Re: [PATCH v2] x86-64: Optimize bzero

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Thu, Feb 24, 2022 at 4:58 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Thu, Feb 24, 2022 at 7:16 AM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
> >
> > Hi Noah,
> >
> > > The performance improvement for the build/link step for varying amount of
> > > small functions per file was consistently in the ~.8% range. Not mind blowing
> > > but I believe its a genuine improvement.
> >
> > I don't see how it's possible to get anywhere near 0.8%. I tried compiling a file with
> > 10000 empty functions, and the latest __memset_exex_unaligned_erms takes about
> > 1.16% of total time.
>
> Smart sanity check I'll start using.
>
> What method are you using for getting total function call overhead?
>
> Using `perf record` + `pref report` and see a fair amount of variance but much
> higher memset overheader (counting `_*unaligned_erms` and `*_unaligned`)
> in `cc1` and `as`.
>
> From average of 3 runs compiling file with 1/10/100/1000 functions I get:
>
> 1: 4.04%
> 10: 3.94%
> 100: 2.86%
> 1000: 2.68%
>
> So its slightly less insane, arguing for the following speedups:
>
> 1: ~15%
> 10: ~15%
> 100: ~25% <--- this makes little to no sense
> 1000: ~15%
>
> personally agree with you that those numbers seem to high though.
>
> In the best case micro-benchmark that stressed the p5 bottleneck
> this is about what we see.
>
> >
> > There are 81.5 million calls to memset in 48 billion cycles for this benchmark. That
> > means 6.8 cycles per memset call on average. A 0.8% speedup would require making
> > each memset 4.7 cycles faster, and that's not possible with bzero.
> >
> > To verify whether vpbroadcastb is a bottleneck I repeated it 16 times. This increased
> > the memset percentage to 1.86%, however the total cycles didn't change measurably.
> >
> > I'm not sure how you're measuring this, but it's clear what you're seeing is not a
> > speedup from bzero.

Bah, youre right. The the non-memsetzero GCC  was choosing my systems
memset implementation (2.31 + avx2).

Sorry, I'll rerun tonight and post an update.
> >
> > Cheers,
> > Wilco