Truffle further speed up (10-40%)

ypicchi-arm commented 5 months ago

Those three commits enable three different optimizations to truffle:

Make the neon path explicit instead of using the supervector (~10% speedup over generic neon path)
Enable hybrid neon/SVE path. This enable the explicit neon code for the comparison while using the SVE code for returning the match index. (~10% speedup over regular SVE)
Optimization for wide SVE (256+ bit wide) by simplifying the algorithm. It have its own code path but provide ~40% speedup compared to the normal SVE path over the same wide machine (graviton 3), making truffle the fastest nfa algorithm in the microbenchmark.

markos commented 5 months ago

All builds failed on CI right now, but even the few ones that were successfull in compiling, failed in the internal tests:

https://buildbot-ci.vectorcamp.gr/#/changes/192

But I'm sceptical about accepting this, it is invasive and breaks the builds for all other arches, plus the extra 10% you get from the Neon rewrite is not because of ditching the SuperVector code. The same effect could be had by just fixing that part. I don't mind adding the extra SVE code as currently it's hard to abstract SVE code. However, that does not apply to ASIMD/Neon. All SIMD code will be converted to use the SuperVector class.

markos commented 4 months ago

This has been merged in #290 closing

VectorCamp / vectorscan

Truffle further speed up (10-40%) #266