NVIDIA, Mentor Graphics May Harm GCC

Yesterday there was news that OpenACC 2.0 parallel programming support was coming to GCC complete with GPU acceleration support for NVIDIA GPUs. While it was exciting on the surface, it appears that this work may be poisonous and could have a very tough time making it upstream.

The news yesterday was about Oak Ridge, Mentor Graphics, and NVIDIA working to add OpenACC 2.0 parallel programming support to the GCC compiler for C and Fortran. GCC right now doesn't have any support for OpenACC, even the older versions of the specification, and the patches thus far haven't fully exploited the GPU potential besides converting OpenACC to OpenCL or another implementation that just runs OpenACC over OpenMP on the CPU. Mentor Graphics is now responsible for bringing OpenACC 2.0 with NVIDIA GPU support to the GNU Compiler Collection.

One of my trusted compiler sources wrote to me today about yesterday's news and that "Mentor Graphics selling out GCC." Dampening the initial OpenACC 2.0 GCC support excitement is that this implementation will likely just be targeting NVIDIA PTX. Parallel Thread Execution is NVIDIA's Assembly-like language used in NVIDIA CUDA that their binary graphics driver then compiles into machine code. Even NVIDIA's own NVCC CUDA Compiler isn't targeting the GPU's hardware instruction set directly but just targets this Assembly-like intermediate representation.

While AMD has documented their various generations of GPU ISA, NVIDIA does not provide any documentation on their instruction set architecture. Intel, ARM, and other processor companies also document their ISA publicly. This source had written in a private email, "NVIDIA refuses to publicly disclose their instruction set. They force any company targeting their hardware to use PTX. They claim it's for portability, but that's a lie. AMD's GPU instruction set changes, but they and *EVERYONE* else in the industry publishes their details no problem. The arguments about protecting their IP is just bs.. How can *everyone* else publish the details and open source drivers, but stay in business?"

With targeting PTX over hardware instructions for OpenACC on GCC, for those hoping this implementation would work on the open-source (Nouveau) driver, it will not -- at least until they're able to go from PTX to hardware instructions themselves with their own compiler. NVIDIA's LLVM back-end also targets PTX compared to the AMD R600 LLVM back-end that generates ISA for the appropriate hardware itself.

My contact went on to say, "If they add OpenACC to gcc that means it will be targeting PTX. This will force multiple closed source dependencies on gcc. CUDA is now all closed since moving away from open64, ptxas is all closed, their runtime are closed and so is their drivers. Why would Mentor Graphics sellout gcc like this? We must act now before Mentor Graphics sells gcc to NVIDIA. They have the worst open source track record and reputation of any other major company."