Merge CheckAlpha into texture decoding #15474
Conversation
hrydgard
changed the title
| int w, int h, int bufw, bool reverseColors, bool useBGRA, | ||
| u32 *alphaSum, u32 *fullAlphaMask) { |
There was a problem hiding this comment.
nit: Maybe we should use a struct. My general rule of thumb is if I have so many arguments that I'm motivated to put some on another line, then I have too many arguments.
-[Unknown]
There was a problem hiding this comment.
Yeah, it's probably time to clean this up.
There was a problem hiding this comment.
Not gonna bother for now.. I got rid of the new parameters from the calling function DecodeTextureLevel though.
hrydgard
force-pushed
the
sum-alpha-early
branch
from
9c51930 to
fe035dc
Compare
|
OK, this is probably nearly ready to go, barring any more silly bugs. Greatly speeds up decoding of large textures in Vulkan, though there's a bit more to eke out here and there. |
hrydgard
force-pushed
the
sum-alpha-early
branch
from
bde5737 to
a5ee188
Compare
|
Thanks for looking, fixed the issues! This currently also includes #15481 as the first commit, rebased on that. But let's merge that one before this one. |
Fixes a recently discovered performance issue with CheckAlpha - on Vulkan and D3D we write decoded texture data directly to video memory, which can be super slow to read from, depending on driver, platform, phase of moon, etc.
This changes the code to "sum up" (AND together) the colors, and then check against the full alpha mask at the end.
Since we can remove CheckAlpha entirely this should improve performance despite the slight extra checking. And even so, it doesn't increase memory traffic, just some extra arithmetic which rarely is the bottleneck in modern out-of-order CPUs.
Disabled the optimization to deswizzle raw textures directly to VRAM, but no big deal as they're not that common and it's pretty fast anyway. Also, I check the input separately in a few cases to avoid writing more special case functions, so there's still room for minor optimizations.