Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into navi3x_md_bgemm_conv_gemmsoftmaxgemm

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -95,6 +95,12 @@ struct Scale
         y = scale_ * x;
     };
 
+    template <>
+    __host__ __device__ void operator()<double, double>(double& y, const double& x) const
+    {
+        y = scale_ * x;
+    };
+
     float scale_;
 };
 

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

@@ -154,6 +154,50 @@ struct BlockToCTileMap_M00_N0_M01Adapt
         index_t idx_M01          = idx_M0 % M01_;
         index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
 
+        /**
+         *                        idxN0
+         *
+         *           |<               mtx   N                 >|
+         *
+         *             NPerBlock   NPerBlock   NPerBlock   NPerBlock
+         *                N_0         N_1        N_2         N_3
+         *       -   |-----------|-----------|-----------|-----|-----|-
+         *       ^   | -   -  0  |/---->  2  |           |     |     |
+         *           | |   |     /     |     |           |     |     |  M_0  MPerBlock
+         *           | M   |    /|     |     |           |     |     |
+         *           |-0---|---/-|-----|-----|-----------|-----|-----|-
+         *           | 1   |  /  |     |     |  blockid  |     |     |
+         * idxM0     | |   | /   |     V     |     5     |     |     |  M_1  MPerBlock
+         *           | -   V   1 |     -  3  |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *    mtx M  |           |           |           |     |     |
+         *           |           |           |           |     |     |  M_2  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *           |           |           |           |     |     |
+         *           |           |           |           |     |     |  M_3  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *       V   |           |           |           |     |     |
+         *       -   |-----------|-----------|-----------|-----|-----|- M_4  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *  Example:
+         *   assume:
+         *      M0 = 5
+         *      N0 = 4
+         *      block_1d_id = 5
+         *      M01 = 2
+         *
+         *   idx_N0 = 1
+         *   idx_M0 = 1
+         *   M01_adapt = 2
+         *   idx_M00 = 0
+         *   idx_M01 = 1
+         *   idx_N0_M01_local = 5
+         *   output {1, 2}
+         */
+
         return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
                           idx_N0_M01_local / M01_adapt);
     }

