adding conv_xdl_bias_relu_add example

example/3_conv_xdl/conv_xdl.cpp

@@ -11,7 +11,6 @@
 #include "host_tensor_generator.hpp"
 #include "device_tensor.hpp"
 #include "tensor_layout.hpp"
-#include "host_conv.hpp"
 #include "device_conv_fwd_xdl.hpp"
 #include "device_conv_fwd_xdl_nhwc_kyxc_nhwk.hpp"
 

example/4_conv_xdl_bias_relu_add/README.md

+# Instructions for ```gemm_xdl_bias_add``` Example
+
+## Docker script
+```bash
+docker run                                                                   \
+-it                                                                          \
+--rm                                                                         \
+--privileged                                                                 \
+--group-add sudo                                                             \
+-w /root/workspace                                                           \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                                \
+rocm/tensorflow:rocm4.3.1-tf2.6-dev                                          \
+/bin/bash
+```
+
+## Build ```gemm_xdl_bias_add```
+```bash
+mkdir build && cd build
+```
+
+```bash
+# Need to specify target ID, example below is gfx908
+cmake                                                                  \
+-D BUILD_DEV=OFF                                                       \
+-D CMAKE_BUILD_TYPE=Release                                            \
+-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 "   \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                              \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                         \
+..
+```
+
+```bash
+ make -j gemm_xdl_bias_add
+```
+
+## Run ```gemm_xdl_bias_add```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
+./example/gemm_xdl_bias_add 0 1 5 3840 4096 4096 4096 4096 4096
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+arg.c0_grid_desc_m_n_{ 3840, 4096}
+arg.c1_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s
+```

example/4_conv_xdl_bias_relu_add/conv_xdl_bias_relu_add.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce.hpp"
+#include "example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce_nhwc_kyxc_nhwk.hpp"
+
+struct PassThrough
+{
+    template <typename T>
+    __host__ __device__ constexpr T operator()(T v) const
+    {
+        return v;
+    }
+};
+
+struct Relu
+{
+    template <typename T>
+    __host__ __device__ constexpr T operator()(T v) const
+    {
+        T tmp = 0.1 * v;
+        return tmp > 0 ? tmp : 0;
+    }
+};
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using WeiLayout = ck::tensor_layout::convolution::KYXC;
+using OutLayout = ck::tensor_layout::convolution::NHWK;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = Relu;
+
+using DeviceConvFwdInstance =
+    // clang-format off
+    //############################################|    NDim|     InData|     WeiData|     OutData|     AccData|       In|       Wei|       Out|          In|         Wei|           Out| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
+    //############################################| Spatial|       Type|        Type|        Type|        Type|   Layout|    Layout|    Layout| Elementwise| Elementwise|   Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| SrcDstVectorDim|       DstScalar| AddExtraM| AddExtraN|
+    //############################################|        |           |            |            |            |         |          |          |   Operation|   Operation|     Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|                |       PerVector|          |          |
+    //############################################|        |           |            |            |            |         |          |          |            |            |              |      |      |      |      |   |     |     |     |     |                |                |               |               |               |               |               |                |                |               |               |              |               |               |                |                |          |          |
+    ck::tensor_operation::device::DeviceConvFwdXdl_two_extra_source_reduce<       2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp,   256,   128,   256,     4,  8,   32,   32,    2,    4,      S<1, 2, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      S<1, 4, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,               7,               1,      true,      true>;
+// clang-format on
+
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp>
+void host_verify(const Tensor<TIn>& in,
+                 const Tensor<TWei>& wei,
+                 Tensor<TOut>& out,
+                 const std::vector<ck::index_t>& conv_strides,
+                 const std::vector<ck::index_t>& conv_dilations,
+                 const std::vector<ck::index_t>& in_left_pads,
+                 const std::vector<ck::index_t>&,
+                 const InElementOp& in_element_op,
+                 const WeiElementOp& wei_element_op,
