You work is awesome and the speed on Apple devices with ANE is blasting fast. However, AFAIK and reading the paper, 3x3 depthwise convs help the ANE to parallelize alot of ops, which are also more tips/tricks to make it fast on the ANE?

Thanks!