cudaErrorMemoryAllocation on Summit when using N_binning = 2
There are occasional issues using PyCUDA on Summit -- this problem is intermittent and manifests itself as cudaErrorMemoryAllocation
when using large inputs. Doing a basic back of the envelope calculation shows that this problem manifests already when the image size is approx 2 GB (V100's have 16 GB => this shouldn't be an out-of-memory issue).
Reproducing
Run spinifel on summit with N_binning hard-coded to:
N_binning = 2