Merge branch 'develop' into lwpck-471

Jenkinsfile

@@ -618,9 +618,9 @@ pipeline {
                 stage('Clang Format') {
                     agent{ label rocmnode("nogpu") }
                     environment{
-                        execute_cmd = "find .. -iname \'*.h\' \
-                                -o -iname \'*.hpp\' \
-                                -o -iname \'*.cpp\' \
+                        execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
+                                -o -not -path \'*.git*\' -iname \'*.hpp\' \
+                                -o -not -path \'*.git*\' -iname \'*.cpp\' \
                                 -o -iname \'*.h.in\' \
                                 -o -iname \'*.hpp.in\' \
                                 -o -iname \'*.cpp.in\' \

client_example/14_instance_id/CMakeLists.txt

+add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp)
+target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_operations)

client_example/14_instance_id/batchnorm_fwd_instance_id.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// In the actual application, the instance index and name are usually from the perf db
+static int instance_index = -1;
+static std::string instance_name;
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    bool found          = false;
+    int best_op_index   = -1;
+    float best_ave_time = std::numeric_limits<float>::max();
+
+    // profile device operation instances and save the best performant instance index and instance
+    // name
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            if(ave_time < best_ave_time)
+            {
+                found         = true;
+                best_op_index = i;
+                best_ave_time = ave_time;
+            }
+        }
+    }
+
+    if(found)
+    {
+        instance_index = best_op_index;
+        instance_name  = op_ptrs[instance_index]->GetTypeIdHashCode();
+    };
+
+    // simulate the execution of the operation when the instance index and name are available
+    const auto op_ptrs_2 = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(instance_index >= 0 && instance_index < op_ptrs_2.size())
+    {
+        auto& op_ptr = op_ptrs_2[instance_index];
+
+        if(op_ptr->GetTypeIdHashCode() == instance_name)
+        {
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                            xyStrides,
+                                                            xyStrides,
+                                                            reduceDims,
+                                                            scaleBiasMeanVarLengths,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            x.GetDeviceBuffer(),
+                                                            scale.GetDeviceBuffer(),
+                                                            bias.GetDeviceBuffer(),
+                                                            epsilon,
+                                                            PassThrough{},
+                                                            y.GetDeviceBuffer(),
+                                                            mean.GetDeviceBuffer(),
+                                                            invVariance.GetDeviceBuffer(),
+                                                            averageFactor,
+                                                            nullptr,
+                                                            nullptr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+                SimpleDeviceMem workspace(workspace_sz);
+
+                op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+                float exec_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+                size_t num_bytes = numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                                   numScaleBiasMeanVarElement *
+                                       (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                        sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+                float gb_per_sec = num_bytes / 1.E6 / exec_time;
+
+                std::cout << "Kernel execution time: " << std::setw(10) << exec_time
+                          << " ms,  effective data transfer bandwidth: " << gb_per_sec << " GB/s"
+                          << std::endl;
+            }
+        };
+    }
+
+    return 0;
+}

doc/markdown/dockerhub.md

+## CK docker hub
+
+[Docker hub](https://hub.docker.com/r/rocm/composable_kernel)
+
+## Why do I need this?
+
+To make our lives easier and bring Composable Kernel dependencies together, we recommend using docker images.
+
+## So what is Composable Kernel?
+
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
+
+To get the CK library
+
+```
+git clone https://github.com/ROCmSoftwarePlatform/composable_kernel.git
+```
+
+run a docker container 
+
+```
+docker run                                                            \
+-it                                                                   \
+--privileged                                                          \
+--group-add sudo                                                      \
+-w /root/workspace                                                    \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                         \
+rocm/composable_kernel:ck_ub20.04_rocm5.3_release                     \
+/bin/bash
+```
+
+and build the CK
+
+```
+mkdir build && cd build
+
+# Need to specify target ID, example below is for gfx908 and gfx90a
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+..
+```
+
+and 
+
+```
+make -j examples tests
+```
+
+To run all the test cases including tests and examples run
+
+```
+make test
+```
+
+We can also run specific examples or tests like
+
+```
+./bin/example_gemm_xdl_fp16
+./bin/test_gemm_fp16
+```
+
+For more details visit [CK github repo](https://github.com/ROCmSoftwarePlatform/composable_kernel), [CK examples](https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/example), [even more CK examples](https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/client_example).
+
+## And what is inside?
+
+The docker images have everything you need for running CK including:
+
+* [ROCm](https://www.amd.com/en/graphics/servers-solutions-rocm)
+* [CMake](https://cmake.org/)
+* [Compiler](https://github.com/RadeonOpenCompute/llvm-project)
+
+## Which image is right for me?
+
+Let's take a look at the image naming, for example "ck_ub20.04_rocm5.4_release". The image specs are:
+
+* "ck" - made for running Composable Kernel
+* "ub20.04" - based on Ubuntu 20.04
+* "rocm5.4" - ROCm platform version 5.4
+* "release" - compiler version is release
+
+So just pick the right image for your project dependencies and you're all set.
+
+## DIY starts here
+
+If you need to customize a docker image or just can't stop tinkering, feel free to adjust the [Dockerfile](https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/Dockerfile) for your needs.
+
+## License
+
+CK is released under the MIT [license](https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/LICENSE).

