LAPACKException on Archlinux 64-bit

[rennis250]

Hey everyone,

I'm receiving this LAPACKException lately:

ERROR: LAPACKException(13567071543623680)
 in gtsv! at linalg/lapack.jl:1235
 in \ at linalg/tridiag.jl:181
 in Woodbury at linalg/woodbury.jl:25
 in _interp_invert_matrix at /home/rje/.julia/Grid/src/interp.jl:663
 in interp_invert! at /home/rje/.julia/Grid/src/interp.jl:637
 in InterpGrid at /home/rje/.julia/Grid/src/interp.jl:50

I'm on Archlinux 64-bit with an Intel i7 processor.

This is for code in Grid.jl, but I've seen it occur on several other occasions. Sometimes, a SingularException occurs, instead of a LAPACKException.

I've tried rebuilding the latest Julia release candidate (Version 0.2.0-rc1+131). Using the system provided BLAS (64-bit) and the Julia-make compiled Suitesparse; I still have yet to try OpenBLAS (my system gives out an incredibly annoying beep sound during compliation, so I will need to wait until everyone has left the office! :/ ). Cross-posting this to the Grid.jl repo, as well (timholy/Grid.jl#8).

Thanks,
Rob

[andreasnoack]

This looks like a problem with integer support in LAPACK. Are you sure that the system BLAS and LAPACK are compiled with 64 bit integer support? Could you try to recompile with USE_BLAS64=0 or with OpenBLAS and run the tests again.

[rennis250]

I'm fairly certain they are, but I will take a closer look.

I'll give the USE_BLAS64=0 option a try in a moment. The OpenBLAS test will have to wait until tonight, unfortunately. If it helps to diagnose the issue, when I try brute-force by re-running the code over and over, eventually the Exception does not occur and the process runs. However, subsequent attempts produce the error again.

[jiahao]

I was unable to reproduce on Ubuntu x86_64.

julia> versioninfo()

Julia Version 0.2.0-rc1+124
Commit dfdbf4a 2013-10-24 03:55:21 UTC
Platform Info:
  System: Linux (x86_64-linux-gnu)
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY)
  LAPACK: libopenblas
  LIBM: libopenlibm

[jiahao]

There is almost certainly a 32-bit/64-bit issue going on here. The LAPACKException value is the INFO variable returned by LAPACK's _gtsv routines.

          INFO is INTEGER
          = 0: successful exit
          < 0: if INFO = -i, the i-th argument had an illegal value
          > 0: if INFO = i, U(i,i) is exactly zero, and the solution
               has not been computed.  The factorization has not been
               completed unless i = N.

It is exceedingly unlikely that you are solving 13567071543623680-dimensional matrix problems or managed to call LAPACK.gtsv! with 9223371607358046208 arguments.

[jiahao]

Please reopen if this continues to be an issue after looking into the BLAS libraries you are linking with.

[rennis250]

Yea, USE_BLAS64=0 fixes it. Thought I had tried that already, but apparently not. This was getting extremely frustrating, so glad to have it fixed! I'll need to get OpenBLAS on here, because 64-bit Integer support could be useful.

Thanks for the assistance,
Rob