Since starting their GPU Technology Conference in 2010, NVIDIA has expanded into several events so that they can hold events in Europe and Asia. The next flagship GTC will be in San Jose in May, but NVIDIA’s #2 conference, GTC Asia, is occurring this week in Beijing. As with GTC America, GTC Asia serves several purposes for the company: a research symposium, a developer training program, and of course a platform for NVIDIA to announce new GPU computing products and initiatives.

The latter is the spotlight today, as NVIDIA is using GTC Asia to announce the latest versions of their development toolkits. Parallel Nsight, NVIDIA’s Visual Studio development toolkit, has just had its second release candidate for version 2.1 released. Meanwhile (and more importantly), CUDA 4.1 is also being released as a release candidate.

CUDA version numbering is based on features so don’t let the .1 numbering undersell CUDA, CUDA 4.1 is a major release for NVIDIA due to what NVIDIA’s been doing to the backend. Previously the CUDA compiler toolchain was developed entirely within NVIDIA as a proprietary product; developers could write tools that could generate PTX code (NVIDIA’s intermediate virtual ISA), but the compiling of PTX to binary code was handled by NVIDIA’s tools. CUDA 4.1 changes this in a big way: the CUDA compiler is now being built against LLVM, the modular compiler.

LLVM in the strictest sense isn’t a true compiler (it doesn’t generate binary code on its own), but as a modular compiler it’s earned quite a reputation for generating efficient intermediate code and for being easy to add new support for new languages and architectures to. If you can generate code that goes into LLVM, then you can get out code for any architecture LLVM supports and it will probably be pretty efficient too. LLVM has been around for quite some time – and is most famously used as the compiler for Mac OS X and iOS starting with Mac OS X 10.6 – but this is the first time it’s been used for a GPU in this fashion.

So why is CUDA's using LLVM a big deal? If nothing else it should result in shorter compile times and slightly faster performing code for CUDA developers. How much any individual developer will benefit remains to be seen, but from what we’ve been hearing the compile times are anywhere between 10% to 50% quicker. Meanwhile the nature of GPU computing means that application/kernel performance won’t improve nearly as much – LLVM can’t parallelize your code for you – but it should be able to generate slightly smarter code, particularly code from non-NVIDIA languages where developers haven’t been able to invest as much time in optimizing their PTX code generation.

However the move to LLVM isn’t just about immediate performance benefits, it also marks the start of a longer transition by NVIDIA. As we mentioned previously LLVM is a modular compiler and was designed that way so that developers could easily add support for new languages (what goes in to LLVM) and architectures (what goes out of LLVM) to the LLVM ecosystem. In NVIDIA’s case moving to LLVM not only allows them to open up GPU computing to additional developers by making it possible to support more languages, but it allows CUDA developers to build CUDA applications for more architectures than just NVIDIA’s GPUs. Currently it’s possible to compile CUDA down to x86 through The Portland Group’s proprietary x86 CUDA compiler, and the move to LLVM would allow NVIDIA to target not just x86, but ARM too. ARM in fact is more than likely the key to all of this – just as how developers want to be able to use CUDA on their x86 + NVGPU clusters, they will want to be able to use CUDA on their Denver (ARM) + NVGPU clusters, and LLVM is how NVIDIA is going to accomplish this.

Finally, with the move to LLVM NVIDIA is also opening up CUDA, if ever so slightly. On a technical level NVIDIA’s CUDA LLVM compiler is a closed fork of LLVM (allowed via LLVM’s BSD-type license), and due to the changes NVIDIA has made it’s not possible to blindly plug in languages and architectures to the compiler. To actually add languages and architectures to CUDA LLVM you need the source code to it, and that’s where CUDA is becoming “open.” NVIDIA will not be releasing CUDA LLVM in a truly open source manner, but they will be releasing the source in a manner akin to Microsoft’s “shared source” initiative – eligible researchers and developers will be able to apply to NVIDIA for access to the source code. This allows NVIDIA to share CUDA LLVM with the necessary parties to expand its functionality without sharing it with everyone and having the inner workings of the Fermi code generator exposed, or having someone (i.e. AMD) add support for a new architecture and hurt NVIDIA’s hardware business in the process.