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.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_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/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : F[M, N0], where N0 is number of blocks along N dimension
+//   output : G[M, N0], where N0 is number of blocks along N dimension
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   F, G = welford(E)
+// Assume:
+//   D0, D1, ... and E have the same layout
+//   Calculate mean & variance along N dimension for E
+template <typename ABDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EMeanVarDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename DsGridDesc_M_N,
+          typename EGridDesc_M_N,
+          typename MeanVarGridDesc_M_NBlock,
+          typename CountGridDesc_M_NBlock,
+          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_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_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 PostShuffleThreadClusterSize_M_N,
+          index_t PostShuffleScalarPerVector,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    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_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 + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_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 DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_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>{});
+    }
+
+    template <typename GridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(const GridDescriptor_M_N& grid_desc_m_n)
+    {
+        const auto M      = grid_desc_m_n.GetLength(I0);
+        const auto NBlock = grid_desc_m_n.GetLength(I1);
+        const auto MBlock = M / MPerBlock;
+
+        const auto grid_desc_mblock_mperblock_nblock = transform_tensor_descriptor(
+            grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_pass_through_transform(NBlock)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}));
+
+        return grid_desc_mblock_mperblock_nblock;
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __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);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            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_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EMeanVarDataType) <= 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 DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using MeanVarGridDescriptor_MBlock_MPerBlock_NBlock      = remove_cvref_t<decltype(
+        MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(MeanVarGridDesc_M_NBlock{}))>;
+    using CountGridDescriptor_MBlock_MPerBlock_NBlock        = remove_cvref_t<decltype(
+        MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(CountGridDesc_M_NBlock{}))>;
+    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{}))>;
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void
+    Run(const ABDataType* __restrict__ p_a_grid,
+        const ABDataType* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        EMeanVarDataType* __restrict__ p_e_grid,
+        EMeanVarDataType* __restrict__ p_welford_mean_grid,
+        EMeanVarDataType* __restrict__ p_welford_var_grid,
+        int32_t* __restrict__ p_welford_count,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            ds_grid_desc_mblock_mperblock_nblock_nperblock,
+        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            e_grid_desc_mblock_mperblock_nblock_nperblock,
+        const MeanVarGridDescriptor_MBlock_MPerBlock_NBlock&
+            mean_var_grid_desc_mblock_mperblock_nblock,
+        const CountGridDescriptor_MBlock_MPerBlock_NBlock& count_grid_desc_mblock_mperblock_nblock,
+        const Block2ETileMap& block_2_etile_map,
+        index_t NRaw)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_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());
+
+        auto mean_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_mean_grid, mean_var_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        auto var_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_var_grid, mean_var_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        auto welford_count_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_count, count_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * 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_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();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                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<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // 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<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        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<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(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 / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(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_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_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, Welford and write out
+        {
+            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{}};
+
+            // space filling curve for threadwise C in VGPR
+            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>,
+                                  false>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_der_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>,
+                                  false>{};
+
+            // LDS c_shuffle_block_desc_mperblock_nperblock
+            constexpr auto c_shuffle_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(PostShuffleThreadClusterSize_M_N::At(I0) *
+                                  PostShuffleThreadClusterSize_M_N::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      PostShuffleThreadClusterSize_M_N::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      PostShuffleThreadClusterSize_M_N::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t PostShuffleThreadSliceSize_M =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                PostShuffleThreadClusterSize_M_N::At(I0);
+
+            constexpr index_t PostShuffleThreadSliceSize_N =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                PostShuffleThreadClusterSize_M_N::At(I1);
+
+            constexpr auto PostShuffleThreadSliceSize_M_N =
+                Sequence<PostShuffleThreadSliceSize_M, PostShuffleThreadSliceSize_N>{};
+
+            // VGPR post_shuffle_thread_desc_m_n
+            constexpr auto post_shuffle_thread_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{},
+                           Number<PostShuffleThreadSliceSize_N>{}));
+
+            auto e_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                post_shuffle_thread_desc_m_n.GetElementSpaceSize());
+
+            // To apply D0, D1, ... and Welford.
+            // threadwise copy from LDS to VGPR
+            constexpr auto post_shuffle_thread_cluster_desc =
+                make_cluster_descriptor(PostShuffleThreadClusterSize_M_N{}, Sequence<0, 1>{});
+
+            const auto post_shuffle_thread_cluster_idx =
+                post_shuffle_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto post_shuffle_thread_data_idx_begin =
+                post_shuffle_thread_cluster_idx * PostShuffleThreadSliceSize_M_N;
+
+            // To apply D0, D1, ... and Welford.
+            // Copy c shuffle from LDS back to VGPR
+            auto post_shuffle_thread_copy_lds_to_vgpr =
+                ThreadwiseTensorSliceTransfer_v2<CShuffleDataType,
+                                                 AccDataType,
+                                                 decltype(c_shuffle_block_desc_mperblock_nperblock),
+                                                 decltype(post_shuffle_thread_desc_m_n),
+                                                 decltype(PostShuffleThreadSliceSize_M_N),
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 PostShuffleScalarPerVector,
+                                                 1,
+                                                 true>{c_shuffle_block_desc_mperblock_nperblock,
+                                                       post_shuffle_thread_data_idx_begin};
+
+            // D0, D1, ..., Dn
+            constexpr auto post_shuffle_thread_desc_I1_mperblock_I1_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<PostShuffleThreadSliceSize_M>{},
+                               I1,
+                               Number<PostShuffleThreadSliceSize_N>{}));
+
+            // FIXME: Decrease usage of VGPR
+            // Apply pointwise lambda function from multi-source (Global and LDS) into VGPR
+            auto ds_thread_buf = generate_tuple(
+                [&](auto) {
+                    return make_static_buffer<AddressSpaceEnum::Vgpr, CShuffleDataType>(
+                        post_shuffle_thread_desc_I1_mperblock_I1_nperblock.GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // Copy D0, D1, ..., Dn from global to VGPR
+            auto ds_thread_copy_global_to_vgpr = generate_tuple(
+                [&](auto I) {
+                    using DDataType = remove_cvref_t<tuple_element_t<I.value, DsDataType>>;
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        AccDataType,
+                        decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock[I]),
+                        decltype(post_shuffle_thread_desc_I1_mperblock_I1_nperblock),
+                        Sequence<I1,
+                                 PostShuffleThreadSliceSize_M,
+                                 I1,
+                                 PostShuffleThreadSliceSize_N>,
+                        Sequence<0, 1, 2, 3>,
+                        3,
+                        PostShuffleScalarPerVector,
+                        1,
+                        true>(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[I],
+                        make_multi_index(
+                            I0,
+                            m_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I0],
+                            I0,
+                            n_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I1]));
+                },
+                Number<NumDTensor>{});
+
+            auto e_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                AccDataType,
+                EMeanVarDataType,
+                decltype(post_shuffle_thread_desc_I1_mperblock_I1_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1,
+                         PostShuffleThreadSliceSize_M,
+                         I1,
+                         PostShuffleThreadSliceSize_N>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                   // DimAccessOrder
+                3,                                      // DstVectorDim
+                PostShuffleScalarPerVector,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            // Welford
+            constexpr auto thread_welford_src_desc_m_k = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{},
+                           Number<PostShuffleThreadSliceSize_N>{}));
+
+            constexpr auto thread_welford_dst_desc_m = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{}));
+
+            using ThreadwiseWelford = ThreadwiseWelford<AccDataType,
+                                                        decltype(thread_welford_src_desc_m_k),
+                                                        decltype(thread_welford_dst_desc_m)>;
+
+            using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                                      BlockSize,
+                                                      PostShuffleThreadClusterSize_M_N,
+                                                      Sequence<0, 1>,
+                                                      false>;
+
+            constexpr int num_shuffleM =
+                MPerBlock / (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl);
+
+            constexpr int num_shuffleN =
+                NPerBlock / (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl);
+
+            using mean_var_vgpr_type =
+                decltype(make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize()));
+
+            using welford_count_vgpr_type =
+                decltype(make_static_buffer<AddressSpaceEnum::Vgpr, int32_t>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize()));
+
+            Array<ThreadwiseWelford, num_shuffleM> threadwise_welfords;
+            Array<mean_var_vgpr_type, num_shuffleM> mean_thread_bufs;
+            Array<mean_var_vgpr_type, num_shuffleM> var_thread_bufs;
+            Array<welford_count_vgpr_type, num_shuffleM> welford_count_thread_bufs;
+
+            int max_count     = PostShuffleThreadSliceSize_N * num_shuffleN;
+            const auto nblock = mean_var_grid_desc_mblock_mperblock_nblock.GetLength(I2);
+
+            // tail block
+            if(block_work_idx[I1] % nblock == nblock - 1)
+            {
+                constexpr index_t NPerShuffleBlock =
+                    CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl;
+
+                int NPerBlockTail = NRaw - NPerBlock * (nblock - 1);
+                int thread_max_len =
+                    PostShuffleThreadSliceSize_N * (post_shuffle_thread_cluster_idx[I1] + 1);
+                int shuffle_step = 0;
+                while(thread_max_len <= NPerBlockTail && shuffle_step < num_shuffleN)
+                {
+                    ++shuffle_step;
+                    thread_max_len += NPerShuffleBlock;
+                }
+
+                int delta = 0;
+                if(thread_max_len - NPerBlockTail > PostShuffleThreadSliceSize_N)
+                    delta = 0;
+                else if(NPerBlockTail > thread_max_len)
+                    delta = PostShuffleThreadSliceSize_N;
+                else
+                    delta = PostShuffleThreadSliceSize_N - thread_max_len + NPerBlockTail;
+
+                max_count = shuffle_step * PostShuffleThreadSliceSize_N + delta;
+            }
+
+            static_for<0, num_shuffleM, 1>{}([&](auto i) {
+                threadwise_welfords(i).max_count_ = max_count;
+                mean_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                var_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                welford_count_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, int32_t>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                static_for<0, PostShuffleThreadSliceSize_M, 1>{}([&](auto j) {
+                    mean_thread_bufs(i)(j)          = type_convert<AccDataType>(0.0f);
+                    var_thread_bufs(i)(j)           = type_convert<AccDataType>(0.0f);
+                    welford_count_thread_bufs(i)(j) = 0;
+                });
+            });
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_der_global.GetNumOfAccess(), "wrong!");
+
+            int shuffleM_index = __builtin_amdgcn_readfirstlane(0);
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each thread shuffle 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 write to LDS
+                block_sync_lds();
+
+                // Get shuffle data from LDS to VGPR
+                post_shuffle_thread_copy_lds_to_vgpr.Run(c_shuffle_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         post_shuffle_thread_desc_m_n,
+                                                         make_tuple(I0, I0),
+                                                         e_thread_buf);
+
+                // Global read D0, D1, ...
+                static_for<0, NumDTensor, 1>{}([&](auto Id) {
+                    auto& d_thread_copy_global_to_vgpr = ds_thread_copy_global_to_vgpr(Id);
+                    d_thread_copy_global_to_vgpr.Run(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[Id],
+                        ds_grid_buf[Id],
+                        post_shuffle_thread_desc_I1_mperblock_I1_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        ds_thread_buf(Id));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on D0, D1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        d_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock[Id], de_global_step);
+                    }
+                });
+
+                // cde_element_op(e, c, d0, d1, ...);
+                static_for<0, post_shuffle_thread_desc_m_n.GetElementSize(), 1>{}([&](auto i) {
+                    const auto c_ds_src_data_refs = concat_tuple_of_reference(
+                        tie(e_thread_buf[i]),
+                        generate_tie(
+                            [&](auto Id) -> const auto& { return ds_thread_buf[Id][i]; },
+                            Number<NumDTensor>{}));
+                    auto e_dst_data_refs = tie(e_thread_buf(i));
+                    unpack2(cde_element_op, e_dst_data_refs, c_ds_src_data_refs);
+                });
+
+                // Global write E
+                e_thread_copy_vgpr_to_global.Run(post_shuffle_thread_desc_I1_mperblock_I1_nperblock,
+                                                 make_tuple(I0, I0, I0, I0),
+                                                 e_thread_buf,
+                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                 e_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    // move on E
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    e_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        e_grid_desc_mblock_mperblock_nblock_nperblock, de_global_step);
+                }
+
+                // Threadwise welford
+                auto& threadwise_welford = threadwise_welfords(shuffleM_index);
+                auto& mean_thread_buf    = mean_thread_bufs(shuffleM_index);
+                auto& var_thread_buf     = var_thread_bufs(shuffleM_index);
+
+                threadwise_welford.Run(e_thread_buf, mean_thread_buf, var_thread_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    constexpr int shuffleMInc =
+                        de_global_step[I1] /
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1);
+                    shuffleM_index = __builtin_amdgcn_readfirstlane(shuffleM_index + shuffleMInc);
+                }
+            }); // copy c, d, e + welford
+
+            // Blockwise welford and write out
+            static_for<0, num_shuffleM, 1>{}([&](auto i) {
+                auto& mean_thread_buf  = mean_thread_bufs(i);
+                auto& var_thread_buf   = var_thread_bufs(i);
+                auto& count_thread_buf = welford_count_thread_bufs(i);
+
+                static_for<0, PostShuffleThreadSliceSize_M, 1>{}([&](auto j) {
+                    block_sync_lds();
+                    count_thread_buf(j) = threadwise_welfords(i).cur_count_;
+                    BlockwiseWelford::Run(
+                        mean_thread_buf(j), var_thread_buf(j), count_thread_buf(j));
+                });
+
+                if(post_shuffle_thread_cluster_idx[I1] == 0)
+                {
+                    constexpr auto thread_welford_desc_I_m_I = make_naive_tensor_descriptor_packed(
+                        make_tuple(I1, Number<PostShuffleThreadSliceSize_M>{}, I1));
+
+                    constexpr int shuffleMPerBlock =
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1);
+
+                    auto mean_var_count_thread_copy_index = make_multi_index(
+                        block_work_idx[I0],                                            // mblock
+                        shuffleMPerBlock * i + post_shuffle_thread_data_idx_begin[I0], // mperblock
+                        block_work_idx[I1]);                                           // nblock
+
+                    auto mean_var_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                        AccDataType,
+                        EMeanVarDataType,
+                        decltype(thread_welford_desc_I_m_I),
+                        decltype(mean_var_grid_desc_mblock_mperblock_nblock),
+                        tensor_operation::element_wise::PassThrough,
+                        Sequence<1, PostShuffleThreadSliceSize_M, 1>,
+                        Sequence<0, 1, 2>,
+                        1,
+                        1,
+                        InMemoryDataOperationEnum::Set,
+                        1,
+                        true>{mean_var_grid_desc_mblock_mperblock_nblock,
+                              mean_var_count_thread_copy_index,
+                              tensor_operation::element_wise::PassThrough{}};
+
+                    mean_var_thread_copy_vgpr_to_global.Run(
+                        thread_welford_desc_I_m_I,
+                        make_tuple(I0, I0, I0),
+                        mean_thread_buf,
+                        mean_var_grid_desc_mblock_mperblock_nblock,
+                        mean_grid_buf); // write mean
+
+                    mean_var_thread_copy_vgpr_to_global.Run(
+                        thread_welford_desc_I_m_I,
+                        make_tuple(I0, I0, I0),
+                        var_thread_buf,
+                        mean_var_grid_desc_mblock_mperblock_nblock,
+                        var_grid_buf); // write variance
+
+                    // Stride of count is [0, 1]. Only the first row in count[0, 0:nblock] need
+                    // to be written.
+                    if(i == 0 && block_work_idx[I0] == 0 &&
+                       post_shuffle_thread_cluster_idx[I0] == 0)
+                    {
+                        auto count_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                            int32_t,
+                            int32_t,
+                            decltype(thread_welford_desc_I_m_I),
+                            decltype(count_grid_desc_mblock_mperblock_nblock),
+                            tensor_operation::element_wise::PassThrough,
+                            Sequence<1, PostShuffleThreadSliceSize_M, 1>,
+                            Sequence<0, 1, 2>,
+                            1,
+                            1,
+                            InMemoryDataOperationEnum::Set,
+                            1,
+                            false>{count_grid_desc_mblock_mperblock_nblock,
+                                   mean_var_count_thread_copy_index,
+                                   tensor_operation::element_wise::PassThrough{}};
+
+                        count_thread_copy_vgpr_to_global.Run(
+                            thread_welford_desc_I_m_I,
