Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into add_fp16_wmma_conv_instance

include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp

@@ -375,10 +375,12 @@ struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
         remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
     using DefaultBGridDesc_BK0_N_BK1 =
         remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
-    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
-        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
-    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
 
     using DefaultBlock2ETileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}, 1))>;

include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType, // FIXME: don't assume A/B have same datatype
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    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 auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(ADataType) +
+                             b_block_space_size_aligned * sizeof(BDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto AK0 = KPad / (KBatch * AK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeType,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeType,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp

@@ -138,8 +138,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
@@ -308,8 +308,9 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
             c_grid_desc_m_n);
     }
 
-    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp

@@ -35,13 +35,17 @@ __global__ void
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
-template <typename GridwiseGemm, typename FloatAB, typename FloatC, bool HasMainKBlockLoop>
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          bool HasMainKBlockLoop>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdl_cshuffle_v1(const FloatAB* __restrict__ p_a_grid,
-                                    const FloatAB* __restrict__ p_b_grid,
+        kernel_gemm_xdl_cshuffle_v1(const FloatA* __restrict__ p_a_grid,
+                                    const FloatB* __restrict__ p_b_grid,
                                     FloatC* __restrict__ p_c_grid,
                                     typename GridwiseGemm::Problem problem)
 {
@@ -61,7 +65,8 @@ __global__ void
 template <typename ALayout,
           typename BLayout,
           typename CLayout,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatGemmAcc,
           typename FloatCShuffle,
           typename FloatC,
@@ -102,7 +107,8 @@ template <typename ALayout,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched,
-          PipelineVersion PipelineVer = PipelineVersion::v1>
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = FloatC>
 struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 {
     static constexpr auto I0 = Number<0>{};
@@ -463,8 +469,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
     // Argument
     struct Argument : public tensor_operation::device::BaseArgument, public Problem
     {
-        __host__ Argument(const FloatAB* p_a_grid_,
-                          const FloatAB* p_b_grid_,
+        __host__ Argument(const FloatA* p_a_grid_,
+                          const FloatB* p_b_grid_,
                           FloatC* p_c_grid_,
                           index_t M_,
                           index_t N_,
@@ -479,14 +485,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         {
         }
 
-        const FloatAB* p_a_grid;
-        const FloatAB* p_b_grid;
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
         FloatC* p_c_grid;
     };
 
     // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
     __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
@@ -541,8 +547,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         constexpr auto c_block_size =
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
-        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
-                             sizeof(FloatAB),
+        return math::max((a_block_space_size_aligned * sizeof(ComputeType) +
+                          b_block_space_size_aligned * sizeof(ComputeType)),
                          c_block_size * sizeof(FloatCShuffle));
     }
 
@@ -676,8 +682,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
     using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
 
     template <bool HasMainKBlockLoop>
-    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
-                               const FloatAB* __restrict__ p_b_grid,
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
                                FloatC* __restrict__ p_c_grid,
                                void* __restrict__ p_shared,
                                const Problem& problem)
@@ -743,8 +749,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 Sequence<AK0Number, MPerBlock, AK1Number>,
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                FloatA,
+                                                ComputeType,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -774,8 +780,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 Sequence<BK0Number, NPerBlock, BK1Number>,
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                FloatB,
+                                                ComputeType,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -805,11 +811,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         // sanity check
         constexpr index_t KPack =
             math::max(math::lcm(AK1Number, BK1Number),
-                      MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            FloatAB,
+            ComputeType,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -827,10 +833,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp

@@ -173,8 +173,8 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
@@ -345,8 +345,9 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
             c_grid_desc_m_n);
     }
 
-    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
 
     using C0GridDescriptor_NBlock_NPerBlock =
         remove_cvref_t<decltype(MakeC0GridDescriptor_NBlock_NPerBlock(C0GridDesc_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp

@@ -330,8 +330,9 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle
         return e_grid_desc_mblock_mperblock_nblock_nperblock;
     }
 
-    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
 
     using DefaultBlock2ETileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp

@@ -259,8 +259,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     // denorm test fix, required to work around fp16 mfma issue
     // we convert fp16->fp32->bf16 and execute bf16 mfma instruction

