Update tuning method

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDTensor>
+struct ContractionDesc
+{
+    std::vector<index_t> a_ms_ks_lengths;
+    std::vector<index_t> a_ms_ks_strides;
+
+    std::vector<index_t> b_ns_ks_lengths;
+    std::vector<index_t> b_ns_ks_strides;
+
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_lengths;
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_strides;
+
+    std::vector<index_t> e_ms_ns_lengths;
+    std::vector<index_t> e_ms_ns_strides;
+};
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ContractionMultiDKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_contraction_multiple_d_xdl_cshuffle(
+            const void CK_CONSTANT_ADDRESS_SPACE* contraction_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto contraction_arg_ptr = reinterpret_cast<const ContractionMultiDKernelArg*>(
+        cast_pointer_to_generic_address_space(contraction_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while((!(block_id >= contraction_arg_ptr[group_id].block_start_ &&
+             block_id < contraction_arg_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < contraction_arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        contraction_arg_ptr[group_id].p_a_grid_,
+        contraction_arg_ptr[group_id].p_b_grid_,
+        contraction_arg_ptr[group_id].p_ds_grid_,
+        contraction_arg_ptr[group_id].p_e_grid_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        contraction_arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        contraction_arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        contraction_arg_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = contraction_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
+    : public DeviceGroupedContractionMultipleD<NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_ms_ks_strides_vec)
+    {
+        assert(a_ms_ks_lengths_vec.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_vec.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_ns_ks_strides_vec)
+    {
+        assert(b_ns_ks_lengths_vec.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_vec.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_vec, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_vec, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_ms_ns_strides_vec)
+    {
+        assert(e_ms_ns_lengths_vec.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_vec.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_vec, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_vec, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths_vec[i],
+                                                         ds_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    struct GroupedContractionBlock2ETileMap
+    {
+        // block-to-e-tile map
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedContractionBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n,
+                                         ck::index_t BlockStart)
+        {
+            default_block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            block_start_               = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return default_block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - block_start_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return default_block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return default_block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap default_block_2_etile_map_;
+        ck::index_t block_start_;
+    };
+
+    struct ContractionMultiDKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // lock-to-e-tile map
+        GroupedContractionBlock2ETileMap block_2_etile_map_;
+
+        ck::index_t block_start_, block_end_;
+    };
+
+    struct ContractionMultiDDeviceArg
+    {
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        // index_t e_mz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                 std::vector<void*> p_e_vec,
+                 std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            group_count_ = contraction_descs.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_e_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/e_vec.size");
+            }
+
+            contraction_multi_d_kernel_args_.reserve(group_count_);
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_e_grid = static_cast<EDataType*>(p_e_vec[i]);
+
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(
+                    contraction_descs[i].a_ms_ks_lengths, contraction_descs[i].a_ms_ks_strides);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(
+                    contraction_descs[i].b_ns_ks_lengths, contraction_descs[i].b_ns_ks_strides);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+                typename GridwiseGemm::DsGridPointer p_ds_grid;
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_ds_vec[i][j]);
+
+                    // D desc
+                    ds_grid_desc_m_n(j) =
+                        DeviceOp::MakeEGridDescriptor_M_N(contraction_descs[i].ds_ms_ns_lengths[j],
+                                                          contraction_descs[i].ds_ms_ns_strides[j]);
+                });
+
+                const auto e_grid_desc_m_n = DeviceOp::MakeEGridDescriptor_M_N(
+                    contraction_descs[i].e_ms_ns_lengths, contraction_descs[i].e_ms_ns_strides);
+
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n);
+                const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n);
+
+                const index_t grid_size_grp =
+                    GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n)
+                        .CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                const auto block_2_etile_map =
+                    GroupedContractionBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                // for sanity check of vector memory access
+                const index_t a_mz_stride = contraction_descs[i].a_ms_ks_strides[NumDimM - 1];
+                const index_t a_kz_stride =
+                    contraction_descs[i].a_ms_ks_strides[NumDimM + NumDimK - 1];
+
+                const index_t b_nz_stride = contraction_descs[i].b_ns_ks_strides[NumDimN - 1];
+                const index_t b_kz_stride =
+                    contraction_descs[i].b_ns_ks_strides[NumDimN + NumDimK - 1];
+
+                std::array<index_t, NumDTensor> ds_nz_stride;
+                for(index_t j = 0; j < NumDTensor; ++j)
+                {
+                    ds_nz_stride[j] =
+                        contraction_descs[i].ds_ms_ns_strides[j][NumDimM + NumDimN - 1];
+                }
+
+                const index_t e_nz_stride =
+                    contraction_descs[i].e_ms_ns_strides[NumDimM + NumDimN - 1];
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    contraction_multi_d_kernel_args_.push_back(
+                        {p_a_grid,
+                         p_b_grid,
+                         p_ds_grid,
+                         p_e_grid,
+                         a_grid_desc_ak0_m_ak1,
+                         b_grid_desc_bk0_n_bk1,
+                         ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                         e_grid_desc_mblock_mperblock_nblock_nperblock,
+                         block_2_etile_map,
+                         BlockStart,
+                         BlockEnd});
+
+                    contraction_multi_d_device_args_.push_back({a_grid_desc_m_k,
+                                                                b_grid_desc_n_k,
+                                                                ds_grid_desc_m_n,
+                                                                e_grid_desc_m_n,
+                                                                a_mz_stride,
+                                                                a_kz_stride,
+                                                                b_nz_stride,
+                                                                b_kz_stride,
+                                                                ds_nz_stride,
+                                                                e_nz_stride});
+                }
+            }
+        }
+
+        std::vector<ContractionMultiDKernelArg> contraction_multi_d_kernel_args_;
+        std::vector<ContractionMultiDDeviceArg> contraction_multi_d_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K =
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(hipMemcpy(arg.p_workspace_,
+                                        arg.contraction_multi_d_kernel_args_.data(),
+                                        arg.contraction_multi_d_kernel_args_.size() *
+                                            sizeof(ContractionMultiDKernelArg),
+                                        hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_contraction_multiple_d_xdl_cshuffle<GridwiseGemm,
+                                                                       ContractionMultiDKernelArg,
+                                                                       AElementwiseOperation,
+                                                                       BElementwiseOperation,
+                                                                       CDEElementwiseOperation,
+                                                                       has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.cde_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto a_grid_desc_m_k_ = arg.contraction_multi_d_device_args_[i].a_grid_desc_m_k_;
+            const auto b_grid_desc_n_k_ = arg.contraction_multi_d_device_args_[i].b_grid_desc_n_k_;
+            const auto ds_grid_desc_m_n_ =
+                arg.contraction_multi_d_device_args_[i].ds_grid_desc_m_n_;
+            const auto e_grid_desc_m_n_ = arg.contraction_multi_d_device_args_[i].e_grid_desc_m_n_;
+            const auto a_grid_desc_ak0_m_ak1_ =
+                arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_;
+            const auto b_grid_desc_bk0_n_bk1_ =
+                arg.contraction_multi_d_kernel_args_[i].b_grid_desc_bk0_n_bk1_;
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+            const auto block_2_etile_map_ =
+                arg.contraction_multi_d_kernel_args_[i].block_2_etile_map_;
+
+            const auto a_mz_stride_  = arg.contraction_multi_d_device_args_[i].a_mz_stride_;
+            const auto a_kz_stride_  = arg.contraction_multi_d_device_args_[i].a_kz_stride_;
+            const auto b_nz_stride_  = arg.contraction_multi_d_device_args_[i].b_nz_stride_;
+            const auto b_kz_stride_  = arg.contraction_multi_d_device_args_[i].b_kz_stride_;
+            const auto ds_nz_stride_ = arg.contraction_multi_d_device_args_[i].ds_nz_stride_;
+            const auto e_nz_stride_  = arg.contraction_multi_d_device_args_[i].e_nz_stride_;
+