+                 const OutElementOp& out_element_op)
+{
+    auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+        double v = 0;
+        for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
+        {
+            for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
+            {
+                int hi = ho * conv_strides[0] + y * conv_dilations[0] - in_left_pads[0];
+                for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
+                {
+                    int wi = wo * conv_strides[1] + x * conv_dilations[1] - in_left_pads[1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        v += in_element_op(static_cast<const double>(in(n, c, hi, wi))) *
+                             wei_element_op(static_cast<const double>(wei(k, c, y, x)));
+                    }
+                }
+            }
+        }
+        out(n, k, ho, wo) = out_element_op(v);
+    };
+
+    make_ParallelTensorFunctor(f_nchw,
+                               out.mDesc.GetLengths()[0],
+                               out.mDesc.GetLengths()[1],
+                               out.mDesc.GetLengths()[2],
+                               out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+}
+
+int main(int argc, char* argv[])
+{
+    if(argc != 4)
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\n");
+        exit(0);
+    }
+
+    const bool do_verification = std::stoi(argv[1]);
+    const int init_method      = std::stoi(argv[2]);
+    const int nrepeat          = std::stoi(argv[3]);
+
+    // Conv shape
+    const ck::index_t N               = 128;
+    const ck::index_t K               = 256;
+    const ck::index_t C               = 192;
+    const ck::index_t Y               = 3;
+    const ck::index_t X               = 3;
+    const ck::index_t Hi              = 71;
+    const ck::index_t Wi              = 71;
+    const ck::index_t conv_stride_h   = 2;
+    const ck::index_t conv_stride_w   = 2;
+    const ck::index_t conv_dilation_h = 1;
+    const ck::index_t conv_dilation_w = 1;
+    const ck::index_t in_left_pad_h   = 1;
+    const ck::index_t in_left_pad_w   = 1;
+    const ck::index_t in_right_pad_h  = 1;
+    const ck::index_t in_right_pad_w  = 1;
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
+    const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
+    const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+    const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+    // tensor layout
+    auto f_host_tensor_descriptor = [](std::size_t N_,
+                                       std::size_t C_,
+                                       std::size_t H,
+                                       std::size_t W,
+                                       auto layout) {
+        if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                     ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                     ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                        std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
+        }
+        else if constexpr(ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::NHWC>::value ||
+                          ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::KYXC>::value ||
+                          ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::NHWK>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                        std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
+        }
+    };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+
+    // do GEMM
+    auto conv     = DeviceConvFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                      static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                      static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                      N,
+                                      K,
+                                      C,
+                                      std::vector<ck::index_t>{{Hi, Wi}},
+                                      std::vector<ck::index_t>{{Y, X}},
+                                      std::vector<ck::index_t>{{Ho, Wo}},
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float ave_time = invoker.Run(argument, nrepeat);
+
+    std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+    std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
+                            sizeof(WeiDataType) * (K * C * Y * X) +
+                            sizeof(OutDataType) * (N * K * Ho * Wo);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        host_verify(in_n_c_hi_wi,
+                    wei_k_c_y_x,
+                    out_n_k_ho_wo_host_result,
+                    conv_filter_strides,
+                    conv_filter_dilations,
+                    input_left_pads,
+                    input_right_pads,
+                    InElementOp{},
+                    WeiElementOp{},
+                    OutElementOp{});
+
+        out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+        check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
+    }
+}

