BC7 determinism about ispctexturecompressor HOT 9 CLOSED

pics or it didn't happen

from ispctexturecompressor.

Few examples as requested:
example.zip

from ispctexturecompressor.

I can't seem to reproduce similar files at all, and I tried all profiles. What exactly did you use?

from ispctexturecompressor.

I used basic profile:
mode_selection: {true, true, true, true}
refineIterations: {2, 2, 2, 2, 2, 2, 2, -858993460}
skip_mode2: true
fastSkipTreshold_mode1: 12
fastSkipTreshold_mode3: 8
fastSkipTreshold_mode7: 0
mode45_channel0: 0
refineIterations_channel: 2
channels: 3

I'm using multiple threads for compression of one texture slicing it to 8 equal parts if possible. Using one thread per texture didn't help.

from ispctexturecompressor.

I've prepared examples done on single thread with no other thread working simultaneously: example_singlethread.zip

from ispctexturecompressor.

I am definitively not getting the same results with the basic profile. Are you sure you are using the last version on the repository?

Also, can you avoid mipmaps in your examples? The main textures in your examples 01, 03 & 04 are exactly the same...

from ispctexturecompressor.

Yes, I'm using the latest version. I've done some changes with initializing local arrays, but it shouldn't make a difference for worse?

As for mipmaps, they just give more examples, I suppose. I've tested that each time mipmaps are generated the same. The difference appears after compression. I'll double check if I didn't miss anything [EDITED: I forgot my own files naming]

Is there any chance that heavy use of multithreading could break something? I looked through the code, and couldn't find anything.

I'll try to prepare some isolated code that will exhibit this nondeterministic behaviour.

from ispctexturecompressor.

I saw one block difference in picture 02, that's it. But what concerns me more is that I can't reproduce your results at all from the .tga file,

Basic profile produces completely different results, with significantly higher quality. For instance:

01_a.dds: RGB psnr = 46.305 dB (01_b is exactly the same data)
basic profile: RGB psnr = 46.778 dB
(using 01.tga as a reference for all)

from ispctexturecompressor.

My bad, I've rebuild the library, but forgot to swap the DLL file. So I was using the old version all along. Everything is fine now.

from ispctexturecompressor.