>looked closely and I think what you did is perfect for GPU
Really? The random series generation is pretty straight forward. But the permutation stage is all out-of-order memory access. And the decrypt stage is very heavy in conditional branching.
>still need more target images
https://pastebin.com/Mj4d1jXM
It doesn't look like spidering through chan sites looking for PK images is going to work. 8chan has started reencoding again. And I tested halfchan: they don't fully reencode, but do add the missing JFIF header. Game forums are likely going to be the same way.
So until Q points at a image and says, "This pic contains a hidden message and was used by No Name to arrange a weapons sale. Have at it boys!" then I don't think there is anything left for us to do with this.
>could you zip both and post?
https://nofile.io/f/RjoJ4qr4dXX/4chan_transmogrification.zip.zip
Same DQT and DHT chunks. And the image scan is (or at least starts out) identical. They must have some little script that slips in the APP0 chunk if it's missing.
Wow. How does this compare with the CPU alone? I know you have a monster of a system.