The mip math in the compositor uses gl style mips, which means for
width = 2k + 1 and mip level 1, it produces width = k;
The JPEG decoder however in this case, supports decoding to
width = k + 1
This mismatch causes us to use the 1 pixels biggers size and scale it down which is not optimal. We need to figure out what to do in this case.
Comment 1 by vmp...@chromium.org
, Sep 1 2017