example/09_convnd_fwd/CMakeLists.txt

@@ -8,3 +8,4 @@ add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp6
 add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
 add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
 add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
+

example/09_convnd_fwd/convnd_fwd_dl_common.hpp

@@ -30,6 +30,7 @@ void print_helper_msg()
 template <ck::index_t NDimSpatial,
           typename InDataType,
           typename WeiDataType,
+          typename DsDataType,
           typename OutDataType,
           typename InElementOp,
           typename WeiElementOp,
@@ -46,8 +47,10 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
                              const WeiElementOp& wei_element_op,
                              const OutElementOp& out_element_op)
 {
+    using DDataType = ck::remove_cvref_t<ck::tuple_element_t<0, DsDataType>>;
     Tensor<InDataType> in(in_g_n_c_wis_desc);
     Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<DDataType> bias(out_g_n_k_wos_desc);
     Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
     Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
 
@@ -59,31 +62,38 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
     {
     case 0: break;
     case 1:
-        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
-        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 3});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 3});
+        bias.GenerateTensorValue(GeneratorTensor_2<DDataType>{-2, 3});
         break;
     case 2:
         in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
         wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
         break;
     default:
         in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
-        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{-1});
+        bias.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
     }
 
     DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
     DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(DDataType) * bias.mDesc.GetElementSpaceSize());
     DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
 
     in_device_buf.ToDevice(in.mData.data());
     wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
 
     std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
     std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
     std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
     std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
-    std::array<ck::index_t, NDimSpatial + 3> c_g_n_k_wos_lengths{};
-    std::array<ck::index_t, NDimSpatial + 3> c_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
     std::array<ck::index_t, NDimSpatial> input_left_pads{};
@@ -95,8 +105,10 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
     copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
     copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
     copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
-    copy(out_g_n_k_wos_desc.GetLengths(), c_g_n_k_wos_lengths);
-    copy(out_g_n_k_wos_desc.GetStrides(), c_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
     copy(conv_param.conv_filter_strides_, conv_filter_strides);
     copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
     copy(conv_param.input_left_pads_, input_left_pads);
@@ -105,25 +117,32 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
     // do Conv
     auto conv     = DeviceConvNDFwdInstance{};
     auto invoker  = conv.MakeInvoker();
-    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
-                                      wei_device_buf.GetDeviceBuffer(),
-                                      out_device_buf.GetDeviceBuffer(),
-                                      a_g_n_c_wis_lengths,
-                                      a_g_n_c_wis_strides,
-                                      b_g_k_c_xs_lengths,
-                                      b_g_k_c_xs_strides,
-                                      c_g_n_k_wos_lengths,
-                                      c_g_n_k_wos_strides,
-                                      conv_filter_strides,
-                                      conv_filter_dilations,
-                                      input_left_pads,
-                                      input_right_pads,
-                                      in_element_op,
-                                      wei_element_op,
-                                      out_element_op);
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
 
     if(!conv.IsSupportedArgument(argument))
     {
+        std::cout << "wrong! device_conv with the specified compilation parameters does not "
+                     "support this Conv problem"
+                  << std::endl;
         return true;
     }
 
@@ -139,28 +158,34 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
 
     if(do_verification)
     {
-        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
-                                                                     InDataType,
-                                                                     WeiDataType,
-                                                                     OutDataType,
-                                                                     InElementOp,
-                                                                     WeiElementOp,
-                                                                     OutElementOp>();
-
-        auto ref_invoker  = ref_conv.MakeInvoker();
-        auto ref_argument = ref_conv.MakeArgument(in,
-                                                  wei,
-                                                  out_host,
-                                                  conv_param.conv_filter_strides_,
-                                                  conv_param.conv_filter_dilations_,
-                                                  conv_param.input_left_pads_,
-                                                  conv_param.input_right_pads_,
-                                                  in_element_op,
-                                                  wei_element_op,
-                                                  out_element_op);
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
+            NDimSpatial,
+            InDataType,
+            WeiDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            ck::tensor_operation::element_wise::PassThrough>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+        auto ref_argument =
+            ref_conv.MakeArgument(in,
+                                  wei,
+                                  out_host,
+                                  conv_param.conv_filter_strides_,
+                                  conv_param.conv_filter_dilations_,
+                                  conv_param.input_left_pads_,
+                                  conv_param.input_right_pads_,
+                                  in_element_op,
+                                  wei_element_op,
+                                  ck::tensor_operation::element_wise::PassThrough{});
 
         ref_invoker.Run(ref_argument);
 
+        // cde_elementwise
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), out_host(idx), bias(idx)); });
+
         out_device_buf.FromDevice(out_device.mData.data());
 
         return ck::utils::check_err(

example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp

@@ -3,13 +3,14 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 
 using InDataType  = ck::half_t;
 using WeiDataType = ck::half_t;
 using AccDataType = float;
+using DsDataType  = ck::Tuple<ck::half_t>;
 using OutDataType = ck::half_t;
 
 template <ck::index_t... Is>
@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
 
 using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
-using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
 
 static constexpr auto ConvSpec =
     ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
 
 template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
 // clang-format off
-using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
-// ######|        NDim|     InData|     WeiData|     OutData|     AccData| InLayout| WeiLayout| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
-// ######|     Spatial|       Type|        Type|        Type|        Type|         |          |          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
-// ######|            |           |            |            |            |         |          |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
-// ######|            |           |            |            |            |         |          |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
-         < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
 // clang-format on
 
 #include "run_convnd_fwd_dl_example.inc"

example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp

@@ -3,13 +3,14 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 
 using InDataType  = float;
 using WeiDataType = float;
 using AccDataType = float;
+using DsDataType  = ck::Tuple<float>;
 using OutDataType = float;
 
 template <ck::index_t... Is>
@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
 
 using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
-using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
 
 static constexpr auto ConvSpec =
     ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
 
 template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
 // clang-format off
-using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
-// ######|        NDim|     InData|     WeiData|     OutData|     AccData| InLayout| WeiLayout| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
-// ######|     Spatial|       Type|        Type|        Type|        Type|         |          |          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
-// ######|            |           |            |            |            |         |          |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
-// ######|            |           |            |            |            |         |          |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
-         < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
 // clang-format on
 
 #include "run_convnd_fwd_dl_example.inc"

example/09_convnd_fwd/convnd_fwd_dl_int8.cpp

@@ -3,13 +3,14 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 
 using InDataType  = int8_t;
 using WeiDataType = int8_t;
 using AccDataType = int32_t;
+using DsDataType  = ck::Tuple<int8_t>;
 using OutDataType = int8_t;
 
 template <ck::index_t... Is>
@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
 
 using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
-using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
 
 static constexpr auto ConvSpec =
     ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
 
 template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
 // clang-format off
-using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
-// ######|        NDim|     InData|     WeiData|     OutData|     AccData| InLayout| WeiLayout| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
-// ######|     Spatial|       Type|        Type|        Type|        Type|         |          |          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
-// ######|            |           |            |            |            |         |          |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
-// ######|            |           |            |            |            |         |          |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
-         < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
 // clang-format on
 
 #include "run_convnd_fwd_dl_example.inc"

example/09_convnd_fwd/run_convnd_fwd_dl_example.inc

@@ -61,6 +61,7 @@ bool run_convnd_fwd_dl_example(int argc, char* argv[])
             ndim_spatial_value,
             InDataType,
             WeiDataType,
+            DsDataType,
             OutDataType,
             InElementOp,
             WeiElementOp,

example/44_elementwise_permute/CMakeLists.txt

 add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)

