Grunnur, a generalized API for CUDA and OpenCL

Grunnur is a thin layer on top of PyCUDA and PyOpenCL that makes it easier to write platform-agnostic programs. It is a reworked cluda submodule of Reikna, extracted into a separate module.

Main features

• For the majority of cases, allows one to write platform-independent code.

• Simple usage of multiple GPUs (in particular, no need to worry about context switching for CUDA).

• A way to split kernel code into modules with dependencies between them (see Module and Snippet).

• Various mathematical functions (with complex numbers support) organized as modules.

• Static kernels, where you can use global/local shapes with any kinds of dimensions without worrying about assembling array indices from blockIdx and gridIdx.

• A temporary buffer manager that can pack several virtual buffers into the same physical one depending on the declared dependencies between them.

Where to get help

Please file issues in the the issue tracker.

Discussions and questions are handled by Github’s discussion board.

Table of contents

• Introduction
• Tutorial: modules and snippets
  • Snippets
  • Modules
  • Other constructors
  • Module and snippet discovery
  • Nontrivial example
• Public API
  • API discovery
  • Platforms
  • Devices
  • Device discovery
  • Contexts
  • Queues
  • Buffers and arrays
  • Programs and kernels
  • Static kernels
  • Utilities
  • Data type utilities
  • Function modules
  • Virtual buffers
  • Kernel toolbox
• Version history
  • 0.3.1 (Unreleased)
  • 0.3.0 (29 Jan 2023)
  • 0.2.0 (10 Mar 2021)
  • 0.1.1 (9 Oct 2020)
  • 0.1.0 (9 Oct 2020)

Indices and tables

• Index

• Module Index

• Search Page