add conv+add fusion

host/driver_offline/CMakeLists.txt

@@ -13,18 +13,21 @@ include_directories(BEFORE
 
 set(CONV_FWD_DRIVER_OFFLINE_SOURCE src/conv_fwd_driver_offline.cpp)
 set(CONV_FWD_DRIVER_OFFLINE_NCHWC_SOURCE src/conv_fwd_driver_offline_nchwc.cpp)
+set(CONV_ADD_FWD_DRIVER_OFFLINE_NCHWC_SOURCE src/conv_add_fwd_driver_offline_nchwc.cpp)
 set(CONV_BWD_DRIVER_OFFLINE_SOURCE src/conv_bwd_driver_offline.cpp)
 set(CONV_WRW_DRIVER_OFFLINE_SOURCE src/conv_wrw_driver_offline.cpp)
 set(GEMM_DRIVER_OFFLINE_SOURCE src/gemm_driver_offline.cpp)
 
 add_executable(conv_fwd_driver_offline ${CONV_FWD_DRIVER_OFFLINE_SOURCE})
 add_executable(conv_fwd_driver_offline_nchwc ${CONV_FWD_DRIVER_OFFLINE_NCHWC_SOURCE})
+add_executable(conv_add_fwd_driver_offline_nchwc ${CONV_ADD_FWD_DRIVER_OFFLINE_NCHWC_SOURCE})
 add_executable(conv_bwd_driver_offline ${CONV_BWD_DRIVER_OFFLINE_SOURCE})
 add_executable(conv_wrw_driver_offline ${CONV_WRW_DRIVER_OFFLINE_SOURCE})
 add_executable(gemm_driver_offline ${GEMM_DRIVER_OFFLINE_SOURCE})
 
 target_link_libraries(conv_fwd_driver_offline PRIVATE host_tensor)
 target_link_libraries(conv_fwd_driver_offline_nchwc PRIVATE host_tensor)
+target_link_libraries(conv_add_fwd_driver_offline_nchwc PRIVATE host_tensor)
 target_link_libraries(conv_bwd_driver_offline PRIVATE host_tensor)
 target_link_libraries(conv_wrw_driver_offline PRIVATE host_tensor)
 target_link_libraries(gemm_driver_offline PRIVATE host_tensor)

host/driver_offline/include/device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp

+#include <unistd.h>
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "driver_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp"
+
+template <typename TInWei,
+          typename TAcc,
+          typename TOut,
+          ck::index_t activ_type,
+          typename InLengths,
+          typename WeiLengths,
+          typename AddLengths,
+          typename OutLengths,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+void device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1(
+    const InLengths& in_n_c0_hi_wi_c1_lengths,
+    const WeiLengths& wei_k_c0_y_x_c1_lengths,
+    const AddLengths& add_n_k0_hox2_wox2_k1_lengths,
+    const OutLengths& out_n_k0_ho_wo_k1_lengths,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const Tensor<TInWei>& in_n_c0_hi_wi_c1,
+    const Tensor<TInWei>& wei_k_c0_y_x_c1,
+    const Tensor<TOut>& add_n_k0_hox2_wox2_k1,
+    Tensor<TOut>& add_n_k0_hox2_wox2_k1_out,
+    Tensor<TOut>& out_n_k0_ho_wo_k1,
+    ck::index_t nrepeat)
+{
+    using namespace ck;
+
+    std::cout << __func__ << std::endl;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+    constexpr auto I4 = Number<4>{};
+
+    const auto N  = out_n_k0_ho_wo_k1_lengths[I0];
+    const auto K0 = out_n_k0_ho_wo_k1_lengths[I1];
+    const auto Ho = out_n_k0_ho_wo_k1_lengths[I2];
+    const auto Wo = out_n_k0_ho_wo_k1_lengths[I3];
+    const auto K1 = out_n_k0_ho_wo_k1_lengths[I4];
+
+    const auto C0 = in_n_c0_hi_wi_c1_lengths[I1];
+    const auto Hi = in_n_c0_hi_wi_c1_lengths[I2];
+    const auto Wi = in_n_c0_hi_wi_c1_lengths[I3];
+    const auto C1 = in_n_c0_hi_wi_c1_lengths[I4];
+
+    const auto K = wei_k_c0_y_x_c1_lengths[I0];
+    const auto Y = wei_k_c0_y_x_c1_lengths[I2];
+    const auto X = wei_k_c0_y_x_c1_lengths[I3];
+
+    const auto Hox2 = add_n_k0_hox2_wox2_k1_lengths[I2];
+    const auto Wox2 = add_n_k0_hox2_wox2_k1_lengths[I3];
+
+    DeviceMem in_n_c0_hi_wi_c1_device_buf(sizeof(TInWei) *
+                                          in_n_c0_hi_wi_c1.mDesc.GetElementSpace());
+    DeviceMem wei_k_c0_y_x_c1_device_buf(sizeof(TInWei) * wei_k_c0_y_x_c1.mDesc.GetElementSpace());
+    DeviceMem add_n_k0_hox2_wox2_k1_device_buf(sizeof(TOut) *
+                                               add_n_k0_hox2_wox2_k1.mDesc.GetElementSpace());
+    DeviceMem out_n_k0_ho_wo_k1_device_buf(sizeof(TOut) *
+                                           out_n_k0_ho_wo_k1.mDesc.GetElementSpace());
+
+    in_n_c0_hi_wi_c1_device_buf.ToDevice(in_n_c0_hi_wi_c1.mData.data());
+    wei_k_c0_y_x_c1_device_buf.ToDevice(wei_k_c0_y_x_c1.mData.data());
+    add_n_k0_hox2_wox2_k1_device_buf.ToDevice(add_n_k0_hox2_wox2_k1.mData.data());
+
+    constexpr index_t InWeiVectorSize = 8;
+
+    if(C1 % InWeiVectorSize != 0)
+    {
+        throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
+    }
+
+#if 0
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t KPerBlock  = 32;
+    constexpr index_t HoPerBlock = 8;
+    constexpr index_t WoPerBlock = 64;
+
+    constexpr index_t E1        = C0 * 9;
+    constexpr index_t E2        = 1;
+    constexpr index_t E1PerBlock = C0;
+
+    constexpr index_t KPerThread  = 16;
+    constexpr index_t HoPerThread = 2;
+    constexpr index_t WoPerThread = 2;
+    constexpr index_t EPerThread  = 1;
+
+    using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2   = Sequence<1, 9, 1, E2>;
+    using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 = Sequence<1, E1PerBlock, KPerBlock, 1>;
+
+    constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
+    constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
+
+    constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
+
+    constexpr index_t CThreadTransferDstScalarPerVector_K = K1;
+#elif 1
+    constexpr auto BlockSize = 64;
+
+    constexpr auto KPerBlock  = 16;
+    constexpr auto HoPerBlock = 8;
+    constexpr auto WoPerBlock = 32;
+
+    constexpr auto E1         = 2 * 9;
+    constexpr auto E2         = 1;
+    constexpr auto K2         = 2;
+    constexpr auto E1PerBlock = 2;
+
+    constexpr auto KPerThread  = 16;
+    constexpr auto HoPerThread = 2;
+    constexpr auto WoPerThread = 2;
+    constexpr auto EPerThread  = 1;
+
+    using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, 1, E2>;
+    using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 =