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp"
+#include "ck/utility/workgroup_barrier.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_streamk(const typename GridwiseGemm::FloatAB* p_a_grid,
+                                   const typename GridwiseGemm::FloatAB* p_b_grid,
+                                   typename GridwiseGemm::FloatC* p_c_grid,
+                                   void* p_workspace,
+                                   index_t M,
+                                   index_t N,
+                                   index_t K,
+                                   index_t StrideA,
+                                   index_t StrideB,
+                                   index_t StrideC,
+                                   typename GridwiseGemm::Block2CTileMap block_mapping)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+
+    __shared__ uint8_t p_shared[shared_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_workspace,
+                      M,
+                      N,
+                      K,
+                      StrideA,
+                      StrideB,
+                      StrideC,
+                      block_mapping,
+                      static_cast<void*>(p_shared));
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_workspace;
+    ignore = M;
+    ignore = N;
+    ignore = K;
+    ignore = StrideA;
+    ignore = StrideB;
+    ignore = StrideC;
+    ignore = block_mapping;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename Block2CTileMap_,
+          typename FloatAB_,
+          typename FloatAcc_,
+          typename FloatC_,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk
+{
+    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 auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1        = Number<K1Value>{};
+    static constexpr auto M01       = 1;
+    static constexpr auto N01       = 1;
+    static constexpr auto KPerBlock = K0PerBlock * K1;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+    using FloatAcc        = FloatAcc_;
+    using FloatCShuffle   = FloatAcc;
+
+    using Block2CTileMap = Block2CTileMap_;
+    using FloatAB        = FloatAB_;
+    using FloatC         = FloatC_;
+
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        const FloatAB* p_a_grid;
+        const FloatAB* p_b_grid;
+        FloatC* p_c_grid;
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        Block2CTileMap block_mapping;
+
+        Argument(const FloatAB* p_a_grid_,
+                 const FloatAB* p_b_grid_,
+                 FloatC* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
+                 uint32_t num_cu,
+                 uint32_t occupancy,
+                 uint32_t num_sk_blocks_)
+            : p_a_grid(p_a_grid_),
+              p_b_grid(p_b_grid_),
+              p_c_grid(p_c_grid_),
+              M(M_),
+              N(N_),
+              K(K_),
+              StrideA(StrideA_),
+              StrideB(StrideB_),
+              StrideC(StrideC_),
+              block_mapping(M, N, K, num_cu, occupancy, num_sk_blocks_)
+        {
+        }
+
+        void Print() const
+        {
+            std::cout << "arg {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << std::endl;
+        }
+    };
+
+    __host__ __device__ static auto CalculateGridSize(const Argument& karg)
+    {
+        return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
+                               math::integer_divide_ceil(karg.M, MPerBlock),
+                               karg.k_batch);
+    }
+
+    __host__ __device__ static auto CalculateK0(index_t KPad) { return KPad / K1; }
+
+    __host__ __device__ static auto
+    MakeAGridDescriptor_K0_M_K1(index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(a_grid_desc_m_kpad,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(M, MPad - M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto
+    MakeBGridDescriptor_K0_N_K1(index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(b_grid_desc_kpad_n,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        return transform_tensor_descriptor(c_grid_desc_m_n,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * K1, K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * K1, K1, I1));
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle().GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat / CShuffleMRepeatPerShuffle>{},
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       Number<NRepeat / CShuffleNRepeatPerShuffle>{},
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto GetClusterLengthReduction()
+    {
+        // TODO: assume C is row major
+        // TODO: we always first loop over N, then M
+        constexpr auto NPerBlockPow2 = math::next_power_of_two<NPerBlock>();
+        constexpr auto NPerBlockReduction =
+            NPerBlockPow2 / CBlockTransferScalarPerVector_NWaveNPerXDL;
+        constexpr auto MPerBlockReduction =
+            (BlockSize + NPerBlockReduction - 1) / NPerBlockReduction;
+        return Sequence<MPerBlockReduction, NPerBlockReduction>{};
+    }
+
+    __host__ __device__ static constexpr auto GetPartialAccBlockDescriptor()
+    {
+        const auto c_partial_acc_block_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(NPerBlock, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(I1, MPerBlock));
+            }
+        }();
+        return c_partial_acc_block_m_n;
+    }
+
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
+
+    __device__ static void Run(const FloatAB* p_a_grid,
+                               const FloatAB* p_b_grid,
+                               FloatC* p_c_grid,
+                               void* p_workspace,
+                               index_t M,
+                               index_t N,
+                               index_t K,
+                               index_t StrideA,
+                               index_t StrideB,
+                               index_t StrideC,
+                               Block2CTileMap block_mapping,
+                               void* __restrict__ p_shared_block)
+    {
+        uint32_t m        = M;
+        uint32_t n        = N;
+        uint32_t k        = K;
+        uint32_t pad_m    = (m + MPerBlock - 1) / MPerBlock * MPerBlock;
+        uint32_t pad_n    = (n + NPerBlock - 1) / NPerBlock * NPerBlock;
+        uint32_t pad_k    = (k + KPerBlock - 1) / KPerBlock * KPerBlock;
+        uint32_t stride_a = StrideA;
+        uint32_t stride_b = StrideB;
+        uint32_t stride_c = StrideC;
+
+        const auto a_k0_m_k1_grid_desc = MakeAGridDescriptor_K0_M_K1(m, pad_m, k, pad_k, stride_a);
+        const auto b_k0_n_k1_grid_desc = MakeBGridDescriptor_K0_N_K1(k, pad_k, n, pad_n, stride_b);
+        const auto c_grid_desc_m_n     = MakeCGridDescriptor_M_N(m, pad_m, n, pad_n, stride_c);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
+        const AElementwiseOperation a_element_op = AElementwiseOperation{};
+        const BElementwiseOperation b_element_op = BElementwiseOperation{};
+        const CElementwiseOperation c_element_op = CElementwiseOperation{};
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
+        FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_v3();
+
+        uint32_t block_idx = block_mapping.get_block_idx();
+        bool is_sk_block   = block_idx < block_mapping.sk_num_blocks;
+        bool is_dp_block   = block_idx >= block_mapping.dp_start_block_idx &&
+                           block_idx < block_mapping.reduction_start_block_idx;
+        bool is_reduction_block = block_idx >= block_mapping.reduction_start_block_idx;
+        bool is_padding_block   = block_idx >= block_mapping.sk_num_blocks &&
+                                block_idx < block_mapping.dp_start_block_idx;
+        uint32_t iter_start, iter_end;
+        block_mapping.get_block_itr(block_idx, iter_start, iter_end);
+        uint32_t total_iter_length = iter_end - iter_start;
+
+        if(is_padding_block)
+            return;
+
+        uint32_t* p_semaphore =
+            reinterpret_cast<uint32_t*>(reinterpret_cast<char*>(p_workspace) +
+                                        block_mapping.get_workspace_size_for_acc(sizeof(FloatAcc)));
+