example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_nhwc_kyxc_nhwk_two_extra_source_reduce.hpp

+#ifndef DEVICE_CONV_FWD_XDL_NHWC_KYXC_NHWK_HPP
+#define DEVICE_CONV_FWD_XDL_NHWC_KYXC_NHWK_HPP
+
+#include <iostream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r3.hpp"
+#include "example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// specialization for 2D conv: in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          bool ABlockLdsAddExtraM,
+          bool BBlockLdsAddExtraN>
+struct DeviceConvFwdXdl_two_extra_source_reduce<
+    2,                                        // ck::index_t NDimSpatial,
+    InDataType,                               // typename InDataType,
+    WeiDataType,                              // typename WeiDataType,
+    OutDataType,                              // typename OutDataType,
+    AccDataType,                              // typename AccDataType,
+    ck::tensor_layout::convolution::NHWC,     // typename InLayout,
+    ck::tensor_layout::convolution::KYXC,     // typename WeiLayout,
+    ck::tensor_layout::convolution::NHWK,     // typename OutLayout,
+    InElementwiseOperation,                   // typename InElementwiseOperation,
+    WeiElementwiseOperation,                  // typename WeiElementwiseOperation,
+    OutElementwiseOperation,                  // typename OutElementwiseOperation,
+    BlockSize,                                // ck::index_t BlockSize,
+    MPerBlock,                                // ck::index_t MPerBlock,
+    NPerBlock,                                // ck::index_t NPerBlock,
+    K0PerBlock,                               // ck::index_t K0PerBlock,
+    K1,                                       // ck::index_t K1,
+    MPerXDL,                                  // ck::index_t MPerXDL,
+    NPerXDL,                                  // ck::index_t NPerXDL,
+    MXdlPerWave,                              // ck::index_t MXdlPerWave,
+    NXdlPerWave,                              // ck::index_t NXdlPerWave,
+    ABlockTransferThreadSliceLengths_K0_M_K1, // typename ABlockTransferThreadSliceLengths_K0_M_K1,
+    ABlockTransferThreadClusterLengths_K0_M_K1, // typename
+                                                // ABlockTransferThreadClusterLengths_K0_M_K1,
+    ABlockTransferThreadClusterArrangeOrder,    // typename ABlockTransferThreadClusterArrangeOrder,
+    ABlockTransferSrcAccessOrder,               // typename ABlockTransferSrcAccessOrder,
+    ABlockTransferSrcVectorDim,                 // ck::index_t ABlockTransferSrcVectorDim,
+    ABlockTransferSrcScalarPerVector,           // ck::index_t ABlockTransferSrcScalarPerVector,
+    ABlockTransferDstScalarPerVector_K1,        // ck::index_t ABlockTransferDstScalarPerVector_K1,
+    BBlockTransferThreadSliceLengths_K0_N_K1, // typename BBlockTransferThreadSliceLengths_K0_N_K1,
+    BBlockTransferThreadClusterLengths_K0_N_K1, // typename
+                                                // BBlockTransferThreadClusterLengths_K0_N_K1,
+    BBlockTransferThreadClusterArrangeOrder,    // typename BBlockTransferThreadClusterArrangeOrder,
+    BBlockTransferSrcAccessOrder,               // typename BBlockTransferSrcAccessOrder,
+    BBlockTransferSrcVectorDim,                 // ck::index_t BBlockTransferSrcVectorDim,
+    BBlockTransferSrcScalarPerVector,           // ck::index_t BBlockTransferSrcScalarPerVector,
+    BBlockTransferDstScalarPerVector_K1,        // ck::index_t BBlockTransferDstScalarPerVector_K1,
+    CThreadTransferSrcDstVectorDim,             // ck::index_t CThreadTransferSrcDstVectorDim,
+    CThreadTransferDstScalarPerVector,          // ck::index_t CThreadTransferDstScalarPerVector,
+    ABlockLdsAddExtraM,                         // bool ABlockLdsAddExtraM,
+    BBlockLdsAddExtraN                          // bool BBlockLdsAddExtraN>
+    > : public BaseOperator
+{
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+        const index_t GemmK    = Y * X * C;
+
+        const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
+
+        assert(GemmK % GemmK1Number == 0);
+
+        const index_t GemmK0 = GemmK / GemmK1Number;
+
+        // A: input tensor
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+            in_n_hip_wip_c_grid_desc,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_gemmk_gemmmraw_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                   make_merge_transform(make_tuple(N, Ho, Wo))),
+                                        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk_gemmmraw_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmMRaw)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+            transform_tensor_descriptor(in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                                        make_tuple(make_pass_through_transform(GemmK0),
+                                                   make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmK1Number)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // B: weight tensor
+        const auto wei_k_yxc_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+        const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+            wei_k_yxc_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+        const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            wei_gemmk_gemmn_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmN)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        // C: output tensor
+        const auto out_nhowo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+        const auto out_gemmmraw_gemmn_grid_desc = transform_tensor_descriptor(
+            out_nhowo_k_grid_desc,
+            make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_gemmm_gemmn_grid_desc =
+            transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                        make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmN)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                          wei_gemmk0_gemmn_gemmk1_grid_desc,
+                          out_gemmm_gemmn_grid_desc);
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // TODO remove these hacks
+    static constexpr auto a_k0_m_k1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{},   // 0+: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 1+: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}),  // 2+: K1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{},   // 0-: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 1-: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto b_k0_n_k1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: N
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: K1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0-: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: N
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
+
+    static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadSliceLengths_K0_M_K1,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        BBlockTransferThreadSliceLengths_K0_N_K1,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false,                            // BThreadTransferSrcResetCoordinateAfterRun,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        decltype(a_k0_m_k1_grid_step_hacks),                   //  AGridStepHacks,
+        decltype(b_k0_n_k1_grid_step_hacks),                   //  BGridStepHacks,
+        decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),   //  CGridStepHacks,
+        decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), //  AGridMoveSliceWindowStepHacks,
+        decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), //  BGridMoveSliceWindowStepHacks,
+        false,                                                 // CAccessOrderMRepeatNRepeat,
+        ABlockLdsAddExtraM,
+        BBlockLdsAddExtraN>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+
+    using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+            const auto descs = DeviceConvFwdXdl::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+                N,
+                K,
+                C,
+                input_spatial_lengths,
+                filter_spatial_lengths,
+                output_spatial_lengths,
+                conv_filter_strides,
+                conv_filter_dilations,
+                input_left_pads,
+                input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceConvFwdXdl::Argument;
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceConvFwdXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceConvFwdXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceConvFwdXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceConvFwdXdl::Block2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceConvFwdXdl::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceConvFwdXdl::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceConvFwdXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceConvFwdXdl::Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce.hpp

