using Pkg
Pkg.add("OpenLibmPorts")As the name says, ports from openlibm (https://github.com/JuliaMath/openlibm). This effort was borne of the need for pure Julia implementations of various special functions for use in automatic differentation. Expect low coverage as I create ports only as need dictates.
Why bother, you might ask? In the construction of Bayesian models, certain special functions may show up in the likelihood (!!! e.g. loggamma in negative binomial), which means that considerable latency can be introduced by marginally slower special functions. One might be inclined to be incredulous at such a statement, but consider that a log-likelihood computation is essentially a sum involving M terms, i.e. ∑ᵢ₌₁ᴹ f(xᵢ). The total cost is linear in the number of terms. If we assume the cost of summation itself to be negligible (it's not), then the cost of the log-likelihood is equal to M times the cost of f. In this light, any reduction to the cost of f translates directly to an equivalent reduction in the log-likelihood computation. As the log-likelihood is most often the dominant cost in a probabilistic model (which does not involve ODEs), reductions thereof translate directly to reductions in overall time required to fit a model.
Note that I am being purposefully vague on what "fitting" means -- we might require both the evaluation of the log-posterior and its gradient, or, just the log-posterior (though, that is rare these days). In either case, the cost of special functions is not to be ignored.
using OpenLibmPorts
julia> OpenLibmPorts.loggamma(2.009904444771008)
0.004219012233950873
julia> OpenLibmPorts.logabsgamma(2.009904444771008)
(0.004219012233950873, 1)
