make check failure on opteron (CVS version)

make check failure on opteron (CVS version)

[Giulio Bottazzi]

[-- Attachment #1: Type: text/plain, Size: 319 bytes --]

Hi,
on a dual opteron (linux kernel 2.6.8, gcc ver. 3.3.4, GNU libc ver.
2.3.4.20040808) the standard sequence


./configure
make
make check

gives a

...
FAIL:   LQ_solve hilbert(4)
...


with the latest (couple of hours ago) CVS version of gsl. The same DID
NOT happen with the official gsl-1.5 release.

Best,

	G.


[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]

Re: make check failure on opteron (CVS version)

[Brian Gough]

Giulio Bottazzi writes:
 > on a dual opteron (linux kernel 2.6.8, gcc ver. 3.3.4, GNU libc ver.
 > 2.3.4.20040808) the standard sequence
 > ..
 > FAIL:   LQ_solve hilbert(4)
 > ..
 > with the latest (couple of hours ago) CVS version of gsl. The same DID
 > NOT happen with the official gsl-1.5 release.

This is new code since 1.5, for LQ decomposition.

Could you look at the failure values and how they compare with the
true values.

I expect it is a tolerance problem only.

Thanks.

-- 
Brian Gough

Re: make check failure on opteron (CVS version)

[Brian Gough]

Giulio Bottazzi writes:
 > It seems you are right. There are four FAILed tests and these are the
 > relevant pieces of output from make check
 > 
 > ..
 > PASS:   LQ_solve hilbert(2)
 > PASS:   LQ_solve hilbert(3)
 >   4[0]:   -63.9999999999698304                      -64
 > FAIL:   LQ_solve hilbert(4)
 > PASS:   LQ_solve hilbert(12)

It's surprising that this fails, since the LQ algorithm should just be
the tranpose of the QR algorithm.

For reference, what version of gcc and compilation options are you
using?

-- 
Brian Gough