cuDecomp: An Adaptive Pencil Decomposition Library for NVIDIA GPUs

These pages contain the documentation for cuDecomp, an adaptive pencil decomposition library for NVIDIA GPUs.

• Disclaimer: This library is currently in a research-oriented state, and has been released as a companion to a paper presented at the PASC22 conference (link). We are making it available here as it can be useful in other applications outside of this study or as benchmarking tool and usage example for various GPU communication libraries to perform transpose and halo communication.

Please contact us or open a GitHub issue if you are interested in using this library in your own solvers and have questions on usage and/or feature requests.

Table of Contents

• Overview
  • What is cuDecomp?
  • Decomposition Layout
• Basic Usage Guide
  • Starting up cuDecomp
  • Creating a grid descriptor
  • Allocate pencil memory
  • Working with pencil data
  • Allocating workspace
  • Transposing the data
  • Updating halo regions
  • Cleaning up and finalizing the library
  • Building and running the example
• Autotuning
  • Autotuning process
  • Autotuning usage
• Working with NVSHMEM-enabled builds
  • Controlling the symmetric heap size
  • MPI compatibility
• cuDecomp API
  • cuDecomp C API
  • cuDecomp Fortran API
• Environment Variables
  • CUDECOMP_ENABLE_NCCL_UBR

Indices and tables

• Index

• Module Index

• Search Page