+                            make_tuple(I0, I0, I0),
+                            count_thread_buf,
+                            count_grid_desc_mblock_mperblock_nblock,
+                            welford_count_grid_buf); // write count
+                    }
+                }
+            });
+
+        } // shuffle C + Ds + welford + write out
+    }     // run
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.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_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/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+
+namespace ck {
+
+template <typename EMeanVarDataType,
+          typename HDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename EHGridDesc_M_N,
+          typename MeanVarGridDesc_M_NBlock,
+          typename CountGridDesc_M_NBlock,
+          typename GammaBetaGridDesc_N,
+          typename HElementwiseOperation,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t NThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t NThreadSliceSize,
+          index_t ESrcVectorSize,
+          index_t HDstVectorSize,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorSize>
+struct GridwiseWelfordSecondHalfLayernorm2d
+{
+    static_assert(NThreadSliceSize % ESrcVectorSize == 0 &&
+                      NThreadSliceSize % GammaSrcVectorSize == 0 &&
+                      NThreadSliceSize % BetaSrcVectorSize == 0,
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NThreadSliceSize % HDstVectorSize == 0,
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using ThreadClusterLengths_M_N   = Sequence<MThreadClusterSize, NThreadClusterSize>;
+    using ThreadBufferDimAccessOrder = Sequence<0, 1>;
+    using ThreadClusterArrangeOrder  = Sequence<0, 1>;
+
+    static constexpr auto thread_cluster_desc_m_n =
+        make_cluster_descriptor(ThreadClusterLengths_M_N{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_N                = Sequence<MThreadSliceSize, NThreadSliceSize>;
+    static constexpr auto thread_buffer_desc_m_n = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<NThreadSliceSize>{}));
+
+    using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+    static constexpr auto thread_buffer_desc_m_1 =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+    using ThreadBufferLengths_N = Sequence<NThreadSliceSize>;
+    static constexpr auto thread_buffer_desc_n =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<NThreadSliceSize>{}));
+
+    using ThreadWelfordSrcDesc_M_1 = decltype(thread_buffer_desc_m_1);
+    using ThreadWelfordDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<ComputeDataType, ThreadWelfordSrcDesc_M_1, ThreadWelfordDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_N,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t N_BlockTileSize = NThreadClusterSize * NThreadSliceSize;
+
+    __device__ static void Run(const EMeanVarDataType* __restrict__ p_e_grid,
+                               const EMeanVarDataType* __restrict__ p_in_welford_mean_grid,
+                               const EMeanVarDataType* __restrict__ p_in_welford_var_grid,
+                               const int32_t* __restrict__ p_in_welford_count_grid,
+                               const GammaDataType* __restrict__ p_gamma_grid,
+                               const BetaDataType* __restrict__ p_beta_grid,
+                               HDataType* __restrict__ p_h_grid,
+                               const EHGridDesc_M_N& e_grid_desc_m_n,
+                               const EHGridDesc_M_N& h_grid_desc_m_n,
+                               const MeanVarGridDesc_M_NBlock& mean_var_grid_desc_m_nblock,
+                               const CountGridDesc_M_NBlock& count_grid_desc_m_nblock,
+                               const GammaBetaGridDesc_N& gamma_grid_desc_n,
+                               const GammaBetaGridDesc_N& beta_grid_desc_n,
+                               index_t numMeanVarCountBlockTileIteration_N,
+                               index_t NBlockClusterLength,
+                               ComputeDataType epsilon,
+                               HElementwiseOperation h_element_op)
+    {
+        // Thread/Block id
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const auto block_work_idx     = make_tuple(block_global_id / NBlockClusterLength,
+                                               block_global_id % NBlockClusterLength);
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc_m_n.CalculateBottomIndex(make_multi_index(thread_local_id));
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_n_cluster_id = thread_cluster_idx[I1];
+
+        // Global Memory
+        const auto e_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_m_n.GetElementSpaceSize());
+
+        const auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_mean_grid, mean_var_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_var_grid, mean_var_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_count_grid, count_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_grid, gamma_grid_desc_n.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_grid, beta_grid_desc_n.GetElementSpaceSize());
+
+        auto h_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_h_grid, h_grid_desc_m_n.GetElementSpaceSize());
+
+        // VGPR
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            e_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            gamma_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            beta_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            h_thread_buf;
+
+        // IO
+        auto threadwise_mean_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             MeanVarGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_var_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             MeanVarGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_count_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             CountGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                count_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_e_load_m_n =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             decltype(e_grid_desc_m_n),
+                                             decltype(thread_buffer_desc_m_n),
+                                             ThreadBufferLengths_M_N,
+                                             ThreadBufferDimAccessOrder,
+                                             1, // SrcVectorDim
+                                             ESrcVectorSize,
+                                             1,
+                                             true>(
+                e_grid_desc_m_n,
+                make_multi_index(
+                    block_work_idx[I0] * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_work_idx[I1] * N_BlockTileSize + thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_gamma_load_n =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             ComputeDataType,
+                                             decltype(gamma_grid_desc_n),
+                                             decltype(thread_buffer_desc_n),
+                                             ThreadBufferLengths_N,
+                                             Sequence<0>, // DimAccessOrder,
+                                             0,           // SrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_n,
+                make_multi_index(block_work_idx[I1] * N_BlockTileSize +
+                                 thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_beta_load_n =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             ComputeDataType,
+                                             decltype(beta_grid_desc_n),
+                                             decltype(thread_buffer_desc_n),
+                                             ThreadBufferLengths_N,
+                                             Sequence<0>, // DimAccessOrder,
+                                             0,           // SrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_n,
+                make_multi_index(block_work_idx[I1] * N_BlockTileSize +
+                                 thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_h_store_m_n =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               HDataType,
+                                               decltype(thread_buffer_desc_m_n),
+                                               decltype(h_grid_desc_m_n),
+                                               HElementwiseOperation,
+                                               ThreadBufferLengths_M_N,
+                                               ThreadBufferDimAccessOrder,
+                                               1, // DstVectorDim
+                                               HDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                h_grid_desc_m_n,
+                make_multi_index(
+                    block_work_idx[I0] * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_work_idx[I1] * N_BlockTileSize + thread_n_cluster_id * NThreadSliceSize),
+                h_element_op);
+
+        // step1: Merge mean and variance
+        constexpr auto mean_var_count_thread_copy_step_I0_n =
+            make_multi_index(I0, NThreadClusterSize);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
+            welford_var_thread_buf(I)   = type_convert<ComputeDataType>(0.0f);
+            welford_count_thread_buf(I) = 0;
+        });
+
+        for(index_t n = 0; n < numMeanVarCountBlockTileIteration_N; ++n)
+        {
+            threadwise_mean_load_m_nblock.Run(mean_var_grid_desc_m_nblock,
+                                              welford_mean_global_val_buf,
+                                              thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              in_welford_mean_thread_buf);
+
+            threadwise_var_load_m_nblock.Run(mean_var_grid_desc_m_nblock,
+                                             welford_var_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_var_thread_buf);
+
+            threadwise_count_load_m_nblock.Run(count_grid_desc_m_nblock,
+                                               welford_count_global_val_buf,
+                                               thread_buffer_desc_m_1,
+                                               make_tuple(I0, I0),
+                                               in_welford_count_thread_buf);
+
+            ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                   in_welford_var_thread_buf,
+                                   in_welford_count_thread_buf,
+                                   welford_mean_thread_buf,
+                                   welford_var_thread_buf,
+                                   welford_count_thread_buf);
+
+            threadwise_mean_load_m_nblock.MoveSrcSliceWindow(mean_var_grid_desc_m_nblock,
+                                                             mean_var_count_thread_copy_step_I0_n);
+            threadwise_var_load_m_nblock.MoveSrcSliceWindow(mean_var_grid_desc_m_nblock,
+                                                            mean_var_count_thread_copy_step_I0_n);
+            threadwise_count_load_m_nblock.MoveSrcSliceWindow(count_grid_desc_m_nblock,
+                                                              mean_var_count_thread_copy_step_I0_n);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        // step2: normalization
+        // h[m, n] = [(e[m, n] - mean[m]) / sqrt(var[m] + eps)] * gamma[n] + beta[n]
+        threadwise_e_load_m_n.Run(e_grid_desc_m_n,
+                                  e_global_val_buf,
+                                  thread_buffer_desc_m_n,
+                                  make_tuple(I0, I0),
+                                  e_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            auto divisor = 1 / ck::math::sqrt(welford_var_thread_buf(m) + epsilon);
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) =
+                    (e_thread_buf(Number<m_n>{}) - welford_mean_thread_buf(m)) * divisor;
+            });
+        });
+
+        threadwise_gamma_load_n.Run(gamma_grid_desc_n,
+                                    gamma_global_val_buf,
+                                    thread_buffer_desc_n,
+                                    make_tuple(I0),
+                                    gamma_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) = h_thread_buf(Number<m_n>{}) * gamma_thread_buf(n);
+            });
+        });
+
+        threadwise_beta_load_n.Run(beta_grid_desc_n,
+                                   beta_global_val_buf,
+                                   thread_buffer_desc_n,
+                                   make_tuple(I0),
+                                   beta_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) = h_thread_buf(Number<m_n>{}) + beta_thread_buf(n);
+            });
+        });
+
+        threadwise_h_store_m_n.Run(thread_buffer_desc_m_n,
+                                   make_tuple(I0, I0),
+                                   h_thread_buf,
+                                   h_grid_desc_m_n,
+                                   h_global_val_buf);
+
+    } // run
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.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_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_v6r1.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/block/blockwise_softmax.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename D0sDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename C1GridDesc_M_N,
+          typename D0sGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+
+    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<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __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(AK0, 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(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * 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;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const C1GridDesc_M_N& c1_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c1_grid_desc_m_n.GetLength(I0) && Gemm1N == c1_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c1_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        const auto M = c1_grid_desc_m_n.GetLength(I0);
+        const auto N = c1_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c1_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, C1GridDesc_M_N>(
+            c1_grid_desc_m_n);
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    static constexpr auto MakeD0sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                return static_cast<const D0DataType*>(nullptr);
+            },
+            Number<NumD0Tensor>{});
+    }
+    // D0 desc for source in blockwise copy
+    template <typename D0GridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
+    {
+        const auto M = d0_grid_desc_m_n.GetLength(I0);
+        const auto N = d0_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            d0_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    // D0s desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    using D0sGridPointer                                  = decltype(MakeD0sGridPointer());
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(C1GridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               D0sGridPointer p_d0s_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const C0DEElementwiseOperation& c0de_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const C1DEElementwiseOperation& c1de_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        const auto d0s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d0s_grid[i],
+                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
+            },
+            Number<NumD0Tensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // 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();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // d0 matrix threadwise copy
+        constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           I1,   // MRepeat
+                                                           I1,   // NRepeat
+                                                           I1,   // MWaveId
+                                                           I1,   // NWaveId
+                                                           I1,   // MPerXdl
+                                                           I1,   // NGroupNum
+                                                           I1,   // NInputNum
+                                                           n4)); // registerNum
+
+        auto d0s_thread_buf = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return StaticBuffer<
+                    AddressSpaceEnum::Vgpr,
+                    D0DataType,
+                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                    true>{};
+            },
+            Number<NumD0Tensor>{});
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MXdlPerWave>{}, Number<NXdlPerWave>{}, n2, n4));
+
+        auto d0s_threadwise_copy = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return ThreadwiseTensorSliceTransfer_v2<
+                    D0DataType,
+                    D0DataType,
+                    decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
+                    decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+                    Sequence<I1, I1, I1, I1, I1, I1, I1, I1, I1, n4>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+                    9,
+                    n4,
+                    1,
+                    false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                           make_multi_index(block_work_idx[I0], // MBlockId
+                                            0,                  // NBlockId
+                                            0,                  // mrepeat
+                                            0,                  // nrepeat
+                                            wave_id[I0],        // MWaveId
+                                            wave_id[I1],        // NWaveId
+                                            wave_m_n_id[I1],    // MPerXdl
+                                            0,                  // group
+                                            wave_m_n_id[I0],    // NInputIndex
+                                            0));                // register number
+            },
+            Number<NumD0Tensor>{});
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+            // multiple d
+            if constexpr(NumD0Tensor)
+            {
+                static_for<0, MXdlPerWave, 1>{}([&](auto mr) {
+                    static_for<0, NXdlPerWave, 1>{}([&](auto nr) {
+                        static_for<0, n2, 1>{}([&](auto groupid) {
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).Run(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    d0s_grid_buf[i],
+                                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                    make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                    d0s_thread_buf(i));
+                            });
+
+                            static_for<0, n4, 1>{}([&](auto i) {
+                                constexpr index_t c_offset = acc0_thread_desc.CalculateOffset(
+                                    make_tuple(mr, nr, groupid, i));
+
+                                // get reference to src data
+                                const auto src_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto iSrc) -> const auto& {
+                                        return d0s_thread_buf[iSrc][i];
+                                    },
+                                    Number<NumD0Tensor>{});
+
+                                // get reference to dst data
+                                auto dst_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto) -> auto& {
+                                        return acc_thread_buf(Number<c_offset>{});
+                                    },
+                                    Number<2>{});
+
+                                unpack2(c0de_element_op, dst_data_refs, src_data_refs);
+                            });
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0));
+                            });
+                        });
+                        static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                            d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0));
+                        });
+                    });
+                    static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                        d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                            d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                            make_multi_index(0, 0, 1, -NXdlPerWave, 0, 0, 0, 0, 0, 0));
+                    });
+                });
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                        make_multi_index(0, 1, -MXdlPerWave, 0, 0, 0, 0, 0, 0, 0));
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { c0de_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(c_thread_lengths),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+                    Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for reduction LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_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<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = 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 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_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_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    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],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                C1DEElementwiseOperation,   // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c1de_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  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_c_global.GetNumOfAccess(), "wrong!");
+
+            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_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              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
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -879,14 +879,6 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
                 }
             } // end gemm1
 