example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_2d.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
+                                                    ck::Tuple<BDataType>,
+                                                    PassThrough,
+                                                    3, // NumDim_M
+                                                    1, // NumDim_N
+                                                    8,
+                                                    8,
+                                                    ck::Sequence<8>,
+                                                    ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        const std::vector<std::size_t>& shape_nchw,
+                        Functor functor)
+{
+    for(std::size_t n = 0; n < shape_nchw[0]; ++n)
+        for(std::size_t c = 0; c < shape_nchw[1]; ++c)
+            for(std::size_t h = 0; h < shape_nchw[2]; ++h)
+                for(std::size_t w = 0; w < shape_nchw[3]; ++w)
+                {
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor(B_nhwc(n, h, w, c), a_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const int N = 120;
+    const int C = 128;
+    const int H = 32;
+    const int W = 1024;
+
+    /**const int N = 120;
+    const int H = 32;
+    const int W = 64;
+
+    const int C = 128;**/
+
+    std::vector<std::size_t> nchw = {N, C, H, W};
+    std::vector<std::size_t> nhwc = {N, H, W, C};
+
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+    // LogRangeAsType<float>(std::cout << "Tensor a  : ", a.mData, ",") << std::endl;
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths{N, H, W, C};
+
+    std::array<ck::index_t, 4> a_strides = {C * H * W, W, 1, H * W};
+    std::array<ck::index_t, 4> b_strides = {H * W * C, W * C, C, 1};
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        // LogRangeAsType<float>(std::cout << "Tensor b  : ", b.mData, ",") << std::endl;
+
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
+            host_b, a, nchw, PassThrough{});
+
+        // LogRangeAsType<float>(std::cout << "Host b  : ", host_b.mData, ",") << std::endl;
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

include/ck/ck.hpp

@@ -25,7 +25,7 @@
 // check GPU target
 #ifdef __HIP_DEVICE_COMPILE__
 #if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
-      defined(__gfx90a__) || defined(__gfx1030__))
+      defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__))
 #error Not supported target
 #endif
 #endif
@@ -38,6 +38,8 @@
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+#elif defined(__gfx1100__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
 #endif
 
 // FMA instruction
@@ -62,6 +64,13 @@
 #define CK_USE_AMD_MFMA_BF16_1K_OP
 #endif
 
+// WMMA instruction
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_WMMA
+#elif defined(__gfx1100__) // for GPU code
+#define CK_USE_AMD_WMMA
+#endif
+
 // buffer load
 #define CK_USE_AMD_BUFFER_LOAD 1
 

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

@@ -5,6 +5,7 @@
 
 #include <cmath>
 #include <string>
+#include <sstream>
 
 #include "ck/stream_config.hpp"
 
@@ -46,6 +47,17 @@ struct BaseOperator
     virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
     virtual std::string GetTypeString() const { return ""; }
 
+    virtual std::string GetTypeIdName() const { return typeid(*this).name(); }
+
+    virtual std::string GetTypeIdHashCode() const
+    {
+        std::ostringstream oss;
+
+        oss << std::hex << typeid(*this).hash_code();
+
+        return oss.str();
+    };
+
     virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
 
     virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const

include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_base.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim_m,
+          index_t NumDim_n,
+          index_t MPerThread,
+          index_t NPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct DeviceElementwise : public DeviceElementwiseBase<InDataTypeTuple,
+                                                        OutDataTypeTuple,
+                                                        ElementwiseOperation,
+                                                        NumDim_m + NumDim_n>
+{
+    static constexpr index_t NumDim = NumDim_m + NumDim_n;
+
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumOutput>{});
+    };
+
+    using InDataTypePointerTuple  = decltype(GenerateInDataTypePointerTuple());
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    template <typename Desc_MN>
+    static auto PadDescriptor_MN_2d(Desc_MN desc_mn,
+                                    index_t gridSize,
+                                    index_t blockSize,
+                                    index_t num_threads_m,
+                                    index_t num_threads_n)
+    {
+        std::ignore               = blockSize;
+        std::ignore               = gridSize;
+        const auto m              = desc_mn.GetLength(I0);
+        const auto n              = desc_mn.GetLength(I1);
+        const index_t loop_step_m = num_threads_m * MPerThread;
+        const index_t loop_step_n = num_threads_n * NPerThread;
+        const auto pad_m          = math::integer_least_multiple(m, loop_step_m) - m;
+        const auto pad_n          = math::integer_least_multiple(n, loop_step_n) - n;
+
+        const auto desc_mn_pad = transform_tensor_descriptor(
+            desc_mn,
+            make_tuple(make_right_pad_transform(m, pad_m), make_right_pad_transform(n, pad_n)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+        return desc_mn_pad;
+    }
+
+    static auto MakeDescriptor_MN(const std::array<index_t, NumDim>& lengths,
+                                  const std::array<index_t, NumDim>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize,
+                                  index_t num_threads_m,
+                                  index_t num_threads_n)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDim_m, 1>::type();
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDim_m, NumDim_m + NumDim_n, 1>::type();
+
+        const auto mLengths = get_container_subset(tupleOfShape, mDimIds);
+        const auto nLengths = get_container_subset(tupleOfShape, nDimIds);
+
+        // merge nd to 2d desc - [s0 * s1 * ...]
+
+        if constexpr(NumDim > 2)
+        {
+            const auto desc_mn = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return PadDescriptor_MN_2d(desc_mn, gridSize, blockSize, num_threads_m, num_threads_n);
+        }
+        else
+            return PadDescriptor_MN_2d(desc, gridSize, blockSize, num_threads_m, num_threads_n);
+    }
+
+    template <index_t TupleSize>
+    static auto GenerateInOutGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple(
+            [&](auto) {
+                if constexpr(NumDim > 2)
+                {
+                    return MakeDescriptor_MN({1, 1}, {1, 1}, 1, 1, 1, 1);
+                }
+                else
+                {
+                    return MakeDescriptor_MN({1}, {1}, 1, 1, 1, 1);
+                };
+            },
+            Number<TupleSize>{});
+    };
+
+    using OutGrid2dDescTuple = decltype(GenerateInOutGrid2dDescTuple(Number<NumOutput>{}));
+    using InGrid2dDescTuple  = decltype(GenerateInOutGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridwiseElementwise = GridwiseElementwise_2D<InGrid2dDescTuple,
+                                                       OutGrid2dDescTuple,
+                                                       InDataTypePointerTuple,
+                                                       OutDataTypePointerTuple,
+                                                       ElementwiseOperation,
+                                                       MPerThread,
+                                                       NPerThread,
+                                                       InScalarPerVectorSeq,
+                                                       OutScalarPerVectorSeq>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+
+            : lengths_(lengths),
+              inStridesArray_(inStridesArray),
+              outStridesArray_(outStridesArray),
+              elementwise_op_(elementwise_op),
+              blockSize_(256),
+              gridSize_(120), // FIXME - Calculate the grid size by number of CU in the future
+              num_threads_m_((gridSize_ * blockSize_) / 16),
+              num_threads_n_(16)
+        {
+            static_assert(NumDim_m > 0, "");
+            static_assert(NumDim_n > 0, "");
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+                    return static_cast<DataType*>(out_dev_buffers[I.value]);
+                },
+                Number<NumOutput>{});
+
+            in_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(lengths,
+                                             inStridesArray[I.value],
+                                             gridSize_,
+                                             blockSize_,
+                                             num_threads_m_,
+                                             num_threads_n_);
+                },
+                Number<NumInput>{});
+
+            out_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(lengths,
+                                             outStridesArray[I.value],
+                                             gridSize_,
+                                             blockSize_,
+                                             num_threads_m_,
+                                             num_threads_n_);
+                },
+                Number<NumOutput>{});
+        }
+
+        InDataTypePointerTuple in_dev_buffers_;
+        OutDataTypePointerTuple out_dev_buffers_;
+        InGrid2dDescTuple in_grid_2d_desc_tuple_;
+        OutGrid2dDescTuple out_grid_2d_desc_tuple_;
+
+        std::array<index_t, NumDim> lengths_;
+        std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
+        std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
+
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+        index_t gridSize_;
+        index_t num_threads_m_;
+        index_t num_threads_n_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel = kernel_elementwise_2d<GridwiseElementwise,
+                                                      InGrid2dDescTuple,
+                                                      OutGrid2dDescTuple,
+                                                      InDataTypePointerTuple,
+                                                      OutDataTypePointerTuple,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(arg.gridSize_),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        arg.in_grid_2d_desc_tuple_,
+                                                        arg.out_grid_2d_desc_tuple_,
+                                                        arg.in_dev_buffers_,
+                                                        arg.out_dev_buffers_,
+                                                        arg.elementwise_op_,
+                                                        arg.num_threads_m_,
+                                                        arg.num_threads_n_);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg == nullptr)
+            return false;
+
+        if(pArg->lengths_.back() % MPerThread != 0)
+            return false;
+
+        auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
+                                          const std::array<index_t, NumDim>& strides,
+                                          index_t scalarPerVector,
+                                          index_t vectorDim) {
+            if(strides[vectorDim] == 1 &&
+               (lengths[vectorDim] % scalarPerVector == 0 ||
+                lengths[vectorDim] % scalarPerVector == lengths[vectorDim]))
+            {
+                return true;
+            }
+            if(strides[vectorDim] != 1 && scalarPerVector == strides[vectorDim])
+            {
+                return true;
+            }
+            return false;
+        };
+
+        bool valid = true;
+        static_for<0, NumInput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->inStridesArray_[I.value],
+                                       InScalarPerVectorSeq::At(I),
+                                       NumDim_m - 1))
+                valid = false;
+        });
+
+        static_for<0, NumOutput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->outStridesArray_[I.value],
+                                       OutScalarPerVectorSeq::At(I),
+                                       NumDim - 1))
+                valid = false;
+        });
+
+        return valid;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          outStridesArray,
+                                          in_dev_buffers,
+                                          out_dev_buffers,
+                                          elementwise_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_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 Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          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_grouped_conv_fwd_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 AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            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 Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_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);
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    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 + c_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 = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = batch_count;
+    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;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetEPtrOffset(0);
+#endif
+}
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          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>
+struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            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>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            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>{}));
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        MakeAGridDescriptor_AK0_M_AK1<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_AK0_M_AK1,
+                                         BGridDesc_BK0_N_BK1,
+                                         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_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using CGridDesc_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,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   e_g_n_k_wos_lengths,
+                                                                   e_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_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_{},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, e_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << e_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        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_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config)
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK 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.num_group_;
+
+            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;
+
+                const auto kernel = kernel_grouped_conv_fwd_dl_multiple_d<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    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::CGridDesc_M0_M10_M11_N0_N10_N11,
+                    DefaultBlock2CTileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    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.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_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.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            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>{});
+            }
+            return 0;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -373,7 +373,8 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
               N01_{N01},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              c_element_op_{c_element_op}
+              c_element_op_{c_element_op},
+              kraw_{K}
         {
             if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
                                            b_grid_desc_k0_n_k1_,
@@ -401,6 +402,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         CElementwiseOperation c_element_op_;
+        index_t kraw_;
     };
 
