LAPACK error on simple SVD

[maximsch2]

X = ones(Float32, (19764, 13620));
svd(X)

throws an error:

 ** On entry to SGESDDPM parameter number 12 had an illegal value
ERROR: ArgumentError: invalid argument JuliaLang/julia#12 to LAPACK call
 in gesdd! at linalg/lapack.jl:1482
 in svdfact! at linalg/svd.jl:17
 in svdfact at linalg/svd.jl:23

This sounds very close to scipy/scipy#5401, maybe we need a similar workaround?

This is on julia 0.4.5 built from source on Ubuntu 14.04.

[andreasnoack]

Ha. That's a funny one. It should be easy to work around, though. Thanks for the SciPy people.

[maximsch2]

Changing places like https://github.com/JuliaLang/julia/blob/master/base/linalg/lapack.jl#L302 to

lwork = BlasInt(nextfloat(real(work[1])))

should be enough, right?

[andreasnoack]

Almost. I think this would give an inexact error for smaller values so something like

lwork = round(BlasInt, nextfloat(real(work[1])))

should work. It would be great if you could make a PR.

[jianghaizhu]

Just tested it with svd(rand(Float32, (9537, 9537))), I still have this error.

** On entry to SGESDDPM parameter number 12 had an illegal value
ERROR: ArgumentError: invalid argument JuliaLang/julia#12 to LAPACK call
in gesdd! at linalg/lapack.jl:1505
in svdfact! at linalg/svd.jl:17
in svdfact at linalg/svd.jl:23

My julia version is

Julia Version 0.4.6
Commit 2e358ce (2016-06-19 17:16 UTC)
Platform Info:
System: Darwin (x86_64-apple-darwin13.4.0)
CPU: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
WORD_SIZE: 64
BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
LAPACK: libopenblas64_
LIBM: libopenlibm
LLVM: libLLVM-3.3

[andreasnoack]

It is fixed on master but it hasn't been backported yet. I forgot to label the fix for backporting but have done that now so it should be included in next bugfix release of 0.4.

[jianghaizhu]

Great! Thanks.