Hi all. I would ask some informations regarding the BAtched API provided in cuSolver. In particular, I observed that for an lda (Leading dimension of two-dimensional array) bigger than n (number of rows of each single matrix to be processed by the batched solver), some NaN values show up in the results. On the other hand, the results are correctly computed for lda = n.
I looked up and read the documentation, but I didn’t find any certain information or reason behind this behavior.
I would ask if my suggestion is correct and possibly the reason behind it.
Hi, can you please tell me what version of cuSOLVER you are using? Where do the NaN’s show up in the result: the eigenvectors or the eigenvalues? We will look into this, but if you have a reproducer, that would be great. Thanks for the report.
Hi. My cuSOLVER version is the 12.0.9. The NaNs values are shown for the eigenvectors. Unfortunately I’m not able to reproduce the wrong values with a script, I have the attached one (converted from .f90 to .txt), but it works also with lda /= n. A peculiarity of my program is that it has OpenMP threads with on top of them Cuda Streams introduced to get a concurrent execution.
Thank you so much for replying.