A previous thread on this topic.
I really like the idea from that thread of using GPUCompiler to emit small “kernels” that run on the microcontroller. It is still a lot of work, but it sure seems better than limiting to microconctrollers that can fit the whole runtime.