-            // workaround compiler issue; see ck/ck.hpp
-            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
-                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
-                         Gemm1NPerBlock == 128)
-            {
-                __builtin_amdgcn_sched_barrier(0);
-            }
-
             constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
                 gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
             constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);

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

@@ -434,7 +434,7 @@ struct GridwiseElementwiseLayernormWelfordVariance_mk_to_mk
             });
 
             static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
                 static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
                     static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
                         constexpr auto offset_m_k =

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+template <typename TileLoadThreadGroup, index_t NumGemmKPrefetchStage>
+struct GridwiseGemmLoadWave;
+
+// 1-stage prefetch
+template <typename TileLoadThreadGroup>
+struct GridwiseGemmLoadWave<TileLoadThreadGroup, 1>
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */)
+    {
+        // TODO: improve applicability
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep>
+    static __device__ void RunLoadWavePipeline(const AGridDesc& a_grid_desc,
+                                               const ABlockDesc& a_block_desc,
+                                               ABlockTransfer& a_blockwise_copy,
+                                               const AGridBuffer& a_grid_buf,
+                                               ABlockBuffer& a_block_buf,
+                                               const ABlockTransferStep& a_block_copy_step,
+                                               const BGridDesc& b_grid_desc,
+                                               const BBlockDesc& b_block_desc,
+                                               BBlockTransfer& b_blockwise_copy,
+                                               const BGridBuffer& b_grid_buf,
+                                               BBlockBuffer& b_block_buf,
+                                               const BBlockTransferStep& b_block_copy_step,
+                                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // move to 1
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                // sync for Load threads()
+                block_sync_lds();
+                // global read i + 1
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                // move to i + 2
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // sync with math threads()
+                block_sync_lds();
+
+                // LDS write i+1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+            // GEMM num_loop - 1
+        }
+    }
+};
+
+template <typename TileMathThreadGroup, index_t NumGemmKPrefetchStage>
+struct GridwiseGemmMathWave;
+// 1- stage prefetch
+template <typename TileMathThreadGroup>
+struct GridwiseGemmMathWave<TileMathThreadGroup, 1>
+{
+
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void RunMathWavePipeline(ABlockBuffer& a_block_buf,
+                                               BBlockBuffer& b_block_buf,
+                                               const BlockwiseGemm& block_gemm,
+                                               CThreadBuffer& c_thread_buf,
+                                               index_t num_loop)
+    {
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                block_sync_lds();
+
+                // GEMM i
+                block_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            block_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.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_waveletmodel.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_v6r1.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 ABDataType,
+          typename FloatGemmAcc,
+          typename EDataTypeShuffle,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename EElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename EGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t TileLoadThreadGroupSize,
+          index_t TileMathThreadGroupSize,
+          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_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_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 CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle
+{
+
+    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>{};
+
+    struct TileLoadThreadGroup
+    {
+        __device__ static constexpr index_t GetNumOfThread() { return TileLoadThreadGroupSize; }
+
+        __device__ static constexpr bool IsBelong()
+        {
+            return (get_thread_local_1d_id() >= TileLoadThreadGroupSize);
+        }
+
+        __device__ static index_t GetThreadId()
+        {
+            return get_thread_local_1d_id() - TileMathThreadGroupSize;
+        }
+    };
+
+    struct TileMathThreadGroup
+    {
+        __device__ static constexpr index_t GetNumOfThread() { return TileMathThreadGroupSize; }
+
+        __device__ static constexpr bool IsBelong()
+        {
+            return get_thread_local_1d_id() < TileMathThreadGroupSize;
+        }
+
+        __device__ static index_t GetThreadId() { return get_thread_local_1d_id(); }
+    };
+
+    using CShuffleBlockTransferThreadGroup = ThisThreadBlock<TileMathThreadGroupSize>;
+
+    // load and math+store Wave pipelines.
+    // TODO: build pipelines blocks scheduling parallel tasks
+    using GridwiseGemmLoad = GridwiseGemmLoadWave<TileLoadThreadGroup, NumGemmKPrefetchStage>;
+    using GridwiseGemmMath = GridwiseGemmMathWave<TileMathThreadGroup, NumGemmKPrefetchStage>;
+
+    __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;
+    }
+
+    __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 + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(EDataTypeShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                  const BGridDesc_N_K& b_grid_desc_n_k,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2ETileMap& /*block_2_etile_map*/)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
+             K == b_grid_desc_n_k.GetLength(I1)))
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmMath::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_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= 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 GridwiseGemmMath::CalculateHasMainLoop(num_loop);
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(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);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M01 = I1;
+        constexpr auto N01 = I1;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M0, M01)),
+                           make_unmerge_transform(make_tuple(N0, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M0, N0, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(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 index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_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;
+    }
+
+    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{}))>;
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const EElementwiseOperation& e_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        // build loadWave and MathWave pipelines
+        // loadWave and MathWave synchronized through LDS
+
+        // 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();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // 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<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(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 / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        // 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 m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        if(TileLoadThreadGroup::IsBelong())
+        {
+
+            // LoadWave
+            const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+            const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<TileLoadThreadGroup,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                    ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    ABDataType,
+                                                    ABDataType,
+                                                    decltype(a_grid_desc_ak0_m_ak1),
+                                                    decltype(a_block_desc_ak0_m_ak1),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    2,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_AK1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true,
+                                                    NumGemmKPrefetchStage>(
+                    a_grid_desc_ak0_m_ak1,
+                    make_multi_index(0, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_block_desc_ak0_m_ak1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // B matrix blockwise copy
+            auto b_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<TileLoadThreadGroup,
+                                                    BElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                    BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    ABDataType,
+                                                    ABDataType,
+                                                    decltype(b_grid_desc_bk0_n_bk1),
+                                                    decltype(b_block_desc_bk0_n_bk1),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    2,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_BK1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true,
+                                                    NumGemmKPrefetchStage>(
+                    b_grid_desc_bk0_n_bk1,
+                    make_multi_index(0, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_block_desc_bk0_n_bk1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            GridwiseGemmLoad::template RunLoadWavePipeline<HasMainKBlockLoop>(
+                a_grid_desc_ak0_m_ak1,
+                a_block_desc_ak0_m_ak1,
+                a_blockwise_copy,
+                a_grid_buf,
+                a_block_buf,
+                a_block_slice_copy_step,
+                b_grid_desc_bk0_n_bk1,
+                b_block_desc_bk0_n_bk1,
+                b_blockwise_copy,
+                b_grid_buf,
+                b_block_buf,
+                b_block_slice_copy_step,
+                num_k_block_main_loop);
+
+            block_sync_lds();
+            block_sync_lds();
+        }
+        else if(TileMathThreadGroup::IsBelong())
+        {
+            // branch early for math wave
+            constexpr index_t KPack =
+                math::max(math::lcm(AK1, BK1),
+                          MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+            auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
+                TileMathThreadGroupSize,
+                ABDataType,
+                FloatGemmAcc,
+                decltype(a_block_desc_ak0_m_ak1),
+                decltype(b_block_desc_bk0_n_bk1),
+                MPerXdl,
+                NPerXdl,
+                MXdlPerWave,
+                NXdlPerWave,
+                KPack>{};
+
+            auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+            auto c_grid_buf   = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            // TODO re-architect LDS+math stages
+            // Writing data to GMEM: only math wave is doing the work in cshuffle
+            GridwiseGemmMath::template RunMathWavePipeline<HasMainKBlockLoop>(
+                a_block_buf, b_block_buf, blockwise_gemm, c_thread_buf, num_k_block_main_loop);
+
+            // 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
+
+            // shuffle C and write out
+            {
+                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<EDataTypeShuffle*>(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<
+                    FloatGemmAcc,
+                    EDataTypeShuffle,
+                    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{}};
+
+                // shuffle: blockwise copy C from LDS to global
+                auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                    CShuffleBlockTransferThreadGroup, // ThreadGroup
+                    EElementwiseOperation,            // ElementwiseOperation,
+                    CGlobalMemoryDataOperation,       // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    EDataTypeShuffle,     // typename SrcData,
+                    EDataType,            // typename DstData,
+                    decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                    3,                                              // index_t VectorDim,
+                    CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                    true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                    {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(0, 0, 0, 0),
+                     e_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                     e_element_op};
+
+                // space filling curve for threadwise C in VGPR
+                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 in global mem
+                constexpr auto sfc_c_global =
+                    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_c_global.GetNumOfAccess(), "wrong!");
+
+                // Different way of getting coalesced writes:
+                // We can get rid of doing cshuffle. Instead of reading A rows in contiguous manner
+                // do it interleaved, then mfma can have nice c-mat layout as below:
+                //
+                // TODO
+                //      We do not need to do LDS swizzle to align global writes writing cache lines:
+                //         v_mfma  cmat, amat, bmat, cmat   - c-mat register layout   are 1xN
+                //                                            elments  (N is vertical or strided
+                //                                            dimension)
+                //         v_mfma  cmat, bmat, amat, cmat   - c-mat register layout   are Mx1
+                //         elments  (M is coalescing
+                //                                            dimension) by enumerating M index in
+                //                                            amat, bmat you can align cmat
+                //                                            register(s) to contiguous M elements
+                //                                            for example
+                //              1st mfma instruction  output space : 0 4 8  12 16 ....
+                //              2nd mfma instruction  output space : 1 5 9  13 17 ....
+                //              3rd mfma instruction  output space : 2 6 10 14 18 ....
+                //              4th mfma instruction  output space : 3 7 11 15 19 ....
+                //              you can pack 4 registers output space into 2WORD and do global write
+                //              (no LDS swizzling required)
+
+                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
+                    c_shuffle_block_copy_lds_to_global.Run(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                        c_shuffle_block_buf,
+                        e_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c_grid_buf);
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                        // move on C
+                        c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                            e_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                    }
+                });
+            }
+        }
+    }
+};
+
+} // namespace ck

