I was just looking through a preview of the newest edition of Laganiere's "OpenCV 4 Computer Vision Application Programming Cookbook" (with added author Shilkrot). It is substantially similar to the 3rd edition, which I've found helpful. The main difference is the last chapter, with info on use of DNN via C++.
In that chapter, there is a section "Enabling Halide to improve efficiency" which makes some claims for possible drastic speed increases. But it requires building LLVM with clang, etc.
The few posts on Halide that I've seen here are mixed reviews, and often are referring to non-Intel platforms.
I was wondering if it's worthwhile to pursue this at all. It sounds like it could involve some effort for a technology that may or may not be useful at present.