tensor view for E/C tensor (from MakeGemmTensorViews)
         const auto& e_tensor_view = [&]() {
             if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     e_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(kargs.stride_E, 1),
                     number<EpiloguePipeline::GetVectorSizeC()>{},
                     number<1>{});
             }
             else
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     e_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(1, kargs.stride_E),
                     number<1>{},
                     number<1>{});
             }
         }();
  
         // Step 2: Create padded view (from MakeGemmPadViews)
         const auto& e_pad_view = [&]() {
             if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
             {
                 return pad_tensor_view(e_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<false, GemmPipeline::kPadN>{});
             }
             else
             {
                 return pad_tensor_view(e_tensor_view,
                                        make_tuple(number<TilePartitioner::MPerBlock>{},
                                                   number<TilePartitioner::NPerBlock>{}),
                                        sequence<GemmPipeline::kPadM, false>{});
             }
         }();
  
         // Step 3: Create tile window (from MakeGemmTileWindows)
         auto e_block_window = make_tile_window(
             e_pad_view,
             make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
             {i_m, i_n});
  
         return e_block_window;
     }
  
     template <bool UseDefaultScheduler = true>
     CK_TILE_DEVICE static void RunGemm(const std::array<const ADataType*, NumATensor>& as_ptr,
                                        const std::array<const BDataType*, NumBTensor>& bs_ptr,
                                        const std::array<const void*, NumDTensor>& ds_ptr,
                                        EDataType* e_ptr,
                                        void* smem_ptr,
                                        const KernelArgs& kargs,
                                        const SplitKBatchOffset& splitk_batch_offset,
                                        const index_t block_idx_m,
                                        const index_t block_idx_n)
     {
         // Create block windows using specialized methods
         const auto& as_block_window =
             MakeABlockWindows(as_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_m);
         const auto& bs_block_window =
             MakeBBlockWindows(bs_ptr, kargs, splitk_batch_offset.splitted_k, block_idx_n);
         const auto& ds_block_window = MakeDBlockWindows(ds_ptr, kargs, block_idx_m, block_idx_n);
  
         const index_t num_loop =
             amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
  
         // Run GEMM cooperatively by whole workgroup.
         const auto& c_block_tile = GemmPipeline{}.template operator()(
             as_block_window, AElementWise{}, bs_block_window, BElementWise{}, num_loop, smem_ptr);
  
         const index_t k_batch = amd_wave_read_first_lane(kargs.k_batch);
         // Run Epilogue Pipeline
         if(k_batch == 1)
         {
             auto c_block_window = MakeCBlockWindows<memory_operation_enum::set>(
                 e_ptr, kargs, block_idx_m, block_idx_n);
             EpiloguePipeline{}(c_block_window, c_block_tile, ds_block_window, smem_ptr);
         }
         else
         {
             auto c_block_window = MakeCBlockWindows<memory_operation_enum::atomic_add>(
                 e_ptr, kargs, block_idx_m, block_idx_n);
             EpiloguePipeline{}(c_block_window, c_block_tile, ds_block_window, smem_ptr);
         }
     }
  
     CK_TILE_DEVICE static auto
     GetTileCoordinates(const KernelArgs& kargs) -> tuple<index_t, index_t>
     {
         index_t iM, iN;
  
         // Regular launch: use 1D block indexing
         const auto blockId          = amd_wave_read_first_lane(blockIdx.x);
         const auto [tile_m, tile_n] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
         iM                          = tile_m;
         iN                          = tile_n;
  
         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);
  
         return make_tuple(i_m, i_n);
     }
  
     // Helper functions
     CK_TILE_DEVICE static auto GetBlockId() -> index_t
     {
         // For 1D regular launch
         return amd_wave_read_first_lane(get_block_id());
     }
  
     CK_TILE_DEVICE static auto GetGridSize() -> index_t
     {
         // For 1D regular launch
         return amd_wave_read_first_lane(get_grid_size());
     }
  
     // Helper to get total number of tiles, handling both dim3 and index_t return types
     template <typename... Args>
     CK_TILE_HOST_DEVICE static auto GetNumTiles(Args&&... args) -> index_t
     {
         auto grid_size = TilePartitioner::GridSize(std::forward<Args>(args)...);
  
         using GridSizeType = decltype(grid_size);
  
         if constexpr(std::is_same_v<GridSizeType, dim3>)
         {
             // GridSize returns dim3: compute total tiles as x * y * z
             return amd_wave_read_first_lane(grid_size.x * grid_size.y * grid_size.z);
         }
         else
         {
             // GridSize returns scalar (index_t): use directly
             return amd_wave_read_first_lane(grid_size);
         }
     }
  
