DirectCompute (Direct3D11 Compute Shaders)

Discussion in 'GPGPU Technology & Programming' started by Scali, Sep 11, 2009.

  1. Broken Hope

    Regular

    Joined:
    Jul 13, 2004
    Messages:
    483
    Likes Received:
    1
    Location:
    England
    Perhaps not, but I still find it impossible to believe that the 5800 series has the same compute power as the 5700 series considering the 5800 series is listed as having 2.72 TeraFLOPs compute power and the 5700 series is listed as having 1.36 TeraFLOPs compute power as per AMD slides.
     
  2. EduardoS

    Newcomer

    Joined:
    Nov 8, 2008
    Messages:
    131
    Likes Received:
    0
    Please answer, what M scores for GPUs are?
     
  3. trinibwoy

    trinibwoy Meh
    Legend

    Joined:
    Mar 17, 2004
    Messages:
    10,436
    Likes Received:
    443
    Location:
    New York
    Given his comments about DC drivers for the 4890 is it fair to assume this was written on his 9800GT? I agree though, would be nice to know what it does.
     
  4. Broken Hope

    Regular

    Joined:
    Jul 13, 2004
    Messages:
    483
    Likes Received:
    1
    Location:
    England
    Would the fact that the shader code is compiled for CS 5.0 affect the performance if the actual shader code is the same?

    Could this be the reason behind the odd scores, Nvidia cards running profile 4.1 vs ATI cards running 5.0? Though I guess that still doesn't answer by 57xx performs the same as 58xx in this 'benchmark'
     
  5. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,257
    Likes Received:
    1,946
    Location:
    Finland
    Isn't nV running 4.0, while HD4 (when the CS support comes) will do 4.1?
     
  6. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    Catalyst 9.12 are posted. These contain DirectCompute 10.1 support for HD 4800 and HD 4700 series (as well as OpenGL3.2 for that matter).
     
  7. Arnold Beckenbauer

    Veteran

    Joined:
    Oct 11, 2006
    Messages:
    1,415
    Likes Received:
    348
    Location:
    Germany
    No DC for me (HD4800). DirectX from August 2009 is installed, my OS is Win7 7100 (RC).
     
  8. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,805
    Likes Received:
    2,067
    Location:
    Germany
    (I'm not asking if there are plans this time... ;))
    Will there ever be Direct Compute support, be it 4.0 or 4.1, for the HD-3000-class of cards?
     
  9. fellix

    fellix Hey, You!
    Veteran

    Joined:
    Dec 4, 2004
    Messages:
    3,494
    Likes Received:
    405
    Location:
    Varna, Bulgaria
    Those don't have LDS memory.
     
  10. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,805
    Likes Received:
    2,067
    Location:
    Germany
    Which could be emulated according to postings here.
     
  11. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    DirectCompute 10/10.1 was a specification added after any of these (3000/4000) products had already been in retail and do not necessarily relate to he original DX10/10.1 spec.
     
  12. fellix

    fellix Hey, You!
    Veteran

    Joined:
    Dec 4, 2004
    Messages:
    3,494
    Likes Received:
    405
    Location:
    Varna, Bulgaria
    That's the case for OpenCL 1.0, though.
     
  13. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,805
    Likes Received:
    2,067
    Location:
    Germany
    So, that means it's not possible edit: for 3000/2000? (Maybe due to the reaseon fellix mentioned?)
     
  14. pcchen

    pcchen Moderator
    Moderator Veteran Subscriber

    Joined:
    Feb 6, 2002
    Messages:
    2,773
    Likes Received:
    153
    Location:
    Taiwan
    I'm using Windows 7 x64 RTM and DirectCompute works on my computer (Radeon HD 4850). OpenCL works too but it's not very different from earlier version.
     
  15. Arnold Beckenbauer

    Veteran

    Joined:
    Oct 11, 2006
    Messages:
    1,415
    Likes Received:
    348
    Location:
    Germany
    It looks like DC is not supported under Win 7 7100 (your RTM has a higher version number).
     
  16. codedivine

    Regular

    Joined:
    Jan 22, 2009
    Messages:
    271
    Likes Received:
    0
    The problem with 3xxx is not LDS (which can be emulated) but rather the ability to form thread groups. Thread groups have 2 required attributes in both OpenCL and in DirectCompute:

    a) You must be able to divide threads into thread groups which are guaranteed to be executed on a single SIMD. Threads within a group should not be distributed among 2 different SIMD. All you can do on HD 3000 is to specify a massive 2d grid of threads (lets say 1024x1024) should be executed across the GPU. The hardware divides into wavefronts of 8x8 and then dynamically distributes those wavefronts across to SIMDs. There is no way to specify that "these 512 threads should all be scheduled on a single SIMD".

    b) Threads within a thread group must be able to able to synchronize. Again this is only possible if the threads execute on the same SIMD as most GPUs do not have a global sync capability. If the thread group of 512 threads is divided into wavefronts and those wavefronts happen to be executing on different SIMD, you will not be able to sync.

    So for these reasons, OpenCL and DC are not available on HD 3000 and will NEVER be available. Its not a software issue. The hardware simply cannot do them. HD 4000 OTOH has "compute shader" support , i.e ability to form thread groups and sync threads within a thread group.

    Well, maybe you can implement the thread group sync by doing a global kernel termination and dividing your kernel into different parts, but that will probably be too expensive to do and am not sure if there are other missing parts as well.
     
  17. fellix

    fellix Hey, You!
    Veteran

    Joined:
    Dec 4, 2004
    Messages:
    3,494
    Likes Received:
    405
    Location:
    Varna, Bulgaria
    Thanks for the clarification! ;)
     
  18. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
  19. rpg.314

    Veteran

    Joined:
    Jul 21, 2008
    Messages:
    4,298
    Likes Received:
    0
    Location:
    /
    Indirect dispatch

    What does indirect dispatch in SM5 mean? And what does it do?

    Can a sm5 compute shader launch another thread block from the present one? With what thread sizes? What are the limits? Is it constrained to run in the the same SIMD/SM? Is it allowed access to the shared mem used by the previous block?
     
  20. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    http://www.legitreviews.com/article/1001/2/

    Not only is the whitepaper that AMD supposedly released not available anywhere, but MSDN is completely useless.

    So, er, that's it I'm afraid.

    Jawed
     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...