cupla [qχɑpˈlɑʔ] is a simple user interface for the platform independent parallel kernel acceleration library alpaka. It follows a similar concept as the NVIDIA® CUDA® API by providing a software layer to manage accelerator devices. alpaka is used as backend for cupla.
Please keep in mind that a first, "find & replace" port from CUDA to cupla(x86) will result in rather bad performance. In order to reach decent performance on x86 systems you just need to add the alpaka element level to your kernels.
(Read as: add some tiling to your CUDA kernels by letting the same thread compute a fixed number of elements (N=4..16) instead of just computing one element per thread. Also, make the number of elements in your tiling a compile-time constant and your CUDA code (N=1) will just stay with the very same performance while adding single-source performance portability for, e.g., x86 targets).
cupla is licensed under LGPLv3 or later.
For more information see LICENSE.md.
- cmake 3.22.0 or higher (depends on the used alpaka version)
- alpaka 1.0.0 or newer
- alpaka is loaded as
git subtreewithin cupla, see INSTALL.md
- alpaka is loaded as
- See our notes in INSTALL.md.
- Checkout the guide how to port your project.
- Checkout the tuning guide for a step further to performance portable code.
- Checkout the interoperability guide to learn more on how to use cupla with software developed with an alpaka compatible interface.
cupla can be used as a header-only library and without the CMake build system
Any pull request will be reviewed by a maintainer.
Thanks to all active and former contributors.
All product names and trademarks are the property of their respective owners. CUDA® is a trademark of the NVIDIA Corporation.