RE: Lion make check woes

[Robinson, Melvin D]

From: Ben Abbott address@hidden
Sent: Saturday, January 14, 2012 10:31 PM
To: Robinson, Melvin D
Cc: octave maintainers mailing list
Subject: Re: Lion make check woes

On Jan 14, 2012, at 5:33 PM, Robinson, Melvin D wrote:

> I'm still having difficulty with the make check and would appreciate just a 
> bit of guidance on how I can make some progress.
>
> With ATLAS installed I have the following problems with make check:
>   src/DLD-FUNCTIONS/betainc.cc ........................... ***MESSAGE FROM 
> ROUTINE INITS IN LIBRARY SLATEC.
>  ***POTENTIALLY RECOVERABLE ERROR, PROG ABORTED, TRACEBACK REQUESTED
>  *  Chebyshev series too short for specified accuracy
>  *  ERROR NUMBER = 1
>  *
>  ***END OF MESSAGE
>  ***JOB ABORT DUE TO UNRECOVERED ERROR.
> 0          ERROR MESSAGE SUMMARY
>  LIBRARY    SUBROUTINE MESSAGE START             NERR     LEVEL     COUNT
>  SLATEC     INITS      Chebyshev series too         1         1         1
>  ***MESSAGE FROM ROUTINE INITS IN LIBRARY SLATEC.
>  ***POTENTIALLY RECOVERABLE ERROR, PROG ABORTED, TRACEBACK REQUESTED
>  *  Chebyshev series too short for specified accuracy
>  *  ERROR NUMBER = 1
>  *
>  ***END OF MESSAGE
>  ***JOB ABORT DUE TO UNRECOVERED ERROR.
> 0          ERROR MESSAGE SUMMARY
>  LIBRARY    SUBROUTINE MESSAGE START             NERR     LEVEL     COUNT
>  SLATEC     INITS      Chebyshev series too         1         1         2
>  PASS    4/6    FAIL 2
>   src/DLD-FUNCTIONS/bsxfun.cc ............................ PASS   73/73
>   src/DLD-FUNCTIONS/cellfun.cc ........................... PASS  121/121
>   src/DLD-FUNCTIONS/chol.cc .............................. PASS   29/29
>   src/DLD-FUNCTIONS/conv2.cc ............................. PASS    4/4
>   src/DLD-FUNCTIONS/convhulln.cc ......................... PASS    2/3    
> FAIL 1
>   src/DLD-FUNCTIONS/dassl.cc ............................. PASS    4/4
>   src/DLD-FUNCTIONS/det.cc ............................... PASS    5/5
>   src/DLD-FUNCTIONS/dlmread.cc ........................... PASS   20/20
>   src/DLD-FUNCTIONS/dmperm.cc ............................ PASS    5/5
>   src/DLD-FUNCTIONS/dot.cc ...............................
> The convhulln test is one that I will have to work on, but I'm not 
> understanding why it would lock up in dot.cc.  This is simply the source code 
> to comptes the dot product of vectors, right?  I'm trying to narrow this down
>
> octave:1> test dot.cc verbose
> >>>>> /Users/melrobin/octave/src/DLD-FUNCTIONS/dot.cc
>   ***** test
>  x = [2, 1; 2, 1];
>  y = [-0.5, 2; 0.5, -2];
>  assert(dot (x, y), [0 0]);
>   ***** test
>  x = [ 1+i, 3-i; 1-i, 3-i];
>  assert(dot (x, x), [4, 20]);
>
> At this point it is locks up.  I verified this behavior by using ./run-octave:
> octave:1>  x = [ 1+i, 3-i; 1-i, 3-i];
> octave:2> dot(x,x)
> and it locks up dotting x with itself.  This smells like it is a bug with 
> ATLAS because I do not have the same issue when using -framework veclib (the 
> default BLAS/LAPACK for Mac OS X).  My problem using -framework veclib is 
> that it segfaults on eigs.cc, but one problem at a time.
>
> I would like to recreate this problem for this test case by calling ATLAS 
> directly.  Can you tell me what call dot(x,x) makes for complex values?  I'm 
> going to try it with zdot, but wanted to make sure that I know what I'm 
> talking about.
>
> Thanks for your help.

How are you satisfying Octave's dependencies ?

Are you using a package manager ?  (i.e. Fink Macports, Homebrew, something 
else?)

Which c/c++ compiler are you using ? ... Which fortran compiler are you using ?

Ben

I'm using a mixture of MacPorts and a few hand compiled libraries (QRupdate and 
FFTW).   I'm using the GCC 4.6 suite and with your sed commands seems to be 
building OK.  I have a hand compiled ATLAS and think that I have found some bug 
in the multithreaded ATLAS that does not occur in the single threaded version 
that I want to try to reproduce.  

There are a couple of problems I'm having, but I wanted to focus on that one 
first.  BTW, it does not seem that your environment LAPACK_LIBS and BLAS_LIBS 
work because without specifying a --with-lapack and a --with-blas option to 
configure, Octave is compiled with -framework veclib which is very puzzling.

Do you know how Octave would go about calling BLAS in dot.cc to perform this 
test: (the BLAS calling sequence)

 ***** test
 x = [ 1+i, 3-i; 1-i, 3-i];
 assert(dot (x, x), [4, 20]);

it locks up at this point in multithreaded ATLAS.