merge develop to attn-train-develop-qloop

include/ck/tensor_operation/gpu/device/device_normalization.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -14,9 +14,9 @@ namespace device {
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
-          typename AccDataType,
+          typename ComputeDataType,
           typename YDataType,
-          typename AccElementwiseOperation,
+          typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
 struct DeviceNormalization : public BaseOperator
@@ -28,14 +28,14 @@ struct DeviceNormalization : public BaseOperator
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
                         const std::vector<index_t> reduceDims,
-                        AccDataType epsilon,
+                        double epsilon,
                         const void* p_x,
                         const void* p_gamma,
                         const void* p_beta,
                         void* p_y,
                         void* p_savedMean,
                         void* p_savedInvVar,
-                        AccElementwiseOperation acc_elementwise_op) = 0;
+                        YElementwiseOperation y_elementwise_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
@@ -43,17 +43,17 @@ struct DeviceNormalization : public BaseOperator
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
-          typename AccDataType,
+          typename ComputeDataType,
           typename YDataType,
-          typename AccElementwiseOperation,
+          typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
 using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
                                                                    GammaDataType,
                                                                    BetaDataType,
-                                                                   AccDataType,
+                                                                   ComputeDataType,
                                                                    YDataType,
-                                                                   AccElementwiseOperation,
+                                                                   YElementwiseOperation,
                                                                    Rank,
                                                                    NumReduceDim>>;
 

include/ck/tensor_operation/gpu/device/device_permute.hpp

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

include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#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 InOutRank,
+          index_t WindowRank,
+          typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DevicePoolFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        std::vector<ck::index_t> input_lengths,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> output_lengths,
+                        std::vector<ck::index_t> input_stride,
+                        std::vector<ck::index_t> output_stride,
+                        std::vector<ck::index_t> indices_stride,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        std::vector<ck::index_t> pooling_dims) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp → include/ck/tensor_operation/gpu/device/device_put_element.hpp

@@ -3,8 +3,7 @@
 
 #pragma once
 
-#include <iostream>
-#include <array>
+#include <vector>
 
 #include "ck/tensor_operation/gpu/device/device_base.hpp"
 #include "ck/utility/reduction_enums.hpp"
@@ -13,28 +12,25 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <ck::ReduceTensorOp ReduceOpId>
-struct DevicePool2dFwd : public BaseOperator
+// output[indices] = input
+template <typename InDataType,
+          typename IndexDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum Op>
+struct DevicePutElement : 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;
+    MakeArgumentPointer(const void* p_input,
+                        const void* p_indices,
+                        void* p_output,
+                        index_t input_length,
+                        index_t output_length,
+                        ElementwiseOperation elementwise_op) = 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

include/ck/tensor_operation/gpu/device/device_reduce.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -13,10 +13,16 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <index_t Rank,
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
           index_t NumReduceDim,
+          typename ReduceOperation,
           typename InElementwiseOperation,
-          typename AccElementwiseOperation>
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
 struct DeviceReduce : public BaseOperator
 {
     static constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
@@ -27,8 +33,8 @@ struct DeviceReduce : public BaseOperator
                         const std::array<index_t, NumOutDim> outLengths,
                         const std::array<index_t, NumOutDim> outStrides,
                         const std::array<int, NumReduceDim> reduceDims,
-                        float alpha,
-                        float beta,
+                        double alpha,
+                        double beta,
                         const void* in_dev,
                         const void* in_index_dev,
                         void* out_dev,
@@ -39,12 +45,26 @@ struct DeviceReduce : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <index_t Rank,
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
           index_t NumReduceDim,
+          typename ReduceOperation,
           typename InElementwiseOperation,
-          typename AccElementwiseOperation>
-using DeviceReducePtr = std::unique_ptr<
-    DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>>;
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+using DeviceReducePtr = std::unique_ptr<DeviceReduce<InDataType,
+                                                     AccDataType,
+                                                     OutDataType,
+                                                     Rank,
+                                                     NumReduceDim,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     PropagateNan,
+                                                     OutputIndex>>;
 
 } // namespace device
 } // namespace tensor_operation

include/ck/tensor_operation/gpu/device/device_softmax.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -18,7 +18,8 @@ template <typename InDataType,
           typename OutDataType,
           typename InElementwiseOp,
           typename AccElementwiseOp,
-          index_t Rank>
+          index_t Rank,
+          index_t NumReduceDim>
 struct DeviceSoftmax : public BaseOperator
 {
     //
@@ -27,10 +28,8 @@ struct DeviceSoftmax : public BaseOperator
     // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
     // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
     // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
-    // @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
-    //                                 value as type AccDataType
-    // @param[in]  beta                Typeless pointer in host memory storing the beta scaling
-    //                                 value as type AccDataType
+    // @param[in]  alpha               double type value
+    // @param[in]  beta                double type value
     // @param[in]  in_dev              Typeless const pointer in device memory storing the input
     //                                 tensor
     // @param      out_dev             Typeless pointer in device memory storing the output tensor
@@ -43,16 +42,14 @@ struct DeviceSoftmax : public BaseOperator
     MakeArgumentPointer(const std::vector<index_t> inLengths,
                         const std::vector<index_t> inStrides,
                         const std::vector<int> reduceDims,
-                        const void* alpha,
-                        const void* beta,
+                        double alpha,
+                        double beta,
                         const void* in_dev,
                         void* out_dev,
                         InElementwiseOp in_elementwise_op,
                         AccElementwiseOp acc_elementwise_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
-    virtual index_t GetRank() const                           = 0;
-    virtual index_t GetNumReduceDim() const                   = 0;
 };
 
 template <typename InDataType,
@@ -60,9 +57,15 @@ template <typename InDataType,
           typename OutDataType,
           typename InElementwiseOp,
           typename AccElementwiseOp,
-          index_t Rank>
-using DeviceSoftmaxPtr = std::unique_ptr<
-    DeviceSoftmax<InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank>>;
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceSoftmaxPtr = std::unique_ptr<DeviceSoftmax<InDataType,
+                                                       AccDataType,
+                                                       OutDataType,
+                                                       InElementwiseOp,
+                                                       AccElementwiseOp,
+                                                       Rank,
+                                                       NumReduceDim>>;
 
 } // namespace device
 } // namespace tensor_operation

include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp

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

include/ck/tensor_operation/gpu/device/gemm_specialization.hpp

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

include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_batched_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_wmma_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+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[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
+//
+// Detail- Packed tensor satisfies
+//   stride_0 = 1
+//   stride_i = stride_{i - 1} * extent_{i - 1}
+// So tensor
+//   [G0, G1, G2, M, N]
+// transposed into tensor
+//   [G0, G2, G1, M, N]
+// with strides
+//   [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
+// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
+// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
+// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
+//
+// Might need to expose dimension order to the interface to fully support
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
+    : public DeviceBatchedContractionMultipleD<NumDimG,
+                                               NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp                      = DeviceBatchedContractionMultipleD_Wmma_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>{};
+    // K1 = Max Vector Access Pixels
+    static constexpr auto K1Number = Number<K1>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + 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[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + 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[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_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 e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_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(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);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_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 e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_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(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_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>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // Gridwise descriptor, mapping to whole given provblem.
+    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({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_A_;
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_B_;
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M_K1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N_K1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(DeviceOp::MakeAGridDescriptor_K0_M_K1(AGridDesc_M_K{}));
+    using BGridDesc_K0_N_K1 = decltype(DeviceOp::MakeBGridDescriptor_K0_N_K1(BGridDesc_N_K{}));
+
+    // GridwiseOp
+    using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
+        // DataType Family
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        // InMemory Data Descriptor
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        // ElementwiseOp Family
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        // Tiling Family
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        // ThreadCluster Family
+        BlockSize,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{},
+              b_grid_desc_n_k_{},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              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_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
+        {
+            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_grid[i]);
+            });
+
+            a_grid_desc_m_k_ =
+                DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+            b_grid_desc_n_k_ =
+                DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+
+            ds_grid_desc_m_n_ =
+                DeviceOp::MakeDsGridDescriptor_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides);
+
+            e_grid_desc_m_n_ =
+                DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(a_grid_desc_m_k_);
+            b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(b_grid_desc_n_k_);
+
+            block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
+
+            ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n_);
+
+            e_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+        }
+
+        // Pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseOp::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Tensor Descriptors
+        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_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+
+        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock;
+        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        // Block to Tile mapping
+        typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // Idle
+        index_t M01_;
+        index_t N01_;
+
+        // ElementwiseOp
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // 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_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        // Batch Offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.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_contraction_multiple_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    DeviceOp::AGridDesc_K0_M_K1,
+                    DeviceOp::BGridDesc_K0_N_K1,
+                    typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseOp::DefaultBlock2CTileMap,
+                    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_,
+                                              G,
+                                              arg.a_grid_desc_k0_m_k1_,
+                                              arg.b_grid_desc_k0_n_k1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            if(GridwiseOp::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // 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() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                      arg.b_grid_desc_k0_n_k1_,
+                                      arg.ds_grid_desc_m_n_,
+                                      arg.e_grid_desc_m_n_,
+                                      arg.block_2_ctile_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(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.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 i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetLength(I3) %
+                         CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I3) %
+                      CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEShuffleBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    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,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ks_lengths,
+                        b_gs_ns_ks_lengths,
+                        ds_gs_ms_ns_lengths,
+                        e_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_strides,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ks_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ks_lengths,
+                                          b_gs_ns_ks_lengths,
+                                          ds_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_strides,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceBatchedContractionMultipleD_Wmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerWMMA << ", "
+            << NPerWMMA << ", "
+            << MRepeat << ", "
+            << NRepeat
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -56,7 +56,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =
@@ -839,7 +840,9 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp

@@ -74,7 +74,8 @@ __global__ void
                                           const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
                                           const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     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);

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -60,7 +60,8 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -588,7 +589,9 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -83,7 +83,8 @@ __global__ void
                                 const Block2ETileMap block_2_etile_map)
 {
 
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     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);
@@ -579,7 +580,9 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \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.
+ *
+ * \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
+ * impl/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.
+ */
+
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dl_multiple_d(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2CTileMap block_2_ctile_map)
+{
+// TODO: Enable for gfx90a after complier fix
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||              \
+    defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
+
+    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 = DsGridDesc_M0_M10_M11_N0_N10_N11::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::Run(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_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = e_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          enable_if_t<
+              is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+                  is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+              bool> = false>
+struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD<ALayout,
+                                                                      BLayout,
+                                                                      DsLayout,
+                                                                      ELayout,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      DsDataType,
+                                                                      EDataType,
+                                                                      AElementwiseOperation,
+                                                                      BElementwiseOperation,
+                                                                      CDEElementwiseOperation>
+
+{
+    using DeviceOp                      = DeviceBatchedGemmMultipleD_Dl;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    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_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using DsGridDesc_M_N    = decltype(MakeDsGridDescriptor_M_N({}, {}, {}));
+    using EGridDesc_M_N     = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       std::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
+        {
+            std::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_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        std::array<ck::index_t, NumDTensor> BatchStrideDs_;
+        index_t BatchStrideE_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_K0_M_K1,
+                                         BGridDesc_K0_N_K1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using EGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                 index_t BatchStrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              K_(K),
+              Batch_(Batch),
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              e_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
+              block_2_ctile_map_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            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_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
+            });
+            e_grid_desc_m_n_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, e_grid_desc_m_n_))
+            {
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        index_t K_;
+
+        // Batch
+        index_t Batch_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // for calculating batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmMultipleD_Dl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{"
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0),
+                                                arg.e_grid_desc_m_n_.GetLength(I1)) *
+                arg.Batch_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
+
+                const auto kernel =
+                    kernel_gemm_dl_multiple_d<GridwiseGemm,
+                                              ADataType,
+                                              typename GridwiseGemm::DsGridPointer,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation,
+                                              DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                              DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                              DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
+                                              DeviceOp::EGridDesc_M0_M10_M11_N0_N10_N11,
+                                              ComputePtrOffsetOfStridedBatch,
+                                              DefaultBlock2CTileMap,
+                                              has_main_loop,
+                                              has_double_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.Batch_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        // 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)
+    {
+        // TODO: Enable for gfx90a after complier fix
+        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
+           ck::get_device_name() == "gfx1030" || ck::get_device_name() == "gfx940" ||
+           ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            bool pass = true;
+            pass      = pass && arg.K_ % K1 == 0;
+
+            pass = pass && GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                                       arg.b_grid_desc_k0_n_k1_,
+                                                       arg.e_grid_desc_m_n_);
+
+            return pass;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    // polymorphic
+    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,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<ck::index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                             index_t BatchStrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        Batch,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideDs,
+                        BatchStrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideDs,
+                                          BatchStrideE,
+                                          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 << "DeviceBatchedGemmMultipleD_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -68,7 +68,8 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -788,9 +789,45 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
         return true;
     }
 