+        if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            if(is_reduction_block)
+            {
+                // descriptors
+                constexpr auto cluster_length_reduce = GetClusterLengthReduction();
+                constexpr auto reduce_desc = make_cluster_descriptor(cluster_length_reduce);
+                const auto reduce_thread_cluster_idx =
+                    reduce_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+                const auto thread_m_cluster_id = reduce_thread_cluster_idx[I0];
+                const auto thread_n_cluster_id = reduce_thread_cluster_idx[I1];
+
+                constexpr auto MReduceIters =
+                    math::integer_divide_ceil(Number<MPerBlock>{}, cluster_length_reduce.At(I0));
+                constexpr auto NReduceIters = math::integer_divide_ceil(
+                    Number<NPerBlock>{},
+                    cluster_length_reduce.At(I1) *
+                        Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{});
+
+                constexpr auto acc_thread_buf_load_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+                constexpr auto acc_thread_buf_store_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, I1, I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+
+                constexpr auto c_partial_acc_block_m_n = GetPartialAccBlockDescriptor();
+
+                constexpr auto partial_acc_load_step_n = make_multi_index(
+                    0, cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_n_reverse =
+                    make_multi_index(0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_m =
+                    make_multi_index(cluster_length_reduce.At(I0), 0);
+
+                constexpr auto partial_acc_store_step_n = make_multi_index(
+                    0,
+                    0,
+                    0,
+                    cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_n_reverse =
+                    make_multi_index(0,
+                                     0,
+                                     0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_m =
+                    make_multi_index(0, cluster_length_reduce.At(I0), 0, 0);
+
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    parcial_acc_buf;
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    acc_buf;
+
+                // start to compute
+                auto reduction_idx = blockIdx.x - block_mapping.reduction_start_block_idx;
+                auto spatial_idx   = block_mapping.tile_to_spatial(reduction_idx, m, n);
+
+                workgroup_barrier wg_barrier(p_semaphore);
+
+                uint32_t tile_acc_offset_start =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx);
+                uint32_t tile_acc_offset_end =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx + 1);
+
+                auto acc_load = ThreadwiseTensorSliceTransfer_v2<
+                    FloatAcc,                                                // SrcData,
+                    FloatAcc,                                                // DstData,
+                    decltype(c_partial_acc_block_m_n),                       // SrcDesc,
+                    decltype(acc_thread_buf_load_desc),                      // DstDesc,
+                    Sequence<1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1>,                                          // DimAccessOrder,
+                    1,                                                       // SrcVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL,              // SrcScalarPerVector,
+                    1,    // SrcScalarStrideInVector,
+                    false // SrcResetCoordinateAfterRun,
+                    >{c_partial_acc_block_m_n,
+                      make_multi_index(thread_m_cluster_id,
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL)};
+
+                auto acc_store = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,                                                // SrcData,
+                    FloatC,                                                  // DstData,
+                    decltype(acc_thread_buf_store_desc),                     // SrcDesc,
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock), // DstDesc,
+                    CElementwiseOperation, // ElementwiseOperation,
+                    Sequence<1, 1, 1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1, 2, 3>,                       // DimAccessOrder,
+                    3,                                          // DstVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // DstScalarPerVector,
+                    InMemoryDataOperationEnum::Set, // InMemoryDataOperationEnum DstInMemOp,
+                    1,                              // DstScalarStrideInVector,
+                    false                           // DstResetCoordinateAfterRun,
+                    >{c_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                       thread_m_cluster_id,
+                                       __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL),
+                      CElementwiseOperation{}};
+
+                // block synchronization
+                wg_barrier.wait_eq(reduction_idx, tile_acc_offset_end - tile_acc_offset_start);
+
+#if 0
+                if(threadIdx.x == 0) {
+                    printf("bid:%d, rid:%d, os:%d,%d, spatial:%d,%d\n", static_cast<int>(blockIdx.x),
+                        reduction_idx, __builtin_amdgcn_readfirstlane(tile_acc_offset_start), __builtin_amdgcn_readfirstlane(tile_acc_offset_end),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I1]));
+                }
+#endif
+
+                using Accumulation = ck::detail::
+                    AccumulateWithNanCheck<false /*PropagateNan*/, reduce::Add, FloatAcc>;
+
+                for(int i_m = 0; i_m < MReduceIters; i_m++)
+                {
+                    static_for<0, NReduceIters, 1>{}([&](auto i_n_reduce) {
+                        acc_buf.Clear();
+                        for(auto i = tile_acc_offset_start; i < tile_acc_offset_end; i++)
+                        {
+                            auto c_partial_acc_buf =
+                                make_dynamic_buffer<AddressSpaceEnum::Global,
+                                                    AmdBufferCoherenceEnum::GLC>(
+                                    reinterpret_cast<FloatAcc*>(p_workspace) +
+                                        i * c_partial_acc_block_m_n.GetElementSpaceSize(),
+                                    c_partial_acc_block_m_n.GetElementSpaceSize());
+
+                            acc_load.Run(c_partial_acc_block_m_n,
+                                         c_partial_acc_buf,
+                                         acc_thread_buf_load_desc,
+                                         make_tuple(I0, I0),
+                                         parcial_acc_buf);
+
+                            static_for<0, CBlockTransferScalarPerVector_NWaveNPerXDL, 1>{}(
+                                [&](auto i_vec) {
+                                    constexpr auto offset =
+                                        acc_thread_buf_load_desc.CalculateOffset(
+                                            make_tuple(0, i_vec));
+                                    Accumulation::Calculate(acc_buf(Number<offset>{}),
+                                                            parcial_acc_buf[Number<offset>{}]);
+                                });
+                        }
+
+                        if(thread_n_cluster_id * CBlockTransferScalarPerVector_NWaveNPerXDL <
+                           NPerBlock)
+                        {
+                            acc_store.Run(acc_thread_buf_store_desc,
+                                          make_tuple(I0, I0, I0, I0),
+                                          acc_buf,
+                                          c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                          c_grid_buf);
+                        }
+                        if constexpr(NReduceIters != 1)
+                        {
+                            if constexpr(i_n_reduce != (NReduceIters - 1))
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n);
