LAPACK/ScaLAPACK Development

I'm trying out the new Jacobi solver, but having a hard time figuring out
the workspace requirements. As a general comment, it would be very
helpful if this routine had the apparatus offered by many others, where
you can give LWORK = -1 and get a response indicating a suitable workspace
size.

But anyway, take a 3x3 input matrix (M = N = 3), and say we want the left or
right singular vectors but not both. By my reading of dgejsv.f, the minimal
LWORK is supposed to be max(2*N+M,7) = 9. But if I give LWORK = 9, my
program aborts with
** On entry to DGEQP3 parameter number 8 had an illegal value

Param 8 is dgeqp3's LWORK, passed by dgejsv; and looking inside
dgejsv, the first such call is

CALL DGEQP3( M,N,A,LDA, IWORK,WORK, WORK(N+1),LWORK-N, IERR )

With N = 3 and LWORK = 9 on input, this sends dgeqp3 an LWORK of 6
= 2*N. Yet inside dgeqp3 we find the requirement that LWORK >= 3*N+1.
So the abort is not surprising: dgejsv is making an illegal call to
dgeqp3. It seems to me that dgejsv has to specify a bigger minimal
LWORK.

To be more specific: I believe that on account of the xGEQP3 calls,
the minimal LWORK for dgejsv, for the case of one or other but
not both of the left- and right singular vectors wanted, should be
max(2*N+M,5*N+1,7), rather than max(2*N+M,7) as given.

Hello again -- any confirmation/denial that this is a bug?

Hello, thanks for your help on the forum and this bug report. Sorry for the delay. Yes there is a bug in DGEJSV, I confirm. When I perform the same experiment as you did, I have the same crash and your fix fixes it. So that's for the confirmation. Zlatko (original developer of the code) should have a look soon. Please use DGEJSV with your fix for the time being and we'll come back to you. Julien.

Thanks for the confirmation! This function seems to work nicely when the
workspace issue is handled correctly.

This has been corrected in our repository.
See http://www.netlib.org/lapack/Errata/index2.html
Thank you