The Gallium3D compute infrastructure, which is the underlying work for supporting OpenCL
over this open-source graphics driver architecture, is on approach for landing in the very near future. This has been one damn good day for open-source Linux graphics drivers following the earlier Nouveau surprise announcements
Francisco Jerez is a Nouveau developer for the past few months has been working on an X.Org EVoC project
for running OpenCL with the open-source Nouveau driver
. His project was based upon the earlier work of the Clover OpenCL state tracker, the OpenCL work done last year for Google's Summer of Code, and other work by various individuals on bringing up OpenCL/GPGPU computing over Mesa/Gallium3D. Francisco was successful with his work and has made it possible to execute OpenCL with Nouveau Gallium3D
(even prior to Radeon Gallium3D doing OpenCL
), but the Mesa/Gallium3D bits have yet to land in the mainline tree. That's now beginning to change with the publishing of the first part of the patch-set for inclusion.
As Francisco Jerez began his e-mail message
, "This patch series is part of the ongoing work to put together a compute stack on top of Gallium3D."
He's broken this work into six parts (or building blocks as he calls it) and consist of the Gallium3D API and TGSI changes, other fixes and additions to the Gallium3D helper libraries, the OpenCL state tracker itself, an implementation of the R600 Gallium3D driver that works with AMD's LLVM back-end, an implementation of this Nouveau NV560 driver that works with another NV50/NC0 code-generator, and then finally is an LLVM back-end that will be used with LLVM's Clamg to translate compute programs from the OpenCL language and into TGSI, the IR used by the Gallium3D drivers.
Francisco has sent off the first component (the Gallium3D API and TGSI changes) as part of this first message while the second block (the other Gallium3D prerequisite changes) has already been sent in too. For the OpenCL state tracker, he's landed all of that work into a new Gallium-Compute branch of Mesa
The remaining portions (the R600g and NV50 driver implementations of the compute infrastructure and then the LLVM back-end for translating the OpenCL into TGSI) is still considered "immature" so he's not trying to push that at the moment, but the non-mainline form of these changes are publicly available.
So it looks like at least for Mesa 8.1
we will have the necessary Gallium3D/TGSI changes land plus the Clover/OpenCL state tracker. Hopefully the bits implementing the compute support for R600g and NV50 (and ideally NVC0) will be ready in time too for the summer update to Mesa. It's terrific to see that nearly four years after OpenCL 1.0 was ratified, it's finally coming about in an open-source way for GPUs on the Linux desktop.
For Radeon Evergreen owners who want to begin playing with open-source OpenCL acceleration, there is a guide available
It's not been just one day great for open-source graphics drivers, but rather an entire week with the release of Intel Haswell code
, initial support for the very interesting PowerVR-killer Intel Valleyview Atom SoC
, and AMD finally having out Radeon HD 7000 series / Trinity Fusion APU
open-source Linux support. The Linux 3.4 kernel pull
also was accepted today with many interesting features and advancements.