softgpu: Round tex coords properly for nearest #8282
Conversation
Fixes off-by-one errors in rendering certain textures.
softgpu: Round tex coords properly for nearest
| int u = (int)(s * width); | ||
| int v = (int)(t * height); | ||
| int u = (int)(s * width + 0.375f); | ||
| int v = (int)(t * height + 0.375f); |
There was a problem hiding this comment.
Turns out this is wrong. In fact, I think we are half a pixel off for bilinear (i.e. a 0,0 -> 1,1 box will apply 50% of texel 1,1 in the center, which is of course wrong.)
-[Unknown]
There was a problem hiding this comment.
That's interesting - I hadn't considered the possibility that the offset might need to be different in each of the filtering modes - who knows how they are implemented.
There was a problem hiding this comment.
Well, AFAICT it's just 0 for nearest (seems to round right at 0.5) and -0.5 for linear (to sample at the centers.) Need to check some throughmode handling still when I have time.
I did try applying a 0.375f / 256.0f offset to correct for rounding issues, which seemed to largely help most everywhere (including Crisis Core.) And with that the Hexyz issue doesn't happen either (probably the same rounding issue.) However, it introduces other issues...
Also tex wrap with negative float uvs was wrong.
-[Unknown]
Fixes #4853 - off-by-one errors in rendering certain textures.
This rounds up at 0.625, which I think is the correct behavior, from my tests of the PSP's GPU in general. It fixes the examples I had collected of this issue, anyway.
Before, it was always rounding down.
-[Unknown]