On Wed, Sep 2, 2015 at 12:39 PM, Marcus Müller <[email protected]>
wrote:
> Hi Andre!
>
> Wow, that's a bit much to read right now. The problem I have with using
> AVX2 would be portability, which would be no issue if we wrapped RNG in
> a VOLK kernel and offered a good baseline competitor in portable C code.
> Point is though that our alternative in RNG seems to be boost::mt19937,
> which we could only "extern C{...}" into VOLK. I don't like that
> architecturally, but I think it's definitely something I'd like to hear
> Nathan's input on.
>
> What I do like, however is the takeaway that,
> "we can safely ignore the risk of overlapping subsequences"
> for two parallelly running mt19937 sequences that were seeded
> differently -- which means that we can well scale random number
> generation to multiple cores, and especially lets us just divide each
> output buffer into n_threads subbuffers and let the different generators
> run parallely on that, without wrecking memory coherency bus bandwidth
> too much. I guess that we'd ideally just spawn as many threads as there
> are 4kB-pages to fill with random samples (meaning, at every 512
> gr_complexes), just a wild guess.
>
> Cheers,
> Marcus
>
>
I agree with Martin here on the speed vs randomness. This is a noise source
that should primarily be used for simulations where it's OK to be rate
limited by your noise source (if you're biggest problem in life is being
rate limited by your noise source then I'm jealous). I'm a huge fan of
Agner Fog and his work is always great, but it uses C++ which can't be used
in VOLK kernels. If someone wants to port that to C and stick it in VOLK
I'd accept it.
Nathan
_______________________________________________
Discuss-gnuradio mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
