Attached is a test case for some Scalapack calls that are exibiting strange behavior that seems like a bug. I also attached screen shots of the output and of the compile output. I'm using Windows 7 (64-bit) , MKL 11.0.05, Intel MPI 3.0, and the 8-byte integer MKL interface and 8-byte integer MPI interface.
Basically, when I perform a QR factorization on individual mpi nodes followed by a matrix inverse on all nodes, I see a crash depending on the matrix size used in the inverse. More particularly, the matrices in the QR factorization are assigned to individual nodes or groups of nodes.
Whether the crash occurs depends on if the matrices assigned to individual nodes are QR factorized with the lapack QR factor (crash occurs) or the scalapack QR factor (crash does not occur). Any matrix assigned to more than one node will always use the scalapack QR factor. In particular, it is the PZUNQGR/ZUNQGR call that results in the crash in the inverse and not the (P)ZGEQRF call. The final inverse is performed on a single matrix distriubted across all nodes in the grid and the crash may or may not occur basd on the size of the matrix to be inverted.
Does anyone know if I am doing something wrong or if this is a Scalapack bug?