Improve the PRNG for the particles sample

[ben-clayton]

The trig / fract approach for PRNG is not portable and can produce remarkably non-random results on IMG GPUs.

Replace this with a bit-xor & shift function, adapted from the cornell sample.

Fixes the particles sample on Pixel 10.

[Kangz]

🎅

[toji]

Thank you! This has been on my list of things to look at for a bit but it's never been a high priority.

For posterity, do you have a brief explanation as to why the previous approach wasn't portable?

[ben-clayton]

For posterity, do you have a brief explanation as to why the previous approach wasn't portable?

TL;DR: my theory is bad trig precision resulted in a terrible hash function.

I haven't done any deep digging, but I would suspect it's about the precision of the trig functions. IIRC (it's been a while since my head has been in this space) the trig functions have quite a lot of leniency WRT the precision of their outputs, and I wouldn't be surprised if some GPUs used crude approximations (LUTs and the like) which scrape-by the API specification minimum thresholds.
If the trig functions end up quantizing nearby input values to the same output value, then pseudo-hashing via rounds of trig -> fract are likely going to quantize to discrete points and likely miss entire output value ranges. This is what was exactly observed with the particle probability map - spatial gaps where no particles were spawned, and instead thousands of particles spawning at the same discrete points on the plane.