FIGURE 2.

(A) Using the Op builder pattern, the user requests a filter.dog (Difference of Gaussians) Op accepting an ordered list of three input parameters—an image and two sigma values—as well as a suitable container into which the output of the computation will be stored. The terminating method call compute() triggers the Ops engine to search for a computer Op plugin capable of fulfilling the request. (B) If the user does not have an output container, they can terminate their Op request with apply() instead of compute(), and the Ops engine will search for a function Op that generates and returns a new output buffer. (C) The matching filter.dog Op declares dependencies on the helper Ops filter.gauss and math.sub as described above, which are then recursively matched.