+            if(!GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                            b_grid_desc_n_k_,
+                                            ds_grid_desc_m_n_,
+                                            e_grid_desc_m_n_,
+                                            block_2_etile_map_))
+            {
+                return false;
+            }
+
+            // check vector access
+            static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                              (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                          "wrong!");
+
+            // vector memory access of A: could be on M or AK1 dimension
+            if constexpr(ABlockTransferSrcVectorDim == 1)
+            {
+                if(!(a_mz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(a_kz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of B: could be on N or BK1 dimension
+            if constexpr(BBlockTransferSrcVectorDim == 1)
+            {
+                if(!(b_nz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(b_kz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of Ds: always on NPerBlock dimension
+            bool valid_d_access = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto j) {
+                if(!(ds_nz_stride_[j] == 1 &&
+                     ds_grid_desc_mblock_mperblock_nblock_nperblock_[j].GetLength(I3) %
+                             CDEBlockTransferScalarPerVector_NPerBlock ==
+                         0))
+                {
+                    valid_d_access = false;
+                }
+            });
+
+            if(valid_d_access == false)
+            {
+                return false;
+            }
+
+            // vector memory access of E: always on NPerBlock dimension
+            if(!(e_nz_stride_ == 1 && e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                                              CDEBlockTransferScalarPerVector_NPerBlock ==
+                                          0))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                             std::vector<void*> p_e_vec,
+                             std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_ds_vec,
+                        p_e_vec,
+                        contraction_descs,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_ds_vec,
+                                          p_e_vec,
+                                          contraction_descs,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ *
+               sizeof(ContractionMultiDKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Conv backward data multiple D:
+//   input : output image A[G, N, K, Ho, Wo]
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : input image E[G, N, C, Hi, Wi],
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvBwdDataMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, NumDTensor>& p_ds,                 // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_lengths, // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_strides,                                        // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a, // input image
+        const void* p_b, // weight
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Grouped Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   output : R0[G, N, Ho, Wo], R1[G, N, Ho, Wo], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleDMultipleR : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor, index_t NumRTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE,
+                                   Array<ck::index_t, NumRTensor> BatchStrideRs)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE),
+          BatchStrideRs_(BatchStrideRs)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    __host__ __device__ constexpr auto GetRsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumRTensor> rs_offset;
+        static_for<0, NumRTensor, 1>{}(
+            [&](auto i) { rs_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideRs_[i]); });
+        return rs_offset;
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+    Array<ck::index_t, NumRTensor> BatchStrideRs_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename RsPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename RsGridDescriptor_MBlock_MPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batch_gemm_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            RsPointer p_rs_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const QsElementwiseOperation qs_element_op,
+            const RsElementwiseOperation rs_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+    const auto rs_batch_offset = compute_ptr_offset_of_batch.GetRsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    RsPointer p_rs_grid_grp;
+
+    static constexpr index_t NumRTensor = RsGridDescriptor_MBlock_MPerBlock::Size();
+
+    static_for<0, NumRTensor, 1>{}(
+        [&](auto i) { p_rs_grid_grp(i) = p_rs_grid[i] + rs_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_rs_grid_grp,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  qs_element_op,
+                                                  rs_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  rs_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_rs_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = rs_grid_desc_mblock_mperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = qs_element_op;
+    ignore = rs_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ReduceAccDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          typename RsGlobalMemoryDataOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleDMultipleR<NDimSpatial,
+                                                    ALayout,
+                                                    BLayout,
+                                                    DELayout,
+                                                    RLayout,
+                                                    ADataType,
+                                                    BDataType,
+                                                    DsDataType,
+                                                    EDataType,
+                                                    RsDataType,
+                                                    AElementwiseOperation,
+                                                    BElementwiseOperation,
+                                                    CDEElementwiseOperation,
+                                                    RsElementwiseOperation,
+                                                    QsElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    template <typename Descriptor>
+    static auto GetPaddedRGridDescriptor(Descriptor descriptor, index_t MRaw)
+    {
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(
+                descriptor,
+                make_tuple(make_right_pad_transform(descriptor, MPad)),
+                make_tuple(Sequence<0>{}),
+                make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return descriptor;
+        }
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::GNW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNDHW>,
+                                      bool>::type = false>
+    static auto
+    MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                          const std::array<index_t, NDimSpatial + 2>& /* r_g_n_wos_strides */)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t NHoWo = N * std::accumulate(r_g_n_wos_lengths.begin() + 2,
+                                                  r_g_n_wos_lengths.begin() + 2 + NDimSpatial,
+                                                  index_t{1},
+                                                  std::multiplies<index_t>());
+
+        const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(NHoWo));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::G_NW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NDHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NHWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NDHWG>,
+                                      bool>::type = false>
+    static auto MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                                      const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t WoStride = r_g_n_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo = N * std::accumulate(r_g_n_wos_lengths.begin() + 2,
+                                                  r_g_n_wos_lengths.begin() + 2 + NDimSpatial,
+                                                  index_t{1},
+                                                  std::multiplies<index_t>());
+
+        const auto r_grid_desc_mraw =
+            make_naive_tensor_descriptor(make_tuple(NHoWo), make_tuple(WoStride));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(
+        MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using EGridDesc_M_N = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<DELayout>({}, {}))>;
+    using RGridDesc_M   = remove_cvref_t<decltype(MakeRGridDescriptor_M<RLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        ReduceAccDataType,
+        RsDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        QsElementwiseOperation,
+        RsElementwiseOperation,
+        ThreadReduceOperations,
+        InMemoryDataOperationEnum::Set,
+        RsGlobalMemoryDataOperation,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        RGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        RThreadTransferDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 std::array<void*, NumRTensor> p_rs,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op,
+                 const QsElementwiseOperation& qs_element_op,
+                 const RsElementwiseOperation& rs_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              p_rs_grid_{}, // FIXME
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<DELayout>(e_g_n_k_wos_lengths,
+                                                                           e_g_n_k_wos_strides)},
+              r_grid_desc_m_{
+                  DeviceOp::MakeRGridDescriptor_M<RLayout>(r_g_n_wos_lengths, r_g_n_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              rs_grid_desc_mblock_mperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              qs_element_op_{qs_element_op},
+              rs_element_op_{rs_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           e_grid_desc_m_n_,
+                                           r_grid_desc_m_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                    // D batch stride
+                    compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                    // D desc
+                    ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DELayout>(
+                        ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            ds_grid_desc_m_n_(i));
+                });
+
+                // populate pointer for Rs
+                static_for<0, NumRTensor, 1>{}([&](auto i) {
+                    using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
+
+                    // R pointer
+                    p_rs_grid_(i) = static_cast<RDataType*>(p_rs[i]);
+
+                    rs_grid_desc_mblock_mperblock_(i) =
+                        GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
+                });
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+        typename GridwiseGemm::RsGridPointer p_rs_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        RGridDesc_M r_grid_desc_m_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        StaticallyIndexedArray<
+            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+            NumDTensor>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
+                                                             // type from E
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
+            rs_grid_desc_mblock_mperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        QsElementwiseOperation qs_element_op_;
+        RsElementwiseOperation rs_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.r_grid_desc_m_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) *
+                arg.a_g_n_c_wis_lengths_[0]; // Group count
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_batch_gemm_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    typename GridwiseGemm::RsGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    QsElementwiseOperation,
+                    RsElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        NumDTensor>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
+                        NumRTensor>,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.p_rs_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.qs_element_op_,
+                                              arg.rs_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.rs_grid_desc_mblock_mperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            // FIXME: layout
+            if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                         is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                         is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                         is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                         is_same_v<DELayout, ctc::NDHWGK>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                     is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                     is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                     is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                     is_same_v<DELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of R
+        if constexpr(!(is_same_v<RLayout, ctc::G_NW> || is_same_v<RLayout, ctc::G_NHW> ||
+                       is_same_v<RLayout, ctc::G_NDHW> || is_same_v<RLayout, ctc::GNW> ||
+                       is_same_v<RLayout, ctc::GNHW> || is_same_v<RLayout, ctc::GNDHW> ||
+                       is_same_v<RLayout, ctc::NWG> || is_same_v<RLayout, ctc::NHWG> ||
+                       is_same_v<RLayout, ctc::NDHWG>))
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.r_grid_desc_m_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        p_rs,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        r_g_n_wos_lengths,
+                        r_g_n_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        qs_element_op,
+                        rs_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          p_rs,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          r_g_n_wos_lengths,
+                                          r_g_n_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          qs_element_op,
+                                          rs_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 << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = remove_cvref_t<decltype(
+        MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
+                         is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_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 << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp

+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct GemmDesc
+{
+    ck::index_t M_, N_, K_;
+    ck::index_t stride_A_, stride_B_, stride_C_;
+
+    std::vector<ck::index_t> stride_Ds_;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemm : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_a,
+                        std::vector<const void*>& p_b,
+                        std::vector<std::array<const void*, NumDTensor>>& p_ds,
+                        std::vector<void*>& p_e,
+                        std::vector<GemmDesc>& gemm_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
+{
+    struct ProblemDesc
+    {
+        // Overall problem shape
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t O;
+        index_t Batch;
+
+        // Stride for A/B0/B1; layout determined by template args
+        index_t StrideA;
+        index_t StrideB0;
+        index_t StrideB1;
+        index_t BatchStrideA;
+        index_t BatchStrideB0;
+        index_t BatchStrideB1;
+
+        // Lengths and strides for output C
+        std::vector<index_t> c_gs_ms_os_lengths;
+        std::vector<index_t> c_gs_ms_os_strides;
+    };
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b0_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GroupKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto arg_ptr = reinterpret_cast<const GroupKernelArg*>(
+        cast_pointer_to_generic_address_space(group_kernel_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while((!(block_id >= arg_ptr[group_id].block_start_ &&
+             block_id < arg_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    // per-group batch offset
+    const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
+    const index_t g_idx                = __builtin_amdgcn_readfirstlane(
+        (block_id - arg_ptr[group_id].block_start_) / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
+        arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset  = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        arg_ptr[group_id].p_a_grid_ + a_batch_offset,
+        arg_ptr[group_id].p_b_grid_ + b_batch_offset,
+        arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
+        arg_ptr[group_id].p_c_grid_ + c_batch_offset,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        acc_element_op,
+        b1_element_op,
+        c_element_op,
+        arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
+        arg_ptr[group_id].block_2_ctile_map_,
+        arg_ptr[group_id].c0_matrix_mask_);
+#else
+    ignore = group_kernel_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <typename ALayout,
+          typename BLayout, // B0Layout
+          typename B1Layout,
+          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<NumDimG, NumDimM, NumDimGemm1N>
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          bool MaskOutUpperTriangle,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
+    : public DeviceGroupedGemmSoftmaxGemmPermute<ALayout,
+                                                 BLayout,
+                                                 B1Layout,
+                                                 CPermuteNumDims_G_M_Gemm1N,
+                                                 ADataType,
+                                                 BDataType,
+                                                 B1DataType,
+                                                 CDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 AccElementwiseOperation,
+                                                 B1ElementwiseOperation,
+                                                 CElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+    using ProblemDesc =
+        typename DeviceGroupedGemmSoftmaxGemmPermute<ALayout,
+                                                     BLayout,
+                                                     B1Layout,
+                                                     CPermuteNumDims_G_M_Gemm1N,
+                                                     ADataType,
+                                                     BDataType,
+                                                     B1DataType,
+                                                     CDataType,
+                                                     AElementwiseOperation,
+                                                     BElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     B1ElementwiseOperation,
+                                                     CElementwiseOperation>::ProblemDesc;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
+    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
+
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    {
+        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
+        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
+        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
+
+        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto c_ms_ns_lengths = to_tuple(
+            c_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto c_ms_ns_strides = to_tuple(
+            c_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(c_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(c_ms_ns_lengths, nDimIds);
+
+        // naive tensor C[M0, M1, M2, ..., N0, N1, N2...]
+        const auto c_grid_desc_ms_ns =
+            make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
+
+        // transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+        const auto c_grid_desc_mraw_nraw = transform_tensor_descriptor(
+            c_grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds, nDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    {
+        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
+        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
+        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
+
+        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto c_gs_ms_ns_lengths =
+            to_tuple(c_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto c_gs_ms_ns_strides =
+            to_tuple(c_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(c_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(c_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(c_gs_ms_ns_lengths, nDimIds);
+
+        // naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+        const auto c_grid_desc_gs_ms_ns =
+            make_naive_tensor_descriptor(c_gs_ms_ns_lengths, c_gs_ms_ns_strides);
+
+        // transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+        // N2 * ...]