+        Sequence<1, E1PerBlock, 1, KPerBlock, 1>;
+
+    constexpr auto ABlockTransferSrcScalarPerVector_E2 = E2;
+    constexpr auto ABlockTransferDstScalarPerVector_E2 = E2;
+
+    constexpr auto BThreadTransferSrcScalarPerVector_E2 = E2;
+
+    constexpr auto CThreadTransferDstScalarPerVector_K = 8;
+#endif
+
+    const auto in_n_c0_hi_wi_c1_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, C1));
+    const auto wei_k_c0_y_x_c1_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(K, C0, Y, X, C1));
+    const auto add_n_k0_hox2_wox2_k1_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N, K0, Hox2, Wox2, K1));
+    const auto out_n_k0_ho_wo_k1_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1));
+
+    static_assert(in_n_c0_hi_wi_c1_desc.IsKnownAtCompileTime(), "");
+    static_assert(wei_k_c0_y_x_c1_desc.IsKnownAtCompileTime(), "");
+    static_assert(add_n_k0_hox2_wox2_k1_desc.IsKnownAtCompileTime(), "");
+    static_assert(out_n_k0_ho_wo_k1_desc.IsKnownAtCompileTime(), "");
+
+    constexpr auto conv_driver =
+        DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_add<
+            BlockSize,
+            typename vector_type<TInWei, InWeiVectorSize>::type,
+            TAcc,
+            TOut,
+            E1,
+            E2,
+            K2,
+            KPerBlock,
+            HoPerBlock,
+            WoPerBlock,
+            E1PerBlock,
+            KPerThread,
+            HoPerThread,
+            WoPerThread,
+            EPerThread,
+            ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
+            ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
+            ABlockTransferSrcScalarPerVector_E2,
+            ABlockTransferDstScalarPerVector_E2,
+            BThreadTransferSrcScalarPerVector_E2,
+            CThreadTransferDstScalarPerVector_K,
+            activ_type>{};
+
+    for(int i = 0; i < 5; i++)
+    {
+
+        const auto ave_time =
+            conv_driver.Run(wei_k_c0_y_x_c1_desc,
+                            in_n_c0_hi_wi_c1_desc,
+                            add_n_k0_hox2_wox2_k1_desc,
+                            out_n_k0_ho_wo_k1_desc,
+                            conv_strides,
+                            conv_dilations,
+                            in_left_pads,
+                            in_right_pads,
+                            static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
+                                wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
+                            static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
+                                in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
+                            static_cast<TOut*>(add_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()),
+                            static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()),
+                            nrepeat);
+
+        {
+            float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C0 * C1 * Y * X) /
+                         (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+            std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
+                      << std::endl;
+        }
+    }
+
+    add_n_k0_hox2_wox2_k1_device_buf.ToDevice(add_n_k0_hox2_wox2_k1.mData.data());
+
+    conv_driver.Run(wei_k_c0_y_x_c1_desc,
+                    in_n_c0_hi_wi_c1_desc,
+                    add_n_k0_hox2_wox2_k1_desc,
+                    out_n_k0_ho_wo_k1_desc,
+                    conv_strides,
+                    conv_dilations,
+                    in_left_pads,
+                    in_right_pads,
+                    static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
+                        wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
+                    static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
+                        in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
+                    static_cast<TOut*>(add_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()),
+                    static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()),
+                    1);
+
+    add_n_k0_hox2_wox2_k1_device_buf.FromDevice(add_n_k0_hox2_wox2_k1_out.mData.data());
+    out_n_k0_ho_wo_k1_device_buf.FromDevice(out_n_k0_ho_wo_k1.mData.data());
+}

