The open source implementation of OpenCL PoCL (Portable Computing Language) is now available for download in version 1.6. The new release of the OpenCL implementation designed for CPUs and accelerators with LLVM backend works with LLVM / Clang 11. While only LLVM 10 is officially supported, developers should also be able to work with variants down to LLVM 6.0.

More power for CUDA The project, originally started under the name Portable OpenCL, covers the entire range of functions of OpenCL 1.2 from, as well as individual parts of OpenCL 2.0. Other new features in PoCL 1.6 include higher performance when used with Nvidia’s CUDA, improved support for OpenCL debugging and expanded options for packaging PoCL binaries. According to the PoCL development team, the optimizations that have been incorporated show performance increases in the CUDA backend, particularly in benchmarks that include local storage (e.g. FFT and GEMM). In many cases PoCL is now on par with Nvidia’s proprietary OpenCL driver.

Developers now have greater flexibility when packing and distributing PoCL binaries. While the build computer and the host machine had to have the same driver support when building the binary files, PoCL 1.6 now checks at runtime whether the devices that match the compiled drivers are available. Therefore, with the package manager Conda, PoCL binary packages with CUDA support for Linux-x 86 _ 64 and Linux-ppc 64 le.

The release notes provide a complete overview of all new features in PoCL 1.6 and more details on the improvements.

(map)