I like the approach for generality, but for something like the word count example, there's absolutely no way that parallelizing at the multi-core level would be

I like the approach for generality, but for something like the word count example, there's absolutely no way that parallelizing at the multi-core level would beat doing this in SIMD. None of those examples are topping at at more than a gigabyte per second, which is walking pace for this sort of problem. Further, I suspect that the extremely regular pattern of the space boundaries is probably giving the branch predictor a leg up. It almost certainly would pick up the 'every 5' pattern if not the 'every 17' pattern. So a realistic example would be far more branch miss heavy. This is still fine work, as the effort of writing a nice C++ function and getting some parallel help on it is tiny compared to writing SIMD code, and some problems don't play nice with SIMD at all.