Rayon+SIMD support #6
Conversation
|
Awesome work!! 🙏 |
|
I don't foresee broadcast support being merged any time soon, and now that I'm using SIMD it's going to be harder to split things up manually (I'd have to re-implement the work stealing), so I'm going to mark this as ready for review. I'm planning to merge this instead of #4 and #5, since it combines both of them for greater performance gains. |
|
In my testing (owenthewizard/i3lockr@fcd8ccf) rayon > simd+rayon > default > simd. |
Did you test with lanes other than 32? For me, 8 lanes is the fastest - lower or higher is slower. YMMV, of course, I'm on 10th gen Intel. |
|
That's what I get for doing this at 5:00 A.M.... |
Interesting, for me rayon+simd is about 5-10% faster. It's a very small speedup in all honesty though, so I'm still on the fence about actually merging in SIMD (I might just go back to the plain |
Using the new SIMD iterators in `imgref-iter`, it's possible to implement SIMD version of stackblur. Right now SIMD doesn't offer too much of a benefit, but it is still a benefit nonetheless, and rayon definitely offers a huge benefit. This has been sitting around in PR hell for far too long, and I don't like managing multiple disparate branches, so it's about time this gets merged and released.
Combines #4 and #5
Closes #3
Using both at the same time now actually outperforms
rayonon its own.This is surprising.
Currently a lane count of 8 is the fastest on my machine, but this is tunable.
This is all thanks to
imgref-iter's experimental SIMD iterators.