+    // check if DsLayout is supported
+    template <typename RefLayout, typename DsLayout, const index_t NumDTensor>
+    static bool CheckDLayout()
+    {
+        static bool valid = true;
+        // iterate over DLayout tuple
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+            // if RefLayout and DLayout are same, keep valid true, otherwise false
+            valid = valid && is_same_v<RefLayout, DLayout>;
+        });
+        return valid;
+    }
+
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
+        {
+            return false;
+        }
+
+        // Check supported layouts
+        // A0 - Row
+        // B0 - Col
+        // D0s - Rows
+        // B1 - Row or Col
+        // D1s - Rows
+        // E1 - Row
+        if(!(is_same_v<tensor_layout::gemm::RowMajor, A0Layout> &&
+             is_same_v<tensor_layout::gemm::ColumnMajor, B0Layout> &&
+             CheckDLayout<tensor_layout::gemm::RowMajor, D0sLayout, NumD0Tensor>() &&
+             (is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ||
+              is_same_v<tensor_layout::gemm::ColumnMajor,
+                        B1Layout>)&&CheckDLayout<tensor_layout::gemm::RowMajor,
+                                                 D1sLayout,
+                                                 NumD1Tensor>() &&
+             is_same_v<tensor_layout::gemm::RowMajor, E1Layout>))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -59,7 +59,8 @@ __global__ void
             const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     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);

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -13,7 +13,7 @@
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp"
 #include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
@@ -25,15 +25,17 @@ namespace device {
 template <typename GridwiseGemm,
           typename FloatAB,
           typename FloatC,
+          typename D0sPointer,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename AccElementwiseOperation,
+          typename C0DEElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation,
+          typename C1DEElementwiseOperation,
           typename AGridDesc_AK0_M_AK1,
           typename BGridDesc_BK0_N_BK1,
           typename B1GridDesc_BK0_N_BK1,
-          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
           typename Block2CTileMap,
           typename ComputeBasePtrOfStridedBatch,
           typename C0MatrixMask,
@@ -47,22 +49,26 @@ __global__ void
             const FloatAB* __restrict__ p_b_grid,
             const FloatAB* __restrict__ p_b1_grid,
             FloatC* __restrict__ p_c_grid,
+            D0sPointer p_d0s_grid,
             const AElementwiseOperation a_element_op,
             const BElementwiseOperation b_element_op,
-            const AccElementwiseOperation acc_element_op,
+            const C0DEElementwiseOperation c0de_element_op,
             const B1ElementwiseOperation b1_element_op,
-            const CElementwiseOperation c_element_op,
+            const C1DEElementwiseOperation c1de_element_op,
             const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
             const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
             const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
-            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
-                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
             const Block2CTileMap block_2_ctile_map,
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -77,20 +83,28 @@ __global__ void
     const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
         static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
 
+    static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx, In)));
+        p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
+    });
+
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
                                                   p_b_grid + b_batch_offset,
                                                   p_b1_grid + b1_batch_offset,
                                                   p_c_grid + c_batch_offset,
+                                                  p_d0s_grid,
                                                   p_shared,
                                                   a_element_op,
                                                   b_element_op,
-                                                  acc_element_op,
+                                                  c0de_element_op,
                                                   b1_element_op,
-                                                  c_element_op,
+                                                  c1de_element_op,
                                                   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,
+                                                  c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
                                                   block_2_ctile_map,
                                                   c0_matrix_mask);
 #else
@@ -98,15 +112,17 @@ __global__ void
     ignore = p_b_grid;
     ignore = p_b1_grid;
     ignore = p_c_grid;
+    ignore = p_d0s_grid;
     ignore = a_element_op;
     ignore = b_element_op;
-    ignore = acc_element_op;
+    ignore = c0de_element_op;
     ignore = b1_element_op;
-    ignore = c_element_op;
+    ignore = c1de_element_op;
     ignore = a_grid_desc_ak0_m_ak1;
     ignore = b_grid_desc_bk0_n_bk1;
     ignore = b1_grid_desc_bk0_n_bk1;
-    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
     ignore = block_2_ctile_map;
     ignore = batch_count;
     ignore = compute_base_ptr_of_batch;
@@ -126,15 +142,15 @@ template <index_t NumDimG,
           typename BDataType,
           typename B1DataType,
           typename CDataType,
-          typename Acc0BiasDataType,
-          typename Acc1BiasDataType,
+          typename D0sDataType,
+          typename D1sDataType,
           typename GemmAccDataType,
           typename CShuffleDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename AccElementwiseOperation,
+          typename C0DEElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation,
+          typename C1DEElementwiseOperation,
           GemmSpecialization GemmSpec,
           TensorSpecialization ASpec,
           TensorSpecialization BSpec,
@@ -181,7 +197,8 @@ template <index_t NumDimG,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           MaskingSpecialization MaskingSpec,
-          LoopScheduler LoopSched = LoopScheduler::Default>
+          int D0sTransferSrcScalarPerVector = 4,
+          LoopScheduler LoopSched           = LoopScheduler::Default>
 struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
     : public DeviceBatchedGemmSoftmaxGemmPermute<NumDimG,
                                                  NumDimM,
@@ -192,23 +209,23 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                                                  BDataType,
                                                  B1DataType,
                                                  CDataType,
-                                                 Acc0BiasDataType,
-                                                 Acc1BiasDataType,
+                                                 D0sDataType,
+                                                 D1sDataType,
                                                  AElementwiseOperation,
                                                  BElementwiseOperation,
-                                                 AccElementwiseOperation,
+                                                 C0DEElementwiseOperation,
                                                  B1ElementwiseOperation,
-                                                 CElementwiseOperation,
+                                                 C1DEElementwiseOperation,
                                                  MaskingSpec>
 {
     static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
                   "Number of dimension must be greater than 0");
 
-    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
-    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
 
