Re: Custom OOT Block Running on USRP + General GNU Radio Questions.

[Kyeong Su Shin]

Hello Mike,

2/

Maybe what Ron Economos suggested is the right way for you ('set_output_multiple'). I think you cannot change that on-the-fly (please correct me if I am wrong), but maybe that still works for you.

3/

The latter thought ("because the UHD: USRP sink samp rate is set to a rate that it can 'handle'") is probably correct.

In a hardware source/sink block (like the "UHD" blocks), sampling rate defines the actual processing rate (sample acquisition / sample transmission) of the block. The block will try its best to run at the defined rate. 

All other blocks, except "Throttle" block, do not care about the configured sampling rate and just compute the incoming/outgoing data as fast as they can. They may still ask you to input the sampling rate, but that is just used for computation purposes (usually to allow you to use Hertz on other arguments, because you will have to deal with Radians or normalized frequencies otherwise).

"Throttle" block is an exceptional case, which is used to intentionally slow down the execution speed for simulation purposes (and must not be used when hardware blocks are present). It intentionally slows down the execution of the slowgraph, to match the processing rate with the actual time.

So, rate_now and rate_ave will _not_ be consistent with your defined sampling rate (..or maybe, by pure luck), when your blocks are not keeping up or when hardware sources/sinks and Throttle blocks are not present in your flowgraph. The rate that you should be seeing is the maximum rate that GNU Radio can sustain (ignoring the overheads introduced by the UHD blocks), if you have deleted the UHD blocks. If it still does, then maybe it is by pure luck, or maybe some other blocks are throttling the speed of the flowgraph somehow (which shouldn't really happen).

Regards,

Kyeong Su Shin

---

보낸 사람: Michael Bassi <address@hidden>
보낸 날짜: 2019년 11월 25일 월요일 오후 2:20
받는 사람: Kyeong Su Shin <address@hidden>; Discuss Gnuradio <address@hidden>
제목: Re: Custom OOT Block Running on USRP + General GNU Radio Questions.

 

Thank you, Kyeong

I really appreciate your reply.

1/

Does anyone have experience using these functions?

would any help with 2/?

2/

Thank you. It looks like the work function input array "output_items" is a numpy array.

I was therefore able to eliminate the for loop.

I am also now return = 0 if the output buffer isn't long enough. My waveform PRI now appears to be consistent upon visual inspection.

        out = output_items[0]

        if not len(out) < self.num_samples_pri:

            out[:self.num_samples_pw] = 1.0

            out[self.num_samples_pw:] = 0.0

            num_out = self.num_samples_pri

        else:

            num_out = 0   

        return num_out

This may just be a quick fix though. My PRI may be longer than expected in some cases, and I should be doing like you suggest by keeping some internal array.

Does anyone know how to implement this properly? Assuming it is necessary?

3/

With the python code improvements, I am still getting under-runs at the same sample rate of 1e6.

However, after reducing the samp rate by a factor of 10 (to 1e5 samp/sec). I am no longer getting under-run...

Would this be because my code is now working with smaller arrays and therefore runs faster?

Or because the UHD: USRP sink samp rate is set to a rate that it can 'handle'?

I still have a shaky understanding of what sampling rate means here.

I have attached Probe Rate -> Message Debug blocks to the output of my custom block, as you suggested. rate_now and rate_ave appear to be consistent with my defined sampling rate. Is there a specific test you do with these blocks? what do you look for?

I have the C++ implementation ready to go, though still having troubles running my custom blocks from the USRP (see 5/)

RFNoC is just plain intimidating.

4/

Thanks I will check out other custom blocks.

5/

Yes, hopefully someone can help!

Thanks again,

Mike

---

From: Kyeong Su Shin <address@hidden>
Sent: 25 November 2019 13:44
To: Michael Bassi <address@hidden>; Discuss Gnuradio <address@hidden>
Subject: Re: Custom OOT Block Running on USRP + General GNU Radio Questions.

 

Hello Mike,

I never used USRP E3xx series boards, so I coud be off on some points, but:

1: Yes, you are correct. If I recall correctly, it uses a ring buffer-like structure. Length of the output buffer is basically amount of data points that you can safely write to the buffer at the moment, and therefore variable.

There are a few functions that can affect the buffer size. See: https://www.gnuradio.org/doc/doxygen/classgr_1_1block.html (assumes C++, though).

2:

If your output buffer is too small, just keep your unwritten data in a some internal array (self.something) and write to the output buffer during the next round or wait (and do not write to the buffer) until the buffer grows large enough (maybe there are better ways to do this - can anyone help?).

By the way, I believe you are using Python to write OOT blocks. Python OOT blocks can be extremely slow, especially if you are using for and if statements and/or running it on an low-power ARM board. Use Python tricks to minimise the number of for and if statments (use something like "a[0:2] = [1] * 2 ").  It is just like writing a Matlab/Octave simulation.

