Re: [PATCH] math: Improve fmod(f) performance

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi Adhemerval,

> So at least with current 'close-exponents' from bench-fmod, which was
> generated from exponents between -10 and 10, the gain is more modest
> (and normal inputs does show a small regression).  This should be ok, 
> but I also think we need to outline that A72 gains might not show on
> different hardware.

On a SkyLake I'm seeing this for fmod:

                  master   patch
subnormals         51.34    45.92 (+11.8%)
normal             436.9    420.5 (+3.9%)
close-exponents    56.44    53.11 (+6.3%)

And on Zen2:

                  master   patch
subnormals         10.83    10.39 (+4.2%)
normal             336.1    335.8 (+0.01%)
close-exponents    14.90    14.11 (+5.6%)

So it shows good improvements across the board. It's odd your results on AMD are
worse than my Zen 2 results - are there large variations between runs? I did quite a
few runs to get a fast result and increased iterations of the math benchmarks 10x.

I can't explain why the gains on AArch64 are so much larger - the reduced instruction
counts and branches for the common cases seem to make a big difference. On x86
there are still many MOVABS instructions which are problematic for decode.

> So maybe also add another bench-fmod set for |x/y| < 2^12 to show
> the potential gains.

I'm not sure how that would improve things - ideally we need more realistic
inputs (ie. actual traces) but we could change the existing inputs into workloads
to give it a more difficult problem. Changing close-exponents into a workload
shows 11.0% lower latency and 11.9% better throughput on my SkyLake. On Zen 2
I see 1% lower latency and 7.4% better throughput. Neoverse V1 shows 25.1%
lower latency and 23.9% better throughput.

Cheers,
Wilco