On Wed, Nov 24, 2021 at 11:45 AM Martin Maechler
<maech...@stat.math.ethz.ch> wrote:
>
> >>>>> Avraham Adler on Thu, 18 Nov 2021 02:18:54 +0000 writes:
>
> > Hello. I have isolated the issue: it is the
> > fused-multiply-add instruction set (FMA on Intel
> > processors). Running -march=skylake -mno-fma not only does
> > not hang, but passes make check-all (using R's native
> > BLAS). My intuition remains that something in the new
> > more precise ebd0 code used in dpois_raw—called by dgamma,
> > called by dchsq, called by dnchisq—is hanging when the
> > assembler uses FMA. Unfortunately, I have come across
> > other cases online where the extra precision and the
> > different assembler code of FMA vs. non-FMA has caused
> > bugs, such as [1]. Page 5 of this paper by Dr. William
> > Kahan sheds some light on why this may be happening [2]
> > (PDF).
>
> > Martin & Morton, having written (PR#15628 [3]) and/or
> > implemented the ebd0 code that is now being used, can
> > either of you think of any reason why it would hang if
> > compiled using FMA?
>
> I vaguely remember I had a version of ebd0(), either Morton
> Welinder's original, or a slight modification of it that needed some
> mending, because in some border case, there was an out of
> array-boundary indexing... but that's just a vague recollection.
>
> I had investigated ebd0()'s behavior quite a bit, also notably
> the version -- together with a pure R code version --
> in my CRAN package DPQ, yesterday updated to version 0.5-0 on CRAN
> {written in Summer, but published to CRAN only yesterday}
> where I have dpois_raw() optionally using several experimental versions of
> bd0(), and both 'pure R' and a C version of ebd0(),
> as DPQ::ebd0() and DPQ::edb0C()
> each with an option 'verbose' which shows you which branches are chosen
> for the given arguments.
>
> So, if you install this version (0.5-0 or newer) from the development
> sources, using the *same* FMA configuration,
> I hope you should see the same "hanging" but would be able to see some
> more.. ?
>
> Can you install it from R-forge
>
> install.packages("DPQ", type = "source",
> repos="http://R-Forge.R-project.org";)
>
> and then experiment?
> I'd be grateful {and we maybe can move "off - mailing list"}
>
> Thank you in advance,
> Martin
>
> Martin Maechler
> ETH Zurich and R Core team
Sure. Responding here simply for closure. Will direct further
questions and output directly to you.
Thank yyou,
Avi
______________________________________________
R-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