     // TODO ANT: implement bias combination
-    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+    static_assert(NumD1Tensor == 0, "Gemm1 Bias addition is unimplemented");
 
 #if 0
     // TODO ANT: use alias
@@ -261,14 +278,40 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             Number<B1K1>{});
     }
 
+    static auto MakeD0sGridDescriptor_M_N(
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return Transform::MakeCGridDescriptor_M_N(acc0_biases_gs_ms_ns_lengths[i],
+                                                          acc0_biases_gs_ms_ns_strides[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    static auto MakeD0sGridDescriptor_G_M_N(
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return Transform::MakeCGridDescriptor_G_M_N(acc0_biases_gs_ms_ns_lengths[i],
+                                                            acc0_biases_gs_ms_ns_strides[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
     using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
     using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
     using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
-    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using C1GridDesc_M_N       = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
     using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
     using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
     using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
-    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+    using C1GridDesc_G_M_N     = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+    using D0sGridDesc_M_N      = decltype(MakeD0sGridDescriptor_M_N({}, {}));
+    using D0sGridDesc_G_M_N    = decltype(MakeD0sGridDescriptor_G_M_N({}, {}));
 
     constexpr static auto make_MaskOutPredicate()
     {
@@ -288,11 +331,13 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
                                      const BGridDesc_G_N_K& b_grid_desc_g_n_k,
                                      const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
-                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+                                     const C1GridDesc_G_M_N& c1_grid_desc_g_m_n,
+                                     const D0sGridDesc_G_M_N& d0s_grid_desc_g_m_n)
             : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
               b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
               b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
-              c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
+              c1_grid_desc_g_m_n_(c1_grid_desc_g_m_n),
+              d0s_grid_desc_g_m_n_(d0s_grid_desc_g_m_n)
         {
         }
 
@@ -313,32 +358,42 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
         __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));
+            return c1_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
+                                                                Number<I> d0_idx) const
+        {
+            return d0s_grid_desc_g_m_n_[d0_idx].CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         private:
         AGridDesc_G_M_K a_grid_desc_g_m_k_;
         BGridDesc_G_N_K b_grid_desc_g_n_k_;
         B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
-        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+        C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
+        D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
     };
 
     // GridwiseGemm
-    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    using GridwiseGemm = GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle<
         ADataType, // TODO: distinguish A/B datatype
         GemmAccDataType,
         CShuffleDataType,
         CDataType,
+        D0sDataType,
         AElementwiseOperation,
         BElementwiseOperation,
-        AccElementwiseOperation,
+        C0DEElementwiseOperation,
         B1ElementwiseOperation,
-        CElementwiseOperation,
+        C1DEElementwiseOperation,
         InMemoryDataOperationEnum::Set,
         AGridDesc_AK0_M_AK1,
         BGridDesc_BK0_N_BK1,
         B1GridDesc_BK0_N_BK1,
-        CGridDesc_M_N,
+        C1GridDesc_M_N,
+        D0sGridDesc_M_N,
         NumGemmKPrefetchStage,
         BlockSize,
         MPerBlock,
@@ -384,7 +439,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         CShuffleBlockTransferScalarPerVector_NPerBlock,
         LoopSched,
         Transform::matrix_padder.PadN,
-        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle,
+        D0sTransferSrcScalarPerVector>;
 
     // Argument
     // FIXME: constness
@@ -395,8 +451,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             const BDataType* p_b_grid,
             const B1DataType* p_b1_grid,
             CDataType* p_c_grid,
-            const std::array<void*, NumAcc0Bias> p_acc0_biases,
-            const std::array<void*, NumAcc1Bias> p_acc1_biases,
+            const std::array<void*, NumD0Tensor> p_acc0_biases,
+            const std::array<void*, NumD1Tensor> p_acc1_biases,
             const std::vector<index_t>& a_gs_ms_ks_lengths,
             const std::vector<index_t>& a_gs_ms_ks_strides,
             const std::vector<index_t>& b_gs_ns_ks_lengths,
@@ -405,44 +461,48 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
             const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
             const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
-            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
-            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
-            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+            const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides,
+            const std::array<std::vector<ck::index_t>, NumD1Tensor>&
                 acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
-            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            const std::array<std::vector<ck::index_t>, NumD1Tensor>&
                 acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
             AElementwiseOperation a_element_op,
             BElementwiseOperation b_element_op,
-            AccElementwiseOperation acc_element_op,
+            C0DEElementwiseOperation c0de_element_op,
             B1ElementwiseOperation b1_element_op,
-            CElementwiseOperation c_element_op)
+            C1DEElementwiseOperation c1de_element_op)
             : p_a_grid_{p_a_grid},
               p_b_grid_{p_b_grid},
               p_b1_grid_{p_b1_grid},
               p_c_grid_{p_c_grid},
+              p_d0s_grid_{},
               a_grid_desc_ak0_m_ak1_{
                   DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
               b_grid_desc_bk0_n_bk1_{
                   DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
               b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
                   b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
-              c_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
-                                                                  c_gs_ms_gemm1ns_strides)},
+              c1_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                   c_gs_ms_gemm1ns_strides)},
               a_grid_desc_g_m_k_{
                   Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
               b_grid_desc_g_n_k_{
                   Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
               b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
                   b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
-              c_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
-                                                                      c_gs_ms_gemm1ns_strides)},
-              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
-              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              c1_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                       c_gs_ms_gemm1ns_strides)},
+              d0s_grid_desc_g_m_n_{DeviceOp::MakeD0sGridDescriptor_G_M_N(
+                  acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)},
+              c1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c1_grid_desc_m_n_)},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              acc_element_op_{acc_element_op},
+              c0de_element_op_{c0de_element_op},
               b1_element_op_{b1_element_op},
-              c_element_op_{c_element_op},
+              c1de_element_op_{c1de_element_op},
               c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
               raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
                                             b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
@@ -456,27 +516,44 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                                 b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
               c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
                                     c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
-              batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
-              compute_base_ptr_of_batch_{
-                  a_grid_desc_g_m_k_, b_grid_desc_g_n_k_, b1_grid_desc_g_n_k_, c_grid_desc_g_m_n_}
+              batch_count_{c1_grid_desc_g_m_n_.GetLength(I0)},
+              compute_base_ptr_of_batch_{a_grid_desc_g_m_k_,
+                                         b_grid_desc_g_n_k_,
+                                         b1_grid_desc_g_n_k_,
+                                         c1_grid_desc_g_m_n_,
+                                         d0s_grid_desc_g_m_n_}
         {
             // TODO ANT: implement bias addition
-            ignore = p_acc0_biases;
             ignore = p_acc1_biases;
-            ignore = acc0_biases_gs_ms_ns_lengths;
-            ignore = acc0_biases_gs_ms_ns_strides;
             ignore = acc1_biases_gs_ms_gemm1ns_lengths;
             ignore = acc1_biases_gs_ms_gemm1ns_strides;
 