Source: NVIDIA Press Center

Comments Locked


View All Comments

  • MySchizoBuddy - Wednesday, December 14, 2011 - link

    OpenCL always had the vendor neutral moniker to back itself. Now CUDA being open sourced. It too will be vendor neutral. Choice between CUDA and OpenCL just got a lot lot harder.
  • MySchizoBuddy - Wednesday, December 14, 2011 - link

    By vendor neutral I mean it works on lot of different platforms from different companies.
    OpenCL Intel, AMD, ARM and CPU, GPU, DSP, FPGA etc
    Cuda Nvidia, PGI with GPU and CPU. this will change from now on.
  • A5 - Wednesday, December 14, 2011 - link

    You can run OpenCL code on Nvidia hardware as well (though it's a little slower in most cases).

    In my experience CUDA is the easier language to write, but if I were doing something commercially it would be really hard to not do it in OpenCL. If I were doing an in-house project where I got to choose the hardware, I'd go CUDA.
  • ltcommanderdata - Wednesday, December 14, 2011 - link

    As the article notes, nVidia is not giving anyone else full control of CUDA. AMD would not be able to modify LLVM to allow it run CUDA programs on their own GPUs.

    Most (all?) OpenCL drivers from all manufacturers already use LLVM as their compiler anyways. It's just another language that is fed in the front-end. Apple has been using LLVM since Leopard to compile OpenGL code to run on x86 CPUs or pass onto Intel/AMD/nVidia GPUs. So LLVM use in GPUs is not new.
  • Noriaki - Wednesday, December 14, 2011 - link

    "LLVM in the strictest sense isn’t a true compiler (it doesn’t generate binary code on its own)"

    Firstly, the strictest sense of compiler doesn't necessarily need anything to do with binary code. A compiler, in the strictest sense, just translates one language to another. Usually a higher level language to assembly, but not always.

    But more importantly, LLVM is generally used as *part* of compiler. Some compiler front end compiles your code to LLVM IF. Clang, for example, is built expressly for generating LLVM IF, but there are others. Then you can do a bunch of optimizations on the IF, and then if you want you can translate that IF into machine code, or a variety of other things (there are LLVM IF interpreters for example).

    So I'm not just trying to make a big fuss, look at how much I know about LLVM, etc. I bring this all up because I'm a little unclear on what they are doing with LLVM and what bits are open. They are forking LLVM, I assume that means from the LLVM IF optimizer back, including binary generation from the IF to something the GPU can understand all remains under wraps.

    So are they dropping PTX and using a closed version of LLVM IF instead? You addressed the front end flexibility in LLVM, but how different is the nVidia version of LLVM from regular? If I took Clang and ran a C program through it to generate some LLVM IF, is there any hope of feeding that into CUDA LLVM? Or does it have to be special nVidia LLVM? In which case we aren't really that far ahead of the PTX days. They are just taking advantage of some of the work that's gone on at the LLVM IF optimizer level, with no real "openness" benefit to speak of.

    Which is fine, as you pointed out the BSD license allows for this. I'm just not sure. Maybe you're not either, but I thought I'd ask.
  • SlyNine - Wednesday, December 14, 2011 - link

    I'm no programmer, But I gotta wonder if Open CL or Cuda can be leveraged to do GPU calculations in a virtual enviroment. Could we have full GPU acceleration on mutliple VMs with this ?

    What would be great if I could do something like OnLive but on a local scale.
  • skiboysteve - Wednesday, December 14, 2011 - link

    if anyone uses this.... A year and a half ago labview 2010 refactored its compiler to use llvm. There is a good write up on it here:

    it gave an average 15% code speed up but actually increased the compile time. However thanks to llvm code speed and compile time are getting better every release. Also porting to new targets like arm is way easier.
  • obsidience - Thursday, December 15, 2011 - link

    Sigh, wanna educate us on what the LLVM acronym stands for?
  • vlado08 - Thursday, December 15, 2011 - link

    LLVM (Low Level Virtual Machine), a compiler framework
  • wildon - Sunday, December 18, 2011 - link


Log in

Don't have an account? Sign up now