host/driver_offline/src/conv_add_fwd_driver_offline_nchwc.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+//#include <half.hpp>
+#include "config.hpp"
+#include "debug.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "conv_common.hpp"
+#include "host_conv.hpp"
+#include "device_tensor.hpp"
+#include "device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp"
+
+#define USE_DYNAMIC_MODE 0
+#define USE_CONV_FWD_V5R1_NCHWC 1
+
+enum ConvForwardAlgo
+{
+    V5R1NCHWC // 0
+};
+
+int main(int argc, char* argv[])
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+    constexpr auto I4 = Number<4>{};
+    constexpr auto I5 = Number<5>{};
+    constexpr auto I6 = Number<6>{};
+    constexpr auto I7 = Number<7>{};
+
+#if USE_DYNAMIC_MODE
+    // dynamic mode
+    if(argc != 23)
+    {
+        printf("arg1 to 5: algo, do_verification, init_method, do_log, nrepeat\n");
+        printf("rest: N, K0, K1, C0, C1, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    constexpr index_t activ_type = 0;
+
+    const ConvForwardAlgo algo = static_cast<ConvForwardAlgo>(std::stoi(argv[1]));
+    const bool do_verification = std::stoi(argv[2]);
+    const int init_method      = std::stoi(argv[3]);
+    const bool do_log          = std::stoi(argv[4]);
+    const int nrepeat          = std::stoi(argv[5]);
+
+    const index_t N  = std::stoi(argv[6]);
+    const index_t K0 = std::stoi(argv[7]);
+    const index_t K1 = std::stoi(argv[8]);
+    const index_t C0 = std::stoi(argv[9]);
+    const index_t C1 = std::stoi(argv[10]);
+    const index_t Y  = std::stoi(argv[11]);
+    const index_t X  = std::stoi(argv[12]);
+    const index_t Hi = std::stoi(argv[13]);
+    const index_t Wi = std::stoi(argv[14]);
+
+    const index_t conv_stride_h   = std::stoi(argv[15]);
+    const index_t conv_stride_w   = std::stoi(argv[16]);
+    const index_t conv_dilation_h = std::stoi(argv[17]);
+    const index_t conv_dilation_w = std::stoi(argv[18]);
+    const index_t in_left_pad_h   = std::stoi(argv[19]);
+    const index_t in_left_pad_w   = std::stoi(argv[20]);
+    const index_t in_right_pad_h  = std::stoi(argv[21]);
+    const index_t in_right_pad_w  = std::stoi(argv[22]);
+
+    const index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+#else
+    // static mode
+    if(argc < 6)
+    {
+        printf("arg1 to 5: algo, do_verification, init_method, do_log, nrepeat\n");
+        exit(1);
+    }
+
+    const ConvForwardAlgo algo = static_cast<ConvForwardAlgo>(std::stoi(argv[1]));
+
+    const bool do_verification = std::stoi(argv[2]);
+    const int init_method      = std::stoi(argv[3]);
+    const bool do_log          = std::stoi(argv[4]);
+    const int nrepeat          = std::stoi(argv[5]);
+
+    constexpr index_t activ_type = 0;
+
+#if 0
+    constexpr auto N             = Number<1>{};
+    constexpr auto Hi            = Number<1080>{};
+    constexpr auto Wi            = Number<1920>{};
+    constexpr auto Y             = Number<3>{};
+    constexpr auto X             = Number<3>{};
+    constexpr auto C0            = Number<2>{};
+    constexpr auto C1            = Number<8>{};
+    constexpr auto K1            = Number<8>{};
+    constexpr auto K0            = Number<8>{};
+#elif 0
+    constexpr auto N  = Number<1>{};
+    constexpr auto Hi = Number<540>{};
+    constexpr auto Wi = Number<960>{};
+    constexpr auto Y  = Number<3>{};
+    constexpr auto X  = Number<3>{};
+    constexpr auto C0 = Number<2>{};
+    constexpr auto C1 = Number<8>{};
+    constexpr auto K1 = Number<8>{};
+    constexpr auto K0 = Number<8>{};
+#elif 0
+    constexpr auto N  = Number<1>{};
+    constexpr auto Hi = Number<270>{};
+    constexpr auto Wi = Number<480>{};
+    constexpr auto Y  = Number<3>{};
+    constexpr auto X  = Number<3>{};
+    constexpr auto C0 = Number<2>{};
+    constexpr auto C1 = Number<8>{};
+    constexpr auto K1 = Number<8>{};
+    constexpr auto K0 = Number<8>{};
+#elif 1
+    constexpr auto N  = Number<1>{};
+    constexpr auto Hi = Number<135>{};
