adding c shuffle

composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r3.hpp

@@ -672,7 +672,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
                               c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks);
         }
     }
-}; // namespace ck
+};
 
 } // namespace ck
 #endif

composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r5.hpp

@@ -649,7 +649,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
                               c1_grid_buf);
         }
     }
-}; // namespace ck
+};
 
 } // namespace ck
 #endif

composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r1.hpp

+#ifndef CK_GRIDWISE_GEMM_XDLOPS_V3R1_HPP
+#define CK_GRIDWISE_GEMM_XDLOPS_V3R1_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_gemm_xdlops.hpp"
+#include "blockwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_set.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared_block,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGridStepHacks,
+          typename BGridStepHacks,
+          typename CGridStepHacks,
+          typename AGridMoveSliceWindowStepHacks,
+          typename BGridMoveSliceWindowStepHacks,
+          bool CAccessOrderMRepeatNRepeat,
+          bool ABlockLdsExtraM,
+          bool BBlockLdsExtraN>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = (K0 / K0PerBlock) > 1;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_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;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto c_blockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_blockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  c_blockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return c_blockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using Block2CTileMap = decltype(MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_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);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            AElementwiseOperation,
+                                            InMemoryDataOperationEnum_t::Set,
+                                            Sequence<K0PerBlock, MPerBlock, K1>,
+                                            ABlockTransferThreadSliceLengths_K0_M_K1,
+                                            ABlockTransferThreadClusterLengths_K0_M_K1,
+                                            ABlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(a_grid_desc_k0_m_k1),
+                                            decltype(a_block_desc_k0_m_k1),
+                                            ABlockTransferSrcAccessOrder,
+                                            Sequence<1, 0, 2>,
+                                            ABlockTransferSrcVectorDim,
+                                            2,
+                                            ABlockTransferSrcScalarPerVector,
+                                            ABlockTransferDstScalarPerVector_K1,
+                                            1,
+                                            1,
+                                            AThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(a_grid_desc_k0_m_k1,
+                                                  make_multi_index(0, m_block_data_idx_on_grid, 0),
+                                                  a_block_desc_k0_m_k1,
+                                                  make_multi_index(0, 0, 0),
+                                                  a_element_op);
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            BElementwiseOperation,
+                                            InMemoryDataOperationEnum_t::Set,
+                                            Sequence<K0PerBlock, NPerBlock, K1>,
+                                            BBlockTransferThreadSliceLengths_K0_N_K1,
+                                            BBlockTransferThreadClusterLengths_K0_N_K1,
+                                            BBlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(b_grid_desc_k0_n_k1),
+                                            decltype(b_block_desc_k0_n_k1),
+                                            BBlockTransferSrcAccessOrder,
+                                            Sequence<1, 0, 2>,
+                                            BBlockTransferSrcVectorDim,
+                                            2,
+                                            BBlockTransferSrcScalarPerVector,
+                                            BBlockTransferDstScalarPerVector_K1,
+                                            1,
+                                            1,
+                                            BThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(b_grid_desc_k0_n_k1,
+                                                  make_multi_index(0, n_block_data_idx_on_grid, 0),
+                                                  b_block_desc_k0_n_k1,
+                                                  make_multi_index(0, 0, 0),
+                                                  b_element_op);
+
+        // 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
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = p_shared_block;
+        FloatAB* p_b_block = p_shared_block + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // hack to control index calculation when iterating over A and B matrix for threadwise copy
+        constexpr auto a_k0_m_k1_grid_step_hacks = AGridStepHacks{};
+        constexpr auto b_k0_n_k1_grid_step_hacks = BGridStepHacks{};
+
+        // hack to control index calculation when move slice window for A and B matrix for
+        // threadwise copy
+        constexpr auto a_k0_m_k1_grid_move_slice_window_step_hack = AGridMoveSliceWindowStepHacks{};
+        constexpr auto b_k0_n_k1_grid_move_slice_window_step_hack = BGridMoveSliceWindowStepHacks{};
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            p_b_block, b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        // preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf, a_k0_m_k1_grid_step_hacks);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n_k1, b_grid_buf, b_k0_n_k1_grid_step_hacks);
+
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
+        }
+
+        // main body
+        index_t k0_block_data_begin = 0;
+
+        c_thread_buf.Clear();
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            do
+            {
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
+                                                    a_block_slice_copy_step,
+                                                    a_k0_m_k1_grid_move_slice_window_step_hack);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n_k1,
+                                                    b_block_slice_copy_step,
+                                                    b_k0_n_k1_grid_move_slice_window_step_hack);
+
+                a_blockwise_copy.RunRead(
+                    a_grid_desc_k0_m_k1, a_grid_buf, a_k0_m_k1_grid_step_hacks);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(
+                    b_grid_desc_k0_n_k1, b_grid_buf, b_k0_n_k1_grid_step_hacks);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
