Predicting memory usage?

[JeremyRand]

If I have a given bin+param file for a model, and I know the shape of the input, is there a straightforward way to compute (or at least get a close estimate) how much memory the inference will consume, without actually executing the inference? If so, is there also a straightforward way to compute the memory consumption impact of enabling/disabling Winograd convolution (or other optimizations that are configurable in ncnn)?