+                            }
+                            else
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n_reverse);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n_reverse);
+                            }
+                        }
+                    });
+                    {
+                        acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                    partial_acc_load_step_m);
+                        acc_store.MoveDstSliceWindow(c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                     partial_acc_store_step_m);
+                    }
+                }
+                return;
+            }
+        }
+
+        // offset for last acc buffer of this block
+        uint32_t block_acc_offset =
+            (block_mapping.get_acc_buffer_offset_from_block(block_idx + 1) - 1) * MPerBlock *
+            NPerBlock;
+
+        while(true)
+        {
+            uint32_t current_iter_length = __builtin_amdgcn_readfirstlane(
+                block_mapping.get_current_iter_length(iter_start, iter_end, total_iter_length));
+            uint32_t tile_idx, iter_offset;
+            block_mapping.get_tile_idx_with_offset(iter_end - 1, tile_idx, iter_offset);
+            iter_offset = __builtin_amdgcn_readfirstlane(iter_offset - current_iter_length + 1);
+            auto spatial_idx = block_mapping.tile_to_spatial(tile_idx, m, n);
+
+            const index_t m_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I0] * MPerBlock);
+
+            const index_t n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I1] * NPerBlock);
+
+            const index_t k0_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(iter_offset * K0PerBlock);
+
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<K0PerBlock, MPerBlock, K1>,
+                                                    ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(a_k0_m_k1_grid_desc),
+                                                    decltype(a_block_desc_k0_m_k1),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    2,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    a_k0_m_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_block_desc_k0_m_k1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // B matrix blockwise copy
+            auto b_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    BElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<K0PerBlock, NPerBlock, K1>,
+                                                    BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(b_k0_n_k1_grid_desc),
+                                                    decltype(b_block_desc_k0_n_k1),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    2,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    b_k0_n_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_block_desc_k0_n_k1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            const index_t num_k_block_main_loop = current_iter_length;
+
+            gridwise_gemm_pipeline.Run(a_k0_m_k1_grid_desc,
+                                       a_block_desc_k0_m_k1,
+                                       a_blockwise_copy,
+                                       a_grid_buf,
+                                       a_block_buf,
+                                       a_block_slice_copy_step,
+                                       b_k0_n_k1_grid_desc,
+                                       b_block_desc_k0_n_k1,
+                                       b_blockwise_copy,
+                                       b_grid_buf,
+                                       b_block_buf,
+                                       b_block_slice_copy_step,
+                                       blockwise_gemm,
+                                       c_thread_buf,
+                                       num_k_block_main_loop);
+
+            // output: register to global memory
+            {
+                constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+                constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+                constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+                constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+                constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+                constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+                constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+                constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+                constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+                constexpr auto c_block_desc_mblock_mpershuffle_nblock_npershuffle =
+                    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle();
+
+                constexpr auto c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle =
+                    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle();
+
+                auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                    reinterpret_cast<FloatCShuffle*>(p_shared_block),
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle.GetElementSpaceSize());
+
+                auto c_partial_acc_buf =
+                    make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                        reinterpret_cast<FloatAcc*>(p_workspace) + block_acc_offset,
+                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle
+                            .GetElementSpaceSize());
+
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                    make_tuple(make_freeze_transform(I0), // freeze mblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleMRepeatPerShuffle,
+                                              M1,
+                                              M2,
+                                              M3,
+                                              M4)),       // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                               make_freeze_transform(I0), // freeze nblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleNRepeatPerShuffle,
+                                              N1,
+                                              N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<>{},
+                               Sequence<0, 2, 4, 5, 6>{},
+                               Sequence<>{},
+                               Sequence<1, 3, 7>{}));
+
+                // calculate origin of thread output tensor on global memory
+                //     blockwise GEMM c matrix starting index
+                const auto c_thread_mtx_on_block =
+                    blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+                const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+                const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+                const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                        make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto m_thread_data_on_block_idx =
+                    m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                        make_multi_index(m_thread_data_on_block));
+
+                const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                        make_tuple(Sequence<0, 1, 2>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto n_thread_data_on_block_idx =
+                    n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                        make_multi_index(n_thread_data_on_block));
+
+                // VGPR to LDS
+                auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,
+                    FloatCShuffle,
+                    decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    ck::tensor_operation::element_wise::PassThrough,
+                    Sequence<CShuffleMRepeatPerShuffle,
+                             CShuffleNRepeatPerShuffle,
+                             I1,
+                             I1,
+                             M2,
+                             I1,
+                             M4,
+                             I1>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                    7,
+                    1,
