Function apply_on_physical_tensor

imperative/src/impl/ops/elemwise.cpp:101–135

}

SmallVector<TensorPtr> apply_on_physical_tensor(
        const OpDef& def, const SmallVector<TensorPtr>& inputs,
        SmallVector<LogicalTensorDesc>& output_descs, const bool& validated) {
    auto comp_node = inputs[0]->comp_node();
    auto dtype = inputs[0]->dtype();
    using Mode = Elemwise::Mode;
    auto&& op_def = def.cast_final_safe<Elemwise>();
    auto mode = op_def.mode;
    TensorShapeArray input_shapes;
    input_shapes.reserve(inputs.size());
    for (auto&& input : inputs) {
        input_shapes.push_back(input->shape());
    }
    // deduce_shape is static and fast
    TensorLayout output_layout{dtype};
    // TODO: deduce_layout directly
    megdnn::Elemwise::deduce_shape(input_shapes, output_layout);
    output_layout.init_contiguous_stride();
    auto output = Tensor::make(output_layout, comp_node);
    if (output_layout.is_empty()) {
        return {output};
    }
    DnnOprCaller<megdnn::Elemwise> dnn_opr(comp_node, op_def.param());
    if (mode == Mode::FUSE_MUL_ADD3 || mode == Mode::FUSE_MUL_ADD4 ||
        dtype.category() == DTypeCategory::QUANTIZED) {
        dnn_opr.call_dnn(
                [&](auto&& inputs, auto&& output) {
                    opr::Elemwise::perform_dnn(comp_node, output, inputs, dnn_opr.op());
                },
                inputs, output);
    } else {
        dnn_opr.exec(inputs, output);
    }
    return {output};
}

MGB_DEFINE_OPR_CLASS(
        ForceInplaceElemwise,