+        const auto c_grid_desc_g_mraw_nraw =
+            transform_tensor_descriptor(c_grid_desc_gs_ms_ns,
+                                        make_tuple(make_merge_transform(gLengths),
+                                                   make_merge_transform(mLengths),
+                                                   make_merge_transform(nLengths)),
+                                        make_tuple(gDimIds, mDimIds, nDimIds),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // this desc is only for calculating batch offset so no padding needed
+        return c_grid_desc_g_mraw_nraw;
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N({}, {}));
+    using CGridDesc_G_M_N      = decltype(MakeCGridDescriptor_G_M_N({}, {}));
+
+    // to track the points which need to be set to -inf on C0
+    // Note: no need to reset M padding value, because they will not be stored out.
+    struct C0MatrixMask
+    {
+        C0MatrixMask(index_t NRaw) : NRaw_(NRaw) {}
+
+        __host__ __device__ bool IsUpperTriangle(index_t m, index_t n) const { return n > m; }
+
+        __host__ __device__ bool IsNOutOfBound(/*index_t m, */ index_t n) const
+        {
+            return n >= NRaw_;
+        }
+
+        __host__ __device__ bool IsMaskedElement(index_t m, index_t n) const
+        {
+            return IsUpperTriangle(m, n) || IsNOutOfBound(n);
+        }
+
+        private:
+        // index_t MRaw_;
+        index_t NRaw_;
+    };
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideB1,
+                                     CGridDesc_G_M_N c_grid_desc_g_m_n)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideB1_(BatchStrideB1),
+              c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideB1_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        matrix_padder.PadN,
+        MaskOutUpperTriangle>;
+
+    using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
+
+    struct GroupKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // batch & stride
+        index_t num_blocks_per_batch_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // block-to-c-tile map
+        Block2CTileMap block_2_ctile_map_;
+
+        index_t block_start_, block_end_;
+    };
+
+    struct GroupDeviceArg
+    {
+        // problem definiton
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t O;
+
+        // Strides for the last dimensions of C for sanity check of vector load/store
+        index_t c_extent_lowest_;
+        index_t c_stride_lowest_;
+
+        CGridDesc_M_N c_grid_desc_m_n_;
+    };
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<const void*> p_b1_vec,
+                 std::vector<void*> p_c_vec,
+                 std::vector<ProblemDesc> problem_desc_vec,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op}
+        {
+            group_count_ = problem_desc_vec.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
+            }
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid  = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid  = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
+                const auto p_c_grid  = static_cast<CDataType*>(p_c_vec[i]);
+
+                const auto a_grid_desc_ak0_m_ak1 = DeviceOp::MakeAGridDescriptor_AK0_M_AK1(
+                    problem_desc_vec[i].M, problem_desc_vec[i].K, problem_desc_vec[i].StrideA);
+                const auto b_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
+                    problem_desc_vec[i].K, problem_desc_vec[i].N, problem_desc_vec[i].StrideB0);
+                const auto b1_grid_desc_bk0_n_bk1 = DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
+                    problem_desc_vec[i].N, problem_desc_vec[i].O, problem_desc_vec[i].StrideB1);
+                const auto c_grid_desc_m_n = DeviceOp::MakeCGridDescriptor_M_N(
+                    problem_desc_vec[i].c_gs_ms_os_lengths, problem_desc_vec[i].c_gs_ms_os_strides);
+
+                const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n);
+
+                const index_t BlockStart     = grid_size_;
+                const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
+                const index_t grid_size_grp = block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) *
+                                              problem_desc_vec[i].Batch;
+                const index_t BlockEnd = grid_size_ + grid_size_grp;
+
+                // batch stride
+                // TODO ANT: only keep batch stride in tensor desc to reduce scalar cache pressure
+                const auto c_grid_desc_g_m_n = DeviceOp::MakeCGridDescriptor_G_M_N(
+                    problem_desc_vec[i].c_gs_ms_os_lengths, problem_desc_vec[i].c_gs_ms_os_strides);
+                const auto compute_base_ptr_of_batch =
+                    ComputeBasePtrOfStridedBatch(problem_desc_vec[i].BatchStrideA,
+                                                 problem_desc_vec[i].BatchStrideB0,
+                                                 problem_desc_vec[i].BatchStrideB1,
+                                                 c_grid_desc_g_m_n);
+
+                // C0 mask
+                const auto c0_matrix_mask = C0MatrixMask(problem_desc_vec[i].N);
+
+                grid_size_ += grid_size_grp;
+
+                group_kernel_args_.push_back({p_a_grid,
+                                              p_b_grid,
+                                              p_b1_grid,
+                                              p_c_grid,
+                                              a_grid_desc_ak0_m_ak1,
+                                              b_grid_desc_bk0_n_bk1,
+                                              b1_grid_desc_bk0_n_bk1,
+                                              c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n),
+                                              compute_base_ptr_of_batch,
+                                              c0_matrix_mask,
+                                              block_2_ctile_map,
+                                              BlockStart,
+                                              BlockEnd});
+
+                group_device_args_.push_back({problem_desc_vec[i].M,
+                                              problem_desc_vec[i].N,
+                                              problem_desc_vec[i].K,
+                                              problem_desc_vec[i].O,
+                                              problem_desc_vec[i].c_gs_ms_os_lengths.back(),
+                                              problem_desc_vec[i].c_gs_ms_os_strides.back(),
+                                              c_grid_desc_m_n});
+            }
+        }
+
+        std::vector<GroupKernelArg> group_kernel_args_;
+        std::vector<GroupDeviceArg> group_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            bool all_has_main_k_block_loop  = true;
+            bool some_has_main_k_block_loop = false;
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K = arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+                const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+                all_has_main_k_block_loop &= y;
+                some_has_main_k_block_loop |= y;
+            }
+
+            hipGetErrorString(hipMemcpy(arg.p_workspace_,
+                                        arg.group_kernel_args_.data(),
+                                        arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
+                                        hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                     GroupKernelArg,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.acc_element_op_,
+                    arg.b1_element_op_,
+                    arg.c_element_op_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(all_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else if(!some_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+            else
+            {
+                throw std::runtime_error("wrong! all gemm problems have to simultaneously meet "
+                                         "has_main_k_block_loop or no_main_k_block_loop");
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        bool all_has_main_k_block_loop  = true;
+        bool some_has_main_k_block_loop = false;
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto& kernel_arg = arg.group_kernel_args_[i];
+            const auto& device_arg = arg.group_device_args_[i];
+
+            // Check if C permute dimension matches GEMM + GEMM shape
+            const index_t c_m       = device_arg.c_grid_desc_m_n_.GetLength(I0);
+            const index_t c_gemm1n  = device_arg.c_grid_desc_m_n_.GetLength(I1);
+            const index_t a_m       = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+            const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+            if(!(c_m == a_m && c_gemm1n == b1_gemm1n))
+            {
+                return false;
+            }
+
+            // Check if having main loop
+            const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                           kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+            const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+            all_has_main_k_block_loop &= y;
+            some_has_main_k_block_loop |= y;
+
+            // Note: we need raw lengths since threadwise copy can not handle vector load when
+            // part of vector is out of bounds
+            const auto MRaw      = device_arg.M;
+            const auto NRaw      = device_arg.N;
+            const auto KRaw      = device_arg.K;
+            const auto Gemm1NRaw = device_arg.O;
+
+            // Check scalar per vector requirement
+            const auto a_extent_lowest =
+                is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
+            const auto b_extent_lowest =
+                is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
+            const auto b1_extent_lowest =
