Function TEST

src/cambricon/test/magicmind_runtime_opr.cpp:336–433

}  // namespace

TEST(TestMagicMindRuntimeOpr, Basic) {
    REQUIRE_CAMBRICON_DEVICE(1);
    auto cn = CompNode::load("cambricon0");
    MMNetwork network(cn, magicmind::DataType::FLOAT32, false);
    size_t dtype_size = magicmind::DataTypeSize(magicmind::DataType::FLOAT32);

    // prepare parameter for addpad and conv
    const int ni = 16, ci = 64, hi = 32, wi = 32;
    const int no = 16, co = 64, ho = 32, wo = 32;

    // count tensor nums
    int conv_input_count = ni * hi * wi * ci;
    int relu_output_count = no * ho * wo * co;

    // prepare cpu origin data
    std::vector<float> conv_input_cpu_data;
    gen_rand_data(conv_input_cpu_data, conv_input_count, 256);
    std::vector<float> add_input_cpu_data;
    gen_rand_data(add_input_cpu_data, relu_output_count, 256);
    std::vector<float> relu_output_cpu_data(relu_output_count);
    std::vector<float> add_output_cpu_data(relu_output_count);

    auto mlu_deleter = [](void* p) { MGB_CNRT_CHECK(cnrtFree(p)); };
    void* conv_input_mlu_ptr;
    void* add_input_mlu_ptr;
    void* relu_output_mlu_ptr;
    void* add_output_mlu_ptr;

    // malloc mlu mem for fusion input and output
    MGB_CNRT_CHECK(cnrtMalloc(&conv_input_mlu_ptr, conv_input_count * dtype_size));
    MGB_CNRT_CHECK(cnrtMalloc(&add_input_mlu_ptr, relu_output_count * sizeof(float)));
    MGB_CNRT_CHECK(cnrtMalloc(&relu_output_mlu_ptr, relu_output_count * sizeof(float)));
    MGB_CNRT_CHECK(cnrtMalloc(&add_output_mlu_ptr, relu_output_count * sizeof(float)));

    // memory copy cpu->mlu
    MGB_CNRT_CHECK(cnrtMemcpy(
            conv_input_mlu_ptr, conv_input_cpu_data.data(),
            conv_input_count * dtype_size, CNRT_MEM_TRANS_DIR_HOST2DEV));
    MGB_CNRT_CHECK(cnrtMemcpy(
            add_input_mlu_ptr, add_input_cpu_data.data(),
            relu_output_count * sizeof(float), CNRT_MEM_TRANS_DIR_HOST2DEV));
    std::unique_ptr<void, decltype(mlu_deleter)> conv_input_holder{
            conv_input_mlu_ptr, mlu_deleter};
    std::unique_ptr<void, decltype(mlu_deleter)> add_input_holder{
            add_input_mlu_ptr, mlu_deleter};
    std::unique_ptr<void, decltype(mlu_deleter)> relu_output_holder{
            relu_output_mlu_ptr, mlu_deleter};
    std::unique_ptr<void, decltype(mlu_deleter)> add_output_holder{
            add_output_mlu_ptr, mlu_deleter};

    network.infer_model(
            {conv_input_mlu_ptr, add_input_mlu_ptr},
            {relu_output_mlu_ptr, add_output_mlu_ptr},
            {Dims{{ni, hi, wi, ci}}, Dims{{no, ho, wo, co}}});

    // result memory copy cnml->cpu
    // memory copy cpu->mlu
    MGB_CNRT_CHECK(cnrtMemcpy(