Open zgyr opened 4 years ago
Yes, that's possible, see here. The problem is that self.all_grids
is precomputed and can become pretty large depending on the number of projections and their size. Pushing it onto the gpu as a whole wouldn't work then. Transfering single grids to the gpu allows larger/more grids but decreases the speed, of course.
Is it possible to unwrap this transformation loop? If it is possible to collect all the slices into one tensor and send for processing, it should be more efficient. And isn't another there a faster way to do this transform?