+                is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
+            const auto c_extent_lowest = device_arg.c_extent_lowest_;
+
+            if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+                 b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+                 b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+                 c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+
+            // Check vector store requirement; assumes last dimension in N to be contiguous
+            if(device_arg.c_stride_lowest_ != 1)
+            {
+                return false;
+            }
+
+            if(!GridwiseGemm::CheckValidity(kernel_arg.a_grid_desc_ak0_m_ak1_,
+                                            kernel_arg.b_grid_desc_bk0_n_bk1_,
+                                            kernel_arg.b1_grid_desc_bk0_n_bk1_,
+                                            device_arg.c_grid_desc_m_n_,
+                                            kernel_arg.block_2_ctile_map_))
+            {
+                return false;
+            }
+        }
+
+        // all gemm problems have to simultaneously meet has_main_k_block_loop or
+        // no_main_k_block_loop
+        if(!(all_has_main_k_block_loop || !some_has_main_k_block_loop))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<const void*> p_b1_vec,
+                             std::vector<void*> p_c_vec,
+                             std::vector<ProblemDesc> problem_desc_vec,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_b1_vec,
+                        p_c_vec,
+                        problem_desc_vec,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                                                      std::vector<const void*> p_b_vec,
+                                                      std::vector<const void*> p_b1_vec,
+                                                      std::vector<void*> p_c_vec,
+                                                      std::vector<ProblemDesc> problem_desc_vec,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      B1ElementwiseOperation b1_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_b1_vec,
+                                          p_c_vec,
+                                          problem_desc_vec,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GroupKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp

+#pragma once
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                const index_t group_count,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CDEElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].BlockStart_ &&
+             block_id < gemm_desc_ptr[group_id].BlockEnd_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].BlockStart_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        gemm_desc_ptr[group_id].a_ptr_,
+        gemm_desc_ptr[group_id].b_ptr_,
+        gemm_desc_ptr[group_id].ds_ptr_,
+        gemm_desc_ptr[group_id].e_ptr_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        gemm_desc_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        gemm_desc_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        gemm_desc_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
+                                                        BLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        ADataType,
+                                                        BDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        AElementwiseOperation,
+                                                        BElementwiseOperation,
+                                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemm_Xdl;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumPrefetch, // NumGemmKPrefetchStage
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    struct GroupedGemmBlock2ETileMap
+    {
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedGemmBlock2ETileMap()
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{});
+            BlockStart_        = -1;
+        }
+
+        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, ck::index_t BlockStart)
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            BlockStart_        = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - BlockStart_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_;
+    };
+
+    struct GemmBiasTransKernelArg
+    {
+        // pointers
+        const ADataType* a_ptr_;
+        const BDataType* b_ptr_;
+        typename GridwiseGemm::DsGridPointer ds_ptr_;
+        EDataType* e_ptr_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        GroupedGemmBlock2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_, BlockEnd_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
+        {
+            grid_size_ = 0;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideC = gemm_descs[i].stride_C_;
+
+                // pointer
+                typename GridwiseGemm::DsGridPointer p_ds_grid{};
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_Ds[i][j]);
+                });
+
+                // tensor descriptors for problem definiton
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(M, K, StrideA);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(K, N, StrideB);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    ds_grid_desc_m_n(j) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                        M, N, gemm_descs[i].stride_Ds_[j]);
+                });
+
+                const auto e_grid_desc_m_n =
+                    DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideC);
+
+                // tensor descriptors for block/thread-wise copy
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const index_t grid_size_grp =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, 0)
+                        .block_2_etile_map_.CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                const auto block_2_etile_map =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    // tensor descriptors for block/thread-wise copy
+                    DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+                    static_for<0, NumDTensor, 1>{}([&](auto j) {
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                ds_grid_desc_m_n[j]);
+                    });
+
+                    const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n);
+
+                    gemm_desc_kernel_arg_.push_back(
+                        GemmBiasTransKernelArg{static_cast<const ADataType*>(p_As[i]),
+                                               static_cast<const BDataType*>(p_Bs[i]),
+                                               p_ds_grid,
+                                               static_cast<EDataType*>(p_Es[i]),
+                                               a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               a_grid_desc_ak0_m_ak1,
+                                               b_grid_desc_bk0_n_bk1,
+                                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               block_2_etile_map,
+                                               BlockStart,
+                                               BlockEnd});
+                }
+            }
+        }
+
+        //  private:
+        index_t group_count_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation c_element_op_;
+
+        std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
+
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+                std::cout << "group: " << i << " arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.e_grid_desc_m_n_{ "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
+                          << std::endl;
+
+                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_m_k_,
+                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_n_k_,
+                                                arg.gemm_desc_kernel_arg_[i].ds_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].block_2_etile_map_))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                const auto K = arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(
+                hipMemcpy(arg.p_workspace_,
+                          arg.gemm_desc_kernel_arg_.data(),
+                          arg.gemm_desc_kernel_arg_.size() * sizeof(GemmBiasTransKernelArg),
+                          hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_grouped_gemm_xdl<GridwiseGemm,
+                                                            GemmBiasTransKernelArg,
+                                                            AElementwiseOperation,
+                                                            BElementwiseOperation,
+                                                            CDEElementwiseOperation,
+                                                            has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.c_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) != arg.group_count_)
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation c_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GemmBiasTransKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_layernorm_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+template <typename GridwiseReduction,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          typename GridDesc_M_K>
+__global__ void kernel_layernorm(const GridDesc_M_K x_grid_desc_m_k,
+                                 const GridDesc_M_K gamma_grid_desc_m_k,
+                                 const GridDesc_M_K beta_grid_desc_m_k,
+                                 const GridDesc_M_K y_grid_desc_m_k,
+                                 index_t num_k_block_tile_iteration,
+                                 AccDataType epsilon,
+                                 const XDataType* const __restrict__ p_x_global,
