More flexible texture matching, bias matching by the bind sequence counters #15857
Conversation
| candidates.push_back(AttachCandidate{ match, entry, framebuffer, RASTER_COLOR, framebuffer->colorBindSeq }); | ||
| } | ||
| if (gstate_c.Supports(GPU_SUPPORTS_DEPTH_TEXTURE) && MatchFramebuffer(entry, framebuffer, texAddrOffset, RASTER_DEPTH, &match)) { | ||
| candidates.push_back(AttachCandidate{ match, entry, framebuffer, RASTER_DEPTH, framebuffer->depthBindSeq }); |
There was a problem hiding this comment.
I think we might want to swap these lines:
I'd prefer to give color slight preference if they were bound right next to each other (I realize you checked for the overlap case, but I think it's a general thing.)
Also, I wonder if we should be trying to bind a null texture on !gstate_c.Supports(GPU_SUPPORTS_DEPTH_TEXTURE) rather than going to RAM... it feels potentially more consistent. Not sure.
-[Unknown]
There was a problem hiding this comment.
Those lines will be separated (and in your preferred order) in #15858 :)
Yes, binding a null texture makes more sense, I'll do that.
|
Someone reported on Discord that this broke the menu (new game / load game / etc.) of Crisis Core. Haven't had time to check into it. -[Unknown] |
Since we now have the bind sequence numbers, which should almost on its own be a fairly solid way of picking the right framebuffer (or its depth buffer) to texture from, I originally intended to simplify things a lot more, but decided to let the pile of heuristics in MatchFramebuffer live on for now, at least until we get deferred depth copies (and some will still be needed).
In addition to adding a match bias from the sequence number (so the most recent texture is more likely to be chosen if two are very similar), this always tries to match against both color and depth, instead of deciding based on the "depth swizzle" bits in the address. Turns out that games like Kuroyou will texture color from a "depth address" in order to apply the depth swizzle. Since we just ignore swizzle entirely, that's just a copy for us, but it still needs to happen. Some more details here.
The reason I want this in first is that without it, I can't fully test #15777 with the dumps from the game - and splitting it out will make bisection easier, as usual.