Function run_permutation

dnn/test/cuda/rng.cpp:240–268

template <typename T>
void run_permutation(Handle* handle) {
    using ctype = typename DTypeTrait<T>::ctype;
    size_t sample_num = std::min(200000, static_cast<int>(DTypeTrait<T>::max()) - 10);

    auto opr = handle->create_operator<PermutationRNG>();
    opr->param().dtype = DTypeTrait<T>::enumv;
    TensorLayout ly{TensorShape{sample_num}, T()};
    Tensor<dt_byte> workspace(
            handle, {TensorShape{opr->get_workspace_in_bytes(ly)}, dtype::Byte()});
    SyncedTensor<ctype> t(handle, ly);

    opr->exec(t.tensornd_dev(), {workspace.ptr(), workspace.layout().total_nr_elems()});

    auto ptr = t.ptr_mutable_host();
    auto size = t.layout().total_nr_elems();

    std::vector<ctype> res(size);
    int not_same = 0;
    for (size_t i = 0; i < size; ++i) {
        if ((ptr[i] - ctype(i)) >= ctype(1))
            not_same++;
        res[i] = ptr[i];
    }
    ASSERT_GT(not_same, 5000);
    std::sort(res.begin(), res.end());
    for (size_t i = 0; i < size; ++i) {
        ASSERT_LE(std::abs(res[i] - ctype(i)), 1e-8);
    }
}

template <typename T>
void run_shuffle(Handle* handle, bool bwd_flag) {