Function run_beta

dnn/test/cuda/rng.cpp:208–237

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

    TensorLayout ly{TensorShape{200000 * 5}, T()};

    SyncedTensor<ctype> out(handle, ly);
    SyncedTensor<ctype> alpha(handle, ly);
    SyncedTensor<ctype> beta(handle, ly);
    auto alpha_ptr = alpha.ptr_mutable_host();
    auto beta_ptr = beta.ptr_mutable_host();
    for (int i = 0; i < 5; ++i) {
        for (int j = 0; j < 200000; ++j) {
            alpha_ptr[i * 200000 + j] = 0.3 * i + 0.1;
            beta_ptr[i * 200000 + j] = 2 * i * 0.3 + 0.1;
        }
    }

    opr->exec(alpha.tensornd_dev(), beta.tensornd_dev(), out.tensornd_dev(), {});

    auto ptr = out.ptr_mutable_host();
    for (int i = 0; i < 5; ++i) {
        float a = 0.3 * i + 0.1, b = 2 * i * 0.3 + 0.1;
        float mean = a / (a + b);
        float std = a * b / ((a + b) * (a + b) * (a + b + 1));
        auto stat = get_mean_var(ptr + i * 200000, 200000, 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_permutation(Handle* handle) {