support lower precisions

[bournejt]

Hi csp team, I noticed that csp reinforces double64 and int64 in the computation core to prioritize numerical stability and consistency. But if we could allow a config to use float32/float16 and int32, which is quite sufficient for vast majority of the use cases, this would speed things up easily. I know this won't be a trivial task. Just think it's a very natural extension. I want to put the idea forward and see how you guys think about it.

[AdamGlustein]

Not sure I agree with you on "this would speed things up easily" (both the 'speed' and the 'easy' part 😄 ) , although I'm interested to hear if you have a specific use case in mind. My understanding is that these lower precision floating point types are great for vectorized (SIMD) operations (i.e. big matrix multiplication) but as the type of an individual stream in csp I don't see much point. Plus, if/when we convert back to a Python float object, we'd need to cast the value to double precision anyways.

You can propagate numpy arrays through a csp graph like any Python object, so if you are doing some heavier ML/linear algebra based workflows in csp and want to use float16 or float32, I'd recommend passing arrays of these values through your graph rather than the individual values themselves.

[bournejt]

Thanks for the response @AdamGlustein . I did some further research and I can totally agree with what you said. Thanks for the help!