This analysis suggests that a 2D convolution filter of stride=2 would be the most compute-efficient for this particular workload… Also, ideally batch up all the grey data and do the calculation all at once.

→ View original post on X — @alexjc