merge develop

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

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

include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp

@@ -871,7 +871,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
             << MPerBlock << ", "
             << NPerBlock << ", "
             << K0PerBlock << ", "
-            << getConvFwdSpecializationStr(ConvForwardSpecialization)
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
             << ">";
         // clang-format on
 

include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp

@@ -711,7 +711,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             << MPerBlock << ", "
             << NPerBlock << ", "
             << K0PerBlock << ", "
-            << getConvFwdSpecializationStr(ConvForwardSpecialization)
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
             << ">";
         // clang-format on
 

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

@@ -996,22 +996,46 @@ struct DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
                     0,
                     arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
                         sizeof(CDataType)));
+                float elapsed_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.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.block_2_ctile_map_);
+
+                hipGetErrorString(hipMemset(
+                    arg.p_c_grid_,
+                    0,
+                    arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(CDataType)));
+
+                launch_and_time_kernel(StreamConfig{nullptr, false},
+                                       kernel,
+                                       dim3(grid_size),
+                                       dim3(BlockSize),
+                                       0,
+                                       arg.p_a_grid_,
+                                       arg.p_b_grid_,
+                                       arg.p_c_grid_,
+                                       arg.a_grid_desc_kbatch_k0_m_k1_,
+                                       arg.b_grid_desc_kbatch_k0_n_k1_,
+                                       arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                       arg.a_element_op_,
+                                       arg.b_element_op_,
+                                       arg.c_element_op_,
+                                       arg.block_2_ctile_map_);
 
-                return 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.a_grid_desc_kbatch_k0_m_k1_,
-                                              arg.b_grid_desc_kbatch_k0_n_k1_,
-                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                              arg.a_element_op_,
-                                              arg.b_element_op_,
-                                              arg.c_element_op_,
-                                              arg.block_2_ctile_map_);
+                return elapsed_time;
             };
 
             // run kernel for bf16 with splitk
@@ -1022,21 +1046,46 @@ struct DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
                     arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
                         sizeof(AccDataType)));
 
-                return launch_and_time_kernel(stream_config,
-                                              kernel,
-                                              dim3(grid_size),
-                                              dim3(BlockSize),
-                                              0,
-                                              arg.p_a_grid_,
-                                              arg.p_b_grid_,
-                                              static_cast<AccDataType*>(arg.p_workspace_),
-                                              arg.a_grid_desc_kbatch_k0_m_k1_,
-                                              arg.b_grid_desc_kbatch_k0_n_k1_,
-                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                              arg.a_element_op_,
-                                              arg.b_element_op_,
-                                              arg.c_element_op_,
-                                              arg.block_2_ctile_map_);
+                float elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           static_cast<AccDataType*>(arg.p_workspace_),
+                                           arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.block_2_ctile_map_);
+
+                hipGetErrorString(hipMemset(
+                    arg.p_workspace_,
+                    0,
+                    arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(AccDataType)));
+
+                launch_and_time_kernel(StreamConfig{nullptr, false},
+                                       kernel,
+                                       dim3(grid_size),
+                                       dim3(BlockSize),
+                                       0,
+                                       arg.p_a_grid_,
+                                       arg.p_b_grid_,
+                                       static_cast<AccDataType*>(arg.p_workspace_),
+                                       arg.a_grid_desc_kbatch_k0_m_k1_,
+                                       arg.b_grid_desc_kbatch_k0_n_k1_,
+                                       arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                       arg.a_element_op_,
+                                       arg.b_element_op_,
+                                       arg.c_element_op_,
+                                       arg.block_2_ctile_map_);
+
+                return elapsed_time;
             };
 
             // kernel for type conversion
@@ -1210,6 +1259,20 @@ struct DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NumDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
         // vector load A/B matrix from global memory
         if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
              arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
@@ -1334,6 +1397,12 @@ struct DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
             << NPerBlock << ", "
             << K0PerBlock
             << ">";
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0){
+
+            str << " Filter1x1Stride1Pad0";
+        }
+
         // clang-format on
 
         return str.str();

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

@@ -1033,7 +1033,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             << MPerBlock << ", "
             << NPerBlock << ", "
             << K0PerBlock << ", "
-            << getConvFwdSpecializationStr(ConvForwardSpecialization)
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
             << ">";
         // clang-format on
 

