Compat: Refactor fwidth/Fine/Coarse for 0 storage buffers. #4128
Conversation
| pass.setPipeline(pipeline); | ||
| pass.setBindGroup(0, group); | ||
| for (let quad = 0; quad < cases.length / vectorWidth; quad++) { | ||
| for (let quad = 0; quad / vectorWidth; quad++) { |
There was a problem hiding this comment.
This loop condition seems wrong? Looks like the loop never executes because 0 / ... is 0 and that gets coerced to false...
There was a problem hiding this comment.
ha! that's embarrassing 😅
greggman
force-pushed
the
fix-fwidth
branch
2 times, most recently
from
97afdbb to
d43a1ce
Compare
|
Here's a fixed version. Sorry about the previous bug. I originally tried to make it work a quad at a time and I got it working but it was taking 10 seconds per case on my M1 mac allocating 1 buffer per quad to copy the texture and then mapping that 1 buffer etc.. I did some optimizations and got it down to 5 seconds but ran into another issue with that path. Eventually I realized I could do a bunch of quads at once and this is the result. I could optimize more as it's not always using all of the quads now. It uses all 256 quad (2x2 texel areas in a 512x2 texture) for f32 but less for each vectorized version (only 64 for vec4). But, it's fast enough as is, maybe 10% slower than before. |
Modified so this test doesn't use storage buffers by having it return values from a fragment shader as rgba32uint
Modified so this test doesn't use storage buffers by having it return values from a fragment shader as rgba32uint