I started writing some utilities to decode ATF files and I spent time learning about the compressed format according to https://www.adobe.com/devnet/archive/flashruntimes/articles/atf-file-format.html. So, I generated two different ATF files from the same PNG image. One was generated with format 2 (ATFRAWCOMPRESSED) and one with format 3 (ATFCOMPRESSED).
Given that this is supposed to be lossless, I should be able to take the PVRTC data from the second file and it should match the data from the first one. So, after getting the LZMA and JPEG-XR working properly, I got it match except on a few locations. After examining https://github.com/firefalcom/dds2atf/blob/master/pvr2atfcore.cpp#L1269, I can tell from the data I recovered that the code is indeed storing the lower bit of color 0 into the PVRTCTopData but then it stores color 1 as 5:5:4. This means that the original data cannot be 100% fully recovered.
According to the PVRTC specification http://cdn.imgtec.com/sdk-documentation/PVRTC%20Specification%20and%20User%20Guide.pdf, color 0 (color Ashould be 5:5:4, not color 5:5:4 (see figure 9 and table 1 below 1).
Any thoughts? Just an odditiy and even though we don't have the source for png2atf, it seems to match the behavior of the old version of dds2atf.
AdChoices