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

@@ -2,42 +2,62 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
+
 #include <iostream>
 #include <vector>
 
-#include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
 struct DeviceGemm : public BaseOperator
 {
-    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
-                                                              const void* p_b,
-                                                              void* p_c,
-                                                              ck::index_t M,
-                                                              ck::index_t N,
-                                                              ck::index_t K,
-                                                              ck::index_t StrideA,
-                                                              ck::index_t StrideB,
-                                                              ck::index_t StrideC,
-                                                              AElementwiseOperation a_element_op,
-                                                              BElementwiseOperation b_element_op,
-                                                              CElementwiseOperation c_element_op,
-                                                              ck::index_t KBatch = 1) = 0;
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
-using DeviceGemmPtr = std::unique_ptr<
-    DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+using DeviceGemmPtr = std::unique_ptr<DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CElementwiseOperation>>;
 
 } // namespace device
 } // namespace tensor_operation

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

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef DEVICE_GEMM_BIAS_ACTIVATION_ADD_HPP
-#define DEVICE_GEMM_BIAS_ACTIVATION_ADD_HPP
-
-#include <iostream>
-#include "device_base.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-
-template <typename AElementwiseOperation,
-          typename BElementwiseOperation,
-          typename CElementwiseOperation>
-struct DeviceGemmBiasActivationAdd : public BaseOperator
-{
-    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
-                                                              const void* p_b,
-                                                              void* p_c,
-                                                              const void* p_c0,
-                                                              const void* p_c1,
-                                                              ck::index_t M,
-                                                              ck::index_t N,
-                                                              ck::index_t K,
-                                                              ck::index_t StrideA,
-                                                              ck::index_t StrideB,
-                                                              ck::index_t StrideC,
-                                                              ck::index_t StrideC1,
-                                                              AElementwiseOperation a_element_op,
-                                                              BElementwiseOperation b_element_op,
-                                                              CElementwiseOperation c_element_op,
-                                                              ck::index_t KBatch = 1) = 0;
-
-    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
-};
-
-template <typename AElementwiseOperation,
-          typename BElementwiseOperation,
-          typename CElementwiseOperation>
-using DeviceGemmBiasActivationAddPtr =
-    std::unique_ptr<DeviceGemmBiasActivationAdd<AElementwiseOperation,
-                                                BElementwiseOperation,
-                                                CElementwiseOperation>>;
-
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck
-#endif

include/ck/tensor_operation/gpu/device/device_gemm_bias.hpp → include/ck/tensor_operation/gpu/device/device_gemm_bias_c_permute.hpp

@@ -3,33 +3,45 @@
 
 #pragma once
 
-#include <iostream>
+#include <array>
 
-#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "device_base.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
+struct DEGridDesc_M0_M1_M2_N0_N1
+{
+    ck::index_t M0_, M1_, M2_, N0_, N1_;
+    ck::index_t stride_M0_, stride_M1_, stride_M2_, stride_N0_, stride_N1_;
+};
+
+// input : A[M, K], B[K, N],
+// input : D[M, N], ...
+// output : E[M, N]
+// C = a_op(A) * b_op(B)
+// E = cde_op(C, D)
 template <typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename CElementwiseOperation>
-struct DeviceGemmBias : public BaseOperator
+          typename CDEElementwiseOperation>
+struct DeviceGemmBiasCPermute : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_a,
                         const void* p_b,
-                        const void* p_bias,
-                        void* p_c,
+                        const void* p_d,
+                        void* p_e,
                         ck::index_t M,
                         ck::index_t N,
                         ck::index_t K,
                         ck::index_t StrideA,
                         ck::index_t StrideB,
-                        ck::index_t StrideC,
+                        DEGridDesc_M0_M1_M2_N0_N1 d_gride_desc,
+                        DEGridDesc_M0_M1_M2_N0_N1 e_gride_desc,
                         AElementwiseOperation a_element_op,
                         BElementwiseOperation b_element_op,
-                        CElementwiseOperation c_element_op) = 0;
+                        CDEElementwiseOperation cde_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
@@ -37,8 +49,8 @@ struct DeviceGemmBias : public BaseOperator
 template <typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
-using DeviceGemmBiasPtr = std::unique_ptr<
-    DeviceGemmBias<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+using DeviceGemmBiasCPermutePtr = std::unique_ptr<
+    DeviceGemmBiasCPermute<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
 
