Pytorch and tensor flow backend pass through the CPU.

[jonasteuwen]

Current implementations of the forward and backward projector wrappers in ODL pass through numpy arrays. Are there any plans to take the GPU context of ASTRA and connect it somehow?

[ozanoktem]

This has been discussed at several occasions and especially so for algorithms that want to make use of automatic differentiation in PyTorch. A closely related issue is #731 (see also #739) and the pull-request (that is still open) #1546.

[kohr-h]

Main issue is that we don't have CUDA support in ODL directly, so any ODL operator will make the data go through CPU memory. Apart from the issues and PRs @ozanoktem referred to, there are two more: #1231 and #1401 (basically the same though). But my current stance is that all the tedious work to make ODL element wrappers behave nicely is waste of time (read: I won't do it) since I want to get away from that concept and use arrays directly, see #1475.

So yes, there are plans, but at least for me the order is #1475, then #1401. If anyone else would like to give it a try independently, go ahead. But it's not trivial.

In the short run @jonasteuwen you could hack something yourself by pulling out the Operator-specific parts in OperatorFunction and implement the calls to ASTRA directly.

[jonasteuwen]

I guess #1401 would likely take bring it much closer as there seems to be a back and forth possible between cupy <-> pytorch.

[kohr-h]

Definitely, you can just hand over device memory pointers.