R devel - Apr 2016 - Optimization bug when byte compiling with gcc 5.3.0 on windows

>If I recall correctly, some eigen vectors had their
>direction flipped (negative values became positive and vice versa).
>Did you notice anything of this kind when running 'make check' and
>'make check recommended' ? It is important to us that numeric
results
>are reproducible between versions of R.
I think that any code that depends on the direction
of an eigenvector should be considered broken.
It is too bad that eigen's output does not have a class
so that an all.equal method that know that the vector
direction is not relevant could be written for it


Bill Dunlap
TIBCO Software
wdunlap tibco.com

On Mon, Apr 4, 2016 at 11:39 AM, Jeroen Ooms <jeroen.ooms at
stat.ucla.edu>
wrote:
> On Sun, Apr 3, 2016 at 6:44 PM, Ray Donnelly <rdonnelly at
continuum.io>
> wrote:
> > I've started to look into building R for Windows using MSYS2 as
both the
> > build environment and tools + libraries provider (where possible).
>
> Thanks for your reply, and for the patches.
>
> Last time I had a look at this (a few months ago) another problem was
> that mingw-w64 v4 was giving different numeric output for some of the
> tests in r-base. If I recall correctly, some eigen vectors had their
> direction flipped (negative values became positive and vice versa).
> Did you notice anything of this kind when running 'make check' and
> 'make check recommended' ? It is important to us that numeric
results
> are reproducible between versions of R.
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
	[[alternative HTML version deleted]]

>>>>> William Dunlap via R-devel <r-devel at r-project.org>
>>>>>     on Mon, 4 Apr 2016 12:26:38 -0700 writes:
    >> If I recall correctly, some eigen vectors had their
    >> direction flipped (negative values became positive and
    >> vice versa).  Did you notice anything of this kind when
    >> running 'make check' and 'make check recommended' ?
It is
    >> important to us that numeric results are reproducible
    >> between versions of R.

    > I think that any code that depends on the direction of an
    > eigenvector should be considered broken.  

I agree [even though I may have fallen into that trap in some of
the regression tests I had written in the long past!].

    > It is too bad that eigen's output does not have a class so
    > that an all.equal method that know that the vector
    > direction is not relevant could be written for it

That's an interesting / neat idea.
So we could give it S3 class "eigen"  iff  only.values is not
true, which would also "solve" the slightly differently
structured result of eigen() 

Are you (or others) willing to provide a prototype for
all.equal.eigen ?

Martin

    > Bill Dunlap TIBCO Software wdunlap tibco.com

    > On Mon, Apr 4, 2016 at 11:39 AM, Jeroen Ooms
    > <jeroen.ooms at stat.ucla.edu> wrote:

    >> On Sun, Apr 3, 2016 at 6:44 PM, Ray Donnelly
    >> <rdonnelly at continuum.io> wrote: > I've started to
look
    >> into building R for Windows using MSYS2 as both the >
    >> build environment and tools + libraries provider (where
    >> possible).
    >> 
    >> Thanks for your reply, and for the patches.
    >> 
    >> Last time I had a look at this (a few months ago) another
    >> problem was that mingw-w64 v4 was giving different
    >> numeric output for some of the tests in r-base. If I
    >> recall correctly, some eigen vectors had their direction
    >> flipped (negative values became positive and vice versa).
    >> Did you notice anything of this kind when running 'make
    >> check' and 'make check recommended' ? It is important
to
    >> us that numeric results are reproducible between versions
    >> of R.
    >> 
    >> ______________________________________________
    >> R-devel at r-project.org mailing list
    >> https://stat.ethz.ch/mailman/listinfo/r-devel
    >> 

    > 	[[alternative HTML version deleted]]

    > ______________________________________________
    > R-devel at r-project.org mailing list
    > https://stat.ethz.ch/mailman/listinfo/r-devel

all.equal.eigen() should also accommodate complex vectors, right?

Ravi
________________________________________
From: R-devel <r-devel-bounces at r-project.org> on behalf of Martin
Maechler <maechler at stat.math.ethz.ch>
Sent: Monday, April 11, 2016 3:08 AM
To: William Dunlap
Cc: r-devel; Jeroen Ooms
Subject: Re: [Rd] Optimization bug when byte compiling with gcc 5.3.0   on     
windows
>>>>> William Dunlap via R-devel <r-devel at r-project.org>
>>>>>     on Mon, 4 Apr 2016 12:26:38 -0700 writes:
    >> If I recall correctly, some eigen vectors had their
    >> direction flipped (negative values became positive and
    >> vice versa).  Did you notice anything of this kind when
    >> running 'make check' and 'make check recommended' ?
It is
    >> important to us that numeric results are reproducible
    >> between versions of R.

    > I think that any code that depends on the direction of an
    > eigenvector should be considered broken.

I agree [even though I may have fallen into that trap in some of
the regression tests I had written in the long past!].

    > It is too bad that eigen's output does not have a class so
    > that an all.equal method that know that the vector
    > direction is not relevant could be written for it

That's an interesting / neat idea.
So we could give it S3 class "eigen"  iff  only.values is not
true, which would also "solve" the slightly differently
structured result of eigen()

Are you (or others) willing to provide a prototype for
all.equal.eigen ?

Martin

    > Bill Dunlap TIBCO Software wdunlap tibco.com

    > On Mon, Apr 4, 2016 at 11:39 AM, Jeroen Ooms
    > <jeroen.ooms at stat.ucla.edu> wrote:

    >> On Sun, Apr 3, 2016 at 6:44 PM, Ray Donnelly
    >> <rdonnelly at continuum.io> wrote: > I've started to
look
    >> into building R for Windows using MSYS2 as both the >
    >> build environment and tools + libraries provider (where
    >> possible).
    >>
    >> Thanks for your reply, and for the patches.
    >>
    >> Last time I had a look at this (a few months ago) another
    >> problem was that mingw-w64 v4 was giving different
    >> numeric output for some of the tests in r-base. If I
    >> recall correctly, some eigen vectors had their direction
    >> flipped (negative values became positive and vice versa).
    >> Did you notice anything of this kind when running 'make
    >> check' and 'make check recommended' ? It is important
to
    >> us that numeric results are reproducible between versions
    >> of R.
    >>
    >> ______________________________________________
    >> R-devel at r-project.org mailing list
    >> https://stat.ethz.ch/mailman/listinfo/r-devel
    >>

    >   [[alternative HTML version deleted]]

    > ______________________________________________
    > R-devel at r-project.org mailing list
    > https://stat.ethz.ch/mailman/listinfo/r-devel

______________________________________________
R-devel at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel