Re: [PATCH v3 5/7] math: Remove powerpc e_hypot

["Paul A. Clarke via Libc-alpha" <libc-alpha@sourceware.org>]

On Mon, Nov 01, 2021 at 05:20:57PM -0300, Adhemerval Zanella via Libc-alpha wrote:
> The powerpc64 does show a big difference (which is mitigated slightly
> by the last path):
> 
> power9 master:
> 
>   "hypot": {
>    "workload-random": {
>     "duration": 5.41027e+08,
>     "iterations": 3.6e+07,
>     "reciprocal-throughput": 14.6467,
>     "latency": 15.4103,
>     "max-throughput": 6.82746e+07,
>     "min-throughput": 6.48916e+07
>    }
>   }
> 
> power9 patched:
> 
>   "hypot": {
>    "workload-random": {
>     "duration": 5.37776e+08,
>     "iterations": 2.4e+07,
>     "reciprocal-throughput": 19.915,
>     "latency": 24.8997,
>     "max-throughput": 5.02135e+07,
>     "min-throughput": 4.01612e+07
>    }
>   }
> 
> power8 master:
> 
>   "hypot": {
>    "workload-random": {
>     "duration": 5.27104e+08,
>     "iterations": 3.2e+07,
>     "reciprocal-throughput": 15.872,
>     "latency": 17.072,
>     "max-throughput": 6.3004e+07,
>     "min-throughput": 5.85755e+07
>    }
>   }
> 
> power8 patched:
> 
>   "hypot": {
>    "workload-random": {
>     "duration": 5.21972e+08,
>     "iterations": 2.2e+07,
>     "reciprocal-throughput": 21.8583,
>     "latency": 25.5937,
>     "max-throughput": 4.57493e+07,
>     "min-throughput": 3.9072e+07
>    }
>   }

The patches see poorer performance on Power10, also:

Power10 unpatched:

  "hypot": {
   "workload-random": {
    "duration": 5.33875e+08,
    "iterations": 4.8e+07,
    "reciprocal-throughput": 8.47089,
    "latency": 13.7739,
    "max-throughput": 1.18051e+08,
    "min-throughput": 7.2601e+07
   }
  }

Power10 patched:

  "hypot": {
   "workload-random": {
    "duration": 5.34469e+08,
    "iterations": 3.8e+07,
    "reciprocal-throughput": 9.80905,
    "latency": 18.3209,
    "max-throughput": 1.01947e+08,
    "min-throughput": 5.45825e+07
   }
  }

PC