GPU programming for research software

Accelerators dominate modern research computing, but the ecosystem is fragmented: vendor-native models, portability layers and language bindings each trade performance against maintainability. This skill turns that landscape into decision support.

Choosing a programming model

• CUDA: NVIDIA-native; largest ecosystem and learning material; locks the code to one vendor.
• HIP: AMD's near-CUDA model; a pragmatic path when AMD hardware is in scope, with source-to-source translation available from CUDA.
• SYCL: single-source C++ across vendors; growing research adoption.
• OpenACC / OpenMP offloading: directive-based; the gentlest port for existing Fortran/C code, at some control cost.

Decide by: target machines (which vendors, for how long), the team's languages, and how much low-level control the kernels genuinely need. Prefer libraries over hand-written kernels wherever an existing GPU library covers the computation.

Calling GPUs from high-level languages

Research code rarely starts in C++: CuPy and PyCUDA (Python) and CUDA.jl (Julia) expose GPU arrays and kernels with far less ceremony. Reach for them before rewriting a pipeline in a systems language, and keep the array-API boundary clean so kernels stay swappable.

Portability layers

Kokkos and Raja abstract over backends for C++ codebases that must outlive any single vendor; they suit infrastructure-tier software with long horizons more than one-off analysis kernels.

Performance work

• Profile before optimizing, with the vendor profilers; measure transfers as well as kernels - data movement dominates many research workloads.
• Auto-tuning tools (for example Kernel Tuner) search launch configurations systematically; prefer them to hand-tuned magic numbers, and record tuned configurations per hardware target.
• Keep a CPU reference path for correctness testing (rseng-testing) and document the hardware requirements (rseng-documentation, rseng-reproducible-environments for the driver/toolkit environment).

Working with this skill

The generated references.md beside this file lists the source material and pointers:

• references.md - verified Learn more pointers

Learn more (verified):

• https://docs.nvidia.com/cuda/ - NVIDIA CUDA Toolkit documentation
• https://rocm.docs.amd.com - AMD ROCm and HIP documentation
• https://docs.cupy.dev - CuPy GPU array library for Python
• https://kokkos.org - Kokkos performance portability ecosystem
• https://enccs.github.io/gpu-programming/ - ENCCS GPU programming lesson

Related skills

Check whether any of these applies before moving on:

• rseng-green-computing - accelerator energy efficiency matching
• rseng-hpc-computing - running GPU jobs on clusters
• rseng-numerical-accuracy - float32 precision consequences
• rseng-performance-profiling - verify GPU is warranted first
• rseng-reproducible-environments - driver and toolkit pinning
• rseng-testing - CPU reference path for correctness