rcocl

An R package for running MCMC samplers on GPUs via OpenCL. Write your posterior function in OpenCL C, and rcocl handles compilation, multi-device dispatch, adaptive proposals, and sample collection.

Installation

devtools::install_github("bamonroe/rcocl")

Requires an OpenCL-capable GPU and drivers.

Features

• GPU-accelerated MCMC: Run thousands of chains in parallel on GPU hardware
• Adaptive proposals: Automatic tuning of proposal covariance matrices with configurable adaptation schedules
• Multi-device support: Distribute chains across multiple OpenCL devices
• Batched parameters: Group parameters into blocks with independent covariance structures
• Binary caching: Compiled OpenCL kernels are cached by MD5 hash for fast re-runs
• Built-in distributions: OpenCL implementations of common distributions (normal, beta, log-normal, etc.)

Usage

library(rcocl)

# Define your posterior in OpenCL C (inline or as a .c file)
kernel_code <- "
float my_posterior(float* pars, ...) {
  // your log-posterior here
}
"

# Configure the sampler
config <- list(
  draws       = 10000,
  skip        = 1000,
  fn          = "my_posterior",
  algo        = "mhmc",
  device_list = list(list(plat_num = 0, dev_num = 0, work_items = 256))
)

# Run
result <- gpu_mcmc(kernel_code, config,
                   init = c(a = 0, b = 0),
                   init_sd = c(0.1, 0.1),
                   dat = list(double = list(y), int = list()))

# result$samples is a data.frame with columns: iter, chain, posterior, a, b

License

MIT