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>{c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                          make_multi_index(0,
+                                           0,
+                                           m_thread_data_on_block_idx[I1],
+                                           n_thread_data_on_block_idx[I1],
+                                           m_thread_data_on_block_idx[I2],
+                                           m_thread_data_on_block_idx[I3],
+                                           m_thread_data_on_block_idx[I4],
+                                           n_thread_data_on_block_idx[I2]),
+                          ck::tensor_operation::element_wise::PassThrough{}};
+
+                // LDS to global
+                auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatC,               // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                      0,
+                                      __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                      0),
+                     c_element_op};
+
+                // LDS to global partial acc
+                auto c_block_copy_lds_to_partial_acc = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatCShuffle,        // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_element_op};
+
+                constexpr auto mxdlperwave_forward_step =
+                    make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+                constexpr auto nxdlperwave_forward_step =
+                    make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+                constexpr auto nxdlperwave_backward_step =
+                    make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+                static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                    constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                    static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                        constexpr bool nxdlperwave_forward_sweep =
+                            (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                        constexpr index_t nxdlperwave_value =
+                            nxdlperwave_forward_sweep
+                                ? nxdlperwave_iter
+                                : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                        constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                        // make sure it's safe to do ds_write
+                        block_sync_lds();
+
+                        // VGPR to LDS
+                        c_thread_copy_vgpr_to_lds.Run(
+                            c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                            make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                            c_thread_buf,
+                            c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                            c_block_buf);
+
+                        // make sure it's safe to do ds_read
+                        block_sync_lds();
+
+                        c_block_copy_lds_to_global.SetSrcSliceOrigin(
+                            c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                            make_tuple(0, 0, 0, 0));
+
+                        // LDS to global
+                        if(is_dp_block)
+                            c_block_copy_lds_to_global.template Run<decltype(c_block_buf),
+                                                                    decltype(c_grid_buf),
+                                                                    InMemoryDataOperationEnum::Set>(
+                                c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                c_block_buf,
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                c_grid_buf);
+                        else if(is_sk_block)
+                        {
+                            if constexpr(Block2CTileMap::ReductionStrategy ==
+                                         StreamKReductionStrategy::Reduction)
+                            {
+                                // constexpr offset
+                                c_block_copy_lds_to_partial_acc.SetSrcSliceOrigin(
+                                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                    make_tuple(0, 0, 0, 0));
+
+                                c_block_copy_lds_to_partial_acc.SetDstSliceOrigin(
+                                    c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                    make_tuple(mxdlperwave.value, 0, nxdlperwave.value, 0));
+
+                                c_block_copy_lds_to_partial_acc
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_partial_acc_buf),
+                                                  InMemoryDataOperationEnum::Set>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                        c_partial_acc_buf);
+                            }
+                            else if constexpr(Block2CTileMap::ReductionStrategy ==
+                                              StreamKReductionStrategy::Atomic)
+                            {
+                                c_block_copy_lds_to_global
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_grid_buf),
+                                                  InMemoryDataOperationEnum::AtomicAdd>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                        c_grid_buf);
+                            }
+                        }
+
+                        // move on nxdlperwave dimension
+                        if constexpr(nxdlperwave_forward_sweep &&
+                                     (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_forward_step);
+                        }
+                        else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_backward_step);
+                        }
+                    });
+
+                    // move on mxdlperwave dimension
+                    if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            mxdlperwave_forward_step);
+                    }
+                });
+
+                if constexpr(Block2CTileMap::ReductionStrategy ==
+                             StreamKReductionStrategy::Reduction)
+                {
+                    if(is_sk_block)
+                    {
+                        // increase the counter for this tile
+                        workgroup_barrier wg_barrier(p_semaphore);
+                        wg_barrier.inc(tile_idx);
+                    }
+                }
+            }
+
+            // exit condition
+            iter_end -= current_iter_length;
+            if(iter_end <= iter_start)
+                break;
+
+            if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+            {
+                block_acc_offset -= MPerBlock * NPerBlock;
+            }
+            // make sure next loop LDS is ready for use
+            block_sync_lds();
+        }
+    }
+
+    template <typename Layout>
+    struct LStr
+    {
+        static std::string Get() { return ""; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::RowMajor>
+    {
+        static std::string Get() { return "R"; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::ColumnMajor>
+    {
+        static std::string Get() { return "C"; }
+    };
+
+    static std::string GetTypeString()
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "GemmXdlStreamK_"
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << "_"
+            << "B" << BlockSize << "_"
+            << "Vec" << ABlockTransferSrcScalarPerVector << "x"
+            << BBlockTransferSrcScalarPerVector << "x"
+            << CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
+            << MPerBlock << "x"
+            << NPerBlock << "x"
+            << K0PerBlock << "x"
+            << K1 ;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -194,7 +194,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
               StrideC{StrideC_},
               MPadded{CalculateMPadded(M_)},
               NPadded{CalculateNPadded(N_)},
-              K0{CalculateK0(K)}
+              K0{CalculateK0(K_)}
         {
         }
 
@@ -247,8 +247,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
         FloatC* p_c_grid;
     };
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
     // denorm test fix, required to work around fp16 mfma issue
     // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
@@ -383,7 +383,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
 
     __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
     {
-        const index_t num_loop = K / (K0PerBlock * K1);
+        const index_t num_loop = math::integer_divide_ceil(K, K0PerBlock * K1);
 
         return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
     }
@@ -840,7 +840,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
             }
         }();
 
-        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
+            const auto KPad  = K0Pad * K1Value;
+
+            const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
         {
             return transform_tensor_descriptor(
                 a_grid_desc_m_k,
@@ -874,7 +892,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
             }
         }();
 
-        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
+            const auto KPad  = K0Pad * K1Value;
+
+            const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+
+        else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
         {
             return transform_tensor_descriptor(
                 b_grid_desc_k_n,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -45,7 +45,8 @@ __global__ void
 }
 
 template <index_t BlockSize,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatAcc,
           typename FloatC,
           typename ALayout,
@@ -85,7 +86,8 @@ template <index_t BlockSize,
           index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
           typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           LoopScheduler LoopSched     = make_default_loop_scheduler(),
-          PipelineVersion PipelineVer = PipelineVersion::v1>
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = FloatC>
 struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 {
     static constexpr auto I0 = Number<0>{};
@@ -108,13 +110,13 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
     struct Argument : public ck::tensor_operation::device::BaseArgument
     {
-        const FloatAB* p_a_grid;
-        const FloatAB* p_b_grid;
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
         FloatC* p_c_grid;
         index_t M;
         index_t N;
@@ -128,8 +130,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         index_t K0;
         index_t k_batch;
 
