Skip to main content
Ctrl K

hipDecomp

hipDecomp is an Adaptive Pencil Decomposition Library for AMD GPUs.

1
contributor
127 commitsLast commit ≈ 4 weeks ago0 stars0 forks

Description

In multi-dimensional MPI-based codes transposes of the domain across all MPI ranks are often needed for solvers, which can be a significant performance bottleneck.

NVIDIA has developed cuDecomp, a library that provides routines for these transposes as well as halo cell communication on GPUs, however it only supports NVIDIA GPUs. Several of the world’s largest supercomputers contain AMD GPUs and thus are excluded from being used by cuDecomp-enabled codes. hipDecomp is a HIP port of cuDecomp, allowing these routines to be used on AMD GPUs as well.

hipDecomp is implemented in C++/HIP, with an optional interface in Fortran. Optionally, the hipDecomp API can use cuDecomp naming, allowing for easy transition in existing codes.

Keywords
Programming languages
  • C++ 61%
  • Fortran 16%
  • C 10%
  • Cuda 5%
  • Python 4%
  • CMake 2%
  • HIP 1%
License
</>Source code

Participating organisations

Netherlands eScience Center

Contributors

Contact person

LO

Leon Oostrum

Research Software Engineer
Netherlands eScience Center
0000-0001-8724-8372Mail Leon
LO
Research Software Engineer
Netherlands eScience Center
0000-0001-8724-8372

Related projects

PPPD

Packaged Performance Portability for cuDecomp

Updated 2 months ago
In progress

Related software

CaNS

CA

Canonical Navier-Stokes is a library for massively-parallel numerical simulations of fluid flows.

Updated 1 hour ago
133 3