Method make

src/gopt/impl/inference.cpp:687–1005

}

std::unique_ptr<ConvertF32ToF16Pass> ConvertF32ToF16Pass::make(bool use_f32_comp) {
#if MEGDNN_DISABLE_FLOAT16
    mgb_throw(SystemError, "float16 disabled at compile time.");
#else
    auto replace_h2d_opr = [](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& h2d_opr = opr->cast_final_safe<opr::Host2DeviceCopy>();
        if (h2d_opr.output(0)->dtype() == dtype::Float32()) {
            auto cvt_var = opr::TypeCvt::make(h2d_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

    auto replace_sdt_opr = [](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& sdt_opr = opr->cast_final_safe<opr::SharedDeviceTensor>();
        if (sdt_opr.output(0)->dtype() == dtype::Float32()) {
            auto cvt_var = opr::TypeCvt::make(sdt_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

    auto replace_multi_sdt_opr = [](OperatorNodeBase* opr,
                                    const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& multi_sdt_opr = opr->cast_final_safe<opr::MultipleDeviceTensorHolder>();

        VarNodeArray cvt_vars;
        cvt_vars.reserve(multi_sdt_opr.output().size());

        for (size_t i = 0; i < multi_sdt_opr.output().size(); ++i) {
            if (multi_sdt_opr.output(i)->dtype() == dtype::Float32()) {
                cvt_vars.append({opr::TypeCvt::make(
                                         multi_sdt_opr.output(i), dtype::Float16(), {})
                                         .node()});
            } else {
                cvt_vars.append({multi_sdt_opr.output(i)});
            }
        }
        return cvt_vars;
    };

    auto replace_multi_sdt_with_format_opr = [](OperatorNodeBase* opr,
                                                const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& multi_sdt_with_format_opr =
                opr->cast_final_safe<opr::MultipleDeviceTensorWithFormatHolder>();

        VarNodeArray cvt_vars;
        cvt_vars.reserve(multi_sdt_with_format_opr.output().size());

        for (size_t i = 0; i < multi_sdt_with_format_opr.output().size(); ++i) {
            if (multi_sdt_with_format_opr.output(i)->dtype() == dtype::Float32()) {
                cvt_vars.append({opr::TypeCvt::make(
                                         multi_sdt_with_format_opr.output(i),
                                         dtype::Float16(), {})