Open source project pocl 4.0 has been presented, which includes an implementation of the OpenCL standard independent of graphic accelerator manufacturers. The release supports the use of different backends to execute OpenCL shaders on various types of graphic and central processors. The project code can be accessed under the MIT license on GitHub.
Pocl 4.0 is designed to work on multiple platforms, including X86_64, MIPS32, ARM V7, AMD HSA APU, NVIDIA GPU, and specialized ASIP (Application-Specific Instructions Set Processor) and TTA processors with VLIW architecture.
The OpenCL nuclei compile is based on LLVM, with Clang as the frontline for OpenCL C. The compiler can generate combined functions that can make the most of hardware resources for parallelisation of code execution. This includes VLIW, scalar superscalar, SIMD, SIMT, multi-stroke and multi-plate. The project includes support for ICD drivers.
New versions of pocl will add support for Clang/LLVM 6-9 and intermediate representations of spir-V shaders for drivers that use the CPU and Cuda. Additionally, a driver for devices using the API Level Zero has been added. The release also includes support for global variables and general address space, as well as functions from CL_KHR_FP16, CL_KHR_SUBGROUPS and CL_INTEL_UNIFIED_SHARED_MEMORY.
More information and downloads are available on the pocl website.
POCL 4.0 Launches with Independent OpenCL Implementation
/Reports, release notes, official announcements.