1.) I am optimising for performance, so i’d like to know how is padding (mode) implemented in the C++ engine - is padding in convolution all implicit, with indexing for “reflect” or “same”, or is every pad performed first, and then convolved upon?
2.) in the same vein, as i assumed that for efficiency in not copying potentially large tensors it is implemented with indexing, is there a “padding_mode” variant of the functional call?
or would the performance be about the same explicitly first calling f.pad, and then f.conv2d?