+
+                k0_block_data_begin += K0PerBlock;
+            } while(k0_block_data_begin < (K0 - K0PerBlock));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+
+#if 1
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                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.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // 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_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks = CGridStepHacks{};
+
+            const auto m_thread_data_on_grid_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_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_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_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf,
+                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks);
+        }
+#else
+        // shuffle and write out
+        {
+            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXdl);
+
+            constexpr index_t MPerBlock_CCopy = MWave * MPerXdl;
+            constexpr index_t NPerBlock_CCopy = NWave * NPerXdl;
+
+            // hacky
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                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.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            constexpr auto c_block_desc_mwavemperxdl_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(Number<MPerBlock_CCopy>{},
+                                                    Number<NPerBlock_CCopy>{});
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mwavemperxdl_nwavenperxdl,
+                make_tuple(make_unmerge_transform(make_tuple(I1, Number<MWave>{}, M2, M3, M4)),
+                           make_unmerge_transform(make_tuple(I1, Number<NWave>{}, N2))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                mdke_tuple(Sequence<0, 2, 4, 5, 6>{}, 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 index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_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_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_grid_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_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr2lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatAcc,
+                                                   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<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum_t::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // hardcoded
+            constexpr index_t MThread_CCopy = 16;
+            constexpr index_t NThread_CCopy = 16;
+
+            constexpr index_t MPerThread_CCopy = MPerBlock_CCopy / MThread_CCopy;
+            constexpr index_t NPerThread_CCopy = NPerBlock_CCopy / NThread_CCopy;
+
+            constexpr auto c_block_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl =
+                make_naive_tensor_descriptor_packed(
+                    I1, I1, Number<MPerBlock_CCopy>{}, I1, I1, Number<NPerBlock_CCopy>{});
+
+            auto c_block_copy = BlockwiseTensorSliceTransfer_v4<
+                BlockSize,                                              // index_t BlockSize,
+                ck::tensor_operation::element_wise::PassThrough,        // SrcElementwiseOperation,
+                CGlobalMemoryDataOperation,                             // DstInMemOp,
+                Sequence<1, 1, MPerBlock_CCopy, 1, 1, NPerBlock_CCopy>, // BlockSliceLengths,
+                Sequence<1, 1, MPerThread_CCopy, 1, 1, NPerThread_CCopy>, // ThreadSliceLengths,
+                Sequence<1, 1, MPerThread, 1, 1, NPerThread>, // typename ThreadClusterLengths,
+                Sequence<0, 1, 2, 3, 4, 5>,                   // typename ThreadClusterArrangeOrder,
+                FloatAcc,                                     // typename SrcData,
+                FloatC,                                       // typename DstData,
+                decltype(c_block_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl),
+                decltype(c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl),
+                Sequence<0, 1, 2, 3, 4, 5>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename DstDimAccessOrder,
+                5,                          // index_t SrcVectorDim,
+                5,                          // index_t DstVectorDim,
+                MThread_CCopy,              // index_t SrcScalarPerVector,
+                NThread_CCopy,              // index_t DstScalarPerVector,
+                1,                          // index_t SrcScalarStrideInVector,
+                1,                          // index_t DstScalarStrideInVector,
+                true,                       // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false>                      // bool ThreadTransferDstResetCoordinateAfterRun>
+            {
+                c_block_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                    make_multi_index(0, 0, 0, 0, 0, 0),
+                    c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                    make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0)
+            }
+
+            constexpr auto mrepeat_forward_step  = make_multi_index(0, 1, 0, 0, 0, 0);
+            constexpr auto nrepeat_forward_step  = make_multi_index(0, 0, 0, 0, 1, 0);
+            constexpr auto nrepeat_backward_step = make_multi_index(0, 0, 0, 0, -1, 0);
+
+            // make sure all ds_read from GEMM is completed
+            block_sync_lds();
+
+            static_for<0, MRepeat, 1>{}([&](auto mrepeat) {
+                static_for<0, NRepeat, 1>{}([&](auto nrepeat) {
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_block_buf);
+
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                        c_block_buf,
+                        c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                        c_global_buf);
+
+                    constexpr bool nrepeat_forward_sweep = mrepeat % 2 == 0;
+
+                    // move on nrepeat dimension
+                    if constexpr(nrepeat_forward_sweep && nrepeat < NRepeat - 1)
+                    {
+                        c_block_copy.MoveDstSliceWindow(
+                            c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                            nrepeat_forward_step);
+                    }
+                    else if constexpr((!nrepeat_forward_sweep) & nrepeat > 1)
+                    {
+                        c_block_copy.MoveDstSliceWindow(
+                            c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                            nrepeat_backward_step);
+                    }
+                });
+
+                // move on mrepeat dimension
+                if constexpr(mrepeat < MRepeat - 1)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_global_desc_mblock_mrepeat_mwaveMPerXdl_nblock_nrepeat_nwaveNPerXdl,
+                        mrepeat_forward_step);
+                }
+            });
+        }
+#endif
+    }
+};
+
+} // namespace ck
+#endif

composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v2.hpp → composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v3r1.hpp

-#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V2_HPP
-#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V2_HPP
+#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R1_HPP
+#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R1_HPP
 
 #include "common_header.hpp"
 #include "tensor_descriptor.hpp"
@@ -608,169 +608,5 @@ struct ThreadwiseTensorSliceTransfer_v3r1
     DstCoord dst_coord_;
 };
 
-// Assume:
-//   1. src:
-//     1. SrcDesc is known at compile-time
-//     2. SrcBuffer is DynamicBuffer
-//     3. src_ref_idx is known at run-time
-//     4. SrcRefToOriginDisplacement is known at compile-time
-//     5. use #-step
-//   2. dst:
-//     1. DstDesc is known at compile-time
-//     2. DstBuffer is StaticBuffer
-//     3. DstOriginIdx is known at compile-time
-//     4. use direct address calculation
-//   3. vector access on src
-template <typename SrcData,
-          typename DstData,
-          typename SrcDesc,
-          typename DstDesc,
-          typename SliceLengths,
-          typename DimAccessOrder,
-          typename SrcVectorTensorLengths,
-          typename SrcVectorTensorContiguousDimOrder,
-          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
-                             bool>::type = false>
-struct ThreadwiseTensorSliceTransfer_v4r1
-{
-    static constexpr auto I0 = Number<0>{};
-    static constexpr auto I1 = Number<1>{};
-
-    static constexpr index_t nDim = SliceLengths::Size();
-
-    using Index = MultiIndex<nDim>;
-
-    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
-
-    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
-
-    __device__ constexpr ThreadwiseTensorSliceTransfer_v4r1(const Index& src_ref_idx)
-        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
-    {
-        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
-                      "wrong! SrcDesc and DstDesc need to known at compile-time");
-
-        static_for<0, nDim, 1>{}([](auto i) {
-            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0, "wrong!");
-        });
-    }
-
-    template <typename SrcRefToOriginDisplacement,
-              typename DstOriginIdx,
-              typename SrcBuffer,
-              typename DstBuffer>
-    __device__ void Run(const SrcDesc&,
-                        const SrcRefToOriginDisplacement&,
-                        const SrcBuffer& src_buf,
-                        const DstDesc&,
-                        const DstOriginIdx&,
-                        DstBuffer& dst_buf) const
-    {
-        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
-                      "wrong! SrcDesc and DstDesc need to known at compile-time");
-
-        static_assert(
-            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
-                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
-            "wrong! SrcBuffer or DstBuffer data type is wrong");
-
-        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
-
-        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
-                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
-                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
-                      "at compile-time");
-
-        // SrcDesc and DstDesc are known at compile-time
-        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
-        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
-
-        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
-        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
-        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
-
-        // tensor descriptor for src_vector
-        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
-
-        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
-            container_reverse_exclusive_scan(
-                container_reorder_given_new2old(src_vector_tensor_lengths,
-                                                SrcVectorTensorContiguousDimOrder{}),
-                math::multiplies{},
-                I1),
-            SrcVectorTensorContiguousDimOrder{});
-
-        constexpr auto src_vector_desc =
-            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
-                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
-
-        // access order and lengths
-        constexpr auto access_lengths = SliceLengths{} / src_vector_tensor_lengths;
-
-        constexpr auto dim_access_order = DimAccessOrder{};
-
-        constexpr auto ordered_access_lengths =
-            container_reorder_given_new2old(access_lengths, dim_access_order);
-
-        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
-            // position in slice window
-            constexpr auto data_to_origin_disp_idx =
-                ordered_access_idx.ReorderGivenOld2New(dim_access_order) *
-                src_vector_tensor_lengths;
-
-            // src coordinate at starting point of src_vector
-            constexpr auto src_ref_to_data_disp_idx =
-                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
-
-            constexpr auto src_ref_to_data_disp_coord_step =
-                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
-
-            auto src_data_coord = src_ref_coord_;
-
-            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
-
-            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
-
-            using src_vector_t = typename decltype(src_vector)::type;
-
-            const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
-                src_desc, src_data_coord);
-
-            // copy data from src_buf into src_vector
-            src_vector.template AsType<src_vector_t>()(I0) =
-                src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
-
-            // copy data from src_vector into dst_buf (also cast from SrcData to DstData)
-            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
-                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
-
-                constexpr index_t src_vector_offset =
-                    src_vector_desc.CalculateOffset(src_vector_idx);
-
-                constexpr index_t dst_offset = dst_desc.CalculateOffset(
-                    dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
-
-                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
-                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
-            });
-        });
-    }
-
-    template <typename SrcSliceMoveStepIdx>
-    __device__ void MoveSrcSliceWindow(const SrcDesc&,
-                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
-    {
-        constexpr auto src_desc = SrcDesc{};
-
-        const auto src_slice_move_step_iter =
-            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
-
-        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
-    }
-
-    private:
-    SrcCoord src_ref_coord_;
-};
-
 } // namespace ck
 #endif

composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v4r1.hpp

+#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V4R1_HPP
+#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V4R1_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4r1(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0, "wrong!");
+        });
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) *
+                src_vector_tensor_lengths;
+
+            // src coordinate at starting point of src_vector
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            constexpr auto src_ref_to_data_disp_coord_step =
+                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                src_desc, src_data_coord);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf (also cast from SrcData to DstData)
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
+
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
+                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+} // namespace ck
+#endif

device_operation/include/device_conv2d_fwd_xdl_output_shuffle_nhwc_kyxc_nhwk.hpp

+#ifndef DEVICE_CONV2D_FWD_XDL_OUTPUT_SHUFFLE_NHWC_KYXC_NHWK_HPP
+#define DEVICE_CONV2D_FWD_XDL_OUTPUT_SHUFFLE_NHWC_KYXC_NHWK_HPP
+
+#include <iostream>
+#include <sstream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv_fwd.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v3r1.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          bool ABlockLdsAddExtraM,
+          bool BBlockLdsAddExtraN>
+struct DeviceConv2dFwdXdl_Output_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+{
+    using DeviceOp =
+        DeviceConv2dFwdXdl_Output_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    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 K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+        const index_t GemmK    = Y * X * C;
+
+        const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
+
+        assert(GemmK % GemmK1Number == 0);
+
+        const index_t GemmK0 = GemmK / GemmK1Number;
+
+        // A: input tensor
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+            in_n_hip_wip_c_grid_desc,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_gemmk_gemmmraw_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                   make_merge_transform(make_tuple(N, Ho, Wo))),
+                                        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk_gemmmraw_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmMRaw)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+            transform_tensor_descriptor(in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                                        make_tuple(make_pass_through_transform(GemmK0),
+                                                   make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmK1Number)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // B: weight tensor
+        const auto wei_k_yxc_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+        const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+            wei_k_yxc_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+        const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            wei_gemmk_gemmn_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmN)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        // C: output tensor
+        const auto out_nhowo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+        const auto out_gemmmraw_gemmn_grid_desc = transform_tensor_descriptor(
+            out_nhowo_k_grid_desc,
+            make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_gemmm_gemmn_grid_desc =
+            transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                        make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmN)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                          wei_gemmk0_gemmn_gemmk1_grid_desc,
+                          out_gemmm_gemmn_grid_desc);
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // TODO remove these hacks
+    static constexpr auto a_k0_m_k1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{},   // 0+: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 1+: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}),  // 2+: K1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{},   // 0-: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 1-: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto b_k0_n_k1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: N
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: K1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0-: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: N
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
+
+    static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadSliceLengths_K0_M_K1,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        BBlockTransferThreadSliceLengths_K0_N_K1,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false,                            // BThreadTransferSrcResetCoordinateAfterRun,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        decltype(a_k0_m_k1_grid_step_hacks),                   //  AGridStepHacks,
+        decltype(b_k0_n_k1_grid_step_hacks),                   //  BGridStepHacks,
+        decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),   //  CGridStepHacks,
+        decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), //  AGridMoveSliceWindowStepHacks,
+        decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), //  BGridMoveSliceWindowStepHacks,
+        false,                                                 // CAccessOrderMRepeatNRepeat,
+        ABlockLdsAddExtraM,
+        BBlockLdsAddExtraN>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+
+    using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceOp::Block2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceOp::Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_Output_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/4_conv2d_fwd_xdl_output_shuffle/README.md