+                                 const GammaDataType* const __restrict__ p_gamma_global,
+                                 const BetaDataType* const __restrict__ p_beta_global,
+                                 YDataType* const __restrict__ p_y_global,
+                                 const AccElementwiseOperation acc_elementwise_op)
+{
+    GridwiseReduction::Run(x_grid_desc_m_k,
+                           gamma_grid_desc_m_k,
+                           beta_grid_desc_m_k,
+                           y_grid_desc_m_k,
+                           num_k_block_tile_iteration,
+                           epsilon,
+                           p_x_global,
+                           p_gamma_global,
+                           p_beta_global,
+                           p_y_global,
+                           acc_elementwise_op);
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = LayerNorm(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XYSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorSize>
+struct DeviceLayernormImpl : public DeviceLayernorm<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    AccDataType,
+                                                    YDataType,
+                                                    AccElementwiseOperation,
+                                                    Rank,
+                                                    NumReduceDim>
+{
+    static_assert(
+        ((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
+         (GammaSrcVectorDim == 1 && KThreadSliceSize % GammaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        ((BetaSrcVectorDim == 0 && MThreadSliceSize % BetaSrcVectorSize == 0) ||
+         (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseReduceLayernormGeneric =
+        GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  AccDataType,
+                                                  AccElementwiseOperation,
+                                                  GridDesc_M_K,
+                                                  BlockSize,
+                                                  MThreadClusterSize,
+                                                  KThreadClusterSize,
+                                                  MThreadSliceSize,
+                                                  KThreadSliceSize,
+                                                  XYSrcVectorDim,
+                                                  XSrcVectorSize,
+                                                  GammaSrcVectorDim,
+                                                  GammaSrcVectorSize,
+                                                  BetaSrcVectorDim,
+                                                  BetaSrcVectorSize,
+                                                  XYSrcVectorDim,
+                                                  YDstVectorSize,
+                                                  false>;
+    using GridwiseReduceLayernormSweepOnce =
+        GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  AccDataType,
+                                                  AccElementwiseOperation,
+                                                  GridDesc_M_K,
+                                                  BlockSize,
+                                                  MThreadClusterSize,
+                                                  KThreadClusterSize,
+                                                  MThreadSliceSize,
+                                                  KThreadSliceSize,
+                                                  XYSrcVectorDim,
+                                                  XSrcVectorSize,
+                                                  GammaSrcVectorDim,
+                                                  GammaSrcVectorSize,
+                                                  BetaSrcVectorDim,
+                                                  BetaSrcVectorSize,
+                                                  XYSrcVectorDim,
+                                                  YDstVectorSize,
+                                                  true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::vector<index_t> xStrides,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 AccElementwiseOperation acc_elementwise_op,
+                 AccDataType epsilon,
+                 const XDataType* p_x,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : epsilon_(epsilon),
+              p_x_(p_x),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              acc_elementwise_op_(acc_elementwise_op)
+        {
+            Lengths_      = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            xStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
+            yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            blkGroupSize_          = 1;
+            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                        M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            isSweeponce_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+        }
+
+        AccDataType epsilon_;
+
+        const XDataType* p_x_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        AccElementwiseOperation acc_elementwise_op_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool isSweeponce_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main = arg.isSweeponce_
+                                         ? kernel_layernorm<GridwiseReduceLayernormSweepOnce,
+                                                            XDataType,
+                                                            GammaDataType,
+                                                            BetaDataType,
+                                                            YDataType,
+                                                            AccDataType,
+                                                            AccElementwiseOperation,
+                                                            GridDesc_M_K>
+                                         : kernel_layernorm<GridwiseReduceLayernormGeneric,
+                                                            XDataType,
+                                                            GammaDataType,
+                                                            BetaDataType,
+                                                            YDataType,
+                                                            AccDataType,
+                                                            AccElementwiseOperation,
+                                                            GridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.p_x_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.acc_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            if(p_arg_->xStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+        };
+
+        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(GammaSrcVectorDim == 0)
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(BetaSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        AccElementwiseOperation acc_elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          xStrides,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          acc_elementwise_op,
+                                          epsilon,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceLayernormImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include <array>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+struct DeviceMultipleReduce : public BaseOperator
+{
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStrides,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+using DeviceMultipleReducePtr = std::unique_ptr<DeviceMultipleReduce<Rank,
+                                                                     NumReduceDim,
+                                                                     NumReduction,
+                                                                     InElementwiseOperationTuple,
+                                                                     AccElementwiseOperationTuple>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_multiple_reduce_multiblock.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceMultiBlock : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr bool CheckDataTypeTuple()
+    {
+        bool flag = true;
+
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+            flag =
+                flag && ck::reduce::InMemoryDataOperatonSupportedOnDataType<OutMemoryDataOperation,
+                                                                            OutDataType>::value;
+        });
+
+        return flag;
+    };
+
+    static_assert(CheckDataTypeTuple(),
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    // So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
+    // later
+    static constexpr bool use_multiblock =
+        (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
+
+    static_assert(
+        ReduceOperation::IsCompatibleInMemoryDataOperation(OutMemoryDataOperation),
+        "The reduction accumulation operation must be compatible with the OutMemoryDataOperation!");
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(
+        std::array<index_t, NumInputDim>{}, std::array<index_t, NumInputDim>{}, 1, 1));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumOutputDim>& outLengths,
+                                                const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto length = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto pad = math::integer_least_multiple(length, BlockSize) - length;
+
+        auto out_grid_desc_m_padded =
+            transform_tensor_descriptor(out_grid_desc_m,
+                                        make_tuple(make_right_pad_transform(length, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple_2()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptorForBufferSet(std::array<index_t, NumOutputDim>{},
