Function benchmark_winograd

dnn/test/x86/conv_bias.cpp:3067–3109

}

void benchmark_winograd(
        const char* algo_name, Handle* handle, size_t kernel, size_t pack_size) {
    auto&& args = get_winograd_benchmark_args(kernel, pack_size);
    using namespace conv_bias;
    constexpr size_t RUN = 10;
    Benchmarker<ConvBias> benchmark(handle);
    benchmark.set_display(false);
    benchmark.set_times(RUN);

    Benchmarker<ConvBias> benchmark_winograd(handle);
    benchmark_winograd.set_display(false);
    benchmark_winograd.set_times(RUN);

    for (auto&& arg : args) {
        TensorLayout dst_layout;
        auto opr = handle->create_operator<ConvBias>();
        opr->param() = arg.param;
        opr->deduce_layout(
                {arg.src, dtype::Float32()}, {arg.filter, dtype::Float32()},
                {arg.bias, dtype::Float32()}, {}, dst_layout);
        //! dst.nr_elems * IC * FH * FW * 2
        float computations = dst_layout.total_nr_elems() * arg.filter[1] *
                             arg.filter[2] * arg.filter[3] * 2.0 * 8.0 /
                             (1024 * 1024 * 1024) * 1e3;

        auto used =
                benchmark.set_param(arg.param).exec({arg.src, arg.filter, {}, {}, {}}) /
                RUN;

        benchmark_winograd.set_param(arg.param);
        auto used_winograd = algo_benchmark<ConvBias>(
                                     benchmark_winograd,
                                     {arg.src, arg.filter, {}, {}, {}}, algo_name) /
                             RUN;

        printf("%s %s: normal: %f ms %f Gflops winograd: %f ms %f GFlops "
               "speedup: "
               "%f\n",
               arg.src.to_string().c_str(), arg.filter.to_string().c_str(), used,
               computations / used, used_winograd, computations / used_winograd,
               used / used_winograd);
    }
}
}  // namespace

TEST_F(X86, BENCHMARK_CONVBIAS_WINOGRAD_F63_8x8) {