-        Argument(const FloatAB* p_a_grid_,
-                 const FloatAB* p_b_grid_,
+        Argument(const FloatA* p_a_grid_,
+                 const FloatB* p_b_grid_,
                  FloatC* p_c_grid_,
                  index_t M_,
                  index_t N_,
@@ -365,7 +367,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         constexpr auto c_block_size =
             GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
 
-        return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
+        return math::max((a_block_space_size + b_block_space_size) * sizeof(ComputeType),
                          c_block_size * sizeof(FloatC));
     }
 
@@ -577,8 +579,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                void* __restrict__ p_shared_block,
                                const Block2CTileMap& block_2_ctile_map)
     {
-        const FloatAB* p_a_grid          = karg.p_a_grid;
-        const FloatAB* p_b_grid          = karg.p_b_grid;
+        const FloatA* p_a_grid           = karg.p_a_grid;
+        const FloatB* p_b_grid           = karg.p_b_grid;
         FloatC* p_c_grid                 = karg.p_c_grid;
         const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
             karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
@@ -698,8 +700,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                 Sequence<1, K0PerBlock, MPerBlock, K1>,
                                                 ABlockTransferThreadClusterLengths_K0_M_K1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                FloatA,
+                                                ComputeType,
                                                 decltype(a_b_k0_m_k1_grid_desc),
                                                 decltype(a_b_k0_m_k1_block_desc),
                                                 ABlockTransferSrcAccessOrder,
@@ -728,8 +730,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                 Sequence<1, K0PerBlock, NPerBlock, K1>,
                                                 BBlockTransferThreadClusterLengths_K0_N_K1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                FloatB,
+                                                ComputeType,
                                                 decltype(b_b_k0_n_k1_grid_desc),
                                                 decltype(b_b_k0_n_k1_block_desc),
                                                 BBlockTransferSrcAccessOrder,
@@ -759,7 +761,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            FloatAB,
+            ComputeType,
             FloatAcc,
             decltype(a_k0_m_k1_block_desc),
             decltype(b_k0_n_k1_block_desc),
@@ -776,8 +778,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         constexpr auto a_block_space_size =
             math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
-        FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
-        FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
+        ComputeType* p_a_block = static_cast<ComputeType*>(p_shared_block);
+        ComputeType* p_b_block = static_cast<ComputeType*>(p_shared_block) + a_block_space_size;
 
         constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
@@ -787,53 +789,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
 
-#if 0
-        // preload data into LDS
-        {
-            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
-            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
-
-            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
-            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
-        }
-
-        // Initialize C
-        c_thread_buf.Clear();
-
-        // main body
-        if constexpr(HasMainKBlockLoop)
-        {
-            index_t k0_block_data_begin = 0;
-
-            do
-            {
-                a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
-                b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
-
-                a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
-
-                block_sync_lds();
-
-                b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
-
-                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
-
-                block_sync_lds();
-
-                a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
-                b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
-
-                k0_block_data_begin += K0PerBlock;
-            } while(k0_block_data_begin < (karg.K0 - K0PerBlock));
-        }
-
-        // tail
-        {
-            block_sync_lds();
-
-            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
-        }
-#else
         // gridwise GEMM pipeline
         const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
             (a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
@@ -856,7 +811,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                                blockwise_gemm,
                                                                c_thread_buf,
                                                                num_k_block_main_loop);
-#endif
 
         // output: register to global memory
         {

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp

@@ -139,8 +139,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
@@ -315,8 +315,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
             c_grid_desc_m_n);
     }
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -634,10 +634,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
                 Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
                 FloatCShuffle,              // typename SrcData,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp

@@ -142,8 +142,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
@@ -323,13 +323,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
     }
 
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C0GridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -654,12 +654,9 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
                 FloatC,                     // typename Src0Data,
                 FloatC,                     // typename Src1Data,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp

@@ -151,8 +151,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
@@ -331,18 +331,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
             c_grid_desc_m_n);
     }
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C0GridDesc_M_N{}))>;
 