+                                                       std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutGridDesc_M_Tuple_2 = decltype(GenerateOutGrid1dDescTuple_2());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<const void*, NumReduction>& alphas,
+                 const std::array<const void*, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = *static_cast<const AccDataType*>(alphas[i]);
+                beta_values_(i)  = *static_cast<const AccDataType*>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(use_multiblock)
+            {
+
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize =
+                        (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                        (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize = 1;
+                numBlockTileIteration =
+                    (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            in_grid_desc_m_k =
+                MakeSrc2dDescriptor(inLengths_, inStrides_, blkGroupSize, numBlockTileIteration);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            out_grid_desc_m_tuple_2 = generate_tuple(
+                [&](auto I) {
+                    return MakeDst1dDescriptorForBufferSet(outLengths, outStridesArray[I]);
+                },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+
+            gridSize_pre =
+                math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+        OutGridDesc_M_Tuple_2 out_grid_desc_m_tuple_2;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        size_t gridSize_pre;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_multiblock<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             OutMemoryDataOperation,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_multiblock<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            if constexpr(use_multiblock)
+            {
+                auto identity_values = generate_tuple(
+                    [&](auto iR) {
+                        using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[iR])>;
+                        return ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
+                            OutMemoryDataOperation);
+                    },
+                    Number<NumReduction>{});
+
+                const auto kernel_pre = kernel_multiple_buffer_set_value<OutGridDesc_M_Tuple_2,
+                                                                         NumReduction,
+                                                                         BlockSize,
+                                                                         OutDataTypePointerTuple,
+                                                                         OutDataTypeTuple>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kernel_pre,
+                                                   dim3(arg.gridSize_pre),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.out_grid_desc_m_tuple_2,
+                                                   arg.out_dev_buffers_,
+                                                   identity_values);
+            };
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(use_multiblock)
+        {
+            for(size_t i = 0; i < pArg->beta_values_.Size(); i++)
+                if(pArg->beta_values_[i] != 0.0f)
+                    return (false);
+        };
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        if constexpr(use_multiblock)
+        {
+            // blkGroupSize of 1 should be handled by Blockwise path using
+            // InMemoryDataOperationEnum::Set
+            if(pArg->blkGroupSize == 1)
+                return (false);
+
+            // This is very strong restriction, but needed to avoid some failure
+            if(pArg->outLengths_[NumOutputDim - 1] % M_BlockTileSize != 0)
+                return (false);
+        }
+        else
+        {
+            // cases with very small reduce_total_length should be handled by ThreadWise kernel
+            if(pArg->reduce_total_length / KThreadSliceSize < 2)
+                return (false);
+        };
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceMultipleReduceBlockWise<" : "DeviceMultipleReduceMultiBlock<") << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_multiple_reduce_threadwise.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceThreadWise : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K =
+            math::integer_least_multiple(reduceLength, K_BlockTileSize) - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(std::array<index_t, NumInputDim>{},
+                                                        std::array<index_t, NumInputDim>{}));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<const void*, NumReduction>& alphas,
+                 const std::array<const void*, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = *static_cast<const AccDataType*>(alphas[i]);
+                beta_values_(i)  = *static_cast<const AccDataType*>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            in_grid_desc_m_k = MakeSrc2dDescriptor(inLengths_, inStrides_);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_threadwise<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_threadwise<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<const void*, NumReduction> alphas,
+        const std::array<const void*, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceMultipleReduceThreadwise<" << BlockSize << ",";
+        str << "M_C" << BlockSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << 1 << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_normalization.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct DeviceNormalization : public BaseOperator
+{
+    // inLengths: input tensor extent(s) from high to low dimension
+    // inStrides: input tensor stride(s) from high to low dimension
+    // reduceDims: the dimension(s) the normalization operation is applied
+    // alpha: typeless pointer in host memory storing the alpha scaling value of type AccDataType
+    // beta: typeless pointer in host memory storing the beta scaling value of type AccDataType
+    // in_dev: typeless const pointer in device memory storing the input tensor
+    // out_dev: typeless pointer in device memory storing the output tensor
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
+                                                              const std::vector<index_t> inStrides,
+                                                              const std::vector<int> reduceDims,
+                                                              const void* alpha,
+                                                              const void* beta,
+                                                              const void* in_dev,
+                                                              void* out_dev) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual index_t GetRank() const = 0;
+
+    virtual index_t GetNumReduceDim() const = 0;
+};
+
+using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization>;
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceLayernorm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        AccElementwiseOperation acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceLayernormPtr = std::unique_ptr<DeviceLayernorm<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           AccDataType,
+                                                           YDataType,
+                                                           AccElementwiseOperation,
+                                                           Rank,
+                                                           NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <cmath>
+#include <memory>
+#include <type_traits>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDim, typename InDataType, typename OutDataType, typename ElementwiseOperation>
+struct DevicePermute : BaseOperator
+{
+    using Lengths = std::array<index_t, NumDim>;
+    using Strides = Lengths;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const Lengths& in_lengths,
+                        const Strides& in_strides,
+                        const Lengths& out_lengths,
+                        const Strides& out_strides,
+                        const void* in_dev_buffer,
+                        void* out_dev_buffer,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::ReduceTensorOp ReduceOpId>
+struct DevicePool2dFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* in_dev,
+                        void* out_dev,
+                        void* out_indices_dev,
+                        ck::index_t N,
+                        ck::index_t C,
+                        std::array<ck::index_t, 2> input_spatial_lengths,
+                        std::array<ck::index_t, 2> window_spatial_lengths,
+                        std::array<ck::index_t, 2> output_spatial_lengths,
+                        std::array<ck::index_t, 2> window_strides,
+                        std::array<ck::index_t, 2> input_left_pads,
+                        std::array<ck::index_t, 2> input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <ck::ReduceTensorOp ReduceOpId>
+using DevicePool2dFwdPtr = std::unique_ptr<DevicePool2dFwd<ReduceOpId>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OuputIndex,
+          ck::index_t BlockSize,
+          ck::index_t ReduceMThreadClusterSize,
+          ck::index_t ReduceKThreadClusterSize,
+          ck::index_t ReduceMThreadSliceSize,
+          ck::index_t ReduceKThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd<ReduceOpId>
+{
+    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>{};
+
+    using IndexDataType = int32_t;
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    static constexpr index_t InSrcOutDstVectorDim =
+        0; // for NHWC, the dim C is the vector Dim for both input and output in memory, which is
+           // not reduced.
