Webgpu variable length buffer - #9086

Continuous Release

CDN link

• https://raw.esm.sh/pr/p5@8432888/lib/p5.min.js

Published Packages

• https://pkg.pr.new/p5@8432888

Commit hash: 8432888

This is an automated message.

@aashu2006 lmk what you think of the design decisions here!

@davepagurek I had a read, the indirect draw path is really nice copying the atomic length straight into instanceCount and never round tripping to the cpu is exactly the right call

two things that look like bugs

instances(storageBuffer) uses .size not .length

3d_primitives.js:2704 does count = count.size, which is the 16 byte aligned byte length, not the element count. .length is right there on the constructor so looks like a slip. The visual test uses a struct buffer so it might still land onscreen, a flat Float32Array one would show it right away

overflow counter reaches the indirect draw unclamped

_p5_push_* does atomicAdd unconditionally and only guards the store, so the counter climbs past maxCapacity even though nothing gets written. read() clamps with Math.min(..., maxCapacity), but the copyBufferToBuffer into instanceCount copies the raw counter. So an overflowing list draws more instances than exist and reads uninit data. read and draw should probably agree here

heads up on #9083

this adds length as the 5th constructor param on StorageBuffer, mine adds arrayType in the same slot. whichever lands second clobbers the other

the spreading fix

that one's on me. the stride used px but the real dispatch width is ceil(px/8)*8, so rows overlapped whenever ceil(sqrt(N)) wasn't a multiple of 8. fix looks right and N=2500 hits it squarely

Thanks @aashu2006, good catch on the .size/.length thing!

ok I finally got around to putting up some fixes, @aashu2006 lmk what you think!