+            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                // D0 pointer
+                p_d0s_grid_(i) = static_cast<const D0DataType*>(p_acc0_biases[i]);
+                // for  check
+                d0s_nl_ns_lengths_strides_[i].push_back(
+                    acc0_biases_gs_ms_ns_lengths[i][NumDimG + NumDimM]);
+                d0s_nl_ns_lengths_strides_[i].push_back(
+                    acc0_biases_gs_ms_ns_strides[i][NumDimG + NumDimM]);
+            });
+
             if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
                                            b_grid_desc_bk0_n_bk1_,
                                            b1_grid_desc_bk0_n_bk1_,
-                                           c_grid_desc_m_n_,
+                                           c1_grid_desc_m_n_,
                                            block_2_ctile_map_))
             {
-                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
-                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                        c_grid_desc_m_n_);
+                c1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c1_grid_desc_m_n_);
+
+                D0sGridDesc_M_N d0s_grid_desc_m_n{DeviceOp::MakeD0sGridDescriptor_M_N(
+                    acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)};
+                d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
+                    GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+                        d0s_grid_desc_m_n);
             }
         }
 
@@ -491,9 +568,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
                       << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
                       << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
-            std::cout << "c_grid_desc_g_m_n_: " << c_grid_desc_g_m_n_.GetLength(I0) << ", "
-                      << c_grid_desc_g_m_n_.GetLength(I1) << ", "
-                      << c_grid_desc_g_m_n_.GetLength(I2) << '\n';
+            std::cout << "c1_grid_desc_g_m_n_: " << c1_grid_desc_g_m_n_.GetLength(I0) << ", "
+                      << c1_grid_desc_g_m_n_.GetLength(I1) << ", "
+                      << c1_grid_desc_g_m_n_.GetLength(I2) << '\n';
         }
 
         // pointers
@@ -501,18 +578,23 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         const BDataType* p_b_grid_;
         const B1DataType* p_b1_grid_;
         CDataType* p_c_grid_;
+        typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
 
         // tensor descriptor
         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_;
-        CGridDesc_M_N c_grid_desc_m_n_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
         AGridDesc_G_M_K a_grid_desc_g_m_k_;
         BGridDesc_G_N_K b_grid_desc_g_n_k_;
         B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
-        CGridDesc_G_M_N c_grid_desc_g_m_n_;
-        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
-            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
+        D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
+
+        typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c1_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
 
         // block-to-c-tile map
         typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
@@ -520,9 +602,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         // element-wise op
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
-        AccElementwiseOperation acc_element_op_;
+        C0DEElementwiseOperation c0de_element_op_;
         B1ElementwiseOperation b1_element_op_;
-        CElementwiseOperation c_element_op_;
+        C1DEElementwiseOperation c1de_element_op_;
 
         // check C0 masking and padding
         C0MatrixMask c0_matrix_mask_;
@@ -533,6 +615,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         std::vector<index_t> b_nz_kz_strides_;
         std::vector<index_t> b1_nz_kz_strides_;
         std::vector<index_t> c_mz_gemm1nz_strides_;
+        std::array<std::vector<ck::index_t>, NumD0Tensor> d0s_nl_ns_lengths_strides_;
 
         index_t batch_count_;
         ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
@@ -551,7 +634,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             }
 
             const index_t grid_size =
-                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c1_grid_desc_m_n_) * arg.batch_count_;
 
             // Gemm0_K
             const auto K =
@@ -564,15 +647,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                     GridwiseGemm,
                     ADataType, // TODO: distiguish A/B datatype
                     CDataType,
+                    typename GridwiseGemm::D0sGridPointer,
                     AElementwiseOperation,
                     BElementwiseOperation,
-                    AccElementwiseOperation,
+                    C0DEElementwiseOperation,
                     B1ElementwiseOperation,
-                    CElementwiseOperation,
+                    C1DEElementwiseOperation,
                     DeviceOp::AGridDesc_AK0_M_AK1,
                     DeviceOp::BGridDesc_BK0_N_BK1,
                     DeviceOp::B1GridDesc_BK0_N_BK1,
-                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
                     typename GridwiseGemm::DefaultBlock2CTileMap,
                     ComputeBasePtrOfStridedBatch,
                     C0MatrixMask,
@@ -587,15 +672,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                                               arg.p_b_grid_,
                                               arg.p_b1_grid_,
                                               arg.p_c_grid_,
+                                              arg.p_d0s_grid_,
                                               arg.a_element_op_,
                                               arg.b_element_op_,
-                                              arg.acc_element_op_,
+                                              arg.c0de_element_op_,
                                               arg.b1_element_op_,
-                                              arg.c_element_op_,
+                                              arg.c1de_element_op_,
                                               arg.a_grid_desc_ak0_m_ak1_,
                                               arg.b_grid_desc_bk0_n_bk1_,
                                               arg.b1_grid_desc_bk0_n_bk1_,
-                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
                                               arg.block_2_ctile_map_,
                                               arg.batch_count_,
                                               arg.compute_base_ptr_of_batch_,
@@ -636,7 +723,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         arg.Print();
 #endif
 
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
         {
             return false;
         }
@@ -644,9 +733,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         // TODO ANT: Check if tensor specialization & strides mismatch
 
         // Check if C permute dimension matches GEMM + GEMM shape
-        const index_t c_g       = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
-        const index_t c_m       = arg.c_grid_desc_m_n_.GetLength(I0);
-        const index_t c_gemm1n  = arg.c_grid_desc_m_n_.GetLength(I1);
+        const index_t c_g       = arg.c1_grid_desc_g_m_n_.GetLength(I0); // unpadded
+        const index_t c_m       = arg.c1_grid_desc_m_n_.GetLength(I0);
+        const index_t c_gemm1n  = arg.c1_grid_desc_m_n_.GetLength(I1);
         const index_t a_m       = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
         const index_t b1_gemm1n = arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
 
