In Road To OpenCL, R600g LLVM Back-End Arrives
Before calling it a week, Tom Stellard at AMD published a Git branch that offers up an LLVM shader back-end for the AMD R600 Gallium3D driver. This is one of the steps in bringing Compute/OpenCL support to the open-source AMD Radeon Linux graphics drivers.
Tom Stellard announced on Friday afternoon this R600g LLVM shader back-end branch to the Mesa developer's mailing list. Tom's r600g-llvm-shader branch of his Mesa repository offers a TGSI (the current Gallium3D IR) to LLVM IR converter, changes to Gallivm to make it portable with the TGSI-to-LLVM converter, and the LLVM back-end for R600g itself. The TGSI-to-LLVM converter will likely be of use to more Gallium3D drivers besides the R600g driver itself, since TGSI is universal to the drivers and not Radeon-specific.
This LLVM back-end is based upon AMD's AMDIL LLVM back-end for OpenCL with various modifications. Tom says this new shader back-end currently passes 99% of the Piglit regression tests that work with the current R600g shader back-end. The failures are coming from unimplemented texture instructions and on a current lack of indirect addressing support. The LLVM back-end also isn't optimized for VLIW at this point plus other optimizations are lacking. "The optimizations in r600_asm.c are able to do some instruction packing, but the resulting code is not yet as good as the current backend."
And here's the key part of his message:
More information coming soon.
Tom Stellard announced on Friday afternoon this R600g LLVM shader back-end branch to the Mesa developer's mailing list. Tom's r600g-llvm-shader branch of his Mesa repository offers a TGSI (the current Gallium3D IR) to LLVM IR converter, changes to Gallivm to make it portable with the TGSI-to-LLVM converter, and the LLVM back-end for R600g itself. The TGSI-to-LLVM converter will likely be of use to more Gallium3D drivers besides the R600g driver itself, since TGSI is universal to the drivers and not Radeon-specific.
This LLVM back-end is based upon AMD's AMDIL LLVM back-end for OpenCL with various modifications. Tom says this new shader back-end currently passes 99% of the Piglit regression tests that work with the current R600g shader back-end. The failures are coming from unimplemented texture instructions and on a current lack of indirect addressing support. The LLVM back-end also isn't optimized for VLIW at this point plus other optimizations are lacking. "The optimizations in r600_asm.c are able to do some instruction packing, but the resulting code is not yet as good as the current backend."
And here's the key part of his message:
The main motivation for this LLVM backend is to help bring compute/OpenCL support to r600g by making it easier to support different compiler frontends. I don't have a concrete plan for integrating this into mainline Mesa yet, but I don't expect it to be in the next release. I would really like to make it compatible with LLVM 3.0 before it gets merged (it only works with LLVM 2.9 now), but if compute support evolves quickly, I might be tempted to push the 2.9 version into the master branch.
More information coming soon.
28 Comments