CuSMC

CuSMC is an R package for drawing random samples for a posterior probability distribution in Bayesian inference. CuSMC stands for CUDA Sequential Monte Carlo. CuSMC supports Metropolis-Hastings sampler and Multivariate Normal and Student-T distributions.

Prerequisites

You do not need a GPU to install and run CuSMC library. However, to enjoy the best perfomance, make sure you are running Linux and an Nvidia GPU (Graphics Processing Unit) that supports CUDA (most if not all modern GPUs do) and the CUDA SDK installed. Details on how to set up the CUDA SDK kit can be found here. Windows and MacOS generally will not support GPU acceleration. See more details on MacOS support below.

Linux

If there is no NVidia GPU and CUDA SDK installed, CuSMC will install with CPU support only. If an NVidia GPU and CUDA SDK are found during installation then full GPU acceleration becomes available. You will need the GCC compiler and tool chain installed to build the library on your computer. If you are on Debian based distro like Ubuntu, you can install the compiler with:

$ sudo apt get install build-essential -y

MacOS

MacOS supports NVidia GPUs and CUDA up to MacOS version 10.13.6. If you have a version of MacOS after 10.13.6 then GPU acceleration will not be available. There are two ways to get a compiler toolchain installed on MacOS. Option 1, is through xcode-select and the other is via Homebrew:

# Xcode
$ xcode-select --install

# Homebrew
$ brew install gcc

Windows

Currently, GPU acceleration is not available on Windows. Before attempting to install, you will need to install RTOOLS40 to get the MSYS2 MINGW64 environment and GCC compiler and toolchain installed and available in R. RTOOLS40 can be foundR here

When you have installed the GCC compiler and toolchain, next you will need to install the devtools, Rcpp and RcppEigen packages from CRAN as a last step before installing CuSMC

## Install the following packages from CRAN
install.packages("devtools")
install.packages("Rcpp")
install.packages("RcppEigen")

## If you are using an old version of R (R 3.6 or older),
## you could also need to install RcppArmadillo package
install.packages("RcppArmadillo")

Installation

You can install CuSMC from this repository with:

devtools::install_github("tkamucheka/cusmc")
library(CuSMC)

Examples

## basic example code