Hector Martin

@marcan42 ... oh nevermind, that's 589824 bytes of coefficients. gods help you

@FioraAeterna Yeah, I tried a thing where I loaded some pixels and some coefs to shared mem to optimize for reuse, but it was MUCH slower.

@marcan42 it sounds like it may have destroyed occupancy because of limited shared memory per thread group

@FioraAeterna Yeah, I think I had 120 workitems per workgroup sharing local mem, which I guess might not have been enough?

@marcan42 my intuition in this case is you'd want to keep register usage low to ensure enough latency-hiding for the endless memory ops

@marcan42 ... but you might already be maxing out bandwidth and be doomed anyways ;-;

@FioraAeterna Well unless the CUDA version does magic there's clearly *some* way of speeding this up...

@marcan42 do you know if the CUDA version is structurally the same (i.e. same way of dividing up the problem)?