OpenMP proof of concept

[LTLA]

Tagging @mtmorgan as per our previous discussions on Slack.

Typical usage:

library(BiocSingular)
library(Matrix)
mat <- rsparsematrix(2e4, 1e5, density=0.01)
BiocSingular:::set_omp_threads(2L)
system.time(X <- BiocSingular:::compute_scale(mat, NULL))

A couple of points:

• Uses set_omp_threads() (stolen from ShortRead) to control the number of OpenMP threads globally. This would be an excellent candidate for inclusion as an option in a BiocParallelParam object somewhere, such that entry into a bp*apply loop also sets the OpenMP threads. This would centralize control of OpenMP parallelization into the BiocParallel framework, and free developers from having to define their own OpenMP-related arguments in every parallelized functions.
• beachmat's getters are maintained in serial sections. This is because they may involve calls back out to the R API, and it would probably be dicey to assume that R is happy with multiple threads calling it at the same time. As a result, this code saturates quite quickly with increasing cores as the loops are bottlenecked at the data loading step. Oh well.