 } // namespace device
 } // namespace tensor_operation

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

@@ -64,8 +64,16 @@ template <
             is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
             is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
         bool> = false>
-struct DeviceGemmDl
-    : public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
+struct DeviceGemmDl : public DeviceGemm<ALayout,
+                                        BLayout,
+                                        CLayout,
+                                        ADataType,
+                                        BDataType,
+                                        CDataType,
+                                        AElementwiseOperation,
+                                        BElementwiseOperation,
+                                        CElementwiseOperation>
+
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
@@ -534,8 +542,7 @@ struct DeviceGemmDl
                                                       index_t StrideC,
                                                       AElementwiseOperation a_element_op,
                                                       BElementwiseOperation b_element_op,
-                                                      CElementwiseOperation c_element_op,
-                                                      index_t /* KBatch */ = 1) override
+                                                      CElementwiseOperation c_element_op) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),

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

@@ -11,17 +11,28 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-// input : A[M, K], B[K, N],
-// input : D0[M, N], D1[M, N], ...
-// output : E[M, N]
-// C = a_op(A) * b_op(B)
-// E = cde_op(C, D0, D1, ...)
-template <ck::index_t NumDTensor,
+// GEMM:
+//   input : A[M, K], B[K, N],
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CDEElementwiseOperation>
 struct DeviceGemmMultipleD : public BaseOperator
 {
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_a,
                         const void* p_b,
@@ -41,14 +52,26 @@ struct DeviceGemmMultipleD : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <ck::index_t NumDTensor,
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename CElementwiseOperation>
-using DeviceGemmMultipleDPtr = std::unique_ptr<DeviceGemmMultipleD<NumDTensor,
+          typename CDEElementwiseOperation>
+using DeviceGemmMultipleDPtr = std::unique_ptr<DeviceGemmMultipleD<ALayout,
+                                                                   BLayout,
+                                                                   DELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   DsDataType,
+                                                                   EDataType,
                                                                    AElementwiseOperation,
                                                                    BElementwiseOperation,
-                                                                   CElementwiseOperation>>;
+                                                                   CDEElementwiseOperation>>;
 
 } // namespace device
 } // namespace tensor_operation

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

@@ -2,13 +2,16 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
+
 #include <iostream>
+
 #include "device_base.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
+// FIXME: DeviceGemmReduce type need to well define the problem
 template <ck::index_t NumDTensor, ck::index_t NumReduce>
 struct DeviceGemmReduce : public BaseOperator
 {

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

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
+
 #include <iostream>
 #include <vector>
 
@@ -11,7 +12,13 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
 struct DeviceGemmSplitK : public BaseOperator
@@ -33,11 +40,24 @@ struct DeviceGemmSplitK : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
-using DeviceGemmSplitKPtr = std::unique_ptr<
-    DeviceGemmSplitK<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
+                                                             BLayout,
+                                                             CLayout,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementwiseOperation,
+                                                             BElementwiseOperation,
+                                                             CElementwiseOperation>>;
 
 } // namespace device
 } // namespace tensor_operation

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

@@ -57,8 +57,15 @@ template <typename ADataType,
           ck::index_t CThreadTransferSrcDstVectorDim,
           ck::index_t CThreadTransferDstScalarPerVector,
           ck::index_t NumPrefetch = 1>
-struct DeviceGemmXdl
-    : public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
+struct DeviceGemmXdl : public DeviceGemm<ALayout,
+                                         BLayout,
+                                         CLayout,
+                                         ADataType,
+                                         BDataType,
+                                         CDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CElementwiseOperation>
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
@@ -487,8 +494,7 @@ struct DeviceGemmXdl
                                                       index_t StrideC,
                                                       AElementwiseOperation a_element_op,
                                                       BElementwiseOperation b_element_op,
-                                                      CElementwiseOperation c_element_op,
-                                                      index_t /* KBatch */ = 1) override
+                                                      CElementwiseOperation c_element_op) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),