@@ -692,11 +781,23 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         {
             return false;
         }
+        for(int i = 0; i < NumD0Tensor; i++)
+        {
+            if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 &&
+               arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+            if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1)
+            {
+                return false;
+            }
+        }
 
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
                                            arg.b_grid_desc_bk0_n_bk1_,
                                            arg.b1_grid_desc_bk0_n_bk1_,
-                                           arg.c_grid_desc_m_n_,
+                                           arg.c1_grid_desc_m_n_,
                                            arg.block_2_ctile_map_);
     }
 
@@ -711,8 +812,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         const BDataType* p_b,
         const B1DataType* p_b1,
         CDataType* p_c,
-        const std::array<void*, NumAcc0Bias> p_acc0_biases,
-        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::array<void*, NumD0Tensor> p_acc0_biases,
+        const std::array<void*, NumD1Tensor> p_acc1_biases,
         const std::vector<index_t>& a_gs_ms_ks_lengths,
         const std::vector<index_t>& a_gs_ms_ks_strides,
         const std::vector<index_t>& b_gs_ns_ks_lengths,
@@ -721,17 +822,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
         const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
         const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
-        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
-        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
-        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
             acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
-        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
             acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
         AElementwiseOperation a_element_op,
         BElementwiseOperation b_element_op,
-        AccElementwiseOperation acc_element_op,
+        C0DEElementwiseOperation c0de_element_op,
         B1ElementwiseOperation b1_element_op,
-        CElementwiseOperation c_element_op)
+        C1DEElementwiseOperation c1de_element_op)
     {
         return Argument{p_a,
                         p_b,
@@ -753,9 +854,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                         acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
                         a_element_op,
                         b_element_op,
-                        acc_element_op,
+                        c0de_element_op,
                         b1_element_op,
-                        c_element_op};
+                        c1de_element_op};
     }
 
     static auto MakeInvoker() { return Invoker{}; }
@@ -767,8 +868,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         const void* p_b,
         const void* p_b1,
         void* p_c,
-        const std::array<void*, NumAcc0Bias> p_acc0_biases,
-        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::array<void*, NumD0Tensor> p_acc0_biases,
+        const std::array<void*, NumD1Tensor> p_acc1_biases,
         const std::vector<index_t>& a_gs_ms_ks_lengths,
         const std::vector<index_t>& a_gs_ms_ks_strides,
         const std::vector<index_t>& b_gs_ns_ks_lengths,
@@ -777,17 +878,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
         const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
         const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
-        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
-        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
-        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
             acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
-        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
             acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
         AElementwiseOperation a_element_op,
         BElementwiseOperation b_element_op,
-        AccElementwiseOperation acc_element_op,
+        C0DEElementwiseOperation c0de_element_op,
         B1ElementwiseOperation b1_element_op,
-        CElementwiseOperation c_element_op) override
+        C1DEElementwiseOperation c1de_element_op) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),
@@ -809,9 +910,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
                                           acc1_biases_gs_ms_gemm1ns_strides,
                                           a_element_op,
                                           b_element_op,
-                                          acc_element_op,
+                                          c0de_element_op,
                                           b1_element_op,
-                                          c_element_op);
+                                          c1de_element_op);
     }
 
     // polymorphic

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -62,7 +62,8 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -612,7 +613,9 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -45,74 +45,46 @@ namespace device {
  * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
  *
  */
-template <typename GridwiseGemm,
-          typename FloatAB,
-          typename FloatC,
-          typename AGridDesc_K0_M_K1,
-          typename BGridDesc_K0_N_K1,
-          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
-          typename AElementwiseOperation,
-          typename BElementwiseOperation,
-          typename CElementwiseOperation,
-          typename ComputePtrOffsetOfBatch,
-          typename Block2CTileMap,
-          bool HasMainKBlockLoop>
+template <typename DeviceOp, typename GridwiseGemm, bool HasMainKBlockLoop>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_batched_gemm_xdlops_v2r3(
-            const FloatAB* __restrict__ p_a_grid,
-            const FloatAB* __restrict__ p_b_grid,
-            FloatC* __restrict__ p_c_grid,
-            const index_t batch_count,
-            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
-            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
-            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
-            const AElementwiseOperation a_element_op,
-            const BElementwiseOperation b_element_op,
-            const CElementwiseOperation c_element_op,
-            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
-            const Block2CTileMap block_2_ctile_map)
+        kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     const index_t num_blocks_per_batch =
-        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+        __builtin_amdgcn_readfirstlane(get_grid_size() / karg.Batch);
     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)));
+        static_cast<long_index_t>(karg.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)));
+        static_cast<long_index_t>(karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
     const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
-        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+        static_cast<long_index_t>(karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
 
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
-    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
-                                                  p_b_grid + b_batch_offset,
-                                                  p_c_grid + c_batch_offset,
+    const auto a_grid_desc_k0_m_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeAGridDescriptor_K0_M_K1(
+            karg.M, karg.MPadded, karg.K, karg.K0, karg.StrideA));
+    const auto b_grid_desc_k0_n_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeBGridDescriptor_K0_N_K1(
+            karg.K, karg.N, karg.NPadded, karg.K0, karg.StrideB));
+    const auto c_grid_desc_m_n = amd_wave_read_first_lane(GridwiseGemm::MakeCGridDescriptor_M_N(
+        karg.M, karg.MPadded, karg.N, karg.NPadded, karg.StrideC));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(karg.p_a_grid + a_batch_offset,
+                                                  karg.p_b_grid + b_batch_offset,
+                                                  karg.p_c_grid + c_batch_offset,
                                                   p_shared,
                                                   a_grid_desc_k0_m_k1,
                                                   b_grid_desc_k0_n_k1,
-                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
-                                                  a_element_op,
-                                                  b_element_op,
-                                                  c_element_op,
-                                                  block_2_ctile_map);
+                                                  c_grid_desc_m_n);
 #else
-    ignore = p_a_grid;
-    ignore = p_b_grid;
-    ignore = p_c_grid;
-    ignore = batch_count;
-    ignore = a_grid_desc_k0_m_k1;
-    ignore = b_grid_desc_k0_n_k1;
-    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
-    ignore = a_element_op;
-    ignore = b_element_op;
-    ignore = c_element_op;
-    ignore = compute_ptr_offset_of_batch;
-    ignore = block_2_ctile_map;
+    ignore = karg;
 #endif
 }
 
