Function run_elemwise_benchmark

dnn/test/arm_common/elemwise.cpp:427–478

#if MEGDNN_WITH_BENCHMARK
namespace {
void run_elemwise_benchmark(
        const TensorShapeArray& shapes, param::Elemwise::Mode mode,
        const char* mode_str, DType type, Handle* handle_bench) {
    auto handle_fallback = create_cpu_handle(1);
    Benchmarker<Elemwise> benchmarker_bench(handle_bench);
    Benchmarker<Elemwise> benchmarker_fallback(handle_fallback.get());

    float throughput = 0;
    SmallVector<TensorLayout> layouts;
    std::string src_strs;
    for (size_t i = 0; i < shapes.size(); i++) {
        layouts.emplace_back(shapes[i], type);
        throughput += layouts.back().span().dist_byte();
        src_strs += layouts.back().to_string();
        if (i != shapes.size() - 1) {
            src_strs += ",";
        }
    }
    constexpr size_t RUN = 50;
    benchmarker_fallback.set_times(RUN).set_display(false);
    benchmarker_bench.set_times(RUN).set_display(false);

    benchmarker_fallback.set_param(mode);
    benchmarker_bench.set_param(mode);

    TensorLayout dst_layout;
    auto opr = handle_bench->create_operator<Elemwise>();
    opr->param() = mode;
    opr->deduce_layout(layouts, dst_layout);

    float computations =
            dst_layout.total_nr_elems() * (std::max<size_t>(shapes.size(), 2) - 1);
    throughput += dst_layout.span().dist_byte();
    computations *= (1e3 / (1024.0 * 1024));
    throughput *= (1e3 / (1024.0 * 1024));

    layouts.emplace_back(dst_layout);
    auto fallback_time = benchmarker_fallback.execl(layouts) / RUN;
    auto bench_time = benchmarker_bench.execl(layouts) / RUN;

    float fallback_flops = computations / fallback_time;
    float bench_flops = computations / bench_time;
    float fallback_thr = throughput / fallback_time;
    float bench_thr = throughput / bench_time;

    printf("%s = %s (type: %s, mode: %s) cpu=%fMFLOPS %fMB/s, bench=%fMFLOPS "
           "%fMB/s "
           "computations: %fx, throughput: %fx\n",
           src_strs.c_str(), dst_layout.to_string().c_str(), type.name(), mode_str,
           fallback_flops, fallback_thr, bench_flops, bench_thr,
           bench_flops / fallback_flops, bench_thr / fallback_thr);
}
}  // namespace

TEST_F(ARM_COMMON, BENCHMARK_NCHW_VS_NHWC) {