Function test_subtensor_record

src/gopt/test/no_memory_copy.cpp:388–433

namespace {
void test_subtensor_record(int level) {
    auto test_graph = TestGraph();
    auto compute_graph = test_graph.m_network->graph;
    compute_graph->options().force_output_use_user_specified_memory = true;
    compute_graph->options().comp_node_seq_record_level = 1;
    if (level == 2) {
        test_graph.create_graph_with_setsubtensor();
    } else if (level == 1) {
        test_graph.create_graph_with_subtensor_forward();
    } else {
        test_graph.create_graph_with_subtensor_relayout();
    }
    HostTensorND truth;
    auto func = test_graph.compile_without_copy();
    auto&& outvar = func->get_output_vars()[0];
    DeviceTensorND tmp(test_graph.m_cn, {2, 8, 7, 7});
    outvar->init_mem_plan(&tmp);
    size_t times = 10;
    for (size_t i = 0; i < times; i++) {
        auto input_tensor = test_graph.input_tensor;
        auto layout = input_tensor->layout();
        size_t length = layout.total_nr_elems();
        auto storage = TensorStorage<HostTensorStorageTrait>(test_graph.m_cn);
        storage.ensure_size(length * sizeof(float));
        float* ptr = storage.ptr()->as<float>();
        for (size_t d = 0; d < length; d++) {
            ptr[d] = i / 5 + 3;
        }
        input_tensor->only_reset_raw_storage(storage);
        DeviceTensorND dv(test_graph.m_cn, {2, 8, 7, 7});
        dv.raw_ptr();

        auto& dev_tensor = outvar->mutable_dev_tensor();
        dev_tensor.only_reset_raw_storage(dv.storage());

        func->execute();
        func->wait();
        if (i % 5 == 0) {
            truth.copy_from(dv).sync();
            continue;
        }
        HostTensorND to_check;
        to_check.copy_from(dv).sync();
        MGB_ASSERT_TENSOR_EQ(to_check, truth);
    }
}
}  // namespace

TEST(TestNoCopy, IONoCopyRecordSubTensor) {