[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Discuss-gnuradio] (GSoC) Using hardware based coprocessors in GNU r
Muhammad Junaid Muzammil
Re: [Discuss-gnuradio] (GSoC) Using hardware based coprocessors in GNU radio
Sat, 8 Feb 2014 17:27:26 +0500
@marcus, yes I have gone through that last year GSOC project, that was some real good work for my likings. I haven't gone through Volk link yet. Well, what I have observed is that the speed limitations over hardware is primary imposed by the software piece. But we can't create a dedicated hardware for each module, so we need generalization as well and thats where general purpose computers are good at. So definitely, there is a requirement of careful analysis for moving modules to FPGAs, whether significant performance improvement will be obtained or not.
@Nathan, yes I meant to refer to GSOC project as I wish to apply for this year. Sorry for not writing it down in the subject and causing a confusion. I was going through the GSOC ideas list available here (http://gnuradio.org/redmine/projects/gnuradio/wiki/GSoC
) and I wish to get some more details what the mentor is expecting to achieve in this project, this summer.
On Saturday, February 8, 2014, West, Nathan <address@hidden
On Saturday, February 8, 2014, Marcus Müller <address@hidden> wrote:
such a system already exists in the shape of a xilinx zynq board.
There's been a talk that (although I was not there to witness it) is
supposed to be quite good:
I'm not sure if videos of that talk are available at this point.
However, there is quite a documentation in the GR wiki:
For usage of specialized DSP instructions (if these can be part of the
normal CPU program) please have a look at VOLK, which is a part of GNU
PS: Have to slightly disagree with you on the "hardware is faster than
software, always" part, since design constraints usually limit the
performance of FPGAs, and moving data guaranteeing coherence comes at
a performance penalty.
On 08.02.2014 10:00, Muhammad Junaid Muzammil wrote:
> I am Junaid, currently a Masters student. The topic pretty much
> inspired me, although I couldn't create a clear picture in my mind.
> The project is about using coprocessors for acceleration, so
> basically the target would be a heterogeneous sytem consisting of a
> RISC microprocessor preferably belonging to ARM family alongwith
> coprocessors which can be either FPGAs or DSPs. The hardware cores
> are always faster than the software cores, there is no doubt in
> that. Similarly DSP due its architecture shows better performance
> towards signal processing functions. Both these coprocessors will
> perform acceleration by taking some workload from uP.
> Now which particular functions/tasks will be taken, this isn't
> clear to me at this stage and I am looking forward to get some
> details over it.
I believe Junaid is referring to a GSoC project to expand the coprocessor abilities of GNU Radio. Systems that allow coprocessor a definitely exist, but they aren't widely used and efficiently using a coprocessor with GNU Radio isn't exactly out of the box functionality.
What you might do for such a GSoC project depends on your interests and abilities. It's likely that if you wanted to make accelerators the first step would be some form of profiling to figure out candidate blocks or functions. OTOH maybe you've seen in some paper or your own experience a good candidate for coprocessor off loading.
I think there's also interest in enabling blocks to handle their own buffers to allow DMA, but someone else should be able to discuss that far better than anything I might add at the moment.
PS- if you're going to discuss a GSoC project please mention that so there's no confusion about the subject. If this isn't about GSoC then I apologize for modifying the subject line.