Re: RFC: Improve hypot performance

["Paul A. Clarke" <pc@us.ibm.com>]

On Thu, Nov 18, 2021 at 09:43:12AM -0600, Paul A. Clarke via Libc-alpha wrote:
> On Thu, Nov 18, 2021 at 12:37:24PM +0000, Wilco Dijkstra wrote:
> > > On Power10, this implementation is still has a large delta compared to the
> > > current implementation:
> > 
> > It's obvious something went wrong here since it is slower than Adhemerval's
> > version - did you look at the generated code?
> > 
> > It should use fma, and no division or function calls (if you see calls to fmin/fmax
> > you need to set FAST_FMINMAX to 0).
> 
> Ugh. That was it. I missed that subtlety. (Why would I not want "FAST" to be
> true?  :-)
> 
> Anyway, with FAST_FMINMAX set to 1, the results are much better, and compare

That should say "with FAST_FMINMAX set to *0*".

> favorably with the current implementation:
> 
> current:
>   "hypot": {
>    "workload-random": {
>     "duration": 5.26088e+08,
>     "iterations": 4.8e+07,
>     "reciprocal-throughput": 8.2599,
>     "latency": 13.6604,
>     "max-throughput": 1.21067e+08,
>     "min-throughput": 7.32042e+07
>    }
>   }
> 
> with Adhemerval's patch and this patch with FAST_FMINMAX=1:
>   "hypot": {
>    "workload-random": {
>     "duration": 5.30541e+08,
>     "iterations": 5.2e+07,
>     "reciprocal-throughput": 7.22461,
>     "latency": 13.1808,
>     "max-throughput": 1.38416e+08,
>     "min-throughput": 7.58679e+07
>    }
>   }
> 
> PC