Parallel Programming in Futhark

No idea about its usage in industry, but this is a classical example of a tool that maps to its domain a magnitude better than the competition, yet can't seem to take off because of lack of the network effects that would act as propellant.

I messed around with some slightly above toy work, implementing expectation-maximization for gaussian mixture models and some other supporting tools for 3D point clouds, and found it really intuitive for someone with a decent amount of FP experience but little GPU experience.

Biggest challenge for me was a lack of profiling tools when running OpenCL on an NVIDIA GPU, I got OK performance but had no way to identify hotspots or opportunities for improvement.