Intel Compute Runtime Continues Building Out Exciting "LEO"

Going in-step with last week's release of the Intel Graphics Compiler 2.38.2, out today is the Intel Compute Runtime 26.27.39122.11. Most intriguing with the new release is Intel continuing to build out the still-experimental LEO feature.

Intel Compute Runtime continues to serve as the principal OpenCL and Level Zero compute stack for Intel integrated/discrete graphics hardware on Windows and Linux systems. Introduced in last month's Compute Runtime release was LEO as "Level Zero Executing OpenCL". LEO comes down to being an experimental re-implementation of the OpenCL API from scratch atop the Intel Level Zero interface. Rather than maintaining two separate driver stacks, Intel engineers are ultimately working to have OpenCL layered atop Level Zero.

LEO has the potential of delivering greater OpenCL performance on Intel graphics hardware with the engineers investing more into tuning Level Zero as well as greater test coverage of Level Zero. With the Intel Compute Runtime 26.27.39122.11 release, LEO is still disabled by default and requires setting the "EnableLEO" release variable to enable. When LEO is enabled, "OpenCL 3.0 LEO" will show up in the OpenCL version information rather than "OpenCL 3.0 NEO". While core functionality is working with LEO, not all OpenCL extensions are yet implemented and performance tuning is still ongoing.

The new Intel Compute Runtime release has new performance tuning in general such as now reducing the unified shared memory (USM) pool size to 2MB, adding new AVX2 and AVX-512 optimized memory copy functions, and enabling the USM pool manager on device USM in OpenCL.

Compute Runtime 26.27.39122.11 also has other new features such as introducing zeGraphGetIdExt for querying the root graph's ID, Sysman adding support for the Fan APIs in Linux, and support for a new Zebin section.

Downloads and more details on today's Intel Compute Runtime open-source update via GitHub.