[hpsdr] Announcement of a CudaSharpDSP package for HPSDR: doing parallel DSP processing on your GPU

[Jeremy McDermond]

On Jun 18, 2010, at 12:24 AM, Hermann wrote:

> On Fri, Jun 18, 2010 at 1:00 AM, Jeremy McDermond
> <mcdermj at xenotropic.com> wrote:
>> ***** High Performance Software Defined Radio Discussion List *****
>> 
>> Thanks for the update Hermann.  It might be interesting to combine your work with the work for doing extra-wide FFTs that was done on PowerSDR since CUDA should give you the extra processing grunt.
>> 
> 
> Jeremy,
> yes, I was thinking about that too, especially for the wideband data
> which are provided by Mercury/Ozy. Unfortunately I can run only one
> kernel on a Cuda device at a time. Running concurrent kernels on the
> GPU, moreover, started from different CPU threads, will eventually be
> one of the future features of Cuda, which is also a feature of future
> Nvidia hardware only (at least not a feature of Nvidia compute
> capabiltiy 1.1, which my graphics adapter, which is the G92, belongs
> to). So in the moment we only can run either DSPCuda or some
> extra-wide FFTs for displaying data.

That's kinda interesting.  I haven't gotten beyond the point of merely initializing and recognizing hardware devices on my OpenCL work, but OpenCL supposedly supports multiple kernels on a single device, you just have to compile each kernel and upload it to the device.  OpenCL then has the idea of a "pipleine" into a computing context (like an OpenGL context).  You have to deal with thread locking for multiple threads using a single pipeline, but multiple pipelines are guaranteed to be separate from a threading standpoint.

> But this is what I want to do next. In the view of the upcoming OzyII
> with fast ethernet

Gigabit Ethernet even!

> it may well be, that using Cuda for computing and
> displaying the data makes more sense than trying to do all the DSP on
> the GPU.

I'm sure it will become more useful still with multiple receivers and wider passbands.

> What I see now is, that you definitely can get some CPU resource
> relief, despite the overhead doing parallel computations on the GPU.

It's great to know.  I'm really excited to see your work and see where it ends up leading.  I know there are also others in the community that I chatted with at Dayton that are watching your work with keen interest.

> 
> Vy 73,
> Hermann
> 
> DL3HVH
> 

--
Jeremy McDermond (NH6Z)
Xenotropic Systems
mcdermj at xenotropic.com




 1277160736.0