     // Invoker
@@ -410,6 +412,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
+#if 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) << ", "
@@ -422,6 +425,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                 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;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,
@@ -528,6 +532,11 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(arg.kraw_ % 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_,

include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp

@@ -549,6 +549,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             float ave_time = 0;
             for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
             {
+#if 0
                 {
                     std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
                               << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
@@ -581,6 +582,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                               << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
                               << " ) " << std::endl;
                 }
+#endif
 
                 if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                 arg.b_grid_desc_k0_n_k1_container_[i],

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

@@ -265,7 +265,8 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
               N01_{N01},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              c_element_op_{c_element_op}
+              c_element_op_{c_element_op},
+              kraw_{K}
         {
             a_grid_desc_k0_m_k1_ = DeviceGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
             b_grid_desc_k0_n_k1_ = DeviceGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
@@ -299,6 +300,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         CElementwiseOperation c_element_op_;
+        index_t kraw_;
     };
 
     // Invoker
@@ -443,6 +445,11 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
             return false;
         }
 
+        if(arg.kraw_ % 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_,

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

@@ -422,7 +422,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
               block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              c_element_op_{c_element_op}
+              c_element_op_{c_element_op},
+              kraw_{KRaw}
         {
             if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
                                            b_grid_desc_bk0_n_bk1_,
@@ -448,6 +449,7 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         CElementwiseOperation c_element_op_;
+        index_t kraw_;
     };
 
     // Invoker
@@ -578,6 +580,15 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
             return false;
         }
 
+        if((arg.kraw_ % AK1 != 0 || arg.kraw_ % BK1 != 0) &&
+           !(GemmSpec == GemmSpecialization::MKPadding ||
+             GemmSpec == GemmSpecialization::NKPadding ||
+             GemmSpec == GemmSpecialization::MNKPadding ||
+             GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
                                            arg.b_grid_desc_bk0_n_bk1_,
                                            arg.c_grid_desc_m_n_,