So for quite some time I'd been writing a noise/fractal library, and today I decided to revisit it after a many-months break. The first thing I realized was that Perlin noise generation, being a pure function, ought to be embarrassingly parallel. Not having any experience whatsoever in parallelism, I get some weird results I'm having trouble fixing:



The artifacts in the left image are randomly placed each run, and with varying degrees of severity.

The code behind my for-loop is pretty straightforward:

Color[] noisegen(Racket.Module basis)
{
Color[] ret = new Color[col.Length];
int colx, coly;
//sequential
float d;
for (int i = 0; i < col.Length; i++)
{
colx = i % img.Width;
coly = i / img.Width;
d = (float)(basis.Get((colx + xtrans) / (double)img.Width, (coly + ytrans) / (double)img.Width));
ret = new Color(d, d, d);
}
//end-sequential
//Parallel.For(0, col.Length, (i) =>
//{
// float d;
// colx = i % img.Width;
// coly = i / img.Width;
// d = (float)(basis.Get((colx + xtrans) / (double)img.Width, (coly + ytrans) / (double)img.Width));
// ret = new Color(d, d, d);
//});
return ret;
}


I'm not going to post the actual perlin noise function (noted here as basis.Get(...) ) for space economy reasons, but I'm pretty sure it's pure functional. It's pretty much a facelift of Perlin's "improved" 2002 implementation.

I really have no idea what's going on when I use my parallel implementation instead of my sequential one. Is it that the function is returning before some of the threads finish execution? Are parallel for loops not guaranteed to evaluate everything? I feel like I'm missing something very simple.

I scaled down the image to avoid covering up the entire forum post. There's like 50 white pixels, mostly located in the top right. There's no real bias for where they are, it only happens to be that way in that particular image. It seems visible to me, even on this crappy laptop, but I'm the one that uploaded it after all. If you want a larger side-by-side, I can upload it tonight.

EDIT: I should add it becomes a lot more noticeable if you click to enlarge it.

Looks like you're sharing colx and coly between all the threads, creating a race condition. Try making those variables local within the lambda.

That's exactly the kind of thing I figured it'd be. I already fixed that with the d float, but hadn't noticed colx and coly. I'll fix those tonight and see if that solves the problem. Thanks.

EDIT: That fixed the problem. I'd figured it was related to some shared data, but I was having trouble finding it. Thanks again for the help!