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

@@ -4,9 +4,8 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
-#include "ck/utility/reduction_common.hpp"
+
 #include "ck/utility/reduction_operator.hpp"
-#include "ck/utility/reduction_functions_accumulate.hpp"
 #include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
 #include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
@@ -19,8 +18,8 @@ template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename YDataType,
-          typename AccDataType,
-          typename AccElementwiseOperation,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
           typename GridDesc_M_K,
           index_t BlockSize,
           index_t MThreadClusterSize,
@@ -46,6 +45,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                       (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
                   "Invalid thread slice sizes and/or vector sizes configuration, please check!");
 
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
     static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
 
     using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
@@ -59,19 +62,23 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
     using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
-        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 
-    using BlockwiseSumReduce = PartitionedBlockwiseReduction<AccDataType,
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction<ComputeDataType,
                                                              BlockSize,
                                                              ThreadClusterLengths_M_K,
                                                              ThreadClusterArrangeOrder,
                                                              reduce::Add,
                                                              true>;
 
-    using ThreadwiseSumReduce = ThreadwiseReduction<AccDataType,
+    using ThreadwiseSumReduce = ThreadwiseReduction<ComputeDataType,
                                                     ThreadReduceSrcDesc_M_K,
                                                     ThreadReduceDstDesc_M,
                                                     reduce::Add,
@@ -81,64 +88,70 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
 
-    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
-    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
 
     __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
                                const GridDesc_M_K& gamma_grid_desc_m_k,
                                const GridDesc_M_K& beta_grid_desc_m_k,
                                const GridDesc_M_K& y_grid_desc_m_k,
                                index_t num_k_block_tile_iteration,
-                               AccDataType epsilon,
+                               ComputeDataType epsilon,
                                const XDataType* const __restrict__ p_x_global,
                                const GammaDataType* const __restrict__ p_gamma_global,
                                const BetaDataType* const __restrict__ p_beta_global,
                                YDataType* const __restrict__ p_y_global,
-                               const AccElementwiseOperation acc_elementwise_op)
+                               const YElementwiseOperation y_elementwise_op)
     {
-        if constexpr(SweepOnce)
-        {
-            num_k_block_tile_iteration = 1;
-        }
-
         // LDS
-        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
-
-        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+        __shared__ ComputeDataType p_reduce_work_buffer[BlockSize];
 
         auto reduce_work_buf =
             make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            x_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            gamma_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr,
-                     AccDataType,
-                     MThreadSliceSize * KThreadSliceSize,
-                     true>& beta_thread_buf = gamma_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            y_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr,
-                     AccDataType,
-                     MThreadSliceSize * KThreadSliceSize,
-                     true>& x_square_thread_buf = y_thread_buf;
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
-            mean_square_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& var_thread_buf =
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& beta_thread_buf = gamma_thread_buf;
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& x_square_thread_buf = y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
             mean_square_thread_buf;
-
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<AccDataType>();
-            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
-        });
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>&
+            var_thread_buf = mean_square_thread_buf;
 
         const index_t thread_local_id = get_thread_local_1d_id();
         const index_t block_global_id = get_block_1d_id();
@@ -149,12 +162,8 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
         const auto thread_m_cluster_id = thread_cluster_idx[I0];
         const auto thread_k_cluster_id = thread_cluster_idx[I1];
 
-        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
-        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
-            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
-
         auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
-                                                                  AccDataType,
+                                                                  ComputeDataType,
                                                                   GridDesc_M_K,
                                                                   decltype(thread_buffer_desc_m_k),
                                                                   ThreadBufferLengths_M_K,
@@ -166,11 +175,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
             x_grid_desc_m_k,
             make_multi_index(block_global_id * M_BlockTileSize +
                                  thread_m_cluster_id * MThreadSliceSize,
-                             thread_k_cluster_id * KThreadSliceSize));
+                             thread_k_cluster_id * XSrcVectorSize));
 
         auto threadwise_gamma_load =
             ThreadwiseTensorSliceTransfer_v2<GammaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -182,11 +191,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 gamma_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize));
+                                 thread_k_cluster_id * GammaSrcVectorSize));
 
         auto threadwise_beta_load =
             ThreadwiseTensorSliceTransfer_v2<BetaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -198,14 +207,14 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 beta_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize));
+                                 thread_k_cluster_id * BetaSrcVectorSize));
 
         auto threadwise_y_store =
-            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
                                                YDataType,
                                                decltype(thread_buffer_desc_m_k),
                                                GridDesc_M_K,
-                                               AccElementwiseOperation,
+                                               YElementwiseOperation,
                                                ThreadBufferLengths_M_K,
                                                ThreadBufferDimAccessOrder,
                                                YDstVectorDim,
@@ -216,13 +225,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 y_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize),
-                acc_elementwise_op);
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
 
-        // Copy x from Cache
-        // one pass: fwd, second pass: bwd
-        constexpr auto thread_copy_fwd_step_m_k =
-            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
         constexpr auto thread_copy_bwd_step_m_k =
             make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
 
@@ -239,121 +245,260 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
         // FIXME: Should not hack the transform from deviceOP
         int reduce_length = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
 
-        index_t reducedTiles = 0;
-        do
-        {
-            threadwise_x_load.Run(x_grid_desc_m_k,
-                                  x_global_val_buf,
-                                  thread_buffer_desc_m_k,
-                                  make_tuple(I0, I0),
-                                  x_thread_buf);
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<ComputeDataType>();
+            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<ComputeDataType>();
+        });
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
-                    x_square_thread_buf(Number<offset_m_k>{}) =
-                        x_thread_buf(Number<offset_m_k>{}) * x_thread_buf(Number<offset_m_k>{});
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                        x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                            x_thread_buf(i)(Number<offset_m_k>{}) *
+                            x_thread_buf(i)(Number<offset_m_k>{});
+                    });
                 });
-            });
 
-            ThreadwiseSumReduce::Reduce(x_thread_buf, mean_thread_buf);
-            ThreadwiseSumReduce::Reduce(x_square_thread_buf, mean_square_thread_buf);
+                ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
 
-            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
 
-            ++reducedTiles;
-        } while(reducedTiles < num_k_block_tile_iteration);
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
 
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            if constexpr(I > 0)
                 block_sync_lds();
 
-            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
-            mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
+            });
 
-            block_sync_lds();
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma & beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
 
-            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
-            mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
 