+
+    static constexpr ck::index_t ReduceM_BlockTileSize =
+        ReduceMThreadClusterSize * ReduceMThreadSliceSize;
+    static constexpr ck::index_t ReduceK_BlockTileSize =
+        ReduceKThreadClusterSize * ReduceKThreadSliceSize;
+
+    static auto MakeABGridDescriptor_A_M_K_B_M(ck::index_t N,
+                                               ck::index_t C,
+                                               std::array<ck::index_t, 2> input_spatial_lengths,
+                                               std::array<ck::index_t, 2> window_spatial_lengths,
+                                               std::array<ck::index_t, 2> output_spatial_lengths,
+                                               std::array<ck::index_t, 2> window_strides,
+                                               std::array<ck::index_t, 2> input_left_pads,
+                                               std::array<ck::index_t, 2> input_right_pads)
+    {
+        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 = window_spatial_lengths[0];
+        const index_t X = window_spatial_lengths[1];
+
+        const index_t ConvStrideH = window_strides[0];
+        const index_t ConvStrideW = window_strides[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 ReduceMRaw = N * Ho * Wo * C;
+        const index_t ReduceMPad =
+            math::integer_least_multiple(ReduceMRaw, ReduceM_BlockTileSize) - ReduceMRaw;
+
+        const index_t ReduceKRaw = Y * X;
+        const index_t ReduceKPad =
+            math::integer_least_multiple(ReduceKRaw, ReduceK_BlockTileSize) - ReduceKRaw;
+
+        // A[ReduceM, ReduceK]
+        const auto in_grid_desc_n_hi_wi_c =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        const auto in_grid_desc_n_hip_wip_c = transform_tensor_descriptor(
+            in_grid_desc_n_hi_wi_c,
+            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_grid_desc_n_y_ho_x_wo_c = transform_tensor_descriptor(
+            in_grid_desc_n_hip_wip_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_embed_transform(make_tuple(Y, Ho), make_tuple(I1, ConvStrideH)),
+                       make_embed_transform(make_tuple(X, Wo), make_tuple(I1, 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_grid_desc_reducemraw_reducekraw =
+            transform_tensor_descriptor(in_grid_desc_n_y_ho_x_wo_c,
+                                        make_tuple(make_merge_transform(make_tuple(N, Ho, Wo, C)),
+                                                   make_merge_transform(make_tuple(Y, X))),
+                                        make_tuple(Sequence<0, 2, 4, 5>{}, Sequence<1, 3>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_grid_desc_reducem_reducek = transform_tensor_descriptor(
+            in_grid_desc_reducemraw_reducekraw,
+            make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad),
+                       make_right_pad_transform(ReduceKRaw, ReduceKPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        // B[ReduceM]
+        const auto out_grid_desc_reducemraw =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo * C));
+
+        const auto out_grid_desc_reducem = transform_tensor_descriptor(
+            out_grid_desc_reducemraw,
+            make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+
+        return make_tuple(in_grid_desc_reducem_reducek, out_grid_desc_reducem);
+    }
+
+    using ABGridDescs = decltype(
+        MakeABGridDescriptor_A_M_K_B_M(1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(ABGridDescs{}[I0])>;
+    using BGridDesc_M   = remove_cvref_t<decltype(ABGridDescs{}[I1])>;
+
+    // TODO
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_dev,
+                 OutDataType* p_out_dev,
+                 int* p_out_indices_dev,
+                 ck::index_t N,
+                 ck::index_t C,
+                 std::array<ck::index_t, 2>& input_spatial_lengths,
+                 std::array<ck::index_t, 2>& window_spatial_lengths,
+                 std::array<ck::index_t, 2>& output_spatial_lengths,
+                 std::array<ck::index_t, 2>& window_strides,
+                 std::array<ck::index_t, 2>& input_left_pads,
+                 std::array<ck::index_t, 2>& input_right_pads)
+            : p_in_dev_{p_in_dev},
+              p_out_dev_{p_out_dev},
+              p_out_indices_dev_{p_out_indices_dev},
+              a_grid_desc_m_k_{},
+              b_grid_desc_m_{}
+        {
+            const auto descs = MakeABGridDescriptor_A_M_K_B_M(N,
+                                                              C,
+                                                              input_spatial_lengths,
+                                                              window_spatial_lengths,
+                                                              output_spatial_lengths,
+                                                              window_strides,
+                                                              input_left_pads,
+                                                              input_right_pads);
+
+            a_grid_desc_m_k_ = descs[I0];
+            b_grid_desc_m_   = descs[I1];
+
+            invariant_lowest_length_ = C;
+            reduce_lowest_length_    = window_spatial_lengths[1];
+
+            int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+            std::tie(in_element_op_, acc_element_op_) =
+                reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+        }
+
+        const InDataType* p_in_dev_;
+        OutDataType* p_out_dev_;
+        int* p_out_indices_dev_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_M b_grid_desc_m_;
+        InElementwiseOperation in_element_op_;
+        AccElementwiseOperation acc_element_op_;
+
+        // for checking vector load/store
+        ck::index_t invariant_lowest_length_;
+        ck::index_t reduce_lowest_length_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using gridwise_reduce =
+                GridwiseReduction_mk_to_m_threadwise<InDataType,
+                                                     OutDataType,
+                                                     AccDataType,
+                                                     IndexDataType,
+                                                     AGridDesc_M_K,
+                                                     BGridDesc_M,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     InMemoryDataOperationEnum::Set,
+                                                     false, // propagate_nan
+                                                     BlockSize,
+                                                     ReduceMThreadSliceSize,
+                                                     ReduceKThreadSliceSize,
+                                                     InSrcOutDstVectorDim,
+                                                     InSrcOutDstVectorSize,
+                                                     InSrcOutDstVectorSize>;
+
+            const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
+                                                         OuputIndex,
+                                                         false, // don't have index input
+                                                         InDataType,
+                                                         OutDataType,
+                                                         AccDataType,
+                                                         IndexDataType,
+                                                         AGridDesc_M_K,
+                                                         BGridDesc_M,
+                                                         InElementwiseOperation,
+                                                         AccElementwiseOperation>;
+
+            ck::index_t ReduceM = arg.a_grid_desc_m_k_.GetLength(I0);
+
+            const index_t grid_size = (ReduceM / ReduceM_BlockTileSize);
+
+            return launch_and_time_kernel(stream_config,
+                                          kernel,
+                                          dim3(grid_size),
+                                          dim3(BlockSize),
+                                          0,
+                                          arg.a_grid_desc_m_k_,
+                                          arg.b_grid_desc_m_,
+                                          arg.in_element_op_,
+                                          arg.acc_element_op_,
+                                          float(1),
+                                          arg.p_in_dev_,
+                                          nullptr,
+                                          float(0),
+                                          arg.p_out_dev_,
+                                          arg.p_out_indices_dev_);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg->invariant_lowest_length_ % InSrcOutDstVectorSize != 0)
+        {
+            return (false);
+        }
+
+        return (true);
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        ck::index_t N,
+                        ck::index_t C,
+                        std::array<ck::index_t, 2> input_spatial_lengths,
+                        std::array<ck::index_t, 2> window_spatial_lengths,
+                        std::array<ck::index_t, 2> output_spatial_lengths,
+                        std::array<ck::index_t, 2> window_strides,
+                        std::array<ck::index_t, 2> input_left_pads,
+                        std::array<ck::index_t, 2> input_right_pads) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_dev),
+                                          static_cast<OutDataType*>(p_out_dev),
+                                          static_cast<int*>(p_out_indices_dev),
+                                          N,
+                                          C,
+                                          input_spatial_lengths,
+                                          window_spatial_lengths,
+                                          output_spatial_lengths,
+                                          window_strides,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    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 << "DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<" << BlockSize << ",";
+        str << "M_C" << ReduceMThreadClusterSize << "_S" << ReduceMThreadSliceSize << ",";
+        str << "K_C" << ReduceKThreadClusterSize << "_S" << ReduceKThreadSliceSize << ",";
+        str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

deps/cget/pkg/ROCmSoftwarePlatform__composable_kernel/install/include/ck/tensor_operation/gpu/device/device_reduce.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include <iostream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation, typename AccElementwiseOperation>
+struct DeviceReduce : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> inLengths,
+                        const std::vector<index_t> inStrides,
+                        const std::vector<index_t> outLengths,
+                        const std::vector<index_t> outStrides,
+                        const std::vector<int> reduceDims,
+                        float alpha,
+                        float beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation, typename AccElementwiseOperation>
+using DeviceReducePtr =
+    std::unique_ptr<DeviceReduce<InElementwiseOperation, AccElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck