Improve extreme color calculation for DXT compress

[kaetemi]

Take average color, then use vector math to get the furthest color from the average, then again find the furthest color from the furthest color.

[kaetemi]

Both DXT1a and DXT5 tested and working.

Before:

After:

[kaetemi]

Slightly better with weighted colors for distance calculation.

[riperiperi]

Thanks for having a look at this issue. Two quick notes before I give it a review later (not at a pc for a while):

• Indentation looks incorrect in the diff view, it should be 4 spaces on this project instead of tabs.
• This code is currently performance critical when the texture recompression is enabled, as it happens on the render thread the first time the texture is accessed. I don’t want to make it any slower, so ideally it shouldn’t exhaustively search for endpoints if it doesn’t already. I wrote this ages ago, so I can’t remember its current complexity.
  • I’d be willing to trade off performance if this were done asynchronously or had some kind of cache.

EDIT: looks like you fixed indentation after i wrote this message, but before it posted 🥴 i’ll take a look at the code when i can.

[kaetemi]

Previous code does one iteration and checks for both minimum and maximum in that pass in each color individually. (Which leads to completely incorrect color mixes when the endpoints are different colors. E.g. green to blue becomes black to cyan.)

New code does one iteration to find the average, one iteration to find the furthest, and another iteration to find the opposing endpoint.

Both then iterate again to find which of the 4 colors in the endpoint gradient is nearest to each pixel.

Agree a cache would be good eventually. Haven't noticed any performance drop, at least, and looks better. Could put a performance timer on it to verify.

[kaetemi]

Somehow faster on my machine. Ryzen 5 2600.

new code: Total time: 598ms Total blocks: 748286 Time per million blocks: 799ms

old code: Total time: 712ms Total blocks: 747132 Time per million blocks: 952ms

New code (second run): Total time: 641ms Total blocks: 746612 Time per million blocks: 858ms

Old code (second run): Total time: 780ms Total blocks: 754988 Time per million blocks: 1033ms

New code (third run): Total time: 747ms Total blocks: 773352 Time per million blocks: 965ms

Old code (third run): Total time: 851ms Total blocks: 778296 Time per million blocks: 1093ms

[kaetemi]

Alright, sounds good.

I don't believe the loops can be combined? The loop depends on the outcome of the previous loop.

On Sun, Dec 31, 2023, at 00:26, riperiperi wrote:

@.**** approved this pull request.

All around great work. I can merge without combining the loops if you don't have the time, just let me know.

In TSOClient/tso.common/Utils/TextureUtils.cs https://github.com/riperiperi/FreeSO/pull/276#discussion_r1438643889:

• if (col.R > maxCol.R) maxCol.R = col.R;
• if (col.G > maxCol.G) maxCol.G = col.G;
• if (col.B > maxCol.B) maxCol.B = col.B;
• // Find color furthest from average
• int leftDist = 0;
• int leftIdx = 0;
• for (int i = 0; i < 16; i++)

This loop can be combined with the next one. That would skip duplicating the loop overhead, and let you do the fetch from block[i] into local variable once instead of (at least) twice.

— Reply to this email directly, view it on GitHub https://github.com/riperiperi/FreeSO/pull/276#pullrequestreview-1799569710, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAMB77KSTTR42V47C6BSTJDYMA6D7AVCNFSM6AAAAABA4PS5J2VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTOOJZGU3DSNZRGA. You are receiving this because you authored the thread.Message ID: @.***>

[riperiperi]

Ah, I misread it - thought left col wasn't used. Looks good to me then.