MegEngine/dnn/test/cuda/convolution.cpp

/**
 * \file dnn/test/cuda/convolution.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */
#include "megdnn/oprs.h"
#include "megdnn/opr_param_defs.h"
#include "test/cuda/fixture.h"
#include "test/common/tensor.h"
#include "test/common/workspace_wrapper.h"
#include "test/common/checker.h"
#include "test/common/convolution.h"
#include "test/common/rng.h"
#include "test/cuda/benchmark.h"

#include "src/cuda/utils.h"

#define V1(x) #x
#define V(x) V1(x)
#define CUDNN_VERSION_STRING \
    "v" V(CUDNN_MAJOR) "." V(CUDNN_MINOR) "." V(CUDNN_PATCHLEVEL)

namespace megdnn {
namespace test {

TEST_F(CUDA, CONVOLUTION_8X8X32)
{
    if (!cuda::is_compute_capability_required(6, 1)) {
        printf("Skip CUDA.CONVOLUTION_8X8X32 test as current device"
               "doesn't support\n");
        return;
    }

    using namespace convolution;
    std::vector<TestArg> args;
    {
        auto v = get_args();
        for (auto &&a: v) {
            args.push_back(std::move(a));
        }
    }
    {
        auto v = get_dilated_args();
        for (auto &&a: v) {
            args.push_back(std::move(a));
        }
    }
    {
        auto v = get_chanwise_args();
        for (auto &&a: v) {
            args.push_back(std::move(a));
        }
    }
    Checker<ConvolutionForward> checker(handle_cuda());
    UniformIntRNG rng(-4, 4);
    for (auto arg: args) {
        arg.param.format = param::Convolution::Format::NHWC;
        arg.src = cvt_src_or_dst_nchw2nhwc(arg.src);
        arg.filter = cvt_filter_nchw2nhwc(arg.filter);
        checker.set_dtype(0, dtype::Int8()).
            set_dtype(1, dtype::Int8()).
            set_dtype(2, dtype::Int32()).
            set_param(arg.param).
            set_rng(0, &rng).
            set_rng(1, &rng).
            execs({arg.src, arg.filter, {}});
    }
}

TEST_F(CUDA, CONVOLUTION_FORWARD)
{
    using namespace convolution;
    std::vector<TestArg> args = get_args();
    Checker<ConvolutionForward> checker(handle_cuda());
    NormalRNG default_rng;
    for (auto &&arg: args) {
        float scale = 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
        UniformFloatRNG rng(scale, 2 * scale);
        checker.
            set_dtype(0, dtype::Float32()).
            set_dtype(1, dtype::Float32()).
            set_dtype(2, dtype::Float32()).
            set_rng(0, &default_rng).
            set_rng(1, &default_rng).
            set_epsilon(1e-3).
            set_param(arg.param).
            execs({arg.src, arg.filter, {}});
        checker.
            set_dtype(0, dtype::Float16()).
            set_dtype(1, dtype::Float16()).
            set_dtype(2, dtype::Float16()).
            set_rng(0, &rng).
            set_rng(1, &rng).
            set_epsilon(1e-1).
            set_param(arg.param).
            execs({arg.src, arg.filter, {}});
        arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        checker.set_dtype(0, dtype::Float16())
                .set_dtype(1, dtype::Float16())
                .set_dtype(2, dtype::Float16())
                .set_rng(0, &rng)
                .set_rng(1, &rng)
                .set_epsilon(1e-1)
                .set_param(arg.param)
                .execs({arg.src, arg.filter, {}});
    }
}

