[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: svds test failures
From: |
Tatsuro MATSUOKA |
Subject: |
Re: svds test failures |
Date: |
Fri, 2 Sep 2011 10:08:26 +0900 (JST) |
Hello
The results perhaps depend on the kind of blas.
Today I built OpenBLAS-a546694 and replace from altas-3.9.19 with
OpenBLAS-a546694.
For
%! randn ('state',42); % Initialize to make normest function reproducible
%! rand ('state', 42)
give
warning: returning fewer singular values than requested
warning: try increasing the value of sigma
***** testif HAVE_UMFPACK
[u2,s2,v2,flag] = svds(A,k,0,opts);
s2 = diag(s2);
assert(flag,!1);
assert(s2, s(k:-1:1), 1e-10);
!!!!! test failed
assert (s2,s (k:-1:1),1e-10) expected
38.060
38.034
38.034
38.015
38.015
38.004
38.004
but got
38.034
38.034
38.015
38.015
38.004
38.004
Dimensions don't match
The seed 42 gave no error for atlas-3.9.19.
****The results are
ATLAS OpenBLAS
35 OKOK
36 OK OK
37 fail OK
38 OK
39 OK
40 OK
41 OK
42 OK
43 fail
44 OK
45 OK
46 OK
47 OK
48 OK
49 fail
50 OK
--- On Tue, 2011/8/9, Marco atzeri wrote:
> On 8/9/2011 6:28 AM, Rik wrote:
> > On 08/08/2011 08:11 PM, Liam Groener wrote:
> >> On Aug 8, 2011, at 10:01 AM, Rik wrote:
> >>
> >>> 8/8/11
> >>>
> >>> Marco, Tatsuro,
> >>>
> >>> I've combined the list of initialization values which are acceptable for
> >>> Cygwin, MinGW, and Linux and it amounts to just [35, 40].
> >>>
> >>> Could someone with a MacOS system try those two values in svds.m? Below
> >>> are the two lines to change.
> >>>
> >>> ------
> >>> %! randn ('state', 35); % Initialize to make normest function
> >>> reproducible
> >>> %! rand ('state', 35)
> >>> ------
> >>>
> >>> After each change run the following code
> >>>
> >>> for i = 1:300
> >>> bm(i) = test ("svds");
> >>> endfor
> >>> sum (bm)
> >>>
> >>> If the sum is not 300 then neither of the values will work.
> >>>
> >>> --Rik
> >> I haven't seen any replies yet, so I tested on OS X 10.7, Octave 3.4.2.
> >> Neither 35 nor 40 worked, 42 did.
> >
> > Well this is great. Apparently there is no single value that works for all
> > platforms. The value 42 did work for Linux, Mac, and MinGW, but not for
> > Cygwin. I am going to revert to 42 since that works with the most
> > platforms and then it will need to be debugged why svds is failing on
> > Cygwin.
> >
> > Thanks Liam for the test on Mac platforms.
> >
> > --Rik
>
> Hi Rik,
>
> 42 is fine. I will ignore the failure and put a note for myself.
>
> As all the platform react differently changing the seed , we can assume
> one of:
>
> 1) cygwin doesn't like
> http://en.wikipedia.org/wiki/The_Hitchhiker%27s_Guide_to_the_Galaxy
>
> 2) Are rand/randn producing different outcome on different platform ?
>
> 3) This test is on the borderline and different rounding/libraries will
> produce different outcome. Should we find a less critical matrix ?
>
> I remember the test in the past was passing sometime with lapack and
> always with atlas. But it was before the arpack merge in octave and with
> previous gcc compiler. Eventually moving to gcc-4.5 the thing will
> change again on cygwin
>
> Regards
> Marco
>
- Re: svds test failures,
Tatsuro MATSUOKA <=