Apparent bug in cusolverDnDDgels_bufferSize

As emerged in:

it seems that there is an error in cuSolver documentation: the parameter ldx is reported to be at least max(m,n), but that leads to erroneous results for a trivial LS problem wherem>n. It seems that the issue can be solved by using ldx=n.

Hope that helps,

I’ve filed a bug internal to NVIDIA (3193032) to have the documentation reviewed.