Method do_exec

dnn/test/common/checker.cpp:222–335

}

void CheckerHelper::do_exec(
        const TensorLayoutArray& user_layouts, const TensorLayoutArray& deduced_layouts,
        const OprExec& exec_naive, const OprExec& exec_opr) {
    m_prev_succ = false;

    // check if user provided layouts are correct
    for (size_t i = 0; i < deduced_layouts.size(); ++i) {
        if (user_layouts[i].ndim > 0) {
            ASSERT_TRUE(deduced_layouts[i].eq_shape(user_layouts[i]))
                    << "User provided shape is "
                    << user_layouts[i].TensorShape::to_string()
                    << "\nExpected shape is "
                    << deduced_layouts[i].TensorShape::to_string();
        }
    }
    auto layouts = user_layouts;
    for (size_t i = 0; i < layouts.size(); ++i) {
        if (layouts[i].ndim == 0) {
            //! in some opr, such as conv_bias has ndim==0
            layouts[i] = deduced_layouts[i];
        }
    }

    // allocate
    m_tensors_naive = alloc_tensors(m_handle_naive.get(), layouts, m_offset);
    auto tensors_cur_storage = alloc_tensors(m_handle_cur, layouts, m_offset);
    auto tensors_cur_host_storage =
            alloc_tensors(m_handle_naive.get(), layouts, m_offset);
    auto&& tensors_naive = *m_tensors_naive;
    auto&& tensors_cur = *tensors_cur_storage;
    auto&& tensors_cur_host = *tensors_cur_host_storage;
    std::shared_ptr<TensorValueArray> tensors_extra_opr_impl;
    if (m_extra_opr_impl) {
        tensors_extra_opr_impl = alloc_tensors(m_handle_naive.get(), layouts, m_offset);
    }

    init_naive_values();

    copy_tensors_to_device(tensors_cur, tensors_naive);
    if (m_extra_opr_impl) {
        copy_tensors(*tensors_extra_opr_impl, tensors_naive);
    }

    // execute

    exec_opr(tensors_cur);
    if (m_expect_exec_fail) {
        m_expect_exec_fail();
        m_expect_exec_fail = {};
        return;
    }
    if (m_stable_check) {
        auto tensors_bak_host_storage =
                alloc_tensors(m_handle_naive.get(), layouts, m_offset);
        auto&& tensors_bak_host = *tensors_bak_host_storage;
        copy_tensors_from_device(tensors_bak_host, tensors_cur);
        for (int i = 0; i < 10; i++) {
            exec_opr(tensors_cur);