[Oisf-devel] GPU Support

[Victor Julien]

Charles Wyble wrote:
> Though the PCI-X bus seems to have enough bandwidth.
> 
> Running on......
>        device 0:GeForce 8400 GS
> Quick Mode
> Host to Device Bandwidth for Pageable memory
> .
> Transfer Size (Bytes)	Bandwidth(MB/s)
>   33554432		2550.1
> 
> Quick Mode
> Device to Host Bandwidth for Pageable memory
> .
> Transfer Size (Bytes)	Bandwidth(MB/s)
>   33554432		2049.5
> 
> Quick Mode
> Device to Device Bandwidth
> .
> Transfer Size (Bytes)	Bandwidth(MB/s)
>   33554432		4583.6
> 
> &&&& Test PASSED
> 
> Press ENTER to exit...
> 

I read here that the bw diff between CUDA and OpenCL on nvidia in Linux
is quite large: http://forums.nvidia.com/index.php?showtopic=96897. But
yeah, the bandwidth shouldn't be the issue. Latency on the other hand...

> 
> Do you see the performance hit with both CUDA and OpenCL? Would be very 
> interested in knowing more details on the bottlenecks, how profiling is 
> being done etc.

So far I'm basing this on the Gnort and Pixelsnort papers, where they
reached this conclusion.

>> I will commit our current OpenCL soon which means you can look at it and
>> run it to see if it has bugs (assuming we get rid of the current bugs,
>> see above) and then complain about the performance :) Later I hope our
>> code will be more useful and then any testing will be greatly appreciated!
> 
> 
> Is there an existing git repo separate from the main one? Would be happy 
> to alpha/beta test.

Not at this stage, I'm planning to just commit it to the master branch
branch, only hide it behind a --enable-opencl configure option so it
will be disabled by default.

Cheers,
Victor

-- 
---------------------------------------------
Victor Julien
http://www.inliniac.net/
PGP: http://www.inliniac.net/victorjulien.asc
---------------------------------------------