+#ifndef DEVICE_CONV_FWD_XDL_TWO_EXTRA_SOURCE_REDUCE_HPP
+#define DEVICE_CONV_FWD_XDL_TWO_EXTRA_SOURCE_REDUCE_HPP
+
+#include <iostream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r3.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          bool ABlockLdsAddExtraM,
+          bool BBlockLdsAddExtraN>
+struct DeviceConvFwdXdl_two_extra_source_reduce;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce_nhwc_kyxc_nhwk.hpp

+#ifndef DEVICE_CONV_FWD_XDL_TWO_EXTRA_SOURCE_REDUCE_NHWC_KYXC_NHWK_HPP
+#define DEVICE_CONV_FWD_XDL_TWO_EXTRA_SOURCE_REDUCE_NHWC_KYXC_NHWK_HPP
+
+#include <iostream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r3.hpp"
+#include "example/4_conv_xdl_bias_relu_add/include/device_conv_fwd_xdl_two_extra_source_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// specialization for 2D conv: in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          bool ABlockLdsAddExtraM,
+          bool BBlockLdsAddExtraN>
+struct DeviceConvFwdXdl_two_extra_source_reduce<
+    2,                                        // ck::index_t NDimSpatial,
+    InDataType,                               // typename InDataType,
+    WeiDataType,                              // typename WeiDataType,
+    OutDataType,                              // typename OutDataType,
+    AccDataType,                              // typename AccDataType,
+    ck::tensor_layout::convolution::NHWC,     // typename InLayout,
+    ck::tensor_layout::convolution::KYXC,     // typename WeiLayout,
+    ck::tensor_layout::convolution::NHWK,     // typename OutLayout,
+    InElementwiseOperation,                   // typename InElementwiseOperation,
+    WeiElementwiseOperation,                  // typename WeiElementwiseOperation,
+    OutElementwiseOperation,                  // typename OutElementwiseOperation,
+    BlockSize,                                // ck::index_t BlockSize,
+    MPerBlock,                                // ck::index_t MPerBlock,
+    NPerBlock,                                // ck::index_t NPerBlock,
+    K0PerBlock,                               // ck::index_t K0PerBlock,
+    K1,                                       // ck::index_t K1,
+    MPerXDL,                                  // ck::index_t MPerXDL,
+    NPerXDL,                                  // ck::index_t NPerXDL,
+    MXdlPerWave,                              // ck::index_t MXdlPerWave,
+    NXdlPerWave,                              // ck::index_t NXdlPerWave,
+    ABlockTransferThreadSliceLengths_K0_M_K1, // typename ABlockTransferThreadSliceLengths_K0_M_K1,
+    ABlockTransferThreadClusterLengths_K0_M_K1, // typename
+                                                // ABlockTransferThreadClusterLengths_K0_M_K1,
+    ABlockTransferThreadClusterArrangeOrder,    // typename ABlockTransferThreadClusterArrangeOrder,
+    ABlockTransferSrcAccessOrder,               // typename ABlockTransferSrcAccessOrder,
+    ABlockTransferSrcVectorDim,                 // ck::index_t ABlockTransferSrcVectorDim,
+    ABlockTransferSrcScalarPerVector,           // ck::index_t ABlockTransferSrcScalarPerVector,
+    ABlockTransferDstScalarPerVector_K1,        // ck::index_t ABlockTransferDstScalarPerVector_K1,
+    BBlockTransferThreadSliceLengths_K0_N_K1, // typename BBlockTransferThreadSliceLengths_K0_N_K1,
+    BBlockTransferThreadClusterLengths_K0_N_K1, // typename
+                                                // BBlockTransferThreadClusterLengths_K0_N_K1,
+    BBlockTransferThreadClusterArrangeOrder,    // typename BBlockTransferThreadClusterArrangeOrder,
+    BBlockTransferSrcAccessOrder,               // typename BBlockTransferSrcAccessOrder,
+    BBlockTransferSrcVectorDim,                 // ck::index_t BBlockTransferSrcVectorDim,
+    BBlockTransferSrcScalarPerVector,           // ck::index_t BBlockTransferSrcScalarPerVector,
+    BBlockTransferDstScalarPerVector_K1,        // ck::index_t BBlockTransferDstScalarPerVector_K1,
+    CThreadTransferSrcDstVectorDim,             // ck::index_t CThreadTransferSrcDstVectorDim,
+    CThreadTransferDstScalarPerVector,          // ck::index_t CThreadTransferDstScalarPerVector,
+    ABlockLdsAddExtraM,                         // bool ABlockLdsAddExtraM,
+    BBlockLdsAddExtraN                          // bool BBlockLdsAddExtraN>
+    > : public BaseOperator
+{
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+        const index_t GemmK    = Y * X * C;
+
+        const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
+
+        assert(GemmK % GemmK1Number == 0);
+
+        const index_t GemmK0 = GemmK / GemmK1Number;
+
+        // A: input tensor
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+            in_n_hip_wip_c_grid_desc,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_gemmk_gemmmraw_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                   make_merge_transform(make_tuple(N, Ho, Wo))),
+                                        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk_gemmmraw_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmMRaw)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+            transform_tensor_descriptor(in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                                        make_tuple(make_pass_through_transform(GemmK0),
