Released earlier this month was the OpenCL 3.1 specification with a focus on enhancing AI and HPC workloads for this long-time Khronos specification. Out today is OpenCL 3.1.1 as a point release with an emphasis on addressing a possible performance regression of OpenCL 3.1.

OpenCL 3.1.1 reverts the short-lived OpenCL 3.1 behavior of clGetEventInfo returning CL_COMPLETE as a host synchronization point. Those wanting a host synchronization point should instead call a function waiting on the OpenCL event instead like with clWaitForEvents. The pull request argued the change in behavior back to its OpenCL 3.0 semantics to avoid a performance regression with the cost of host synchronization:

"This PR changes the behavior of clGetEventInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) returning CL_COMPLETE back to the behavior in OpenCL 3.0. This avoids a potential performance regression when the stronger host synchronization point is not needed, for example to determine if the event is CL_COMPLETE to query event profiling data."

OpenCL 3.1.1 also reserves some enum blocks for forthcoming Intel and Qualcomm extensions. Plus a few other minor fixes but the main change and motivating this quick point release is for reverting the clGetEventInfo behavior.

The OpenCL 3.1.1 spec can be found on GitHub.