@@ -170,93 +142,6 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
 
     static constexpr auto K1Number = Number<K1>{};
 
-    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
-    {
-        assert(K % K1 == 0);
-
-        const index_t K0 = K / K1;
-
-        const auto a_grid_desc_m_k = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
-            }
-        }();
-
-        const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
-
-        const auto a_grid_desc_k0_mp_k1 =
-            transform_tensor_descriptor(a_grid_desc_m_k,
-                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
-                                                   make_right_pad_transform(M, PadM)),
-                                        make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-        return a_grid_desc_k0_mp_k1;
-    }
-
-    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
-    {
-        assert(K % K1 == 0);
-
-        const index_t K0 = K / K1;
-
-        const auto b_grid_desc_k_n = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
-            }
-        }();
-
-        const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
-
-        const auto b_grid_desc_k0_np_k1 =
-            transform_tensor_descriptor(b_grid_desc_k_n,
-                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
-                                                   make_right_pad_transform(N, PadN)),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-        return b_grid_desc_k0_np_k1;
-    }
-
-    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
-    {
-        const auto c_grid_desc_m_n = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
-            }
-        }();
-
-        const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
-        const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
-
-        const auto c_grid_desc_mp_np = transform_tensor_descriptor(
-            c_grid_desc_m_n,
-            make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
-            make_tuple(Sequence<0>{}, Sequence<1>{}),
-            make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-        return c_grid_desc_mp_np;
-    }
-
-    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
-    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
-    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
-
     struct ComputePtrOffsetOfStridedBatch
     {
         ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
@@ -288,121 +173,82 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
     };
 
     // GridwiseGemm
-    using GridwiseGemm =
-        GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
-                                                ADataType, // TODO: distinguish A/B datatype
-                                                AccDataType,
-                                                CDataType,
-                                                InMemoryDataOperationEnum::Set,
-                                                AGridDesc_K0_M_K1,
-                                                BGridDesc_K0_N_K1,
-                                                CGridDesc_M_N,
-                                                AElementwiseOperation,
-                                                BElementwiseOperation,
-                                                CElementwiseOperation,
-                                                MPerBlock,
-                                                NPerBlock,
-                                                K0PerBlock,
-                                                MPerXDL,
-                                                NPerXDL,
-                                                K1,
-                                                MXdlPerWave,
-                                                NXdlPerWave,
-                                                ABlockTransferThreadClusterLengths_K0_M_K1,
-                                                ABlockTransferThreadClusterArrangeOrder,
-                                                ABlockTransferSrcAccessOrder,
-                                                ABlockTransferSrcVectorDim,
-                                                ABlockTransferSrcScalarPerVector,
-                                                ABlockTransferDstScalarPerVector_K1,
-                                                false, // AThreadTransferSrcResetCoordinateAfterRun,
-                                                ABlockLdsAddExtraM,
-                                                BBlockTransferThreadClusterLengths_K0_N_K1,
-                                                BBlockTransferThreadClusterArrangeOrder,
-                                                BBlockTransferSrcAccessOrder,
-                                                BBlockTransferSrcVectorDim,
-                                                BBlockTransferSrcScalarPerVector,
-                                                BBlockTransferDstScalarPerVector_K1,
-                                                false, // BThreadTransferSrcResetCoordinateAfterRun,
-                                                BBlockLdsAddExtraN,
-                                                Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
-                                                CThreadTransferSrcDstVectorDim,
-                                                CThreadTransferDstScalarPerVector,
-                                                NumGemmKPrefetchStage,
-                                                LoopSched,
-                                                PipelineVer>;
-
-    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
-        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
-    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpecialization::MNPadding,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        NumGemmKPrefetchStage,
+        LoopSched,
+        PipelineVer>;
+
+    using Problem = typename GridwiseGemm::Problem;
 
     // Argument
-    struct Argument : public BaseArgument
+    struct Argument : public Problem, public BaseArgument
     {
-        Argument(const ADataType* p_a_grid,
-                 const BDataType* p_b_grid,
-                 CDataType* p_c_grid,
-                 index_t M,
-                 index_t N,
-                 index_t K,
-                 index_t StrideA,
-                 index_t StrideB,
-                 index_t StrideC,
+        Argument(const ADataType* p_a_grid_,
+                 const BDataType* p_b_grid_,
+                 CDataType* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
                  index_t BatchStrideA,
                  index_t BatchStrideB,
                  index_t BatchStrideC,
-                 index_t Batch,
-                 index_t M01,
-                 index_t N01,
-                 AElementwiseOperation a_element_op,
-                 BElementwiseOperation b_element_op,
-                 CElementwiseOperation c_element_op)
-            : p_a_grid_{p_a_grid},
-              p_b_grid_{p_b_grid},
-              p_c_grid_{p_c_grid},
-              Batch_(Batch),
-              a_grid_desc_k0_m_k1_{
-                  DeviceBatchedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA)},
-              b_grid_desc_k0_n_k1_{
-                  DeviceBatchedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB)},
-              c_grid_desc_m_n_{DeviceBatchedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC)},
-              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
-              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideC},
-              block_2_ctile_map_{
-                  GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01)},
-              M01_{M01},
-              N01_{N01},
-              a_element_op_{a_element_op},
-              b_element_op_{b_element_op},
-              c_element_op_{c_element_op},
-              kraw_{K}
+                 index_t Batch_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_},
+              Batch(Batch_),
+              compute_ptr_offset_of_batch{BatchStrideA, BatchStrideB, BatchStrideC}
         {
-            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
-                                           b_grid_desc_k0_n_k1_,
-                                           c_grid_desc_m_n_,
-                                           block_2_ctile_map_))
-            {
-                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
-                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
-            }
         }
 
