PoCL 6.0 OpenCL Implementation Brings OpenMP For CPU Driver, More Remote Driver Features

The Portable Computing Language "PoCL" that started off as a CPU-based OpenCL implementation has grown to support multiple hardware targets from NVIDIA PTX to Intel Level Zero to AMD ROCm and other innovations like a recent remote driver for transparent OpenCL across networked systems. PoCL 6.0 was released today for delivering the latest enhancements to this independent OpenCL compute implementation and continuing to enhance support for its different hardware targets.

With PoCL 6.0 for its remote OpenCL driver there is now support for coarse-grained Shared Virtual Memory (SVM), clCompileProgram/clLinkProgram functionality, vsock, and other features for OpenCL across the LAN.

PoCL 6.0 also introduces "cpu-tbb" as a new CPU back-end implementation that targets the Intel oneTBB (Threaded Building Blocks) library for scheduling. PoCL's CPU driver has also added support for OpenMP threading as well as supporting more OpenCL extensions.

In addition to the new Intel oneTBB CPU driver, the existing Intel oneAPI Level Zero driver for targeting Intel iGPUs/dGPUs now has a host synchronization optimization and can support buffers larger than 4GB.

PoCL 6.0 with NVIDIA CUDA now supports a larger CL_DEVICE_MAX_MEM_ALLOC_SIZE. This PoCL update also supports building against the LLVM Clang 18.0 compiler stack. There is also experimental support for the cl_ext_buffer_device_address extension.

Downloads and more details on the PoCL 6.0 release via GitHub.