3:

"Sampling Rate" in the GNU Radio blocks, except in some hardware sources/sinks (including "UHD: USRP sink") and the "Throttle" block (which is for simulation purposes), is just a concept. It does not actually affect the processing speed of the code, and is there to solely make your calculations easier (by allowing you to think in Hertz, instead of in Radians). Actual processing speed is defined by the sampling rate of the hardware sources/sink blocks in your flow graph (the "UHD: USRP Sink" block, in this case), and all other blocks will just try to generate as many samples as they can, regardless of their configured sampling rates.

So, if your codes can generate samples fast enough, and if you do not have conflicting settings (more than two instances of hardware sources/sinks , or "Throttle" blocks, with conflicting sampling rates), you should not see any underruns.

I think your codes are indeed running too slow, especially since the processing power of E313 is very limited. To verify this, replace your "UHD: USRP Sink" with a "Null Sink" block and attach "Probe Rate" -> "Message Debug" to your code (I've never done this on a E313, but I think it should still work). It will show you the maximum throughput of your flowgraph. If this is indeed the case, you must try optimising your codes, or re-write them in C++ (Or even better, see if you can push things to RFNoC).

4:

It depends (there are some cases where you can write fast Python blocks), but yes, C++ blocks are usually _way_ faster, and you are better off with C++ blocks, especially if you are dealing with low-power ARM chips. Poorly written C++ blocks can still run very slowly, so please read OOT blocks that other people wrote before moving on ( http://cgran.hopto.org/ ).

RFNoC is also an option for USRP E3xx / X3xx series boards. The Cortex-A9 CPUs in your USRP E313 have very limited computational power, so you may HAVE to use RFNoC. That requires some HDL experiences, however (FPGA development required).

5: (maybe someone else can answer this?)

Regards,

Kyeong Su Shin

---

보낸 사람: Michael Bassi <address@hidden> 대신 Discuss-gnuradio <discuss-gnuradio-bounces+ksshin=address@hidden>
보낸 날짜: 2019년 11월 25일 월요일 오전 9:32
받는 사람: Discuss Gnuradio <address@hidden>
제목: Custom OOT Block Running on USRP + General GNU Radio Questions.

 

Hi guys,

Firstly, I feel obliged to explain that I'm a young engineer relatively new to Linux, SDR, GNU radio and USRP, and currently overwhelmed with the task assigned to me! Many thanks in advance to any help provided. First time posting to the

I have written a custom OOT block for GNU Radio that reads in information from a file (in this case, pulse width (PW), and pulse repetition interval (PRI)), and outputs a corresponding waveform with matching characteristics. I am attempting to run this from a USRP E312 device. It is currently partially working, and I have several questions:

1/ In the work function, an output array is provided as follows "out = output_items[0]". The length of this array seems to be maxed at 8192, however some times it's much less. Is this because the scheduler doesn't always have enough resources to do the send the full 8192 - I found this statement GNU website somewhere, is it related to this 'phenomenon'?

/* By default, GNU Radio runs a scheduler that attempts to optimize throughput. Using a dynamic scheduler, blocks in a flowgraph pass chunks of items from sources to sinks.The sizes of these chunks will vary depending on the speed of processing. For each block, the number of items it can process is dependent on how much space it has in its output buffer(s) and how many items are available on the input buffer(s). */

Could someone please explain the output_items object to me?

2/ This is how i'm setting the output, out array within my custom block work function:

       self.num_samples_pri = int(self.sample_rate * float(self.pri))

       self.num_samples_pw = int(self.sample_rate * float(self.pw))

       for i  in range(self.num_samples_pri):        

            if i < len(out):

                if i < self.num_samples_pw:

                    out[i] = 1

                else:

                    out[i] = 0

If the out array length is < num_samples_pw, my waveform is cut short, and I get uneven spacing between some of the waveform pulses.

My question is, how do I deal with this issue? I have tried to appending more elements to out array, but they seem to be ignored.

3/ I have built a GRC project with my custom block as source, and UHD: USRP sink block as output so that I can run my project soley on the USRP device with no connection to the PC.

When I run the compiled file on my USRP device, I am getting continuous under-run errors. I know that under-run occurs when a block is not receiving enough samples. In my case, the sample rate used in my code matches that entered in the UHD USRP sink block. D

Is the underrun occurring because my USRP device isn't running my custom code fast enough?

How do I deal with this?

4/ Should I be using custom blocks written in C++ instead of python, when the intent is to run on the USRP device. I have heard that C++ is faster.

5/ I was able to run my custom PYTHON block from the USRP by copying the gr-custom_block folder to the device, and building using cmake, etc. however the C++ implementation of my custom block was not recognized.

AttributeError: 'module' object has no attribute 'readfilesource_cpp'

I got some advice on stack overflow about setting up a cross compilation tool-chain, and recommended Ettus procedures to follow AN-311 and AN-315. However, It was a struggle.

I would love some more advice on this!

Cheers,

Mike