This is the mail archive of the
mailing list for the glibc project.
Re: [PATCH 00/10] Optimized math routines
On 09/07/2018 09:15, Szabolcs Nagy wrote:
> On 06/07/18 18:17, Szabolcs Nagy wrote:
>> On 06/07/18 17:27, Carlos O'Donell wrote:
>>> On 07/06/2018 11:46 AM, Szabolcs Nagy wrote:
>>>> On 06/07/18 13:43, Carlos O'Donell wrote:
>>>>> On 07/06/2018 04:47 AM, Szabolcs Nagy wrote:
>>>>>> Optimized exp, exp2, log, log2, pow, sinf, cosf and sincosf
>>>>> Is it your intent to have these included in 2.28?
>>>> (resending as my previous mail seems to be lost)
>>>> yes, i'd like to add it to the 'desirable in 2.28' list
>>>> if Joseph is ok with the code, but i see he is not available
>>>> right now for review.
>>>> i don't know how other maintainers feel about such change,
>>>> there needs to be an ulp update (i'm willing to do that for
>>>> targets i can access hw for testing).
>>> Where there any unanswered questions in your v4 review?
>>> Do you think v4 is basically as good as it will get?
>>> Who were the people who signed off on the review?
>> Joseph Myers started the review of both the sinf, cosf, sincosf
>> changes and the exp, exp2, log, log2, pow changes.
>> I think I addressed all of his comments in an acceptable way,
>> but i don't know if he had other concerns or if parts of the
>> code he has not reviewed yet.
>> Since the glibc tests pass on 3 different targets (and
>> build-many-glibcs.py) i think there is no danger of the
>> patch being completely broken. Wilco and I tested the patches
>> in detail outside of glibc so it is the glibc integration where
>> I expected most of the issues.
>> I don't expect performance regression on any target, but it
>> was not measured e.g. on powerpc (only aarch64 and x86_64)
>> which might have different behaviour (previous sincosf was
>> optimized on that target hence it might make sense to retest
>> the new code to be sure).
>> I think the patches are in a good quality state now.
>> (The ABI changing part needs further work so i didn't post that.)
> built and tested on a power8 machine now, glibc math
> tests pass (except for an unrelated fmal failure),
> benchmark improvements are consistent with aarch64/x86_64,
> but it was a shared access machine so i won't post exact
> numbers, sincosf improved a bit too, sinf/cosf didn't
> (apparently powerpc has its own implementation).
PowerPC sinf/cosf uses the same algorithm used on x86, I presume
it would be a gain to generic implementation as well.