TEST_F(CUDA, CONV_FORWARD_MATMUL_NCHW4) {
    if (!cuda::is_compute_capability_required(6, 1))
        return;
    using namespace convolution;
    Checker<Convolution> checker(handle_cuda());
    UniformIntRNG int_rng{-127, 127};
    Convolution::Param param;
    param.format = Convolution::Param::Format::NCHW4;

    checker.set_dtype(0, dtype::QuantizedS8(0.132f))
            .set_dtype(1, dtype::QuantizedS8(0.0239f))
            .set_dtype(2, dtype::QuantizedS32(0.132f * 0.0239f))
            .set_rng(0, &int_rng)
            .set_rng(1, &int_rng)
            .set_param(param);

    checker.set_before_exec_callback(AlgoChecker<Convolution>(
            ConvBiasForward::algo_name<ConvBiasForward::MatmulParam>(
                    "MATMUL8X8X32", {})
                    .c_str()));

    param.sparse = Convolution::Param::Sparse::DENSE;
    param.pad_h = param.pad_w = 1;
    param.stride_h = param.stride_w = 1;
    checker.set_param(param);
    checker.exec({{8, 4, 10, 10, 4}, {16, 4, 3, 3, 4}, {}});
    checker.exec({{1, 4, 2, 2, 4}, {16, 4, 3, 3, 4}, {}});
    checker.exec({{8, 64, 12, 12, 4}, {256, 64, 3, 3, 4}, {}});
}

TEST_F(CUDA, CONVOLUTION_1X1_FORWARD)
{
    using namespace convolution;
    std::vector<TestArg> args = get_1x1_args();
    Checker<ConvolutionForward> checker(handle_cuda());
    NormalRNG default_rng;
    for (auto &&arg: args) {
        float scale = 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
        UniformFloatRNG rng(scale, 2 * scale);
        checker.
            set_dtype(0, dtype::Float32()).
            set_dtype(1, dtype::Float32()).
            set_rng(0, &default_rng).
            set_rng(1, &default_rng).
            set_epsilon(1e-3).
            set_param(arg.param).
            execs({arg.src, arg.filter, {}});
    }
}

TEST_F(CUDA, BENCHMARK_CONVOLUTION_1X1_FORWARD)
{
    using namespace convolution;
    std::vector<TestArg> args = get_1x1_args();
    Benchmarker<ConvolutionForward> marker(handle_cuda());
    NormalRNG default_rng;
    for (auto &&arg: args) {
        float scale = 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
        UniformFloatRNG rng(scale, 2 * scale);
        marker.set_dtype(0, dtype::Float32()).
            set_dtype(1, dtype::Float32()).
            set_rng(0, &default_rng).
            set_rng(1, &default_rng).
            set_param(arg.param).
            execs({arg.src, arg.filter, {}});
    }
}

TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA)
{
    using namespace convolution;
    std::vector<TestArg> args = get_args_cuda_conv_bwd_data();
    Checker<ConvolutionBackwardData> checker(handle_cuda());
    NormalRNG default_rng;
    for (auto &&arg: args) {
        float scale =
                64.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
        UniformFloatRNG rng(scale, 2 * scale);
        auto src = TensorLayout(arg.src, dtype::Float32());
        auto filter = TensorLayout(arg.filter, dtype::Float32());
        TensorLayout dst;
        {
            auto opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = arg.param;
            opr->deduce_layout(src, filter, dst);
        }
        src.dtype = dst.dtype = filter.dtype = dtype::Float32();
        checker.set_rng(0, &default_rng)
                .set_rng(1, &default_rng)
                .set_epsilon(1e-3)
                .set_param(arg.param)
                .exec(TensorLayoutArray{filter, dst, src});
        if (!cuda::is_compute_capability_required(6, 0)) {
            src.dtype = dst.dtype = filter.dtype = dtype::Float16();
            checker.set_rng(0, &rng)
                    .set_rng(1, &rng)
                    .set_epsilon(1e-1)
                    .set_param(arg.param)
                    .exec(TensorLayoutArray{filter, dst, src});
            arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
            checker.set_rng(0, &rng)
                    .set_rng(1, &rng)
                    .set_epsilon(1e-1)
                    .set_param(arg.param)
                    .exec(TensorLayoutArray{filter, dst, src});
        }
    }
}

TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5)
{
    // BRAIN-481 failed on architectures 7.0, remove the following if statement,
    // when cudnn fixed the problem.
    if (cuda::is_compute_capability_required(7, 0))
        return;
    using namespace convolution;
    std::vector<TestArg> args = get_args_cudnn_7_5_failures();
    Checker<ConvolutionBackwardData> checker(handle_cuda());
    NormalRNG default_rng;
    for (auto &&arg: args) {
        float scale = 128.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
        scale = std::max(scale, 1.f);
        UniformFloatRNG rng(scale, 2 * scale);
        auto src = TensorLayout(arg.src, dtype::Float32());
        auto filter = TensorLayout(arg.filter, dtype::Float32());
        TensorLayout dst;
        {
            auto opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = arg.param;
            opr->deduce_layout(src, filter, dst);
        }
        src.dtype = dst.dtype = filter.dtype = dtype::Float32();
        checker.
            set_rng(0, &default_rng).
            set_rng(1, &default_rng).
            set_epsilon(1e-3).
            set_param(arg.param).
            exec(TensorLayoutArray{filter, dst, src});
        src.dtype = dst.dtype = filter.dtype = dtype::Float16();
        checker.
            set_rng(0, &rng).
            set_rng(1, &rng).
            set_epsilon(1e-1).
            set_param(arg.param).
            exec(TensorLayoutArray{filter, dst, src});
        arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        checker.set_rng(0, &rng)
                .set_rng(1, &rng)
                .set_epsilon(1e-1)
                .set_param(arg.param)
                .exec(TensorLayoutArray{filter, dst, src});
    }
}

TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER)
{
    using namespace convolution;
    std::vector<TestArg> args = get_args();
    Checker<ConvolutionBackwardFilter> checker(handle_cuda());
    bool f16_checked = false;
    for (auto &&arg: args) {
        auto src = TensorLayout(arg.src, dtype::Float32());
        auto filter = TensorLayout(arg.filter, dtype::Float32());
        TensorLayout dst;
        {
            auto opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = arg.param;
            opr->deduce_layout(src, filter, dst);
        }
        float scale = 1.0f / sqrt(dst[2] * dst[3]);
        UniformFloatRNG rng(scale, 2 * scale);
        src.dtype = dst.dtype = filter.dtype = dtype::Float32();
        checker.
            set_rng(0, &rng).
            set_rng(1, &rng).
            set_epsilon(1e-3).
            set_param(arg.param).
            exec(TensorLayoutArray{src, dst, filter});

        // reduce on large f16 array may introduce significant error
        if (dst.total_nr_elems() >= 1000 && f16_checked)
            continue;

        f16_checked = true;
        src.dtype = dst.dtype = filter.dtype = dtype::Float16();
        checker.
            set_rng(0, &rng).
            set_rng(1, &rng).
            set_epsilon(1e-1).
            set_param(arg.param).
            exec(TensorLayoutArray{src, dst, filter});
        arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        checker.set_rng(0, &rng)
                .set_rng(1, &rng)
                .set_epsilon(1e-1)
                .set_param(arg.param)
                .exec(TensorLayoutArray{src, dst, filter});
    }
}

TEST_F(CUDA, CONV_CONFIG_COMBINATIONS) {
    auto eps_getter = [](bool f16, int stage, const char *name) -> float {
        if (f16) {
            return stage == 2 ? 0.5 : 0.2;
        }
        if (strstr(name, "WINOGRAD_NONFUSED"))
            return 0.3;
        return 1e-3;
    };
    convolution::test_conv_config_combinations(handle_cuda(), false, true, true,
            eps_getter, true);
}

TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_1) {
    if (cuda::is_compute_capability_required(7, 0))
        return;
    using namespace convolution;
    Checker<ConvolutionBackwardData> checker(handle_cuda());
    checker.set_before_exec_callback(AlgoChecker<ConvolutionBackwardData>(
            "CUDNN_CONVOLUTION_BWD_DATA_ALGO_1" CUDNN_VERSION_STRING));
    NormalRNG default_rng;
    TensorShape s_filter = TensorShape{8, 8, 2, 2},
                s_src = TensorShape{2, 8, 18, 18};
    float scale = 1.0f / sqrt(s_filter[0] * s_filter[2] * s_filter[3]);
    UniformFloatRNG rng(scale, 2 * scale);
    auto src = TensorLayout(s_src, dtype::Float16());
    auto filter = TensorLayout(s_filter, dtype::Float16());
    TensorLayout dst;
    param::Convolution param;
    param.pad_h = param.pad_w = 2;
    param.stride_h = param.stride_w = 2;
    {
        auto opr = handle_cuda()->create_operator<Convolution>();
        opr->param() = param;
        opr->deduce_layout(src, filter, dst);
    }
    src.dtype = dst.dtype = filter.dtype = dtype::Float16();
    param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
    checker.set_rng(0, &rng)
            .set_rng(1, &rng)
            .set_epsilon(0.2)
            .set_param(param)
            .exec(TensorLayoutArray{filter, dst, src});
}

#if MEGDNN_WITH_BENCHMARK
TEST_F(CUDA, CONV_FWD_BENCHMARK) {
    auto run = [&](size_t N, size_t OC, size_t IC, size_t IH, size_t IW, size_t SH=1,
            size_t SW=1, size_t FH=1, size_t FW=1, size_t PH=0, size_t PW=0, bool fp16io_c32=false) {
        auto benchmarker = Benchmarker<ConvolutionForward>(handle_cuda());
        benchmarker.set_dtype(0, dtype::Float16())
            .set_dtype(1, dtype::Float16())
            .set_dtype(2, dtype::Float16());
        ConvolutionForward::Param param;
        param.stride_h = SH;
        param.stride_w = SW;
        param.pad_h = PH;
        param.pad_w = PW;
        if (fp16io_c32) {
            param.compute_mode = ConvolutionForward::Param::ComputeMode::FLOAT32;
        }
        benchmarker.set_param(param);
        std::unique_ptr<OprProxy<ConvolutionForward>> proxy{new OprProxy<ConvolutionForward>{true}};
        benchmarker.set_proxy(proxy);
        size_t OH = (IH - FH + 2 * PH) / SH + 1;
        size_t OW = (IW - FW + 2 * PW) / SW + 1;
        auto time = benchmarker.execs({
                {N, IC, IH, IW}, {OC, IC, FH, FW}, {N, OC, OH, OW}});
        time /= 1000.0 * 10.0;
        auto flo = (double) N * OC * IC * OH * OW * FH * FW * 2;
        auto flops = flo / time / 1e12;
        printf("comp_type %s: ", fp16io_c32 ? "32" : "16");
        printf("%.3fG FLO, flops %.3fTFLOPS\n", flo/1e9, flops);
    };
    run(32, 512, 256, 56, 56, 1, 1, 1, 1, 0, 0, false);
    run(32, 512, 256, 56, 56, 1, 1, 1, 1, 0, 0, true);
}

TEST_F(CUDA, CONVOLUTION_FWD_BENCHMARK) {
    CUBenchmarker<ConvolutionForward> bench{handle_cuda()};
    std::unique_ptr<OprProxy<ConvolutionForward>> proxy{new OprProxy<ConvolutionForward>{true}};
    size_t RUNS = 10;
    bench.set_proxy(proxy).set_times(RUNS);

