In this extended scenario, the sub-sampling is no longer over contiguous memory blocks, so the data streaming approach will also require script-level looping and will probably not be particularly efficient.

On my setup, looping over Slice2 calls is still ~20% faster than the the single call to ExprSize, but that may be because my computer has an old processor. I did notice, however, that the timing results varied a great deal and seemed to be connected to foreground activity in support of the Display floating window. For optimum and consistent timing results, I think it is important to either delay calls to ShowImage until the very end of the script, or to make sure the Display floating window is closed. Of course, these issues are side-stepped entirely if one runs the script on the foreground thread, instead of in the background, as you currently seem to do.