CuSMC

An R package for Sequential Monte Carlo sampling, accelerated with CUDA on NVIDIA GPUs.

C++, CUDA, R
GPL-3.0 license
github.com/tkamucheka/CuSMC

CuSMC is an R package for drawing samples from a posterior probability distribution in Bayesian inference. Its name stands for CUDA Sequential Monte Carlo. It provides a Metropolis-Hastings sampler and multivariate normal and Student-t distributions.

A GPU is not required: without an NVIDIA GPU and the CUDA toolkit, CuSMC is built for the CPU only. On Linux with an NVIDIA GPU and the CUDA toolkit, it is built with GPU acceleration. macOS supports CUDA only up to macOS 10.13.6, and Windows builds are CPU-only.

Installation

CuSMC is built from source, which requires a C++ toolchain (on Windows, Rtools) and the R packages devtools, Rcpp, and RcppEigen:

install.packages(c("devtools", "Rcpp", "RcppEigen"))
devtools::install_github("tkamucheka/cusmc")
library(CuSMC)

Platform-specific requirements are described in the README.

Publication

Kamucheka, T., Mai, Q., Wei, G., Huang, M., & Liu, X. CuSMC: Fast Parallel Implementation for Sequential Monte Carlo on GPU, R package.