It was a bit tricky to work with Halide. The Halide code itself is beautiful, but it depends on LLVM which takes forever to build and link, and I ran into some snags with Windows support. I originally wanted to write the whole algorithm in Halide and run it on the GPU, but it turned out that the Halide language isn't expressive enough to implement a Hough transform. I had to write that bit in C and I'm currently only using Halide to do some more straightforward image processing tasks like edge detection filters, blurring, and resampling. It works really well for that though, and performance-wise it runs rings around the naive nested C for loops that I would otherwise have to write.

Thanks, been interested in the language from a distance for a while now.

Can't claim to know what a Hough transform is, but do you think the lack of expressiveness can be fixed without interfering with the tricks that make it work? It's still a language under development after all.

I think the language could be made expressive enough, and the developers thought so too when I asked them about it. I think they're focusing on other things for now though.

A Hough transform is kind of a weird beast; not really your typical image processing problem. For anything that looks like a series of convolutions or other simple image operations, Halide is going to be the fastest thing around; even faster than a series of calls to the Intel IPP library (because of the ability to tune memory access patterns across multiple stages at once).