On Friday, November 13, 2015 at 12:34:41 PM UTC-8, Bill Hart wrote: > > > > On 13 November 2015 at 19:10, Jean-Pierre Flori <jpf...@gmail.com > <javascript:>> wrote: > >> >> >> On Friday, November 13, 2015 at 9:40:27 AM UTC-8, Bill Hart wrote: >>> >>> Not in the published divide-and-conquer algorithm, no. >>> >>> There was a bug in the implementation of the new basecase algorithm, >>> which was not published, nor very important. On most machines it was >>> switched off. And even on the machines where it was used (with the >>> exception of itanium where it seemed to be used a lot), it was only used >>> for *very* small operands. >>> >>> To be verbose, do you mean the basecase schoolbook multiprecision >> division using 3by2 small division using a new 1 word precomputed inverse >> is broken? >> > > Yes it was a 3by2 version using a new 1 word precomputed inverse > (unrelated to the 3by2 division originally used in MPIR and GMP). > So do you still trust the 3by2 algorithm in your preprint?
> > >> So is the code from BDSNT broken as wel? >> https://github.com/wbhart/bsdnt/blob/master/helper.c#L186 >> https://github.com/wbhart/bsdnt/blob/master/nn_quadratic.c#L240 >> >> > Doubtful. The implementation in MPIR was much, much more complex. > > The algorithm itself is quite straightforward. > > On the other hand, bsdnt does not have many users, so if bugs exist there, > it is much less likely they will be detected, so beware. > > Yup, I did an implementation myself for another project using your precomputed reciprocal and a naive schoolbook long division. I tested the values mentioned on http://trac.sagemath.org/ticket/19280 and get a correct remainder. Note that I don't store the quotient but it should be correct as well. If you are aware of any other values triggered wrong answer in MPIR, I'd be more than happy to pass them to my implementation. -- 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.
