Multi-threading Support

library(rwig) |> suppressPackageStartupMessages()

The log methods for sinkhorn() and barycenter() both require row-by-row and column-by-column soft-minimum operations for each iteration of the algorithm, and therefore suffer from slow computation time compared to the vanilla/parallel version.

If the dimensions of \(M\) and \(N\) are large (recall that cost matrix C is of size \(M \times N\)), we can use multi-threading to process the rows and columns simultaneously.

This can be done by setting n_threads to an integer bigger than 0. By default it is 0, and it means that threading is disabled. Also, setting n_threads for “vanilla” or “parallel” methods will be ignored automatically.

But you might ask: if multi-threading is so wonderful, why don’t you set threading as the default? This is because threading comes with an overhead, and sometimes for small problems, it can even be slower than serial processing. So be sure to benchmark your code and see if threading actually helps.

BLAS threads

rwig calls the BLAS library that R is linked against for its matrix products. Some BLAS libraries (OpenBLAS, MKL) run multi-threaded by default. Because rwig already parallelizes the expensive log-domain kernels through n_threads, and because most of its BLAS calls are small, letting the BLAS spawn its own threads on top usually oversubscribes the cores and slows things down. We therefore recommend a single BLAS thread while using rwig.

rwig does not change this setting for you (it is a session-wide setting that also affects every other package). If the optional package RhpcBLASctl is installed, the startup message reports the current BLAS thread count, and you can set it for the session with

RhpcBLASctl::blas_set_num_threads(1)

The exception is wdl()/wig() on the CPU with an optimized BLAS: their cost is dominated by large matrix products, so a multi-threaded BLAS can help there. Benchmark both settings on your own problem sizes.

See Also

See also vignette("sinkhorn"), vignette("barycenter").

References

Peyré, G., & Cuturi, M. (2019). Computational Optimal Transport: With Applications to Data Science. Foundations and Trends® in Machine Learning, 11(5–6), 355–607. https://doi.org/10.1561/2200000073

Xie, F. (2025). Deriving the Gradients of Some Popular Optimal Transport Algorithms (No. arXiv:2504.08722). arXiv. https://doi.org/10.48550/arXiv.2504.08722