December 13, 2011

NVIDIA Opens Up CUDA Compiler

Michael Feldman

GPU maker NVIDIA is going to make its CUDA compiler runtime source code, and internal representation format public, opening up the technology for different programming languages and processor architectures. The announcement was made on Wednesday at the kick-off of the GPU Technology Conference Asia in Beijing, China.

The company says it will use the LLVM compiler infrastructure as the vehicle for the public CUDA source code. LLVM is a open source project that maintains source code collections of various of compile, runtimes, and other development tools. The new LLVM-based CUDA source, will be available in the latest release of the CUDA Toolkit, version 4.1, which was also launched this week.

The CUDA open source set-up does not, however, mean NVIDIA will arbitrarily accept changes and enhancements to its compiler technology from other developers. The company still intends to retain complete control of its source code.  Tool developers will be able to modify the standard compiler and runtime for their own customized needs, but little of this is likely to be folded back into NVIDIA’s code base

The main idea is to allow software tool makers to port the CUDA compiler to other environments that NVIDIA or its commercial partners are not interested in pursuing on their own. In the case of programming languages, there are already compilers for C, C++, and Fortran, which are the big three for high performance computing. But as the market for GPU computing expands, NVIDIA foresees the need for other languages such as Python or Java, as well as domain specific languages.

As far as CUDA compiler targets, there is a lot of room for interesting ports to other platforms. The prime candidate here is the AMD/ATI GPU platform. Even though CUDA is the most widespread programming environment for GPU computing, it only currently works on NVIDIA GPUs (and x86 multicore via a PGI compiler implementation). There are likely to be plenty of users with CUDA-based applications that are now interested in running their applications on AMD GPUs/APUs, or at least are interested in the prospect that their codes can do so at some future date.

AMD is still pushing its OpenCL strategy for GPU computing. OpenCL, a non-vendor-specific open standard for parallel computing, is supported by NVIDIA as well, but has not yet managed to attract a lot applications. By offering to open up CUDA, NVIDIA has probably blunted some of the appeal of OpenCL, that is, assuming a compiler vendor or an academic research group builds an CUDA-ized AMD GPU compiler.

Since CUDA is a general-purpose parallel computing technology, essentially any multicore/manycore architecture would be a potential target. Other possible architectures for CUDA include Intel’s upcoming Many Integrated Core (MIC) coprocessor, Power CPUs, multicore ARM chips (especially for future 64-bit implementations), and even more exotic fare, like Texas Instruments’ new floating-point capable DSPs.

The academic community most likely to take early advantage of an open CUDA compiler.  For example, at Georgia Tech, the Ocelot project is focused on applying CUDA C to different processors, including AMD GPUs and x86-CPUs. The project lead there, Sudhakar Yalamanchili, says the opening up of the CUDA technology is “a significant step.”
Even compiler vendors who already have special arrangements with NVIDIA will be able to take advantage of the new open source strategy. In the press release, The Portland Group (PGI) director Doug Miles says “This initiative enables PGI to create native CUDA Fortran and OpenACC compilers that leverage the same device-level optimization technology used by NVIDIA CUDA C/C++. It will enable seamless debugging and profiling using existing tools, and allow PGI to focus on higher-level optimizations and language features.”

NVIDIA will not always directly benefit from its new open source stance. Certainly, if some enterprising team ports CUDA to AMD chips, that could cut into Tesla GPU sales. But for the greater good of attracting customers to its own hardware, NVIDIA realized that a closed platform discourages plenty of users who don’t want to be locked into a single hardware platform or rely on a sole vendor. As with NVIDIA’s recent endorsement of the OpenACC directives, the opening of CUDA  seems to be part of a strategy designed to broaden the appeal of GPU computing rather than just NVIDIA products. It appears the GPU maker has calculated that expanding the pie will get them further in the long run than just trying to maximize their slice of it.