Operation definitions for Bessel functions

[jumerckx]

An operation is defined for each function listed here: https://specialfunctions.juliamath.org/stable/functions_overview/#Bessel-Functions

For now, all function arguments are mapped onto operation operands, as opposed to constant attributes.
Some of the variants are special cases of more general ones:

besselj(nu,z)   # Bessel function of the first kind of order nu at z

besselj0(z)     # besselj(0,z)
besselj1(z)     # besselj(1,z)

[wsmoses]

Is there a reason for us to specialize j0 here? If we have the same implementation it lowers into anyways, no need for now

[jumerckx]

I imagined those might have custom approximation lowerings but I'm not sure about that. Will see if I can find something

[jumerckx]

for one, mpfr seems to have separate implementations for the different bessels https://github.com/JuliaMath/SpecialFunctions.jl/blob/1743a8b7ac1565213e87de418765c594720929b6/src/bessel.jl#L682C1-L709C4
I'm not sure yet what the lowering in stablehlo would look like.

[wsmoses]

in julia right now it just looks like they call custom c library functinos [which we can't do ourselves here anyways]

[wsmoses]

id probably just start with the more general one, and add the specialized one if we find a lowering for it later

[jumerckx]

SameOperandsAndResultType I'll have to change this as some arguments can be complex.

[github-actions]

EnzymeJAX Benchmarks

Benchmark suite	Current: 45e39ae	Previous: 7054610	Ratio
scatter_sum / JaX / cpu / Primal	0.000004400375899967912 s	0.000004342221499973675 s	1.01
scatter_sum / JaXPipe / cpu / Primal	0.000004491189899999881 s	0.000004368584500025463 s	1.03
scatter_sum / JaX / tpu / Primal	0.0001501408806 s	0.0001383197025999 s	1.09
scatter_sum / JaXPipe / tpu / Primal	0.0001491218746999 s	0.0001325478068999 s	1.13

This comment was automatically generated by workflow using github-action-benchmark.

[jumerckx]

The ops seem to be working, I've got them hooked up to Reactant locally.