Platform inconsistency in power calculation about futhark

[athas]

I get the same result on NixOS, so it's probably just how glibc is implemented. I also get that result in the Futhark interpreter, which uses Haskell's pow (but which is probably implemented via the C library). macOS uses a C library developed as part of the LLVM project. I have no reason to know which is more accurate.

Anyway, Futhark intentionally doesn't implement any of these numerical primitives itself, but just dispatches to whatever the platform does. That is overall a good idea, but sometimes this leads to small inconsistencies. You'll see even larger differences once you start running this on GPUs - they really push the allowable bounds on accuracy.

I'll close the issue as I don't think Futhark is doing anything wrong.

[athas]

Are you sure it's a discrepancy in ** and not the other operations (which I guess would have to be the square root)? In any case, Futhark compiles those to pow in the platform C library, which I suppose can vary between platforms.

[Woogachaka]

I would have to assume its in ** as that is the only difference in calculations between xyf2 and xyf2SQ

[athas]

What happens if you say xi*xi+yi*yi instead of xi**2+yi**2, and use f64.sqrt instead of **0.5?

[athas]

I really think it's just a tiny variance in the precision of pow() across different implementation (and I don't know which one is most accurate). Computing

map(\(xi,yi) -> ((xi**2+yi**2)**0.5-((0.5 + tol2)*diameter))) (zip x y)

I see some numbers very close to zero, and it wouldn't take much nudging to push them over the line.

[Woogachaka]

I suspect you're correct in that. When testing your request, I see consistency with the other sets that seem to match. It appears that the failure case (the inconsistency with xf2) is an accumulated error/inconsistency in pow() in ubuntu 22.04

(venv) zerbeba1@TALOS:~/futhark-tracer/examples$ gcc --version
gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.