Re: NEWS: NVIDIA Open-Sources Its CUDA Compiler

At 02:02 15/12/2011, O. Hartmann wrote:
>Just read this on
>
>phoronix.com
>
>Is this finally a chance to get GPGPU on FreeBSD natively supported?
>
>nVidia has a binary driver, supporting well their higher end graphics
>cards on FreeBSD 64bit natively.
>
>I do not understand much about the compiler itself, it's "nvcc" as far
>as I know, and it is also doing well OpenCL (with some serious bugs we
>revealed).
>
>What would be needed to bring FreeBSd finally back to the HPC scenario
>with being capable of dealing natively with GPGPU stuff on nVidia
>graphics cards? There are libraries installed by the driver or the SDK.
>With a OpenSource compiler it should also be possible for nVidia,
>assumed the compiler works with freeBSD natively, to provide OpenCL
>stuff as well as CUDA stuff.
>Please correct me and destroy me "dreams" having FreeBSD in my lab
>working on GPUs ...

I'm currently trying to install and use the Cuda 4.1 sdk on FreeBSD 
8.2., not only the runtime. Not much luck for now, i'm not a system 
architect, but a developer with some sysadmin duties and doing it 
off-work. Check posts from past weeks on performance list.

>The decission sounds like some pitfall in a contract. Is nVidia dropping
>CUDA in favour of OpenCL or is the CUDA compiler only a tiny piece of
>the whole thing that could be easily considered open source without
>changing the "great restricted Linux-only" picture?

No, nVidia is not dropping cuda in favor opencl. Currently, nVIdia is 
developing OpenACC with Cray, PGI and others. Using pragmas you can 
get gpu acceleration of your code automagically, similar to OpenMP. 
But again, Cuda is its gpgpu flagship.

>Maybe LLVM, now part of FreeBSD's backbone, is capable of taking
>advantage of the opening of the CUDA compiler so we will see a
>combination of CLANG/OpenCL/CUDA soon on FreeBSD introduced by LLVM?

The last sdk nvc compiler is a llvm with proprietary extensions, 
closed source. What can be done with the new sdk now? You can use any 
language that llvm is able to compile with Cuda, in theory. You can 
develop an opensource Cuda/OpenCL/YourGpuLanguage to PTX compiler. 
But you can't compile from PTX to gpu binary code, it's done with the 
closed source part of the sdk. PGI for example has it's own Cuda to 
PTX compiler/translator.

>Well, well, this is awesome ... ;-)
>
>Oliver