+    constexpr auto Wi = Number<240>{};
+    constexpr auto Y  = Number<3>{};
+    constexpr auto X  = Number<3>{};
+    constexpr auto C0 = Number<2>{};
+    constexpr auto C1 = Number<8>{};
+    constexpr auto K1 = Number<8>{};
+    constexpr auto K0 = Number<8>{};
+#endif
+
+    constexpr auto conv_stride_h   = I1;
+    constexpr auto conv_stride_w   = I1;
+    constexpr auto conv_dilation_h = I1;
+    constexpr auto conv_dilation_w = I1;
+    constexpr auto in_left_pad_h   = I1;
+    constexpr auto in_left_pad_w   = I1;
+    constexpr auto in_right_pad_h  = I1;
+    constexpr auto in_right_pad_w  = I1;
+
+    constexpr auto YEff = (Y - I1) * conv_dilation_h + I1;
+    constexpr auto XEff = (X - I1) * conv_dilation_w + I1;
+
+    constexpr auto Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + I1;
+    constexpr auto Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + I1;
+
+    constexpr auto Hox2 = Number<Ho * 2>{};
+    constexpr auto Wox2 = Number<Wo * 2>{};
+
+#endif
+
+#if 0
+    using in_data_t  = float;
+    using acc_data_t = float;
+    using out_data_t = float;
+#elif 1
+    using in_data_t     = half_t;
+    using acc_data_t    = float;
+    using out_data_t    = half_t;
+#elif 1
+    using in_data_t  = int8_t;
+    using acc_data_t = int32_t;
+    using out_data_t = int8_t;
+#endif
+
+    std::vector<std::size_t> in_lengths_host(5), wei_lengths_host(5), out_lengths_host(5),
+        add_lengths_host(5);
+
+    in_lengths_host[0] = static_cast<std::size_t>(N);
+    in_lengths_host[1] = static_cast<std::size_t>(C0);
+    in_lengths_host[2] = static_cast<std::size_t>(Hi);
+    in_lengths_host[3] = static_cast<std::size_t>(Wi);
+    in_lengths_host[4] = static_cast<std::size_t>(C1);
+
+    wei_lengths_host[0] = static_cast<std::size_t>(K0 * K1);
+    wei_lengths_host[1] = static_cast<std::size_t>(C0);
+    wei_lengths_host[2] = static_cast<std::size_t>(Y);
+    wei_lengths_host[3] = static_cast<std::size_t>(X);
+    wei_lengths_host[4] = static_cast<std::size_t>(C1);
+
+    out_lengths_host[0] = static_cast<std::size_t>(N);
+    out_lengths_host[1] = static_cast<std::size_t>(K0);
+    out_lengths_host[2] = static_cast<std::size_t>(Ho);
+    out_lengths_host[3] = static_cast<std::size_t>(Wo);
+    out_lengths_host[4] = static_cast<std::size_t>(K1);
+
+    add_lengths_host[0] = static_cast<std::size_t>(N);
+    add_lengths_host[1] = static_cast<std::size_t>(K0);
+    add_lengths_host[2] = static_cast<std::size_t>(Hox2);
+    add_lengths_host[3] = static_cast<std::size_t>(Wox2);
+    add_lengths_host[4] = static_cast<std::size_t>(K1);
+
+    Tensor<in_data_t> in(in_lengths_host);
+    Tensor<in_data_t> wei(wei_lengths_host);
+    Tensor<in_data_t> add(add_lengths_host);
+    Tensor<out_data_t> out_host(out_lengths_host);
+    Tensor<out_data_t> out_device(out_lengths_host);
+    Tensor<in_data_t> add_device(add_lengths_host);
+    Tensor<in_data_t> add_host(add_lengths_host);
+
+    ostream_HostTensorDescriptor(in.mDesc, std::cout << "in: ");
+    ostream_HostTensorDescriptor(wei.mDesc, std::cout << "wei: ");
+    ostream_HostTensorDescriptor(add.mDesc, std::cout << "add: ");
+    ostream_HostTensorDescriptor(out_host.mDesc, std::cout << "out: ");
+
+    print_array("InLeftPads", make_tuple(in_left_pad_h, in_left_pad_w));
+    print_array("InRightPads", make_tuple(in_right_pad_h, in_right_pad_w));
+    print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
+    print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    switch(init_method)
+    {
+    case 0:
+        // no initialization
+        break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        add.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 2:
+        in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        add.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 3:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        add.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
+        break;
+    case 4:
