Function TEST_F

dnn/test/cuda/chanwise_convolution.cpp:248–283

constexpr auto M = Convolution::Mode::CROSS_CORRELATION;

TEST_F(CUDA, CHANWISE_CONVOLUTION_FORWARD) {
    Checker<Convolution> checker(handle_cuda());
    bool require_algo = false;
    checker.set_before_exec_callback(AlgoChecker<ConvolutionForward>(
            ExecutionPolicyAlgoName{
                    "DEFAULT",
                    {{ConvBiasForward::algo_name<ConvBiasForward::DirectParam>(
                              "CHANNEL_WISE", {})
                              .c_str(),
                      {}}}},
            &require_algo));

    for (auto dtype : std::vector<DType>{dtype::Float32(), dtype::Float16()}) {
        checker.set_dtype(0, dtype).set_dtype(1, dtype).set_dtype(2, dtype);
        if (dtype.enumv() == DTypeEnum::Float16)
            checker.set_epsilon(2e-2);

        // simple case
        // clang-format off
        for (uint32_t s : {1, 2})
        for (uint32_t p : {0, 1, 2, 3})
        for (size_t f : {2, 3, 5, 7})
        for (size_t ocpg : {1, 3}) {
            checker.set_param(gconv_param({M, p, p, s, s}))
                    .execs({{2, 3, 16, 16}, {3, ocpg, 1, f, f}, {}});
        }
        // clang-format on

        checker.set_param(gconv_param({M, 2, 3, 2, 1}))
                .execs({{32, 12, 20, 10}, {12, 2, 1, 4, 5}, {}});

        // padding larger than kern
        checker.set_param(gconv_param({M, 20, 30, 4, 5}))
                .execs({{32, 12, 20, 10}, {12, 2, 1, 4, 5}, {}});
    }
}

TEST_F(CUDA, CHANWISE_CONVOLUTION_FORWARD_SMALL) {
    Checker<Convolution> checker(handle_cuda());