nvptx-rs

Parallel programming with Rust and CUDA

This repository contains applications to demonstrate the capability of Rust to program and run CUDA kernels. It leverages the LLVM nvptx64 backend as the Rust compiler rustc does already support it by its nightly toolchain.

It was built as a project for the module Fachpraktikum Parallel Programming at the FernUniversität in Hagen during winter semester 2024/25.

Prerequisites

To built and run the application it is necessary

• to have a CUDA capable NVIDIA GPU
• to have CUDA installed
• to have a nightly Rust toolchain with some additional tools installed

For the Rust part the easiest way is to install these by the Rust hosted tool rustup.

Installing rustup

To install rustup please visit https://rustup.rs/ and execute the shown command in a shell of your choice.

Installing a nightly toolchain

To build for the nvptx64 backend a nightly toolchain is necessary. It can be installed with the following command:

rustup toolchain install nightly

Installing necessary components

For building and linking some llvm tools are necessary. To install these for the nightly toolchain you can use the following commands:

rustup +nightly target add nvptx64-nvidia-cuda
rustup +nightly component add llvm-tools llvm-bitcode-linker

Installing CUDA

CUDA can be installed by following the instructions on https://developer.nvidia.com/cuda-downloads.

Building and running

If everything is set it is possible to build and run the applications. First step is to navigate to the rust directory. For example if the fapra-project folder is placed directly under ~: cd ~/fapra-project/rust.

The command cargo xtask build --release builds the device code and the host code and places the binaries in fapra-project/rust/target/release.

The applications can simply be run with

target/release/stencil
target/release/reduction
target/release/matrix_multiply

stencil, reduction and matrix_multiply have parameters that can be set from command line. To get some information about these simply call the applications with --help.

License

Licensed under either of

Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.