Salute,

I experienced a queer problem with DGESVD (using the ACML library on an opteron machine).

When using JOBU='N' and JOBVT='N', DGESVD exits in certain cases with info=2, although the corresponding call with JOBU='S' and JOBVT='S' (for the same matrix) works just fine (info=0).

Im working with relatively large numerical matrices, which I cannot post here. In some cases the 'S' and the 'N' calls yield correctly the same singular values. In some cases the 'N' call exits with info=2. For a certain 401-times-401 matrix, where it gave info=2, I replaced for a check all matrix elemnts by 2 - and the 'N' call suddenly worked then, giving the correct singular values. So I guess, I have the parameters correctly.

This is strange.

It's not obvious why the convergence should depend on the JOB parameters.

Does anyone know whether DGESVD does something qualitatively different, when 'N' is given. Or could this be some bug in the library?

thanks and cheers, Thomas

PS: The same problem occurs when using ZGESVD