Function TEST_F

dnn/test/naive/convolution.cpp:15–60

#if MEGDNN_WITH_BENCHMARK

TEST_F(NAIVE, BENCHMARK_CONVOLUTION_BACKWARD_DATA) {
    using Param = ConvolutionBackwardData::Param;
    auto run = [&](const TensorLayoutArray& tensors, Param param) {
        Benchmarker<ConvolutionBackwardData> benchmarker_naive(handle());
        size_t RUN = 500;
        auto tfloat = benchmarker_naive.set_display(false)
                              .set_dtype(0, dtype::Float32{})
                              .set_dtype(1, dtype::Float32{})
                              .set_times(RUN)
                              .set_param(param)
                              .exec(tensors);
        size_t IC = tensors[0][1];
        size_t FH = tensors[0][2];
        size_t FW = tensors[0][3];
        printf("fp32 flops: %.3f mflops\n",
               (IC * tensors[1].total_nr_elems() * FH * FW * 2) /
                       (tfloat / RUN * 1000));
    };

    auto profile = [&](size_t n, size_t ic, size_t oh, size_t ow, size_t oc, size_t fh,
                       size_t fw, size_t stride = 1, size_t padding = 0) {
        Param param;
        param.pad_h = param.pad_w = padding;
        param.stride_h = param.stride_w = stride;
        printf("oc: %zd ic: %zd w: %zd h: %zd stride: %zd kernel_size: %zd\n", oc, ic,
               ow, oh, stride, fh);

        TensorLayout diff = TensorLayout{{n, oc, oh, ow}, dtype::Float32()};
        TensorLayout filter = TensorLayout{{oc, ic, fh, fw}, dtype::Float32()};
        TensorLayout grad;
        {
            auto opr = handle()->create_operator<ConvolutionBackwardData>();
            opr->param() = param;
            opr->deduce_layout(filter, diff, grad);
        }
        run(TensorLayoutArray{filter, diff, grad}, param);
    };

    profile(1, 1, 2, 2, 1, 3, 3);
    profile(1, 1, 4, 4, 1, 3, 3);
    profile(1, 1, 8, 8, 1, 3, 3);
    profile(1, 1, 16, 16, 1, 3, 3);
    profile(1, 1, 32, 32, 1, 3, 3);
    profile(1, 1, 64, 64, 1, 3, 3);
    profile(1, 1, 128, 128, 1, 3, 3);
}

#endif