Bug in xggsvd? "Parameter 5 was incorrect on entry to DORMQR"

[Andreas_Jensen]

When running the program in single threaded mode I get the error message "Parameter 5 was incorrect on entry to DORMQR". The message is not present if I either run the program with multiple threads or if I compile it with reference LAPACK. My compiltervar settings are intel64 ilp64 and the program is compiled with

ifort testgsvd.f90 -lmkl_rt

testgsvd.f90:

program testgsvd

  implicit none
  integer*8 :: k, l, alpha(10), beta(10), iwork(20), info
  double precision :: A(10,20), B(10,20), U, V, Q, work(80)

  open(1, file="a.txt", action = "read")
  read(1, *) A
  open(2, file="b.txt", action = "read")
  read(2, *) B
  close(1)
  close(2)

  call dggsvd('N', 'N', 'N', 10_8, 20_8, 10_8, k, l, A, 10_8, B, 10_8, alpha, beta, U, 1_8, V, 1_8, Q, 1_8, work, iwork, info)

end program testgsvd

The matrices A and B are attached.

[Gennady_F_Intel]

by another words, the issue exists when you link with sequential mode of MKL?

the example contains the ?ggsvd, but the title is about domgr.

why parameter info is INTEGER* 8?

when you link with mkl_rt - you use LP64 interface by the default.

[Andreas_Jensen]

Thanks for the fast reply.

by another words, the issue exists when you link with sequential mode of MKL?

No. The difference is just between

./a.out

and

OMP_NUM_THREADS=1 ./a.out 

the example contains the ?ggsvd, but the title is about domgr.

Yes. The call is to dggsvd, but the error message comes from dormqr. I can see that reference LAPACK doesn't even call that function so thats the reason why I suspect it to be an MKL problem.

why parameter info is INTEGER* 8?
when you link with mkl_rt - you use LP64 interface by the default.

Before I compile and execute, I run source PATHTOINTEL/bin/compilervars.sh intel64 ilp64 and I thought that would cause ilp64 mode when running programs linked with lmkl_rt. Since you asked, I've also tried to set export MKL_INTERFACE_LAYER=ilp64 but I still get the same error.

[Gennady_F_Intel]

Ok, thanks Andreas. I see the similar behaviour on my side too. we will check the cause of the problem.