+# Instructions for ```conv2d_fwd_xdl``` Example
+
+## Docker script
+```bash
+docker run                                                                   \
+-it                                                                          \
+--rm                                                                         \
+--privileged                                                                 \
+--group-add sudo                                                             \
+-w /root/workspace                                                           \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                                \
+rocm/tensorflow:rocm4.3.1-tf2.6-dev                                          \
+/bin/bash
+```
+
+## Build ```conv2d_fwd_xdl```
+```bash
+mkdir build && cd build
+```
+
+```bash
+# Need to specify target ID, example below is gfx908
+cmake                                                                  \
+-D BUILD_DEV=OFF                                                       \
+-D CMAKE_BUILD_TYPE=Release                                            \
+-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 "   \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                              \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                         \
+..
+```
+
+```bash
+ make -j conv2d_fwd_xdl
+```
+
+## Run ```conv2d_fwd_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
+./example/conv2d_fwd_xdl 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_{216, 165888, 8}
+arg.b_grid_desc_k0_n_k1_{216, 256, 8}
+arg.c_grid_desc_m_n_{ 165888, 256}
+launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.43206 ms, 102.486 TFlops, 232.947 GB/s
+```

example/4_conv2d_fwd_xdl_output_shuffle/conv2d_fwd_xdl_output_shuffle.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "device_operation/include/device_conv2d_fwd_xdl_output_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using WeiLayout = ck::tensor_layout::convolution::KYXC;
+using OutLayout = ck::tensor_layout::convolution::NHWK;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceConvFwdInstance = ck::tensor_operation::device::
+    DeviceConv2dFwdXdl_Output_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    // clang-format off
+//##|    InData|     WeiData|     OutData|     AccData|          In|         Wei|           Out| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
+//##|      Type|        Type|        Type|        Type| Elementwise| Elementwise|   Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| SrcDstVectorDim|       DstScalar| AddExtraM| AddExtraN|
+//##|          |            |            |            |   Operation|   Operation|     Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|                |       PerVector|          |          |
+//##|          |            |            |            |            |            |              |      |      |      |      |   |     |     |     |     |                |                |               |               |               |               |               |                |                |               |               |              |               |               |                |                |          |          |
+    <InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp,   256,   128,   256,     4,  8,   32,   32,    2,    4,      S<1, 2, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      S<1, 4, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,               7,               1,      true,      true>;
+// clang-format on
+
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp>
+void host_verify(const Tensor<TIn>& in,
+                 const Tensor<TWei>& wei,
+                 Tensor<TOut>& out,
+                 const std::vector<ck::index_t>& conv_strides,
+                 const std::vector<ck::index_t>& conv_dilations,
+                 const std::vector<ck::index_t>& in_left_pads,
+                 const std::vector<ck::index_t>&,
+                 const InElementOp& in_element_op,
+                 const WeiElementOp& wei_element_op,
+                 const OutElementOp& out_element_op)
+{
+    auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+        double v = 0;
+        for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
+        {
+            for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
+            {
+                int hi = ho * conv_strides[0] + y * conv_dilations[0] - in_left_pads[0];
+                for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
+                {
+                    int wi = wo * conv_strides[1] + x * conv_dilations[1] - in_left_pads[1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        v += in_element_op(static_cast<const double>(in(n, c, hi, wi))) *
+                             wei_element_op(static_cast<const double>(wei(k, c, y, x)));
+                    }
+                }
+            }
+        }
+        out(n, k, ho, wo) = out_element_op(v);
+    };
+
+    make_ParallelTensorFunctor(f_nchw,
+                               out.mDesc.GetLengths()[0],
+                               out.mDesc.GetLengths()[1],
+                               out.mDesc.GetLengths()[2],
