TensorFlow Lite for AI inference on mobile devices now has support for making use of OpenCL on Android devices. In doing so, the TFLite performance presents around a 2x speed-up over the existing OpenGL back-end.

To little surprise, the TensorFlow developers are finding their new OpenCL back-end for TFLite to be much faster than their OpenGL back-end for mobile inference. Thanks to better performance profiling abilities, native FP16 support, constant memory, and OpenCL being better designed for compute than OpenGL ES

Continuing on with all of the OpenCL Mesa work that's been going on by Red Hat developers in recent time, LLVM's libclc library now has support for targeting Mesa SPIR-V.

The new target added to LLVM's libclc is for emitting SPIR-V geared for Mesa's OpenCL support. Libclc is the LLVM sub-project focused on providing a library for OpenCL C run-time usage. Existing targets have included AMDGPU/AMDGN, R600, and NVIDIA NVPTX while SPIR-

It landed sooner than anticipated but the LLVMpipe patches enabling OpenGL 4.5 support were merged to Mesa 20.3-devel today and are also marked for back-porting to the Mesa 20.2 series soon to be promoted to stable.

The Mesa 20.2 feature code was already set to take this software-based OpenGL driver from GL 3.3 to GL 4.3 after all the work carried out by Red Hat's David Airlie in recent months. But now his work on

For years Mesa's LLVMpipe software rasterizer has been bound to OpenGL 3.3 support but finally in recent months Red Hat's David Airlie has been tackling OpenGL 4.x support... Right now in Mesa it's at OpenGL 4.3 but there are patches seemingly to be mainlined very soon that will take it up to OpenGL 4.5.

As it stands right now, Mesa 20.2 will ship in a few weeks with OpenGL 4.3 support for LLVMpipe rather than OpenGL

For a while now Intel has been quietly been working on "mOS" as the "multi-OS" that is a modified version of the Linux kernel that in turn is running lightweight kernels for high-performance computing purposes.

Intel mOS has been seldom talked about (or incredibly rare, based on public searches) as it's still largely a research project but showing much potential in the area of high performance computing for delivering better scalability and reliability of HPC workloads. In fact, mOS can already be used on

It looks like Intel will soon be tagging their oneAPI Level Zero specification as version 1.0.

At the end of last year Intel published the oneAPI Level Zero specification as a low-level API for direct-to-metal interfaces for offload accelerators like FPGAs and GPUs. In the months since they have continued advancing the Level Zero interface and implementation within the Intel software stack (along with the other oneAPI components at large) while it's looking like Level Zero v1.0 is around the corner.

Currently