Re: [Discuss-gnuradio] Latest Hackfest

[Tom Rondeau]

On Mon, Apr 7, 2014 at 11:52 AM, Miklos Maroti <address@hidden> wrote:

Hi Tom,

On Fri, Apr 4, 2014 at 12:21 AM, Tom Rondeau <address@hidden> wrote:
>

> On Thu, Apr 3, 2014 at 3:05 PM, Miklos Maroti <address@hidden>
> wrote:
>>
>> Dear Tom,
>>
>> I have read about the FEC work done during the Hackfest, so this would
>> be an appropriate time to bring this up. We have participated in the
>> Darpa Spectrum Challenge, and developed some good and fast codes.
>> Specifically, we have a generic convolutional code encoder and decoder
>> (any R and K is supported, tailbiting can be enabled), with fast SSE
>> implementation. In the attached BER figure we have used R=1/8, K=9 and
>> length = 3 bytes = 24 bits for encoding headers. The speed is not much
>> worse than the spiral generated one. More importantly, we have a
>> specially crafted version of a repeat accumulate coder/decoder that
>> can be quickly decoded with SSE. On the figure we have used rates 1/3,
>> 1/2, 1/3 and 4/5 with a payload size of 1440 bytes. On the orbit-lab
>> machines (core-i7) our SSE code can decode more than 1000 packets per
>> second on a single core.
>>
>> I plan to release this code under the GPL (and maybe under an MIT
>> license as well) after I have a paper written up about it. So the
>> question would be:
>
>
> Miklos,
>
> This sounds like a fantastic result.
>
>>
>> 1) how to incorporate this into gnuradio (separate OOT or built in)
>
>
> If it's to go into GNU Radio, it would only be license GPLv3, like the rest
> of the code base. Also, we require a copyright transfer from you. We can
> discuss and work on that on the side.
>
> Also, with the use of SSE, we'd want to separate the code into a VOLK kernel
> with a generic and SSE implementation. That way, we're not restricting
> ourselves to certain platforms. And then, we can extend it to other SIMD
> architectures.

I have thought about it. I know that all SSE code is currently
separated into a VOLK kernel (with the exception of FFT which is
coming from a separate library). Of course you can do that, but then
most of the gnuradio blocks will be just wrappers around the VOLK
kernels.

Yeah, so what?

Seriously, this is a good thing, not a problem. The more we can wrap up into volk, the easier it is to use on multiple platforms and reuse in other situations. We won't put any SIMD code into GNU Radio at this point that doesn't come through VOLK since VOLK handles the mechanics of dealing with multiple platforms and different available SIMD architectures.

 

>> 2) how to deal with puncturing (we have some good generic solution)
>
>
> We let the FEC API encoder and decoder blocks handle the puncturing and
> depuncturing separate from the coders themselves. I think they are generic
> enough, but we'd probably want to update them if they don't fit every
> scenario.

I have a generic interleaver / puncturer that takes chunks of memory
blocks and reorders (permutes) them and extends them with zero at
arbitrary places. Basically it takes a map (integer list between -1
and N-1) where -1 means "zero out" and 0 through N-1 means some
element of the input. Probably it is harder to explain than to see it.
Anyhow, the puncturing design is important and you have to omit
encoded (unsigned char) hard bits and insert soft (float) zeros.

Having the block interleaver would be nice. We have the generic puncturer block, as I said, but if you have them roped together, seems like (I'm guessing) it should be easy to use this block as an interleaver and/or a puncture block.

I'm assuming you have the deinterleaver/depuncture block as well?

 

>> 3) how to deal with variable packet lengths and variable rates (have some
>> ides)
>
>
> The FEC API encoder/decoder blocks don't care about the packet lengths as
> long as the coders themselves can handle it. Right now, our working example
> uses fixed packet length, but (I don't think) that's a requirement. Might
> have to double-check this. Same should go for the rate.
>
> In the FEC API world, there is an FEC encoder and decoder block that sets
> just a few parameters. One of them is the variable that actually defines the
> encoder and decoder code, and those variables are designed to work with the
> API. So most of this is on how these variable blocks are built.

Do you have a working coder/decoder pair with this API?

https://github.com/trondeau/gnuradio/tree/fecapi

We're still messing around with things right now, so this is all work-in-progress right now (in other words, it's a bit messy).

 

>> 4) how mature is the FEC API, can it be modified if needed?
>
>
> It's mature in that we've tested it out-of-tree with a number of situations.
> But as we're moving it into GNU Radio, there's a window here to update it.
> Best done before we merge it into master, though, which will set the API
> until 3.8 of GNU Radio. It'd be great if you could take my current github
> repo branch "fecapi" and see about fitting your coder into it. That'll get
> us some feedback before we merge into master.

Probably the best would be to put my code (separate it from the rest)
into a OOT module for you to see it, and then we pick up the
discussion from there.

Miklos

Sure. If you can fit your encoder/decoder blocks into FEC API-style variables, it'll be easy to use within our system and would make vetting it for incorporation easier, too.

Tom