Hi Philippe,

rather than having too much speculation here, what about adding a quick 
OpenCL-to-CUDA translator (just string substitution, you don't need 
more) to the generator? Put the best kernels for Fermi and Kepler into a 
compilation unit and then hopefully Denis or Evan will give it a try? 
You can verify the proper execution on our GTX 470 machine.

Best regards,
Karli


On 08/20/2013 10:30 AM, Philippe Tillet wrote:
> Yes, it is :)
> The generator can find a kernel at 630GFLOP/s for the GTX470, and it
> cannot find a kernel at more than 550GFLOP/s for the GTX680, which means
> that the best Fermi Kernel performs very poorly on the Kepler
> architecture, while NVidia claim themself the architecture is similar,
> and that kernels should show improved performance without any code
> modification...
> Now, it may come from Apple's implementation, but I'm doubtful here.
> Since NVidia has releasing any OpenCL-related stuff since 2010 (OpenCL
> 1.1), even dropping OpenCL tools in CUDA5, it would not surprising that
> they just ensure correctness since then, leading to poor performance on
> post 2010 GPUs (ie Kepler)...
>
> Philippe
>
>
> 2013/8/20 Evan Bollig <bol...@gmail.com <mailto:bol...@gmail.com>>
>
>     You suspect NVidia is throttling CL kernels?
>
>     -E
>
>     On Tue, Aug 20, 2013 at 10:10 AM, Philippe Tillet
>     <phil.til...@gmail.com <mailto:phil.til...@gmail.com>> wrote:
>      > Hi again,
>      >
>      > Actually, I think that it would be worth it to convert the best
>     GEMM kernel
>      > from OpenCL to CUDA, and to observe whether the performance is
>     still bad or
>      > not. It may be even worth it to also do this with the best Kepler
>     kernel
>      > (pretty different from the best Fermi kernel!).
>      >
>      > Best regards,
>      > Philippe
>      >
>      >
>      > 2013/8/20 Philippe Tillet <phil.til...@gmail.com
>     <mailto:phil.til...@gmail.com>>
>      >>
>      >> Hi everybody,
>      >>
>      >> Thanks to Evan and Denis, I can confirm that our GEMM kernels
>     exhibit bad
>      >> performance on Kepler (~25% of the peak for both the GTX680mx
>     and the Tesla
>      >> K20)
>      >> What is weird here is that Magma's Fermi kernel seems to work
>     well enough
>      >> on Kepler, while our Fermi GEMM kernel clearly doesn't work well on
>      >> Kepler... which doesn't make sense to me.
>      >>
>      >> Has anybody ever heard of a good OpenCL kernel for the Kepler
>     architecture
>      >> ?
>      >> I'm aware of what big companies may want to do to promote their own
>      >> solution, and I want to consider every option, especially since
>     the Kepler
>      >> was released at the time OpenCL 1.2 - which is still not
>     supported by NVidia
>      >> - came out.
>      >> I have some links to support my paranoia:
>      >> http://parallelis.com/kepler-underperform-on-gpgpu-gtx680/
>      >>
>      >>
>     
> http://www.tonymacx86.com/343-os-x-10-8-3-nvidia-6xx-opencl-benchmarks.html
>      >>
>      >> Best regards,
>      >> Philippe
>      >
>      >
>      >
>      >
>     
> ------------------------------------------------------------------------------
>      > Introducing Performance Central, a new site from SourceForge and
>      > AppDynamics. Performance Central is your source for news, insights,
>      > analysis and resources for efficient Application Performance
>     Management.
>      > Visit us today!
>      >
>     
> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
>      > _______________________________________________
>      > ViennaCL-devel mailing list
>      > ViennaCL-devel@lists.sourceforge.net
>     <mailto:ViennaCL-devel@lists.sourceforge.net>
>      > https://lists.sourceforge.net/lists/listinfo/viennacl-devel
>      >
>
>
>
>     --
>     -Evan Bollig
>     bol...@gmail.com <mailto:bol...@gmail.com>
>     bol...@scs.fsu.edu <mailto:bol...@scs.fsu.edu>
>
>
>
>
> ------------------------------------------------------------------------------
> Introducing Performance Central, a new site from SourceForge and
> AppDynamics. Performance Central is your source for news, insights,
> analysis and resources for efficient Application Performance Management.
> Visit us today!
> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
>
>
>
> _______________________________________________
> ViennaCL-devel mailing list
> ViennaCL-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/viennacl-devel
>


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and 
AppDynamics. Performance Central is your source for news, insights, 
analysis and resources for efficient Application Performance Management. 
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
ViennaCL-devel mailing list
ViennaCL-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/viennacl-devel

Reply via email to