+                               out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    int nrepeat          = 5;
+
+    // Conv shape
+    ck::index_t N               = 128;
+    ck::index_t K               = 256;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t conv_stride_h   = 2;
+    ck::index_t conv_stride_w   = 2;
+    ck::index_t conv_dilation_h = 1;
+    ck::index_t conv_dilation_w = 1;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = std::stoi(argv[3]);
+    }
+    else if(argc == 19)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = std::stoi(argv[3]);
+
+        N               = std::stoi(argv[4]);
+        K               = std::stoi(argv[5]);
+        C               = std::stoi(argv[6]);
+        Y               = std::stoi(argv[7]);
+        X               = std::stoi(argv[8]);
+        Hi              = std::stoi(argv[9]);
+        Wi              = std::stoi(argv[10]);
+        conv_stride_h   = std::stoi(argv[11]);
+        conv_stride_w   = std::stoi(argv[12]);
+        conv_dilation_h = std::stoi(argv[13]);
+        conv_dilation_w = std::stoi(argv[14]);
+        in_left_pad_h   = std::stoi(argv[15]);
+        in_left_pad_w   = std::stoi(argv[16]);
+        in_right_pad_h  = std::stoi(argv[17]);
+        in_right_pad_w  = std::stoi(argv[18]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
+    const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
+    const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+    const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+    // tensor layout
+    auto f_host_tensor_descriptor = [](std::size_t N_,
+                                       std::size_t C_,
+                                       std::size_t H,
+                                       std::size_t W,
+                                       auto layout) {
+        if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                     ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                     ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                        std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
+        }
+        else if constexpr(ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::NHWC>::value ||
+                          ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::KYXC>::value ||
+                          ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::NHWK>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                        std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
+        }
+    };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+
+    // do GEMM
+    auto conv     = DeviceConvFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                      static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                      static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                      N,
+                                      K,
+                                      C,
+                                      std::vector<ck::index_t>{{Hi, Wi}},
+                                      std::vector<ck::index_t>{{Y, X}},
+                                      std::vector<ck::index_t>{{Ho, Wo}},
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float ave_time = invoker.Run(argument, nrepeat);
+
+    std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+    std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
+                            sizeof(WeiDataType) * (K * C * Y * X) +
+                            sizeof(OutDataType) * (N * K * Ho * Wo);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        host_verify(in_n_c_hi_wi,
+                    wei_k_c_y_x,
+                    out_n_k_ho_wo_host_result,
+                    conv_filter_strides,
+                    conv_filter_dilations,
+                    input_left_pads,
+                    input_right_pads,
+                    InElementOp{},
+                    WeiElementOp{},
+                    OutElementOp{});
+
+        out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+        check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
+    }
+}

example/CMakeLists.txt

@@ -14,17 +14,20 @@ include_directories(BEFORE
 set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
 set(GEMM_XDL_BIAS_RELU_ADD_SOURCE 3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp)
 set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
+set(CONV2D_FWD_XDL_OUTPUT_SHUFFLE_SOURCE 4_conv2d_fwd_xdl_output_shuffle/conv2d_fwd_xdl_output_shuffle.cpp)
 set(CONV2D_FWD_XDL_BIAS_RELU_SOURCE 5_conv2d_fwd_xdl_bias_relu/conv2d_fwd_xdl_bias_relu.cpp)
 set(CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
 add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
 add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
+add_executable(conv2d_fwd_xdl_output_shuffle ${CONV2D_FWD_XDL_OUTPUT_SHUFFLE_SOURCE})
 add_executable(conv2d_fwd_xdl_bias_relu ${CONV2D_FWD_XDL_BIAS_RELU_SOURCE})
 add_executable(conv2d_fwd_xdl_bias_relu_add ${CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)
+target_link_libraries(conv2d_fwd_xdl_output_shuffle PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl_bias_relu PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl_bias_relu_add PRIVATE host_tensor)