From: Andrew B. <an...@ca...> - 2012-11-28 14:32:24
|
On 28 November 2012 11:15, Sylvain Munaut <24...@gm...> wrote: >> Is it possible to point to particular code in SVN, or would there be a >> reasonable amount of work in just identifying this? I guess if just >> one or two of the most compute intensive parts could be identified >> that would be a start. > > The Tranceiver/ directory essentially. > > But it's not just one big method, it's the whole process of > demodulation. When optimizing for SSE/NEON you can "just" optimize the > correlation or convolution function for example, but for parallela you > can't do that because the cost of transfering data back and forth is > going to be too big. You have to put the whole demod ( so send IQ > samples, get softbits out ) and process several bursts in // in the > different cores. That's how I'd do it at least. Thank you for the clarification! It would be great if by identifying some small parts of OpenBTS to enable to use Epiphany it meant that OpenBTS could run on the Parallella board (I'm guessing the dual-core ARM would not be sufficient alone — maybe I'm wrong). But may be that a single processing block or two from GNU Radio would make a better target, or rather be easier to modify to take advantage of the Epiphany chip. That said, Best, Andrew -- Andrew Back http://carrierdetect.com |