-            // var(x) = E[x^2] - E[x]^2
-            var_thread_buf(I) =
-                mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
-        });
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+            });
 
-        // y = (x - E[x]) / sqrt(var[x] + epsilon)
-        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
 
-        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
 
-        reducedTiles = 0;
-        do
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
         {
-            if constexpr(!SweepOnce)
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
             {
-                threadwise_x_load.Run(x_grid_desc_m_k,
-                                      x_global_val_buf,
-                                      thread_buffer_desc_m_k,
-                                      make_tuple(I0, I0),
-                                      x_thread_buf);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+                    static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                            x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                                x_thread_buf(i)(Number<offset_m_k>{}) *
+                                x_thread_buf(i)(Number<offset_m_k>{});
+                        });
+                    });
+
+                    ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                    ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
+                });
             }
 
-            threadwise_gamma_load.Run(gamma_grid_desc_m_k,
-                                      gamma_global_val_buf,
-                                      thread_buffer_desc_m_k,
-                                      make_tuple(I0, I0),
-                                      gamma_thread_buf);
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
-
-                    // normalize
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        (x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) /
-                        sqrt(var_thread_buf(iM) + epsilon);
-
-                    // gamma
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{});
-                });
-            });
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
 
-            threadwise_beta_load.Run(beta_grid_desc_m_k,
-                                     beta_global_val_buf,
-                                     thread_buffer_desc_m_k,
-                                     make_tuple(I0, I0),
-                                     beta_thread_buf);
+                block_sync_lds();
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
 
-                    // beta
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{});
-                });
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
             });
 
-            threadwise_y_store.Run(thread_buffer_desc_m_k,
-                                   make_tuple(I0, I0),
-                                   y_thread_buf,
-                                   y_grid_desc_m_k,
-                                   y_global_val_buf);
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
 
             threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
 
-            ++reducedTiles;
-        } while(reducedTiles < num_k_block_tile_iteration);
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
     }
 };
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
+
+namespace ck {
+template <typename GridwiseReduction,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K>
+__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
+                                     const GridDesc_M_K gamma_grid_desc_m_k,
+                                     const GridDesc_M_K beta_grid_desc_m_k,
+                                     const GridDesc_M_K y_grid_desc_m_k,
+                                     index_t num_k_block_tile_iteration,
+                                     ComputeDataType epsilon,
+                                     const XDataType* const __restrict__ p_x_global,
+                                     const GammaDataType* const __restrict__ p_gamma_global,
+                                     const BetaDataType* const __restrict__ p_beta_global,
+                                     YDataType* const __restrict__ p_y_global,
+                                     const YElementwiseOperation y_elementwise_op)
+{
+    GridwiseReduction::Run(x_grid_desc_m_k,
+                           gamma_grid_desc_m_k,
+                           beta_grid_desc_m_k,
+                           y_grid_desc_m_k,
+                           num_k_block_tile_iteration,
+                           epsilon,
+                           p_x_global,
+                           p_gamma_global,
+                           p_beta_global,
+                           p_y_global,
+                           y_elementwise_op);
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool UseWelford>
+auto NormalizationKernelSelector(bool isSweepOnce)
+{
+    using GridwiseNormalizationGenericNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    false>;
+    using GridwiseNormalizationSweepOnceNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    true>;
+    using GridwiseNormalizationGenericWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      false>;
+    using GridwiseNormalizationSweepOnceWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      true>;
+
+    if constexpr(UseWelford)
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+    else
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+}
+
+} // namespace ck

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

@@ -16,8 +16,8 @@ template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename YDataType,
-          typename AccDataType,
-          typename AccElementwiseOperation,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
           typename GridDesc_M_K,
           index_t BlockSize,
           index_t MThreadClusterSize,
@@ -43,6 +43,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                       (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
                   "Invalid thread slice sizes and/or vector sizes configuration, please check!");
 
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
     static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
 
     using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
@@ -56,15 +60,19 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
     using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
         make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 
     using ThreadwiseWelford =
-        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+        ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
 
-    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
                                               BlockSize,
                                               ThreadClusterLengths_M_K,
                                               ThreadClusterArrangeOrder>;
@@ -77,10 +85,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
     static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
     static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
 
-    static constexpr auto XThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto BetaThreadBufferNumber  = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto YThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
 
     __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
                                         int thread_k_cluster_id)
@@ -93,7 +98,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         if(kPerBlockTail > 0)
         {
-            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 int thread_max_len =
                     (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
                 int delta = thread_max_len - kPerBlockTail;
@@ -110,59 +115,41 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                                const GridDesc_M_K& beta_grid_desc_m_k,
                                const GridDesc_M_K& y_grid_desc_m_k,
                                index_t num_k_block_tile_iteration,
-                               AccDataType epsilon,
+                               ComputeDataType epsilon,
                                const XDataType* const __restrict__ p_x_global,
                                const GammaDataType* const __restrict__ p_gamma_global,
                                const BetaDataType* const __restrict__ p_beta_global,
                                YDataType* const __restrict__ p_y_global,
-                               const AccElementwiseOperation acc_elementwise_op)
+                               const YElementwiseOperation y_elementwise_op)
     {
-        if constexpr(SweepOnce)
-        {
-            num_k_block_tile_iteration = 1;
-        }
-
         auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
 
         auto x_thread_buf = generate_tuple(
             [&](auto) {
                 return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
+                                    ComputeDataType,
                                     MThreadSliceSize * XSrcVectorSize,
                                     true>{};
             },
-            Number<XThreadBufferNumber>{});
+            Number<ThreadBufferNumber>{});
 
         auto gamma_thread_buf = generate_tuple(
             [&](auto) {
                 return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
+                                    ComputeDataType,
                                     MThreadSliceSize * GammaSrcVectorSize,
                                     true>{};
             },
-            Number<GammaThreadBufferNumber>{});
-
-        auto beta_thread_buf = generate_tuple(
-            [&](auto) {
-                return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
-                                    MThreadSliceSize * BetaSrcVectorSize,
-                                    true>{};
-            },
-            Number<BetaThreadBufferNumber>{});
+            Number<ThreadBufferNumber>{});
 
-        auto y_thread_buf = generate_tuple(
-            [&](auto) {
-                return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
-                                    MThreadSliceSize * YDstVectorSize,
-                                    true>{};
-            },
-            Number<YThreadBufferNumber>{});
+        auto& beta_thread_buf = gamma_thread_buf;
+        auto& y_thread_buf    = x_thread_buf;
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            var_thread_buf;
 
         const index_t thread_local_id = get_thread_local_1d_id();
         const index_t block_global_id = get_block_1d_id();
@@ -173,12 +160,8 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
         const auto thread_m_cluster_id = thread_cluster_idx[I0];
         const auto thread_k_cluster_id = thread_cluster_idx[I1];
 
-        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, XSrcVectorSize>;
-        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
-            make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
-
         auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
-                                                                  AccDataType,
+                                                                  ComputeDataType,
                                                                   GridDesc_M_K,
                                                                   decltype(thread_buffer_desc_m_k),
                                                                   ThreadBufferLengths_M_K,
@@ -194,7 +177,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         auto threadwise_gamma_load =
             ThreadwiseTensorSliceTransfer_v2<GammaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -210,7 +193,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         auto threadwise_beta_load =
             ThreadwiseTensorSliceTransfer_v2<BetaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -225,11 +208,11 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                                  thread_k_cluster_id * BetaSrcVectorSize));
 
         auto threadwise_y_store =
-            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
                                                YDataType,
                                                decltype(thread_buffer_desc_m_k),
                                                GridDesc_M_K,
-                                               AccElementwiseOperation,
+                                               YElementwiseOperation,
                                                ThreadBufferLengths_M_K,
                                                ThreadBufferDimAccessOrder,
                                                YDstVectorDim,
@@ -241,7 +224,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
                                  thread_k_cluster_id * YDstVectorSize),
-                acc_elementwise_op);
+                y_elementwise_op);
 
         constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
         constexpr auto thread_copy_bwd_step_m_k =
@@ -260,67 +243,47 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
         threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
 
         static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
-            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
         });
 
-        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
         {
-            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_x_load.Run(x_grid_desc_m_k,
                                       x_global_val_buf,
                                       thread_buffer_desc_m_k,
                                       make_tuple(I0, I0),
                                       x_thread_buf(i));
-                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
-                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
-            });
-        }
-
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            if constexpr(I > 0)
-                block_sync_lds();
 
-            int count = threadwise_welford.cur_count_;
-            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
-        });
-
-        auto thread_copy_tail_m_k =
-            (num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
-
-        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
-
-        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
-        {
-            if constexpr(!SweepOnce)
-            {
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
-                    threadwise_x_load.Run(x_grid_desc_m_k,
-                                          x_global_val_buf,
-                                          thread_buffer_desc_m_k,
-                                          make_tuple(I0, I0),
-                                          x_thread_buf(i));
-                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
-                });
-            }
-
-            static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_gamma_load.Run(gamma_grid_desc_m_k,
                                           gamma_global_val_buf,
                                           thread_buffer_desc_m_k,
                                           make_tuple(I0, I0),
                                           gamma_thread_buf(i));
 
-                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
-                                                         thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
             });
 
             static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
                     static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
                         constexpr auto offset_m_k =
                             thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
@@ -330,7 +293,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                             (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
                             divisor;
 
-                        // gamma
+                        // gamma & beta
                         y_thread_buf(iK0)(Number<offset_m_k>{}) =
                             y_thread_buf(iK0)(Number<offset_m_k>{}) *
                             gamma_thread_buf(iK0)(Number<offset_m_k>{});
@@ -338,18 +301,20 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 });
             });
 
-            static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_beta_load.Run(beta_grid_desc_m_k,
                                          beta_global_val_buf,
                                          thread_buffer_desc_m_k,
                                          make_tuple(I0, I0),
                                          beta_thread_buf(i));
-                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
-                                                        thread_copy_fwd_step_m_k);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
             });
 
             static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
                     static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
                         constexpr auto offset_m_k =
                             thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
@@ -362,22 +327,134 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 });
             });
 
-            static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_y_store.Run(thread_buffer_desc_m_k,
                                        make_tuple(I0, I0),
                                        y_thread_buf(i),
                                        y_grid_desc_m_k,
                                        y_global_val_buf);
-                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
+        {
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+                });
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
             });
 
-            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
-            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
-                                                     2 * thread_copy_bwd_step_m_k);
-            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
-                                                    2 * thread_copy_bwd_step_m_k);
-            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
-        }
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
     }
 };
 

include/ck/tensor_operation/gpu/grid/gridwise_sparse_embedding3_forward_layernorm.hpp → include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp

@@ -17,33 +17,24 @@ template <typename GridwiseSparseEmbedding,
           typename BetaDataType,
           typename AccDataType,
           typename OutType,
-          typename OutGridDesc>
+          typename OutGridDesc,
+          typename EmbElementwiseOperation,
+          ck::index_t NumEmbeddings>
 #if CK_USE_LAUNCH_BOUNDS
 __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-    __global__ void kernel_sparse_embedding3_forward_layernorm(OutType* p_out,
-                                                               const EmbType* p_emb_a,
-                                                               const EmbType* p_emb_b,
-                                                               const EmbType* p_emb_c,
-                                                               const IndexType* p_index_a,
-                                                               const IndexType* p_index_b,
-                                                               const IndexType* p_index_c,
-                                                               const GammaDataType* p_gamma,
-                                                               const BetaDataType* p_beta,
-                                                               const OutGridDesc out_grid_desc,
-                                                               const AccDataType epsilon)
+    __global__ void kernel_sparse_embeddings_forward_layernorm(
+        OutType* p_out,
+        const ck::Array<EmbType*, NumEmbeddings> p_embs,
+        const ck::Array<IndexType*, NumEmbeddings> p_indexes,
+        const GammaDataType* p_gamma,
+        const BetaDataType* p_beta,
+        const OutGridDesc out_grid_desc,
+        const AccDataType epsilon,
+        const EmbElementwiseOperation emb_elementwise_op)
 {
-    GridwiseSparseEmbedding::Run(p_out,
-                                 p_emb_a,
-                                 p_emb_b,
-                                 p_emb_c,
-                                 p_index_a,
-                                 p_index_b,
-                                 p_index_c,
-                                 p_gamma,
-                                 p_beta,
-                                 out_grid_desc,
-                                 epsilon);
+    GridwiseSparseEmbedding::Run(
+        p_out, p_embs, p_indexes, p_gamma, p_beta, out_grid_desc, epsilon, emb_elementwise_op);
 }
 
 template <typename EmbType,
@@ -53,14 +44,16 @@ template <typename EmbType,
           typename AccDataType,
           typename OutType,
           typename OutGridDesc,
+          typename EmbElementwiseOperation,
           ck::index_t BlockSize,
           ck::index_t DimClusterSize,
           ck::index_t RowClusterSize,
           ck::index_t DimPerBlock,   // Row x Dim, along Dim
           ck::index_t RowPerBlock,   // Row x Dim, along Row
           ck::index_t DimThreadSize, // this is actually not vector, but number of registers
-          ck::index_t RowVectorSize>
-struct GridwiseSparseEmbedding3ForwardLayernorm
+          ck::index_t RowVectorSize,
+          ck::index_t NumEmbeddings>
+struct GridwiseSparseEmbeddingsForwardLayernorm
 {
     static constexpr auto I0          = Number<0>{};
     static constexpr auto I1          = Number<1>{};
@@ -97,23 +90,17 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
         BlockwiseWelford<AccDataType, BlockSize, ThreadClusterLength, Sequence<0, 1>>;
 
     __device__ static void Run(OutType* p_out,
-                               const EmbType* p_emb_a,
-                               const EmbType* p_emb_b,
-                               const EmbType* p_emb_c,
-                               const IndexType* p_index_a,
-                               const IndexType* p_index_b,
-                               const IndexType* p_index_c,
+                               const ck::Array<EmbType*, NumEmbeddings> p_embs,
+                               const ck::Array<IndexType*, NumEmbeddings> p_indexes,
                                const GammaDataType* p_gamma,
                                const BetaDataType* p_beta,
                                const OutGridDesc,
-                               const AccDataType epsilon)
+                               const AccDataType epsilon,
+                               const EmbElementwiseOperation emb_elementwise_op)
     {
         const index_t thread_local_id = get_thread_local_1d_id();
         const index_t block_global_id = get_block_1d_id();
 
-        // const auto index_length = out_grid_desc.GetLength(I0);
-        // const auto emb_dim      = out_grid_desc.GetLength(I1);
-
         constexpr auto thread_cluster_desc =
             make_cluster_descriptor(Sequence<DimClusterSize, RowClusterSize>{}, Sequence<0, 1>{});
 
@@ -141,13 +128,11 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
         constexpr auto gamma_beta_buf_desc =
             make_naive_tensor_descriptor_packed(make_tuple(RowSubBlocks, RowVectorSize));
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_a;
-        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_b;
-        StaticBuffer<AddressSpaceEnum::Vgpr, EmbType, thread_buf_size, true> in_thread_buf_c;
-
-        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_a;
-        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_b;
-        StaticBuffer<AddressSpaceEnum::Sgpr, IndexType, DimPerBlock, true> index_buf_c;
+        ck::Array<StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, thread_buf_size, true>,
+                  NumEmbeddings>
+            in_thread_bufs;
+        ck::Array<StaticBuffer<AddressSpaceEnum::Vgpr, IndexType, DimPerBlock, true>, NumEmbeddings>
+            index_bufs;
 
         StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, thread_buf_size, true> acc_thread_buf;
 
@@ -160,42 +145,31 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
         StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, mean_var_buf_size, true> var_thread_buf;
 
         auto load_current_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
-            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_a;
-            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_b;
-            vector_type_maker_t<EmbType, RowVectorSize> emb_vector_c;
-
-            using src_vector_t = typename decltype(emb_vector_a)::type;
+            ck::Array<vector_type_maker_t<EmbType, RowVectorSize>, NumEmbeddings> emb_vectors;
+            auto emb_a         = emb_vectors[0];
+            using src_vector_t = typename decltype(emb_a)::type;
             static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
                 constexpr auto current_dim = i_dim_sub_ * DimPerSubBlock + i_dim_vec_;
-                IndexType index_a          = index_buf_a[Number<current_dim>{}];
-                IndexType index_b          = index_buf_b[Number<current_dim>{}];
-                IndexType index_c          = index_buf_c[Number<current_dim>{}];
 
                 auto thread_offset = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
                                      sizeof(EmbType) * RowVectorSize;
+                static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                    IndexType index = index_bufs[i_embedding_][Number<current_dim>{}];
 
-                int32x4_t emb_res_a =
-                    make_wave_buffer_resource_with_default_range(p_emb_a + index_a * RowPerBlock);
-                int32x4_t emb_res_b =
-                    make_wave_buffer_resource_with_default_range(p_emb_b + index_b * RowPerBlock);
-                int32x4_t emb_res_c =
-                    make_wave_buffer_resource_with_default_range(p_emb_c + index_c * RowPerBlock);
-                emb_vector_a.template AsType<src_vector_t>()(I0) =
-                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_a, thread_offset, 0);
-                emb_vector_b.template AsType<src_vector_t>()(I0) =
-                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_b, thread_offset, 0);
-                emb_vector_c.template AsType<src_vector_t>()(I0) =
-                    amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res_c, thread_offset, 0);
+                    int32x4_t emb_res = make_wave_buffer_resource_with_default_range(
+                        p_embs[i_embedding_] + index * RowPerBlock);
+                    emb_vectors(i_embedding_).template AsType<src_vector_t>()(I0) =
+                        amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res, thread_offset, 0);
+                });
 
                 static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
                     constexpr auto register_offset = thread_buf_desc.CalculateOffset(
                         make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
-                    in_thread_buf_a(Number<register_offset>{}) =
-                        emb_vector_a.template AsType<EmbType>()[i_row_vec_];
-                    in_thread_buf_b(Number<register_offset>{}) =
-                        emb_vector_b.template AsType<EmbType>()[i_row_vec_];
-                    in_thread_buf_c(Number<register_offset>{}) =
-                        emb_vector_c.template AsType<EmbType>()[i_row_vec_];
+                    static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                        in_thread_bufs(i_embedding_)(Number<register_offset>{}) =
+                            ck::type_convert<AccDataType>(
+                                emb_vectors[i_embedding_].template AsType<EmbType>()[i_row_vec_]);
+                    });
                 });
             });
         };
@@ -205,14 +179,15 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
                 static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
                     constexpr auto register_offset = thread_buf_desc.CalculateOffset(
                         make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
-                    AccDataType va =
-                        ck::type_convert<AccDataType>(in_thread_buf_a(Number<register_offset>{}));
-                    AccDataType vb =
-                        ck::type_convert<AccDataType>(in_thread_buf_b(Number<register_offset>{}));
-                    AccDataType vc =
-                        ck::type_convert<AccDataType>(in_thread_buf_c(Number<register_offset>{}));
-
-                    acc_thread_buf(Number<register_offset>{}) += va + vb + vc;
+                    auto in_data_refs = generate_tie(
+                        [&](auto i_embedding_) -> const auto& {
+                            return in_thread_bufs(i_embedding_)(Number<register_offset>{});
+                        },
+                        Number<NumEmbeddings>{});
+                    auto out_data_refs = generate_tie(
+                        [&](auto) -> auto& { return acc_thread_buf(Number<register_offset>{}); },
+                        Number<1>{});
+                    unpack2(emb_elementwise_op, out_data_refs, in_data_refs);
                 });
             });
         };
@@ -242,7 +217,8 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
 