-        //  private:
-        const ADataType* p_a_grid_;
-        const BDataType* p_b_grid_;
-        CDataType* p_c_grid_;
-        index_t Batch_;
-        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
-        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
-        CGridDesc_M_N c_grid_desc_m_n_;
-        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
-        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
-        Block2CTileMap block_2_ctile_map_;
-        index_t M01_;
-        index_t N01_;
-        AElementwiseOperation a_element_op_;
-        BElementwiseOperation b_element_op_;
-        CElementwiseOperation c_element_op_;
-        index_t kraw_;
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        CDataType* p_c_grid;
+        index_t Batch;
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch;
     };
 
     // Invoker
@@ -410,107 +256,39 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
     {
         using Argument = DeviceBatchedGemmXdl::Argument;
 
-        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if DEBUG_LOG
+            if(stream_config.log_level_ > 0)
             {
-                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
-                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
-                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
-
-                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
-                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
-                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
-
-                std::cout << "arg.c_grid_desc_m_n_{" << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
-                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+                karg.Print();
             }
-#endif
 
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
-                                            arg.b_grid_desc_k0_n_k1_,
-                                            arg.c_grid_desc_m_n_,
-                                            arg.block_2_ctile_map_))
+            if(!GridwiseGemm::CheckValidity(karg))
             {
                 throw std::runtime_error(
-                    "wrong! GridwiseBatchedGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext has invalid setting");
             }
 
-            const index_t grid_size =
-                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
-
-            const auto K =
-                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+            auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(karg.M, karg.N);
+            gdx *= karg.Batch;
 
             float ave_time = 0;
 
-            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(karg.K))
             {
-                const auto kernel = kernel_batched_gemm_xdlops_v2r3<
-                    GridwiseGemm,
-                    ADataType, // TODO: distiguish A/B datatype
-                    CDataType,
-                    remove_reference_t<DeviceBatchedGemmXdl::AGridDesc_K0_M_K1>,
-                    remove_reference_t<DeviceBatchedGemmXdl::BGridDesc_K0_N_K1>,
-                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
-                    AElementwiseOperation,
-                    BElementwiseOperation,
-                    CElementwiseOperation,
-                    ComputePtrOffsetOfStridedBatch,
-                    remove_reference_t<Block2CTileMap>,
-                    true>;
-
-                ave_time = launch_and_time_kernel(stream_config,
-                                                  kernel,
-                                                  dim3(grid_size),
-                                                  dim3(BlockSize),
-                                                  0,
-                                                  arg.p_a_grid_,
-                                                  arg.p_b_grid_,
-                                                  arg.p_c_grid_,
-                                                  arg.Batch_,
-                                                  arg.a_grid_desc_k0_m_k1_,
-                                                  arg.b_grid_desc_k0_n_k1_,
-                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
-                                                  arg.a_element_op_,
-                                                  arg.b_element_op_,
-                                                  arg.c_element_op_,
-                                                  arg.compute_ptr_offset_of_batch_,
-                                                  arg.block_2_ctile_map_);
+                const auto kernel =
+                    kernel_batched_gemm_xdlops_v2r3<DeviceBatchedGemmXdl, GridwiseGemm, true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
             }
             else
             {
-                const auto kernel = kernel_batched_gemm_xdlops_v2r3<
-                    GridwiseGemm,
-                    ADataType, // TODO: distiguish A/B datatype
-                    CDataType,
-                    remove_reference_t<DeviceBatchedGemmXdl::AGridDesc_K0_M_K1>,
-                    remove_reference_t<DeviceBatchedGemmXdl::BGridDesc_K0_N_K1>,
-                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
-                    AElementwiseOperation,
-                    BElementwiseOperation,
-                    CElementwiseOperation,
-                    ComputePtrOffsetOfStridedBatch,
-                    remove_reference_t<Block2CTileMap>,
-                    false>;
-
-                ave_time = launch_and_time_kernel(stream_config,
-                                                  kernel,
-                                                  dim3(grid_size),
-                                                  dim3(BlockSize),
-                                                  0,
-                                                  arg.p_a_grid_,
-                                                  arg.p_b_grid_,
-                                                  arg.p_c_grid_,
-                                                  arg.Batch_,
-                                                  arg.a_grid_desc_k0_m_k1_,
-                                                  arg.b_grid_desc_k0_n_k1_,
-                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
-                                                  arg.a_element_op_,
-                                                  arg.b_element_op_,
-                                                  arg.c_element_op_,
-                                                  arg.compute_ptr_offset_of_batch_,
-                                                  arg.block_2_ctile_map_);
+                const auto kernel =
+                    kernel_batched_gemm_xdlops_v2r3<DeviceBatchedGemmXdl, GridwiseGemm, false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
             }
 
             return ave_time;
@@ -530,17 +308,14 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    static bool IsSupportedArgument(const Problem& problem)
     {
-        if(arg.kraw_ % K1 != 0)
+        if(problem.K % K1 != 0)
         {
             return false;
         }
 
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
-                                           arg.b_grid_desc_k0_n_k1_,
-                                           arg.c_grid_desc_m_n_,
-                                           arg.block_2_ctile_map_);
+        return GridwiseGemm::CheckValidity(problem);
     }
 
     // polymorphic
@@ -561,10 +336,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                              index_t BatchStrideA,
                              index_t BatchStrideB,
                              index_t BatchStrideC,
-                             index_t Batch,
-                             AElementwiseOperation a_element_op,
-                             BElementwiseOperation b_element_op,
-                             CElementwiseOperation c_element_op)
+                             index_t Batch)
     {
         return Argument{p_a,
                         p_b,
@@ -578,12 +350,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                         BatchStrideA,
                         BatchStrideB,
                         BatchStrideC,
-                        Batch,
-                        1,
-                        1,
-                        a_element_op,
-                        b_element_op,
-                        c_element_op};
+                        Batch};
     }
 
     static auto MakeInvoker() { return Invoker{}; }
@@ -602,9 +369,9 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                                                       index_t BatchStrideB,
                                                       index_t BatchStrideC,
                                                       index_t Batch,
-                                                      AElementwiseOperation a_element_op,
-                                                      BElementwiseOperation b_element_op,
-                                                      CElementwiseOperation c_element_op) override
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),
@@ -618,12 +385,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                                           BatchStrideA,
                                           BatchStrideB,
                                           BatchStrideC,
-                                          Batch,
-                                          1,
-                                          1,
-                                          a_element_op,
-                                          b_element_op,
-                                          c_element_op);
+                                          Batch);
     }
 
     // polymorphic

include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp

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