Re: [PATCH 7/9] math: Remove powerpc e_hypot

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

On Wed, Oct 06, 2021 at 02:51:50PM -0500, Paul E Murphy via Libc-alpha wrote:
> On 10/6/21 2:46 PM, Adhemerval Zanella wrote:
> > On 06/10/2021 16:43, Paul E Murphy wrote:
> > > On 10/6/21 1:05 PM, Adhemerval Zanella via Libc-alpha wrote:
> > > > power9 master:
> > > > 
> > > >     "hypot": {
> > > >      "workload-random": {
> > > >       "duration": 5.20744e+08,
> > > >       "iterations": 5.6e+07,
> > > >       "reciprocal-throughput": 9.27604,
> > > >       "latency": 9.32194,
> > > >       "max-throughput": 1.07805e+08,
> > > >       "min-throughput": 1.07274e+08
> > > >      }
> > > >     }
> > > > 
> > > > power9 patched:
> > > > 
> > > >     "hypot": {
> > > >      "workload-random": {
> > > >       "duration": 5.23124e+08,
> > > >       "iterations": 5.4e+07,
> > > >       "reciprocal-throughput": 9.51892,
> > > >       "latency": 9.85606,
> > > >       "max-throughput": 1.05054e+08,
> > > >       "min-throughput": 1.0146e+08
> > > >      }
> > > >     }
> > > > 
> > > > Checked on powerpc64-linux-gnu (power8) and powerpc64le-linux-gnu
> > > > (power9).
> > > > ---
> > > 
> > > This also removes the ppc specific hypotf too. Have you had a chance to measure the difference?
> > > 
> > > I think it is OK if the performance difference is negligible. Though, I don't speak for the other ppc contributors in that regard. They may want to double-check the results.
> > 
> > I would expect the performance difference to be similar to 64-bit,
> > at least on the hardware I checked (since most of the time would
> > be spent in FPU calculations).
> > 
> > In any case, I don't see much gain in keeping a 32-bit powerpc
> > implementation that deviates from generic implementation.
> > 
> 
> Oops, I read this patch in isolation to the series. There probably isn't any
> difference. Sorry for the noise.
> 
> Anyhow, in the context of the series, this looks OK to me.

[Accidentally dropped the mailing list from this post. Resending...]

The results above showed over 5% performance degradation.
I'm not sure how rigorous those results are, and while arguably "close",
they were a bit concerning.

I ran the results on a local Power9 system, and see much better performance
with the new code. I thought I should share in case anyone else had similar
concerns.

glibc-master on Power9 with patch 1/9 (to get "workload-random"):
  "hypot": {
   "workload-random": {
    "duration": 5.1525e+08,
    "iterations": 8e+07,
    "reciprocal-throughput": 6.2956,
    "latency": 6.58564,
    "max-throughput": 1.58841e+08,
    "min-throughput": 1.51846e+08
   }
  }

glibc-master on Power9 with patches 1-9/9:
  "hypot": {
   "workload-random": {
    "duration": 5.20747e+08,
    "iterations": 9.8e+07,
    "reciprocal-throughput": 5.10696,
    "latency": 5.52053,
    "max-throughput": 1.95811e+08,
    "min-throughput": 1.81142e+08
   }
  }

Tested-by: Paul A. Clarke <pc@us.ibm.com>

PC