Method scn_do_execute

src/serialization/impl/extern_c_opr.cpp:383–430

}

void ExternCOprRunner::scn_do_execute() {
    SmallVector<MGBTensor> c_inp(input().size()), c_out(output().size());
    SmallVector<HostTensorND> cpu_inp, cpu_out;
    check_param();

    bool need_copy = false;
    if (comp_node().device_type() == CompNode::DeviceType::CPU) {
        for (size_t i = 0; i < input().size(); ++i) {
            c_inp[i] = tensor_to_c(input(i)->dev_tensor());
        }
        for (size_t i = 0; i < output().size(); ++i) {
            c_out[i] = tensor_to_c(output(i)->dev_tensor());
        }
    } else {
        need_copy = true;
        mgb_log_debug(
                "copy is needed to execute extern C "
                "opr `%s' on comp node `%s'",
                cname(), comp_node().to_string().c_str());
        cpu_inp.resize(input().size());
        cpu_out.resize(output().size());
        for (size_t i = 0; i < input().size(); ++i) {
            cpu_inp[i].copy_from(input(i)->dev_tensor());
            c_inp[i] = tensor_to_c(cpu_inp[i]);
        }
        for (size_t i = 0; i < output().size(); ++i) {
            cpu_out[i]
                    .comp_node(comp_node())
                    .dtype(output(i)->dtype())
                    .resize(output(i)->shape());
            c_out[i] = tensor_to_c(cpu_out[i]);
        }
    }

    if (need_copy) {
        comp_node().sync();
        m_desc->execute(m_desc.get(), c_inp.data(), c_out.data());

        for (size_t i = 0; i < output().size(); ++i)
            output(i)->dev_tensor().copy_from_fixlayout(cpu_out[i]).sync();
    } else {
        CompNodeEnv::from_comp_node(comp_node())
                .cpu_env()
                .dispatch([this, c_inp, c_out]() mutable {
                    m_desc->execute(m_desc.get(), c_inp.data(), c_out.data());
                });
    }
}

void ExternCOprRunner::add_input_layout_constraint() {
    for (auto i : input())