+                                                   make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmK1Number)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // B: weight tensor
+        const auto wei_k_yxc_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+        const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+            wei_k_yxc_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+        const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            wei_gemmk_gemmn_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmN)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        // C: output tensor
+        const auto out_nhowo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+        const auto out_gemmmraw_gemmn_grid_desc = transform_tensor_descriptor(
+            out_nhowo_k_grid_desc,
+            make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_gemmm_gemmn_grid_desc =
+            transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                        make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                   make_pass_through_transform(GemmN)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                          wei_gemmk0_gemmn_gemmk1_grid_desc,
+                          out_gemmm_gemmn_grid_desc);
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // TODO remove these hacks
+    static constexpr auto a_k0_m_k1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{},   // 0+: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 1+: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}),  // 2+: K1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{},   // 0-: K0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 1-: M
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto b_k0_n_k1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: N
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: K1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0-: K0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: N
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: K1
+
+    static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
+
+    static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadSliceLengths_K0_M_K1,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        BBlockTransferThreadSliceLengths_K0_N_K1,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false,                            // BThreadTransferSrcResetCoordinateAfterRun,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        decltype(a_k0_m_k1_grid_step_hacks),                   //  AGridStepHacks,
+        decltype(b_k0_n_k1_grid_step_hacks),                   //  BGridStepHacks,
+        decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),   //  CGridStepHacks,
+        decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), //  AGridMoveSliceWindowStepHacks,
+        decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), //  BGridMoveSliceWindowStepHacks,
+        false,                                                 // CAccessOrderMRepeatNRepeat,
+        ABlockLdsAddExtraM,
+        BBlockLdsAddExtraN>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+
+    using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+            const auto descs = DeviceConvFwdXdl_two_extra_source_reduce::
+                MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                K,
+                                                                C,
+                                                                input_spatial_lengths,
+                                                                filter_spatial_lengths,
+                                                                output_spatial_lengths,
+                                                                conv_filter_strides,
+                                                                conv_filter_dilations,
+                                                                input_left_pads,
+                                                                input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceConvFwdXdl_two_extra_source_reduce::Argument;
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::
+                                           CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::Block2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::
+                                           CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<DeviceConvFwdXdl_two_extra_source_reduce::Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op)
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/CMakeLists.txt

@@ -14,11 +14,14 @@ include_directories(BEFORE
 set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
 set(GEMM_XDL_BIAS_ADD_SOURCE 2_gemm_xdl_bias_add/gemm_xdl_bias_add.cpp)
 set(CONV_XDL_SOURCE 3_conv_xdl/conv_xdl.cpp)
+set(CONV_XDL_BIAS_RELU_ADD_SOURCE 4_conv_xdl_bias_relu_add/conv_xdl_bias_relu_add.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
 add_executable(gemm_xdl_bias_add ${GEMM_XDL_BIAS_ADD_SOURCE})
 add_executable(conv_xdl ${CONV_XDL_SOURCE})
+add_executable(conv_xdl_bias_relu_add ${CONV_XDL_BIAS_RELU_ADD_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_add PRIVATE host_tensor)
 target_link_libraries(conv_xdl PRIVATE host_tensor)
+target_link_libraries(conv_xdl_bias_relu_add PRIVATE host_tensor)