    auto run = [&](size_t N, size_t OC, size_t IC, size_t IH, size_t IW,
                   size_t FH, size_t SH, size_t PH) {
        bench.set_dtype(0, dtype::Float32())
                .set_dtype(1, dtype::Float32())
                .set_dtype(2, dtype::Float32());
        param::Convolution param;
        param.stride_h = param.stride_w = SH;
        param.pad_h = param.pad_w = PH;
        param.compute_mode = param::Convolution::ComputeMode::DEFAULT;
        bench.set_param(param);
        bench.proxy()->target_algo = nullptr;
        TensorLayout src{{N, IC, IH, IW}, dtype::Float32()},
                filter{{OC, IC, FH, FH}, dtype::Float32()};
        TensorLayout dst;
        {
            auto&& opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = param;
            opr->deduce_layout(src, filter, dst);
        }
        auto time_ms_fp32 = bench.execl({src, filter, dst}) / RUNS;
        src.dtype = filter.dtype = dst.dtype = dtype::Float16();
        bench.proxy()->target_algo = nullptr;
        bench.set_dtype(0, dtype::Float16())
                .set_dtype(1, dtype::Float16())
                .set_dtype(2, dtype::Float16());
        auto time_ms_true_fp16 = bench.execl({src, filter, dst}) / RUNS;
        param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        bench.proxy()->target_algo = nullptr;
        bench.set_param(param);
        auto time_ms_pseudo_fp16 = bench.execl({src, filter, dst}) / RUNS;
        float flo = 2.0 * N * OC * IC * dst[2] * dst[3] * FH * FH;
        printf("inp=%s, kern=%s, dst=%s ", src.to_string().c_str(),
               filter.to_string().c_str(), dst.to_string().c_str());
        printf("time_fp32=%.2fms, flops=%.3fTFLOPS\ntime_true_fp16=%.2fms, "
               "flops=%.3fTFLOPS\ntime_pseudo_fp16=%.2fms, flops=%.3fFLOPS\n",
               time_ms_fp32, (flo / (time_ms_fp32 * 1e9)), time_ms_true_fp16,
               (flo / (time_ms_true_fp16 * 1e9)), time_ms_pseudo_fp16,
               (flo / (time_ms_pseudo_fp16 * 1e9)));
        printf("speedup (true_fp16/fp32)=%.2f, (true_fp16/pseudo_fp16)=%.2f\n",
               time_ms_fp32 / time_ms_true_fp16,
               time_ms_pseudo_fp16 / time_ms_true_fp16);
    };
    run(32, 64, 3, 224, 224, 7, 2, 3);
    run(32, 128, 128, 28, 28, 3, 1, 1);
    run(32, 256, 256, 14, 14, 3, 1, 1);
    run(32, 512, 512, 7, 7, 3, 1, 1);
    run(32, 64, 64, 56, 56, 3, 1, 1);
    run(32, 512, 256, 56, 56, 1, 2, 0);
    run(32, 1024, 512, 28, 28, 1, 2, 0);
    run(32, 2048, 1024, 14, 14, 1, 2, 0);
    run(32, 512, 128, 28, 28, 1, 1, 0);
    run(32, 128, 512, 28, 28, 1, 1, 0);
    run(32, 1024, 256, 14, 14, 1, 1, 0);
    run(32, 256, 1024, 14, 14, 1, 1, 0);
    run(32, 2048, 512, 7, 7, 1, 1, 0);
    run(32, 512, 2048, 7, 7, 1, 1, 0);
    run(32, 256, 64, 56, 56, 1, 1, 0);
    run(32, 64, 256, 56, 56, 1, 1, 0);
    run(32, 128, 256, 56, 56, 1, 2, 0);
    run(32, 256, 512, 28, 28, 1, 2, 0);
    run(32, 512, 1024, 14, 14, 1, 2, 0);
    run(32, 64, 64, 56, 56, 1, 1, 0);
}

TEST_F(CUDA, CONVOLUTION_BWD_DATA_BENCHMARK) {
    CUBenchmarker<ConvolutionBackwardData> bench{handle_cuda()};
    std::unique_ptr<OprProxy<ConvolutionBackwardData>> proxy{
            new OprProxy<ConvolutionBackwardData>{true}};
    size_t RUNS = 10;
    bench.set_proxy(proxy).set_times(RUNS);

