Mesa Compiler Stacks, A Hard Dependency On LLVM

Tom Stellard, the former Google Summer of Code student who worked on R300 GLSL improvements and a new register allocator, is now working for AMD and his work is focused on bringing up open-source OpenCL / GPGPU support in the Radeon Linux driver.

The OpenCL implementation for the open-source Radeon Linux driver will be relying upon the Mesa Gallium3D "Clover" OpenCL state tracker. This will rely heavily on the LLVM compiler with the Clang front-end. LLVM is used for its optimization library, back-end generators for x86/ARM/PPC, etc. The plan of attack is to go from OpenCL C to compile with Clang to LLVM ASM and then pass it to the LLVM back-end and to then generate the targeted machine code. The hope is also to do some code sharing between OpenCL and GLSL.

What's still to be determined is whether to integrate the LLVM back-end into each Gallium3D hardware driver or to write an LLVM back-end that simply generates TGSI. If going with the approach of having an LLVM back-end for each Gallium3D driver, it's easier to write GPU-specific optimizations and makes a better trivial shader compiler, but might need to use LLVM for shaders and each driver would need to write its own LLVM back-end to generate code for the respective target/architecture. With the later approach of just having one LLVM back-end that can generate TGSI instructions, which any Gallium3D driver can understand, they can re-use the current code emitters and only one back-end would be needed, but TGSI would need to be extended to support OpenCL functionality.

If going with a TGSI to LLVM converter, this already exists for the most part thanks to the LLVMpipe Gallium3D driver that is already doing these translations. Tom Stellard has been working on the TGSI to LLVM converter support and an LLVM back-end for R600g. John Bridgman also shared during the XDC2011 talk that Tom is playing internally with using AMD IL and some other things as part of this open-source Radeon compute upbringing.

In a separate talk, Ian Romanick talked about the Low-Level Virtual Machine in Mesa. It was pretty much agreed upon that LLVM will be made a hard dependency when building Mesa. LLVM is already required when building the older Radeon drivers, but it's been optional for other uses up to this point. Going forward, LLVM is a hard requirement for this open-source graphics library.