Skip to contents

anvl

Package website: release | dev

Composable code transformation framework for R, allowing you to run numerical programs at the speed of light. It currently implements JIT compilation for very fast execution and reverse-mode automatic differentiation. Programs can run on various hardware backends, including CPU and GPU.

Installation

{anvl}1 can be installed from GitHub or r-universe. During runtime, we require libprotobuf. Source installation requires a C++20 compiler and protoc (protobuf compiler).

# Install release
# from r-universe (prebuilt binary)
install.packages("anvl", repos = c("https://cloud.r-project.org", "https://r-xla.r-universe.dev"))
# from GitHub (installation from source)
pak::pak("r-xla/anvl@*release")
# Install dev version
pak::pak("r-xla/anvl")
# Install CUDA support (linux x86_64): only requires a compatible driver
install.packages("cuda12.8", repos = "https://mlverse.r-universe.dev")

Prebuilt Docker images are also available. See the Installation vignette on the package website for more details.

Quick Start

Below, we create an R function and jit() it. Then, we call the resulting function on AnvlArrays, which will compile and subsequently execute it.

library(anvl)
f <- function(a, b, x) {
  a * x + b
}
f_jit <- jit(f)

a <- nv_scalar(1.0, "f32")
b <- nv_scalar(-2.0, "f32")
x <- nv_scalar(3.0, "f32")

f_jit(a, b, x)
#> AnvlArray
#>  1
#> [ CPUf32{} ]

Through automatic differentiation, we can also obtain the gradient of the above function.

g_jit <- jit(gradient(f, wrt = c("a", "b")))
g_jit(a, b, x)
#> $a
#> AnvlArray
#>  3
#> [ CPUf32{} ] 
#> 
#> $b
#> AnvlArray
#>  1
#> [ CPUf32{} ]

For more complex examples, such as implementing a Gaussian Process, see the package website.

Main Features

  • Automatic Differentiation:
    • Gradients for functions with scalar outputs are supported.
  • Fast:
    • Code is JIT compiled into a single kernel.
    • Runs on different hardware backends, including CPU and GPU.
    • Asyncronous allocation and execution, allowing the accelerator to do it’s job while R interpretes.
  • Extendable:
    • It is possible to add new primitives, transformations, and (with some effort) new backends.
    • The package is written almost entirely in R.
  • Multi-backend:
    • The backend supports execution via XLA as well as an experimental {quickr}-based Fortran backend (CPU only).

Platform Support

  • Linux (x86_64)
    • ✅ CPU backend is fully supported.
    • ✅ CUDA (NVIDIA GPU) backend is fully supported.
  • Linux (ARM)
    • ✅ CPU backend is fully supported.
    • ❌ GPU is not supported.
  • Windows
    • ✅ CPU backend is fully supported.
    • ⚠️ GPU is only supported via Windows Subsystem for Linux (WSL2).
  • macOS (ARM)
    • ✅ CPU backend is supported.
    • ⚠️ Metal (Apple GPU) backend is available but not fully functional.
  • macOS (x86_64)
    • ❌ Not supported.

Acknowledgments

  • This work is supported by MaRDI.
  • The design of this package was inspired by and borrows from:
  • For JIT compilation, we leverage the OpenXLA project.