    auto run = [&](size_t N, size_t OC, size_t IC, size_t IH, size_t IW,
                   size_t FH, size_t SH, size_t PH) {
        bench.set_dtype(0, dtype::Float32())
                .set_dtype(1, dtype::Float32())
                .set_dtype(2, dtype::Float32());
        param::Convolution param;
        param.stride_h = param.stride_w = SH;
        param.pad_h = param.pad_w = PH;
        param.compute_mode = param::Convolution::ComputeMode::DEFAULT;
        bench.set_param(param);
        bench.proxy()->target_algo = nullptr;
        TensorLayout src{{N, IC, IH, IW}, dtype::Float32()},
                filter{{OC, IC, FH, FH}, dtype::Float32()};
        TensorLayout dst;
        {
            auto&& opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = param;
            opr->deduce_layout(src, filter, dst);
        }
        auto time_ms_fp32 = bench.execl({filter, dst, src}) / RUNS;
        src.dtype = filter.dtype = dst.dtype = dtype::Float16();
        bench.proxy()->target_algo = nullptr;
        bench.set_dtype(0, dtype::Float16())
                .set_dtype(1, dtype::Float16())
                .set_dtype(2, dtype::Float16());
        auto time_ms_true_fp16 = bench.execl({filter, dst, src}) / RUNS;
        param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        bench.proxy()->target_algo = nullptr;
        bench.set_param(param);
        auto time_ms_pseudo_fp16 = bench.execl({filter, dst, src}) / RUNS;
        float flo = 2.0 * N * OC * IC * dst[2] * dst[3] * FH * FH;
        printf("inp=%s, kern=%s, dst=%s ", src.to_string().c_str(),
               filter.to_string().c_str(), dst.to_string().c_str());
        printf("time_fp32=%.2fms, flops=%.3fTFLOPS\ntime_true_fp16=%.2fms, "
               "flops=%.3fTFLOPS\ntime_pseudo_fp16=%.2fms, flops=%.3fFLOPS\n",
               time_ms_fp32, (flo / (time_ms_fp32 * 1e9)), time_ms_true_fp16,
               (flo / (time_ms_true_fp16 * 1e9)), time_ms_pseudo_fp16,
               (flo / (time_ms_pseudo_fp16 * 1e9)));
        printf("speedup (true_fp16/fp32)=%.2f, (true_fp16/pseudo_fp16)=%.2f\n",
               time_ms_fp32 / time_ms_true_fp16,
               time_ms_pseudo_fp16 / time_ms_true_fp16);
    };
    run(32, 64, 3, 224, 224, 7, 2, 3);
    run(32, 128, 128, 28, 28, 3, 1, 1);
    run(32, 256, 256, 14, 14, 3, 1, 1);
    run(32, 512, 512, 7, 7, 3, 1, 1);
    run(32, 64, 64, 56, 56, 3, 1, 1);
    run(32, 512, 256, 56, 56, 1, 2, 0);
    run(32, 1024, 512, 28, 28, 1, 2, 0);
    run(32, 2048, 1024, 14, 14, 1, 2, 0);
    run(32, 512, 128, 28, 28, 1, 1, 0);
    run(32, 128, 512, 28, 28, 1, 1, 0);
    run(32, 1024, 256, 14, 14, 1, 1, 0);
    run(32, 256, 1024, 14, 14, 1, 1, 0);
    run(32, 2048, 512, 7, 7, 1, 1, 0);
    run(32, 512, 2048, 7, 7, 1, 1, 0);
    run(32, 256, 64, 56, 56, 1, 1, 0);
    run(32, 64, 256, 56, 56, 1, 1, 0);
    run(32, 128, 256, 56, 56, 1, 2, 0);
    run(32, 256, 512, 28, 28, 1, 2, 0);
    run(32, 512, 1024, 14, 14, 1, 2, 0);
    run(32, 64, 64, 56, 56, 1, 1, 0);
}

TEST_F(CUDA, CONVOLUTION_BWD_FILTER_BENCHMARK) {
    CUBenchmarker<ConvolutionBackwardFilter> bench{handle_cuda()};
    std::unique_ptr<OprProxy<ConvolutionBackwardFilter>> proxy{
            new OprProxy<ConvolutionBackwardFilter>{true}};
    size_t RUNS = 10;
    bench.set_proxy(proxy).set_times(RUNS);

    auto run = [&](size_t N, size_t OC, size_t IC, size_t IH, size_t IW,
                   size_t FH, size_t SH, size_t PH) {
        bench.set_dtype(0, dtype::Float32())
                .set_dtype(1, dtype::Float32())
                .set_dtype(2, dtype::Float32());
        param::Convolution param;
        param.stride_h = param.stride_w = SH;
        param.pad_h = param.pad_w = PH;
        param.compute_mode = param::Convolution::ComputeMode::DEFAULT;
        bench.set_param(param);
        bench.proxy()->target_algo = nullptr;
        TensorLayout src{{N, IC, IH, IW}, dtype::Float32()},
                filter{{OC, IC, FH, FH}, dtype::Float32()};
        TensorLayout dst;
        {
            auto&& opr = handle_cuda()->create_operator<Convolution>();
            opr->param() = param;
            opr->deduce_layout(src, filter, dst);
        }
        auto time_ms_fp32 = bench.execl({src, dst, filter}) / RUNS;
        src.dtype = filter.dtype = dst.dtype = dtype::Float16();
        bench.proxy()->target_algo = nullptr;
        bench.set_dtype(0, dtype::Float16())
                .set_dtype(1, dtype::Float16())
                .set_dtype(2, dtype::Float16());
        auto time_ms_true_fp16 = bench.execl({src, dst, filter}) / RUNS;
        param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
        bench.proxy()->target_algo = nullptr;
        bench.set_param(param);
        auto time_ms_pseudo_fp16 = bench.execl({src, dst, filter}) / RUNS;
        float flo = 2.0 * N * OC * IC * dst[2] * dst[3] * FH * FH;
        printf("inp=%s, kern=%s, dst=%s ", src.to_string().c_str(),
               filter.to_string().c_str(), dst.to_string().c_str());
        printf("time_fp32=%.2fms, flops=%.3fTFLOPS\ntime_true_fp16=%.2fms, "
               "flops=%.3fTFLOPS\ntime_pseudo_fp16=%.2fms, flops=%.3fFLOPS\n",
               time_ms_fp32, (flo / (time_ms_fp32 * 1e9)), time_ms_true_fp16,
               (flo / (time_ms_true_fp16 * 1e9)), time_ms_pseudo_fp16,
               (flo / (time_ms_pseudo_fp16 * 1e9)));
        printf("speedup (true_fp16/fp32)=%.2f, (true_fp16/pseudo_fp16)=%.2f\n",
               time_ms_fp32 / time_ms_true_fp16,
               time_ms_pseudo_fp16 / time_ms_true_fp16);
    };
    run(32, 64, 3, 224, 224, 7, 2, 3);
    run(32, 128, 128, 28, 28, 3, 1, 1);
    run(32, 256, 256, 14, 14, 3, 1, 1);
    run(32, 512, 512, 7, 7, 3, 1, 1);
    run(32, 64, 64, 56, 56, 3, 1, 1);
    run(32, 512, 256, 56, 56, 1, 2, 0);
    run(32, 1024, 512, 28, 28, 1, 2, 0);
    run(32, 2048, 1024, 14, 14, 1, 2, 0);
    run(32, 512, 128, 28, 28, 1, 1, 0);
    run(32, 128, 512, 28, 28, 1, 1, 0);
    run(32, 1024, 256, 14, 14, 1, 1, 0);
    run(32, 256, 1024, 14, 14, 1, 1, 0);
    run(32, 2048, 512, 7, 7, 1, 1, 0);
    run(32, 512, 2048, 7, 7, 1, 1, 0);
    run(32, 256, 64, 56, 56, 1, 1, 0);
    run(32, 64, 256, 56, 56, 1, 1, 0);
    run(32, 128, 256, 56, 56, 1, 2, 0);
    run(32, 256, 512, 28, 28, 1, 2, 0);
    run(32, 512, 1024, 14, 14, 1, 2, 0);
    run(32, 64, 64, 56, 56, 1, 1, 0);
}
#endif

#undef CUDNN_VERSION_STRING
#undef V
#undef V1


} // namespace test
} // namespace megdnn

// vim: syntax=cpp.doxygen