Oh, now I'm really, really confused. It's table 16 in that paper. It clearly shows GMP neck and neck with their code. Where did they get this factor of 6 from at Pasco? (Not a claim of the authors of Mathemagix, but someone else.)
Why is it that all these massive gains that get rumoured always vanish when you look into them? At least I expressed due skepticism about the 3x speedup on integer multiplication over GMP 5 made by someone else. Bill. On 22 July 2015 at 14:53, Bill Hart <goodwillh...@googlemail.com> wrote: > Hmm, now I'm confused. That must be the paper. But right now I can't find > the comparison with the GMP integer multiplication they were talking about. > > Anyway, it's very clear we can get a factor of 3 at least, probably 6. > > Bill. > > On 22 July 2015 at 14:51, Bill Hart <goodwillh...@googlemail.com> wrote: > >> Sorry, wrong paper. There is another paper by the same guys somewhere >> which explicitly mentions an FFT. I think Fredrik posted it a while back. >> That must be the one the guys at Pasco were referring to. I'll try to find >> it. >> >> Bill. >> >> On 22 July 2015 at 14:48, Bill Hart <goodwillh...@googlemail.com> wrote: >> >>> Oh yeah, that's years old. I don't know to what extent they had >>> vectorisation in mind when writing that. >>> >>> A better resource would be: >>> >>> http://arxiv.org/ftp/arxiv/papers/1407/1407.3383.pdf >>> >>> It was presented at ISSAC 2015, which I happened to attend. It was also >>> specifically referred to as beating GMP/MPIR by a factor of up to 6 by >>> someone at PASCO (which I also attended). Someone else had beaten GMP 5 by >>> a factor of 3 with just implicit (compiler driven) vectorisation, again >>> with a small primes FFT. They used 6 carefully chosen primes if I recall >>> correctly. >>> >>> Bill. >>> >>> On 22 July 2015 at 14:44, Jean-Pierre Flori <jpfl...@gmail.com> wrote: >>> >>>> It is quoted here as a branch with a small prime FFT: >>>> https://gmplib.org/devel/GMPng.html >>>> >>>> On Wednesday, July 22, 2015 at 2:30:00 PM UTC+2, Bill Hart wrote: >>>>> >>>>> What was that commit? >>>>> >>>>> On 22 July 2015 at 14:19, Jean-Pierre Flori <jpf...@gmail.com> wrote: >>>>> >>>>>> >>>>>> >>>>>> On Wednesday, July 22, 2015 at 1:34:37 PM UTC+2, Bill Hart wrote: >>>>>>> >>>>>>> >>>>>>> * AVX code from GMP 6 >>>>>>> * additional functions added to the GMP interface in GMP 6 >>>>>>> * small primes FFT which makes heavy use of AVX when available (will >>>>>>> likely replace the FFT) -- there's a factor of 6x speedup to be had here >>>>>>> >>>>>>> All but the last one, we can use from the GMP project itself, with >>>>>>> minor modifications. I added all the build system stuff for modern >>>>>>> processors in 2.7.0, so at least that part is done. The last item above, >>>>>>> we'll have to write ourselves. But it shouldn't be as big a project as >>>>>>> writing an FFT. >>>>>>> >>>>>> There is https://gmplib.org/repo/gcd-nisse/file/0d591aa7e02c >>>>>> Not sure in what state it is though. >>>>>> >>>>>> -- >>>>>> You received this message because you are subscribed to the Google >>>>>> Groups "mpir-devel" group. >>>>>> To unsubscribe from this group and stop receiving emails from it, >>>>>> send an email to mpir-devel+...@googlegroups.com. >>>>>> To post to this group, send email to mpir-...@googlegroups.com. >>>>>> Visit this group at http://groups.google.com/group/mpir-devel. >>>>>> For more options, visit https://groups.google.com/d/optout. >>>>>> >>>>> >>>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "mpir-devel" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to mpir-devel+unsubscr...@googlegroups.com. >>>> To post to this group, send email to mpir-devel@googlegroups.com. >>>> Visit this group at http://groups.google.com/group/mpir-devel. >>>> For more options, visit https://groups.google.com/d/optout. >>>> >>> >>> >> > -- You received this message because you are subscribed to the Google Groups "mpir-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to mpir-devel+unsubscr...@googlegroups.com. To post to this group, send email to mpir-devel@googlegroups.com. Visit this group at http://groups.google.com/group/mpir-devel. For more options, visit https://groups.google.com/d/optout.
