Re: SLATEC error: Chebyshev series too short for specified accuracy

[Jarno Rajahalme]

On Feb 25, 2011, at 3:30 , ext Ben Abbott wrote:

> On Feb 24, 2011, at 7:57 PM, Camillo wrote:
> 
>> Hello,
>> 
>> I am working on a Mac build of Octave 3.4.0. It's almost complete: it passes 
>> 6908 tests, has 2 known failures, and 5 unexpected ones. Four of these five 
>> are due to the same issue, an exception inside SLATEC: "Chebyshev series too 
>> short for specified accuracy". I don't know Fortran and I'm not sure how to 
>> best debug the issue. The compilers I am using are gcc and gfortran 4.2.1.
>> 
>> I first posted this question on the general mailing list, where Ben Abbot 
>> asked for more information, so here it is:
>> 
>>> Where did you get your gfortran?
>> 
>> http://r.research.att.com/gfortran-42-5664.pkg
>> 
>>> What dependencies did you install?
>> 
>> All, except for gnuplot and ghostscript.
>> 
>>> What was the result of running ./configure? (it gives a summary at the end)
>>> 
>>> Finally, please copy-n-paste the actual error.
>> 
>> I'll attach the configure summary, and here are the insteresting parts of 
>> the output of "make check":
>> 
>> --------------------------------------------------------------------------------
>> src/DLD-FUNCTIONS/besselj.cc ........................... PASS  180/180 
>> 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/gammainc.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         3
>> 
>> ***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         4
>> 
>> PASS    2/4    FAIL 2
>> --------------------------------------------------------------------------------
>> 
>> This is the information about the fifth error from fntests.log:
>> 
>> --------------------------------------------------------------------------------
>>>>>>> processing 
>>>>>>> /private/tmp/homebrew-octave-3.4.0-kEbx/octave-3.4.0/src/mappers.cc
>> ***** assert(arg (single(-1)), single(pi));
>> !!!!! test failed
>> assert (arg (single (-1)),single (pi)) expected
>> 3.1416
>> but got
>> 3.1416
>> values do not match
>> --------------------------------------------------------------------------------
>> 
>> I know about the problem with Apple's BLAS and gfortran, and I am using a 
>> small wrapper library for the affected routines (as suggested in 
>> <http://developer.apple.com/hardwaredrivers/ve/errata.html#fortran_conventions>);
>>  this is the "dotwrp" library you can see in the configure summary. I was 
>> told that a different wrapper has been proposed on this list, replacing some 
>> functions from LAPACK as well; I was wondering if anybody has a simple test 
>> case for some of those functions, so I could quickly check if the library 
>> I'm using is sufficient.
>> 
>> Best regards,
>> Camillo Lugaresi
>> 
>> <configure summary>
> 
> I am also using 4.2.1 and the gfortran from AT&T.
> 
> Iv'e compiled for 32bit and do not see the SLATEC problem.
> 
> Jarno / Richard, might this be related to the 64bit vecLib problem?
> 

I recall seeing similar error messages in the past. I recommend trying with the 
patch at http://savannah.gnu.org/patch/index.php?7468.

That patch provider gfortran native calling convention stubs for Apple BLAS and 
LAPACK calls. However, it does not otherwise replace any LAPACK routines. 

 Jarno