At the end of December was the last X.Org EVoC OpenCL update. Read that article for the background of this project, which is being funded by the X.Org Foundation and that late-2011 Phoronix posting also covers the background of the Clover state tracker for OpenCL, how this EVoC work builds upon the Google Summer of Code OpenCL work from last year, etc.
Work done so far -- since the project began in mid-October -- included driver-specific changes to Nouveau to make it more OpenCL/compute friendly, and extending the TGSI IR and Gallium3D API to allow for GPGPU (General Purpose) computing. This latest period, from the end of December through earlier this week, was all about hacking on Clover / the OpenCL Gallium3D state tracker. This work is being done by Francisco Jerez.
This past month of Gallium3D compute work included redesigning and simplifying the Clover state tracker, which in the process removed its intermediate device abstraction layer since the Gallium3D API is already fitting well to talk directly to the device -- whether it be an actual GPU or a virtual case like in the case of the LLVMpipe CPU-based driver or the SVGA virtual adapter for VMware virtual machines.
Most of Clover was rewritten during this past month, and in the process of doing so, is now extensively utilizing C++11 (the latest C++ standard previously known as C++0x) and its new functionality (lambda expressions, type inference, STL functionality, variadic templates, range-for loops, etc). Turning to more C++11 usage within the state tracker turned out very well for the developer, but for now this means GCC 4.6 or later must be used in order to handle the functionality used within this latest C++ ISO standard.
The Clover library at the moment is relying upon TGSI IR as the compute input, but an LLVM back-end that will handle the translation to the TGSI instructions is currently being developed. There's also some other work left to the library like handling some of the less-used API functionality, buffer sharing between devices (hey DMA-BUF!), and other work.
Francisco Jerez's EVoC project is set to officially end by the middle of February with the final few weeks meant for final cleaning, documentation, and other miscellaneous changes. Hopefully he will continue to stick around in the Nouveau land -- where he was known before the project began -- and continue seeing the OpenCL support through in the open-source stack.
While seeing progress in the Nouveau / open-source NVIDIA land, Tom Stellard at AMD is continuing to work on the OpenCL support there. A significant portion of the work being done by Francisco is also relevant to all Gallium3D drivers wishing to utilize compute support. The Intel OSTC developers aren't actively working on any of the OpenCL code (or to get LLVMpipe for open-source OpenCL running on Intel CPUs), but you can ask them about it in this very lengthy discussion.
The notes by Francisco Jerez for the X.Org EVoC compute work can be found on the X.Org Wiki. The code at present isn't going into Mesa master, but rather this GitHub repository, but ideally we will see all of the Clover work hit before the Mesa 8.1 release in six months.