     using C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C1GridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -674,14 +674,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
                 FloatC,                     // typename Src1Data,
                 FloatC,                     // typename Src2Data,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector,
+          typename Block2ETileMap>
+struct GridwiseImageToColumn
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __device__ static void Run(const InputGridDesc& in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc& out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap& block_2_tile_map)
+    {
+        const auto block_work_idx =
+            block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t k_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * KPerBlock);
+
+        // Global Memory
+        const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_desc.GetElementSpaceSize());
+        auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_global, out_grid_desc.GetElementSpaceSize());
+
+        auto copy_global_to_global = ThreadGroupTensorSliceTransfer_v7<
+            ThisThreadBlock,
+            Tuple<InputDataType>,
+            Tuple<OutputDataType>,
+            decltype(tie(in_grid_desc)),
+            decltype(tie(out_grid_desc)),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<MPerBlock, KPerBlock>,
+            ThreadClusterLengths,
+            Sequence<0, 1>,
+            Sequence<0, 1>,
+            I1,
+            ScalarPerVector,
+            Sequence<true>,
+            Sequence<true>>{
+            in_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            out_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            tensor_operation::element_wise::PassThrough{}};
+
+        copy_global_to_global.Run(
+            tie(in_grid_desc), tie(in_global_buf), tie(out_grid_desc), tie(out_global_buf));
+    }
+
+    __host__ static constexpr bool CheckValidity(const InputGridDesc& in_grid_desc,
+                                                 const OutputGridDesc& out_grid_desc)
+    {
+        if(in_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           in_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        if(out_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           out_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        return true;
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp

@@ -78,8 +78,8 @@ struct GridwiseSparseEmbeddingsForwardLayernorm
     using ThreadwiseWolfordDesc2D = decltype(make_naive_tensor_descriptor_packed(make_tuple(
         Number<DimSubBlocks * DimThreadSize>{}, Number<RowSubBlocks * RowVectorSize>{})));
 
-    using ThreadwiseWolfordDescReduce = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
+    using ThreadwiseWolfordDescReduce = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
 
     using ThreadwiseWelford =
         ThreadwiseWelford<AccDataType, ThreadwiseWolfordDesc2D, ThreadwiseWolfordDescReduce>;

include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp

@@ -78,17 +78,18 @@ struct GridwiseNormalizationSplitK1st
     static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
 
     __device__ static int
-    GetKPerThread(int kRaw, int kGridSize, int block_k_cluster_id, int thread_k_cluster_id)
+    GetKPerThread(int k, int kRaw, int kGridSize, int block_k_cluster_id, int thread_k_cluster_id)
     {
         bool is_rightmost_block = block_k_cluster_id == kGridSize - 1;
 
         if(is_rightmost_block)
         {
-            int left_kPerBlock = math::integer_divide_ceil(kRaw, kGridSize);
-            int kPerBlock      = kRaw % kGridSize == 0 ? left_kPerBlock : kRaw % left_kPerBlock;
-            int kPerThread =
-                kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
-            int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+            int left_kPerBlock  = math::integer_divide_ceil(k, kGridSize);
+            int kRightmostBlock = kRaw - left_kPerBlock * (kGridSize - 1);
+            int kPerThread      = kRightmostBlock < K_BlockTileSize
+                                      ? 0
+                                      : KThreadSliceSize * (kRightmostBlock / K_BlockTileSize);
+            int kPerBlockTail   = kRightmostBlock - kPerThread * KThreadClusterSize;
 
             if(kPerBlockTail > 0)
             {
@@ -105,7 +106,7 @@ struct GridwiseNormalizationSplitK1st
         }
         else
         {
-            int kPerBlock = math::integer_divide_ceil(kRaw, kGridSize);
+            int kPerBlock = math::integer_divide_ceil(k, kGridSize);
             return KThreadSliceSize * (kPerBlock / K_BlockTileSize);
         }
     }
@@ -193,10 +194,13 @@ struct GridwiseNormalizationSplitK1st
         auto var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_variance_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
 
-        auto threadwise_welford = ThreadwiseWelford();
-        int kRaw                = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
-        threadwise_welford.max_count_ =
-            GetKPerThread(kRaw, k_grid_size, block_k_cluster_id, thread_k_cluster_id);
+        auto threadwise_welford       = ThreadwiseWelford();
+        int kRaw                      = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k.GetLength(I1),
+                                                      kRaw,
+                                                      k_grid_size,
+                                                      block_k_cluster_id,
+                                                      thread_k_cluster_id);
 
         static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
             mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

@@ -129,6 +129,9 @@ struct ThreadwiseTensorSliceTransfer_v3r1
 
         constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
 
+        static_assert(SliceLengths::At(SrcVectorDim) % SrcScalarPerVector == 0,
+                      "SliceLengths[SrcVectorDim] must be divisible by SrcScalarPerVector");
+
         constexpr auto src_dim_access_order = SrcDimAccessOrder{};
 
         constexpr auto ordered_src_access_lengths =

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp

@@ -104,13 +104,13 @@ struct ThreadwiseTensorSliceTransfer_v6r1
 
             // apply pointwise operation
             static_for<0, ScalarPerVector, 1>{}([&](auto i) {
-                SrcData v;
+                DstData v;
 
                 // apply element-wise operation
                 element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
 
                 // apply type convert
-                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+                dst_vector_container.template AsType<DstData>()(i) = v;
             });
 
             const bool is_dst_valid =

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1r2(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, InMemoryDataOperationEnum DstInMemOp>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck