Function run_gamma

dnn/test/cuda/rng.cpp:13–42

template <typename T>
void run_gamma(Handle* handle) {
    using ctype = typename DTypeTrait<T>::ctype;
    auto opr = handle->create_operator<GammaRNG>();

    TensorLayout ly{TensorShape{2000000 * 5}, T()};

    SyncedTensor<ctype> out(handle, ly);
    SyncedTensor<ctype> shape(handle, ly);
    SyncedTensor<ctype> scale(handle, ly);
    auto shape_ptr = shape.ptr_mutable_host();
    auto scale_ptr = scale.ptr_mutable_host();
    for (int i = 0; i < 5; ++i) {
        for (int j = 0; j < 2000000; ++j) {
            shape_ptr[i * 2000000 + j] = 2 * 0.3 * i + 0.3;
            scale_ptr[i * 2000000 + j] = i * 0.2 + 0.1;
        }
    }

    opr->exec(shape.tensornd_dev(), scale.tensornd_dev(), out.tensornd_dev(), {});

    auto ptr = out.ptr_mutable_host();
    for (int i = 0; i < 5; ++i) {
        float a = 2 * 0.3 * i + 0.3, b = i * 0.2 + 0.1;
        float mean = a * b;
        float std = a * (b * b);
        auto stat = get_mean_var(ptr + i * 2000000, 2000000, ctype(mean));
        ASSERT_LE(std::abs(stat.first - mean), 0.01);
        ASSERT_LE(std::abs(stat.second - std), 0.01);
    }
}

template <typename T>
void run_poisson(Handle* handle) {