Re: [Discuss-gnuradio] GSOC ideas [802.11]

[Tom Rondeau]

On Thu, Apr 25, 2013 at 11:15 AM, Martin Braun (CEL)
<address@hidden> wrote:
> On Thu, Apr 25, 2013 at 04:38:19PM +0200, address@hidden wrote:
>> Hi guys,
>> I've a question(maybe this is spam): I'm thinking about my proposal,where I
>
> As long as you don't want an investment of 1000000 from us, it's
> probably not spam :)
>
>> want move some part on cuda( in this way,you can also see/verify the 
>> difference
>> in term of performance). Is cuda okay for you, or do you prefer openCL?
>
> I'm not sure you will be able to do either for this project. Have you
> checked the mailing list archives? Getting GPU-based processing into a
> GNU Radio flowgraph is quite hard, given the structure of the
> scheduler. Also, which elements are you hoping to accelerate?
>
> To answer your question, the more portable, the better, so my gut says
> OpenCL > CUDA.
>
> MB

There is the gr-gpu project (it's a CGRAN project) that is based on
using CUDA for the development of the signal processing blocks.
Mostly, this was done because OpenCL was too new at the time and was
not really usable for this purpose. On the other hand, CUDA is
incompatible with GNU Radio's license, so while it's fine to have
gr-gpu use CUDA and GNU Radio, it's not a solution we can integrate.
If OpenCL can now solve the problem, I would suggest/support using
that.

And, an aside on co-processor projects like gr-gpu. My opinion on this
is that we wouldn't carry gr_block's that are written for
co-processors in GNU Radio, anyways. So the license doesn't matter so
much for that. What we will hopefully be able to do is support the
interface between GNU Radio and the co-processor, but that for a given
co-processor, someone would have an external project that defines
blocks for use with that (like a CUDA-gpu, OpenCL-gpu, etc., etc.).
This is a developing concept, so we'll be figuring it out as more of
these projects exist.

Tom