LAPACK/ScaLAPACK Development

I am trying to use the lapack routine ZGGEV to comput the eigen value lamada of A*X=lamba*B*X with A and B complex*16 non-symmetric matrix. But the error info indicate The QZ iteration failed for some sets of A and B.

In my problem, I calculate 148 sets of different A and B for lamada. If I use the NAG subroutine F02GJF(), then all sets of A and B return correct eigen values. But if I tried the lapack routine ZGGEV (by Intel MKL), then only 36 sets return correct eigen value, and the others have error info which indicate failure in the QZ interation. I also tried the lapack routine ZGGEV provided by AMD's ACML library, then 72 sets of A and B return correct eigenvalue and the others also return the error info which indicate failure in the QZ interation.

I tried to adjust the work space size to optimal condiation, but no improvment for the perfomance.

Any suggestion to solve this problem?

Thank you very much!!

Milton

Hi,

it seems that you hit some of the rare examples, where the QZ
algorithm can fail to converge. (NAG might use an advanced
exceptional shift strategy which overcomes failure of convergence
for your examples, but it might still fail to converge for others.)
We are currently redesigning all QZ codes in LAPACK. One of the
issues is to address all known convergence failures. May I therefore
ask you to send me a program that generates these test sets or at
least some of the test sets for which some of the QZ codes encounter
problems?

Thanks,
Daniel

Yeah, sure. Please open the attachment that I uploaded at the follwoing URL:

http://softwareforums.intel.com/ids/board/message?board.id=MKL&message.id=1157

The file name is: Lapack-test.zip which is located at my reply to that topic at MKL forum.