There are two things that I haven't resolved with this commit.

The first one is that I haven't figure out is how bifrost.orville and bifrost.romein fit together. Orville is effectively a sub-set of Romein with a few key changes:

• kernels - Romein expects 4 or 5-D kernels (baselines/pols/size/size) vs the symmetric 1-D kernels used by Orville
• gridding coordinates - Romein expects integers, Orville floats
• weights - Orville has an additional "weights" input
• gridding planes - Romein grids everything onto a single grid, while Orville can use multiple grids for w-projection

There might be other differences but I think those are the key ones.

The second is whether or not any of the additional computational steps in bifrost.orville.Orville.execute() can be moved into orville.cu. Current the .cu file does the gridding and then the Python module handles the summation across gridding planes, image weighthing, the inverse Fourier transform, and the image plane correction for the gridding kernel. Those all end up being calls back into the Bifrost library so moving things into the .cu would save on Python ➡️ C++/CUDA ➡️ Python transitions. I had attempted to do this in an early version but I ended up with problems with bfFft. Maybe that should be revisited.