     // Non-persistent kernel entry point
     template <bool U = !PersistentKernel, typename = std::enable_if_t<U>>
     CK_TILE_DEVICE void operator()(KernelArgs kargs) const
     {
         const auto blockId  = amd_wave_read_first_lane(blockIdx.x);
         const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
         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 SplitKBatchOffset splitk_batch_offset(kargs);
  
         // options
         std::array<const ADataType*, NumATensor> as_ptr;
         static_for<0, NumATensor, 1>{}([&](auto i) {
             as_ptr[i] = static_cast<const ADataType*>(kargs.as_ptr[i]) +
                         splitk_batch_offset.as_k_split_offset[i];
         });
  
         std::array<const BDataType*, NumBTensor> bs_ptr;
         static_for<0, NumBTensor, 1>{}([&](auto i) {
             bs_ptr[i] = static_cast<const BDataType*>(kargs.bs_ptr[i]) +
                         splitk_batch_offset.bs_k_split_offset[i];
         });
  
         // Calculate output offset from tile partitioner and apply to output pointer
         EDataType* e_ptr = static_cast<EDataType*>(kargs.e_ptr);
         if constexpr(has_tile_partitioner_output_offset)
         {
             const index_t output_offset = TilePartitioner::GetOutputOffset(kargs, blockIdx.z);
             e_ptr += output_offset;
         }
  
         // allocate LDS
         __shared__ char smem_ptr[GetSmemSize()];
  
         constexpr auto scheduler_type =
             GemmPipeline::DoubleSmemBuffer || (GemmPipeline::NumWaveGroups == 1);
         RunGemm<scheduler_type>(
             as_ptr, bs_ptr, kargs.ds_ptr, e_ptr, smem_ptr, kargs, splitk_batch_offset, i_m, i_n);
     }
  
     // Persistent kernel entry point
     template <bool U = PersistentKernel, typename = std::enable_if_t<U>, typename = void>
     CK_TILE_DEVICE void operator()(KernelArgs kargs) const
     {
         const auto grid_size = amd_wave_read_first_lane(get_grid_size());
         const auto num_tiles =
             amd_wave_read_first_lane(TilePartitioner::GridSize(kargs.M, kargs.N));
         const auto num_work = amd_wave_read_first_lane(num_tiles * kargs.k_batch);
         auto block_id       = amd_wave_read_first_lane(get_block_id());
  
         while(block_id < num_work)
         {
             s_waitcnt_barrier();
             // Get the tile index for this block
             const auto tile_idx = amd_wave_read_first_lane(block_id % num_tiles);
             const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(tile_idx);
             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);
  
             // Get the SplitK offset for this block
             const auto k_batch = amd_wave_read_first_lane(block_id / num_tiles);
             const SplitKBatchOffset splitk_batch_offset(kargs, k_batch);
  
             std::array<const ADataType*, NumATensor> as_ptr;
             static_for<0, NumATensor, 1>{}([&](auto i) {
                 as_ptr[i] = static_cast<const ADataType*>(kargs.as_ptr[i]) +
                             splitk_batch_offset.as_k_split_offset[i];
             });
  
             std::array<const BDataType*, NumBTensor> bs_ptr;
             static_for<0, NumBTensor, 1>{}([&](auto i) {
                 bs_ptr[i] = static_cast<const BDataType*>(kargs.bs_ptr[i]) +
                             splitk_batch_offset.bs_k_split_offset[i];
             });
  
             // Calculate output offset from tile partitioner and apply to output pointer
             EDataType* e_ptr = static_cast<EDataType*>(kargs.e_ptr);
             if constexpr(has_tile_partitioner_output_offset)
             {
                 const index_t output_offset = TilePartitioner::GetOutputOffset(kargs, k_batch);
                 e_ptr += output_offset;
             }
  
             // allocate LDS
             __shared__ char smem_ptr[GetSmemSize()];
             // Run the GEMM
  
             RunGemm(as_ptr,
                     bs_ptr,
                     kargs.ds_ptr,
                     e_ptr,
                     smem_ptr,
                     kargs,
                     splitk_batch_offset,
                     i_m,
                     i_n);
  
             // Advance to the next work item
             block_id += grid_size;
             if(block_id >= num_work)
             {
                 break;
             }
         }
     }
 };
 } // namespace ck_tile