Request: add log1p in math prelude

[nbos]

log1p(x) computes accurately log(1 + x) even when x is very small.

It looks like OpenCL and CUDA have it as primitives, although I can't find reference to a 16-bit CUDA log1p.

I can give it a go if it turns out just to require appending definitions to the prelude.

[athas]

Sure, have at it. The main challenge is getting it done for all the backends, but if you add a test to tests/primitive, CI will tell you the ones you missed. For the 16 bit version, I would just use 32 bit arithmetic - we do that with many others. Let me know if you need more pointers.

[nbos]

Will do, thanks!

[nbos]

How bad is it if results vary by 1 or 2 ULPs between backends? E.g. Haskell's log1p 2.0 gives 1.0986122886681096 to the interpreter while my implementation for ISPC gives 1.0986122886681098. Do I just omit those cases in tests?

[athas]

The testing tool uses a fairly generous tolerance when comparing.