                 constexpr auto mean_var_offset =
                     mean_var_buf_desc.CalculateOffset(make_tuple(i_dim_sub_, i_dim_vec_));
-
+                auto divisor =
+                    1 / __builtin_amdgcn_sqrtf(var_thread_buf(Number<mean_var_offset>{}) + epsilon);
                 static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
                     constexpr auto register_offset = thread_buf_desc.CalculateOffset(
                         make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
@@ -250,9 +226,8 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
                         gamma_beta_buf_desc.CalculateOffset(make_tuple(i_row_sub_, i_row_vec_));
 
                     auto acc_val = acc_thread_buf[Number<register_offset>{}];
-                    acc_val      = (acc_val - mean_thread_buf(Number<mean_var_offset>{})) /
-                              sqrt(var_thread_buf(Number<mean_var_offset>{}) + epsilon);
-                    acc_val = acc_val * gamma_thread_buf[Number<gamma_beta_offset>{}] +
+                    acc_val      = (acc_val - mean_thread_buf(Number<mean_var_offset>{})) * divisor;
+                    acc_val      = acc_val * gamma_thread_buf[Number<gamma_beta_offset>{}] +
                               beta_thread_buf[Number<gamma_beta_offset>{}];
 
                     out_vector.template AsType<OutType>()(Number<i_row_vec_>{}) =
@@ -273,9 +248,10 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
         // first load index
         ck::static_for<0, DimPerBlock, 1>{}([&](auto i_idx_) {
             // prefer use s_load
-            index_buf_a(i_idx_) = p_index_a[index_start + i_idx_.value];
-            index_buf_b(i_idx_) = p_index_b[index_start + i_idx_.value];
-            index_buf_c(i_idx_) = p_index_c[index_start + i_idx_.value];
+            ck::static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                index_bufs(i_embedding_)(i_idx_) =
+                    p_indexes[i_embedding_][index_start + i_idx_.value];
+            });
         });
 
         // load gamma/beta
@@ -329,7 +305,6 @@ struct GridwiseSparseEmbedding3ForwardLayernorm
             static_for<0, mean_var_buf_size, 1>{}([&](auto I) {
                 if constexpr(I > 0)
                     block_sync_lds();
-
                 BlockwiseWelford::Run(
                     mean_thread_buf(I), var_thread_buf(I), threadwise_welford.cur_count_);
             });

include/ck/utility/math_v2.hpp

@@ -83,6 +83,11 @@ static inline __host__ bool isnan(int4_t x)
 };
 #endif
 
+static inline __host__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(std::sqrt(static_cast<float>(x)));
+};
+
 static inline __host__ float sqrt(float x) { return std::sqrt(x); };
 
 static inline __host__ double sqrt(double x) { return std::sqrt(x); };
@@ -158,9 +163,14 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
-static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
+static inline __device__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x)));
+};
+
+static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
 
-static inline __device__ double sqrt(double x) { return ::sqrt(x); };
+static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
 
 } // namespace math
 } // namespace ck

include/ck/utility/reduction_operator.hpp

@@ -251,27 +251,27 @@ constexpr T GetIdentityValueForInMemoryDataOperation(InMemoryDataOperationEnum o
 };
 
 template <InMemoryDataOperationEnum Operation, typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType
+struct InMemoryDataOperationSupportedOnDataType
 {
     static constexpr bool value = false;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value ||
@@ -280,7 +280,7 @@ struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Set, D
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value ||

include/ck/utility/synchronization.hpp

@@ -18,6 +18,7 @@ __device__ void block_sync_lds()
     __syncthreads();
 #endif
 }
+
 __device__ void s_nop()
 {
 #if 1

library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp

@@ -90,10 +90,13 @@ struct ReferenceLayernorm : public device::BaseOperator
 
             for(int m = 0; m < M; ++m)
             {
+                AccDataType divisor =
+                    static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
+
                 for(int n = 0; n < N; ++n)
                 {
                     auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
-                    auto y_val = (x_val - mean(m)) / sqrt(var(m) + arg.epsilon_);
+                    auto y_val = (x_val - mean(m)) * divisor;
                     y_val      = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
                     arg.acc_elementwise_op_(y_val, y_val);
                     arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);

library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/ck.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+struct ReferenceReduce : public device::DeviceReduce<InDataType,
+                                                     AccDataType,
+                                                     OutDataType,
+                                                     Rank,
+                                                     NumReduceDim,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     PropagateNan,
+                                                     OutputIndex>
+{
+    using IndexDataType = int32_t;
+
+    static constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_host,
+                 OutDataType* out_host,
+                 IndexDataType* out_index_host,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : reduceDims_(reduceDims),
+              outLengths_(outLengths),
+              outStrides_(outStrides),
+              in_host_(in_host),
+              out_host_(out_host),
+              out_index_host_(out_index_host),
+              in_elementwise_op_(in_elementwise_op),
+              acc_elementwise_op_(acc_elementwise_op)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // get invariant_dims[] and invariant_lengths[]
+                for(int dim = 0, i = 0; dim < Rank; dim++)
+                    if(std::none_of(
+                           reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                    {
+                        invariantDims_[i]     = dim;
+                        invariant_lengths_[i] = inLengths[dim];
+                        i++;
+                    };
+            };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = inLengths[dim];
+            };
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // check invariant_lengths_ and outLengths
+                for(int i = 0; i < NumInvariantDim; i++)
+                    if(invariant_lengths_[i] != outLengths_[i])
+                        throw std::runtime_error("Invalid lengths parameters!");
+            }
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                for(int j = 0, i = 0; j < NumInvariantDim; j++)
+                {
+                    int dim                  = invariantDims_[j];
+                    in_invariant_strides_[i] = inStrides[dim];
+                    i++;
+                };
+            };
+
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim               = reduceDims_[j];
+                in_reduce_strides_[i] = inStrides[dim];
+                i++;
+            };
+
+            if constexpr(NumInvariantDim > 0)
+                invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+
+            reduce_index_set_ = get_index_set<NumReduceDim>(reduce_lengths_);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+        };
+
+        const std::array<int, NumReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumDstDim> outLengths_;
+        const std::array<index_t, NumDstDim> outStrides_;
+
+        std::array<index_t, NumInvariantDim> in_invariant_strides_;
+        std::array<index_t, NumReduceDim> in_reduce_strides_;
+
+        const InDataType* in_host_;
+        OutDataType* out_host_;
+        IndexDataType* out_index_host_;
+        const InElementwiseOperation in_elementwise_op_;
+        const AccElementwiseOperation acc_elementwise_op_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumReduceDim>> reduce_index_set_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            ignore = stream_config;
+
+            using ck::float_equal_one;
+            using ck::float_equal_zero;
+            using ck::type_convert;
+            using ck::host_common::get_index_set;
+            using ck::host_common::get_offset_from_index;
+
+            if constexpr(OutputIndex)
+            {
+                using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                                ReduceOperation,
+                                                                                AccDataType,
+                                                                                IndexDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+                    IndexDataType accuIndex = 0;
+
+                    for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(
+                            arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        auto currIndex = static_cast<IndexDataType>(i);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0]       = type_convert<OutDataType>(accuVal);
+                    arg.out_index_host_[0] = accuIndex;
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+                        IndexDataType accuIndex = 0;
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            auto currIndex = static_cast<IndexDataType>(i);
+
+                            Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset]       = type_convert<OutDataType>(accuVal);
+                        arg.out_index_host_[dst_offset] = accuIndex;
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            }
+            else
+            {
+                using Accumulation =
+                    ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                    for(const auto& reduce_index : arg.reduce_index_set_)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(arg.in_reduce_strides_,
+                                                                             reduce_index);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0] = type_convert<OutDataType>(accuVal);
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(const auto& reduce_index : arg.reduce_index_set_)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, reduce_index);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            Accumulation::Calculate(accuVal, currVal);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset] = type_convert<OutDataType>(accuVal);
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            };
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        ignore = p_arg;
+
+        return true;
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_host,
+                        const void* in_index_host,
+                        void* out_host,
+                        void* out_index_host,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        ignore = in_index_host;
+
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_host),
+                                          static_cast<OutDataType*>(out_host),
+                                          static_cast<IndexDataType*>(out_index_host),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_Reduce<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp

@@ -24,11 +24,14 @@ struct ReferenceSoftmax : public device::BaseOperator
     {
         Argument(const Tensor<InDataType>& in,
                  Tensor<OutDataType>& out,
-                 AccDataType alpha,
-                 AccDataType beta,
+                 double alpha,
+                 double beta,
                  const std::vector<index_t> sm_reduce_dims)
-            : in_(in), out_(out), alpha_(alpha), beta_(beta), sm_reduce_dims_(sm_reduce_dims)
+            : in_(in), out_(out), sm_reduce_dims_(sm_reduce_dims)
         {
+            alpha_ = static_cast<AccDataType>(alpha);
+            beta_  = static_cast<AccDataType>(beta);
+
             // std::cout << "debug: scalar dims: ";
             for(size_t i = 0; i < in.mDesc.GetNumOfDimension(); i++)
             {
@@ -143,8 +146,8 @@ struct ReferenceSoftmax : public device::BaseOperator
 
     static auto MakeArgument(const Tensor<InDataType>& in,
                              Tensor<OutDataType>& out,
-                             AccDataType alpha,
-                             AccDataType beta,
+                             double alpha,
+                             double beta,
                              const std::vector<index_t> sm_reduce_dims)
     {
         return Argument{in, out, alpha, beta, sm_reduce_dims};

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -89,7 +89,10 @@ using Scale          = ck::tensor_operation::element_wise::Scale;
 using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
 using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
 using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
+using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;