+        in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        add.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
+        break;
+    case 5:
+        in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
+        add.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
+        wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
+
+        auto gen_wei = [](auto... is) {
+            return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
+        };
+        wei.GenerateTensorValue(gen_wei, num_thread);
+    }
+
+    auto f_make_for_device_nchwc = [&]() {
+        const auto in_lengths_dev     = make_tuple(N, C0, Hi, Wi, C1);
+        const auto wei_lengths_dev    = make_tuple(K0 * K1, C0, Y, X, C1);
+        const auto add_lengths_dev    = make_tuple(N, K0, Hox2, Wox2, K1);
+        const auto out_lengths_dev    = make_tuple(N, K0, Ho, Wo, K1);
+        const auto conv_strides_dev   = make_tuple(conv_stride_h, conv_stride_w);
+        const auto conv_dilations_dev = make_tuple(conv_dilation_h, conv_dilation_w);
+        const auto in_left_pads_dev   = make_tuple(in_left_pad_h, in_left_pad_w);
+        const auto in_right_pads_dev  = make_tuple(in_right_pad_h, in_right_pad_w);
+
+        return make_tuple(in_lengths_dev,
+                          wei_lengths_dev,
+                          add_lengths_dev,
+                          out_lengths_dev,
+                          conv_strides_dev,
+                          conv_dilations_dev,
+                          in_left_pads_dev,
+                          in_right_pads_dev);
+    };
+
+#if USE_CONV_FWD_V5R1_NCHWC
+    if(algo == ConvForwardAlgo::V5R1NCHWC)
+    {
+        const auto tmp = f_make_for_device_nchwc();
+
+        device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1<in_data_t,
+                                                                                        acc_data_t,
+                                                                                        out_data_t,
+                                                                                        activ_type>(
+            tmp[I0], // in_lengths_dev
+            tmp[I1], // wei_lengths_dev
+            tmp[I2], // add_lengths_dev
+            tmp[I3], // out_lengths_dev
+            tmp[I4], // conv_strides_dev
+            tmp[I5], // conv_dilations_dev
+            tmp[I6], // in_left_pads_dev
+            tmp[I7], // in_right_pads_dev
+            in,
+            wei,
+            add,
+            add_device,
+            out_device,
+            nrepeat);
+    }
+#endif
+
+    if(do_verification)
+    {
+        host_direct_convolution_add_nchwc(in,
+                                          wei,
+                                          add,
+                                          add_host,
+                                          out_host,
+                                          make_tuple(conv_stride_h, conv_stride_w),
+                                          make_tuple(conv_dilation_h, conv_dilation_w),
+                                          make_tuple(in_left_pad_h, in_left_pad_w),
+                                          make_tuple(in_right_pad_h, in_right_pad_w),
+                                          activ_type);
+
+        check_error(out_host, out_device);
+        check_error(add_host, add_device);
+
+        if(do_log)
+        {
+            // LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
+            // LogRangeAsType<float>(std::cout << "wei: ", wei.mData, ",") << std::endl;
+            // LogRangeAsType<float>(std::cout << "out_host  : ", out_host.mData, ",") << std::endl;
+            // LogRangeAsType<float>(std::cout << "out_device: ", out_device.mData, ",") <<
+            // std::endl;
+            // LogRangeAsType<float>(std::cout << "add_device: ", add_device.mData, ",") <<
+            // std::endl;
+            LogRangeAsType<float>(std::cout << "add_host: ", add_host.mData, ",") << std::endl;
+        }
+    }
+}

