This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH v2.1] Use long for mantissa for generic mp code
On 15 March 2013 19:45, Richard Henderson <rth@twiddle.net> wrote:
> For the record, this seems to have sped up the benchmarks
> by a factor of 10-20 on arm:
>
> Before
>>
>> exp: ITERS:100000: TOTAL:19.2022s, MAX:266889ns, MIN:190590ns, 5207.73
>> iter/s
>> pow: ITERS:100000: TOTAL:47.5425s, MAX:482874ns, MIN:474051ns, 2103.38
>> iter/s
>
>
> After
>>
>> exp: ITERS:100000: TOTAL:2.6179s, MAX:26460.7ns, MIN:26068.7ns, 38198.5
>> iter/s
>> pow: ITERS:100000: TOTAL:2.1143s, MAX:21386.5ns, MIN:21019.5ns, 47296.9
>> iter/s
>
>
> which looks fantastic, except the pow test running faster than the exp test,
> which seems implausible.
Is Before and After on the same code with the different benchmark
patches (one I posted vs what I checked in) or is it a comparison
between 2.17 and latest upstream? If it's the latter than it's the
patches I've put in so far on libm that could be the reason. They
make things ~5x faster on x86_64.
> Did you say something about unused outputs in the benchmark between v2 and
> v2.1? Did you check in the wrong version?
I am pretty sure I checked in the right version because I rebased my
WIP branch right after and it didn't complain of any inconsistency.
Also, with the call optimized out, each run should take less than 1ns
to complete (at least that's what happened to me on x86_64). This
doesn't seem to be the same thing.
Siddhesh
--
http://siddhesh.in