Closed telegraphic closed 1 year ago
Indeed, no speed up.
(Aim is to allow reading from disk for next gulp while processing.)
Some background: https://github.com/h5py/h5py/issues/837
In short, h5py does not (yet) support async I/O, and HDF5 has only recently added support. So no performance gains possible without actual pipelining a la bifrost
.
Test using
asyncio
within pipeline.py. Not sure what h5py and cupy do when they are processing (are they blocking?)