host/driver_offline/src/conv_fwd_driver_offline_nchwc.cpp

@@ -45,6 +45,8 @@ int main(int argc, char* argv[])
         exit(1);
     }
 
+    constexpr index_t activ_type = 0;
+
     const ConvForwardAlgo algo = static_cast<ConvForwardAlgo>(std::stoi(argv[1]));
     const bool do_verification = std::stoi(argv[2]);
     const int init_method      = std::stoi(argv[3]);
@@ -90,17 +92,19 @@ int main(int argc, char* argv[])
     const bool do_log          = std::stoi(argv[4]);
     const int nrepeat          = std::stoi(argv[5]);
 
-#if 1
-    constexpr auto N           = Number<1>{};
-    constexpr auto Hi          = Number<1080>{};
-    constexpr auto Wi          = Number<1920>{};
-    constexpr auto Y           = Number<3>{};
-    constexpr auto X           = Number<3>{};
-    constexpr auto C0          = Number<2>{};
-    constexpr auto C1          = Number<8>{};
-    constexpr auto K1          = Number<8>{};
-    constexpr auto K0          = Number<8>{};
-#elif 0
+    constexpr index_t activ_type = 0;
+
+#if 0
+    constexpr auto N             = Number<1>{};
+    constexpr auto Hi            = Number<1080>{};
+    constexpr auto Wi            = Number<1920>{};
+    constexpr auto Y             = Number<3>{};
+    constexpr auto X             = Number<3>{};
+    constexpr auto C0            = Number<2>{};
+    constexpr auto C1            = Number<8>{};
+    constexpr auto K1            = Number<8>{};
+    constexpr auto K0            = Number<8>{};
+#elif 1
     constexpr auto N  = Number<1>{};
     constexpr auto Hi = Number<540>{};
     constexpr auto Wi = Number<960>{};
@@ -250,8 +254,6 @@ int main(int argc, char* argv[])
                           in_right_pads_dev);
     };
 
-    constexpr index_t activ_type = 0;
-
 #if USE_CONV_FWD_V5R1_NCHWC
     if(algo == ConvForwardAlgo::V5R1NCHWC)
     {

host/host_tensor/include/host_conv.hpp

@@ -156,6 +156,70 @@ void host_direct_convolution_nchwc(const Tensor<TIn>& in,
                                out.mDesc.GetLengths()[4])(std::thread::hardware_concurrency());
 }
 
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+void host_direct_convolution_add_nchwc(const Tensor<TIn>& in,
+                                       const Tensor<TWei>& wei,
+                                       const Tensor<TOut>& add,
+                                       Tensor<TOut>& add_host,
+                                       Tensor<TOut>& out_host,
+                                       const ConvStrides& conv_strides,
+                                       const ConvDilations& conv_dilations,
+                                       const InLeftPads& in_left_pads,
+                                       const InRightPads&,
+                                       const ck::index_t activ_type = 0)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    auto f_nchw = [&](auto n, auto k0, auto ho, auto wo, auto k1) {
+        double v = 0;
+        for(int c0 = 0; c0 < wei.mDesc.GetLengths()[1]; ++c0)
+        {
+            for(int c1 = 0; c1 < wei.mDesc.GetLengths()[4]; ++c1)
+            {
+                for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
+                {
+                    int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
+                    for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
+                    {
+                        int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
+                        if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                           wi < in.mDesc.GetLengths()[3])
+                        {
+                            v += static_cast<const double>(in(n, c0, hi, wi, c1)) *
+                                 static_cast<const double>(
+                                     wei(k0 * out_host.mDesc.GetLengths()[4] + k1, c0, y, x, c1));
+                        }
+                    }
+                }
+            }
+        }
+
+        v = activ(v, activ_type);
+
+        out_host(n, k0, ho, wo, k1)         = v;
+        add_host(n, k0, ho, wo, k1)         = v + add(n, k0, ho, wo, k1);
+        add_host(n, k0, ho, wo + 1, k1)     = v + add(n, k0, ho, wo + 1, k1);
+        add_host(n, k0, ho + 1, wo, k1)     = v + add(n, k0, ho + 1, wo, k1);
+        add_host(n, k0, ho + 1, wo + 1, k1) = v + add(n, k0, ho + 1, wo + 1, k1);
+    };
+
+    make_ParallelTensorFunctor(f_nchw,
+                               out_host.mDesc.GetLengths()[0],
+                               out_host.mDesc.GetLengths()[1],
+                               out_host.mDesc.GetLengths()[2],
+                               out_host.mDesc.GetLengths()[3],
+                               out_host.mDesc.GetLengths()[4])(std::thread::hardware_concurrency());
+}
+
 template <typename TIn, typename TWei, typename TOut, typename InLeftPads, typename InRightPads>
 void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
                                    const Tensor<TWei>& wei_kcyx,