Results 1 to 4 of 4

Thread: NVIDIA Publishes New PTX Back-End For LLVM

  1. #1
    Join Date
    Jan 2007
    Posts
    15,429

    Default NVIDIA Publishes New PTX Back-End For LLVM

    Phoronix: NVIDIA Publishes New PTX Back-End For LLVM

    NVIDIA this week announced their release of the "NVPTX" back-end for LLVM with the hope to replace the existing PTX (Parallel Thread Execution) back-end inside this compiler infrastructure. This open-source code coming out of NVIDIA is based upon their internal sources...

    http://www.phoronix.com/vr.php?view=MTA5MzM

  2. #2
    Join Date
    Nov 2007
    Posts
    1,024

    Default

    Just to be clear, this still is reliant on the proprietary libraries from NVIDIA to actually compile the PTX code for the hardare, and not a complete Open solution for compiling code for NVIDIA hardware.

  3. #3
    Join Date
    Feb 2012
    Posts
    15

    Default Useless!

    Well, useless isn't the proper term, but this more PR related move of nVidia is useless for most of the users of CUDA and even OpenCL. I prefer *BSD UNIX as the platform performing my scientific work and lack in the ability to produce GPGPU based code, even using LLVM 3.1 right now as the OS's backend compiler. Using Linux for those tasks is the only way to get out of this dilemma - but as the former poster stated - you need still closed source libraries to compile proper machine code for the target hardware.

    So, what is nVidia aiming for? Since I see only a very narrow corridor of usefulness of their move, this is a PR action and therefore considered "nasty" in my opinion. They fake the fact being "open" by opening something that is completely inappropriate. As long as there is no widespread, broad platform-spanning solution for their CUDA or OpenCL, even this PTX backend is questinable. May those who listen to nVidias "words" be wrapped up in this lullaby song of "OpenSource" blabla.

    I made last year my experiences with another promis of being opensource from PathScale. Since then, nothing useful has been emitted although they claimed having something just to be ready for being published.

    nVidia doesn't support the OpenSource community. All the open drivers which are supposed to be capable of executing OpenCL code for GPGPU applications (number crunching et cetera) and for my knowledge based on the Gallium3D/Clover backend, are reverse engineered. This is for years now. And suddenly, they drop a corn in front of the front of the henns ... and claim to be supporting opensource ... when not openly, but this is an implicite claim.

    Sorry, since the first announcement that nVidia is opensourcing the PTX backend, not much useful has happened then and this is a while ago. nVidia tries to outsource efforts to those who work for free and make systems working even on closed source platforms.

    At the moment, and this is very clear, nVidia provides the only working non-OpenCL driver for the FreeBSD BSD UNIX operating system. This OS is thanks to nVidia now capable of having a decent fast graphics. My experience with AMD are really bad - the opensource driver never worked correctly with the HD48XX series and we stopped purchasing those crap. Their efforts in opensource are better - they do more than nothing, not like nVidia, but still their support is questionable since even AMD is not working towards an open standard that can make OpenCL running on every GPU. But I see chances to get more from AMD, even with their new GraphicsCoreNext approach

  4. #4

    Default

    Quote Originally Posted by Eisnefaust View Post
    Well, useless isn't the proper term, but this more PR related move of nVidia is useless for most of the users of CUDA and even OpenCL. I prefer *BSD UNIX as the platform performing my scientific work and lack in the ability to produce GPGPU based code, even using LLVM 3.1 right now as the OS's backend compiler. Using Linux for those tasks is the only way to get out of this dilemma - but as the former poster stated - you need still closed source libraries to compile proper machine code for the target hardware.

    So, what is nVidia aiming for? Since I see only a very narrow corridor of usefulness of their move, this is a PR action and therefore considered "nasty" in my opinion. They fake the fact being "open" by opening something that is completely inappropriate. As long as there is no widespread, broad platform-spanning solution for their CUDA or OpenCL, even this PTX backend is questinable. May those who listen to nVidias "words" be wrapped up in this lullaby song of "OpenSource" blabla.

    I made last year my experiences with another promis of being opensource from PathScale. Since then, nothing useful has been emitted although they claimed having something just to be ready for being published.
    Umm.. What exactly did PathScale promise to open source and didn't? [libcxxrt - opened and used by BSD and in production - Other things as well, but not GPGPU related]

    While our FreeBSD (*BSD) support isn't stellar we are working on it.

    We've made recent progress on porting PSCNV (our NVIDIA driver) to FreeBSD.
    https://github.com/pathscale/pscnv/

    So tell me why you're mad and flaming us

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •