Re: [Discuss-gnuradio] Avoiding divide by zero in case of input being zero.

[Marcus D. Leech]

On 03/31/2018 01:18 AM, Anshul Thakur wrote:

Hi Gilad,

You are correct about getting zeros when using history and I am using history. But, I am accounting for that by using the offset:

const gr_complex *current = (const gr_complex *)&((const gr_complex *)input_items[0])[(history() -1)];

Also, if that is the case, shouldn't the behaviour be consistent across unit-tests and flowgraph usage?

Regarding getting 0s from the device, I went through the discussion. I agree to it. So, I'll have to account for the zero values. Any ideas on how to do that? 

The reason why I don't want to use an 'if' block is simply because the situation of 0 values is expected to share a very small fraction of the entire run and

to handle that small case, the regular cases will also necessarily have to go through that additional 'if' check. I wanted to avoid that.

Regards,

Anshul

Assuming that your input voltages are uniformly distributed about zero, and we'll say for the sake of argument that we're talking a 12-bit sampler,
   and we model the input as a purely-random process (it isn't of course) then the density of numerical zeros will be about 1/2048 in the data stream.

That's a lot of zeros.

On 31 March 2018 at 10:37, Gilad Beeri (ApolloShield) <address@hidden> wrote:

Disclosure: I haven't looked at your code.

0 values can be presented in GNU Radio when you use history, because if your history is N, the first N-1 items are going to be zeros.

Anyway, regarding your comment "it is not expected that a device/stream would ever spit out zero values.",

I did have 0 values from a USRP device, see discussion in http://lists.ettus.com/pipermail/usrp-users_lists.ettus.com/2017-October/026851.html.

On Sat, Mar 31, 2018 at 6:52 AM Anshul Thakur <address@hidden> wrote:

Michael, Marcus,

Right now, the code is a work in progress so I haven't made a git repository out of it. However, I have it on dropbox. Here's the link to the source folder(p1_detector_impl.cc is the source in question):

https://www.dropbox.com/sh/blfmxsaidrkh28t/AAArp8IHavzCGFlJs6E6-Hrca?dl=0

As for Marcus's question regarding why use a circular buffer?

It isn't exactly a circular buffer now, but more of a shift register. The reasons are as follows:

1. I needed running sums for correlations in B-Branch and C-Branch correlators, and Power Sums (for average power) to normalize them. Then, I also needed a finite delay buffer to delay the C-Branch before it gets multiplied with the B-Branch.

2. It kind of carried over from the last implementation attempt:

Assertion: If a peak is detected after the multiplication, the signal boundary is 1024 samples behind that index. 

Once the correlations crossed a threshold (the code entered state=1), instead of looking back, I then needed to look forward to see if it were a false alarm or not. So, I compute the correlations across all available current inputs and try to find a peak. If a peak is found, enter state=3 where we do a correlation with the carrier distribution sequence after FFT of all signals of interest. So, here, I not only needed just the single value (the running sum), but the entire state of the delay register and the B-Branch correlator.

I hope I am able to convey the reason for implementing one myself.

In the current implementation, I make an assumption that the threshold is so high that only the desired signals would cross it (100-150 times the average). So I skip the state=1 logic and directly go into state=2 logic of aggressively doing a FFT and correlation with the CDS.

However, I don't think this has a binding on the incoming values. Use of buffers is internal to the implementation, I am just printing out the current values as they arrive.

For example, when I use the test file in 'make test', the values fed in are non-zero from t=1. However, when using gnuradio-companion, t=56 line is where the file source starts yielding proper inputs to my block. The stdout prints of the initial values in both GRC and make tests are attached. The gnuradio-companion version has my first 55 samples zeroed and the 56th input onward is then same for both.

P.S.: The source stream is a 1.2 Gigs file, so haven't uploaded it. If you'd like I can do that too. It was generated by using a DVB-T2 Tx block and writing the output into a file sink.

Warm regards,

Anshul Thakur

On 31 March 2018 at 02:27, Müller, Marcus (CEL) <address@hidden> wrote:

Hi Anshul,

you shouldn't have to have your own buffer for a running sum – can you
explain why you're doing that?
A running sum can trivially be implemented with the IIR filter block
with Feed-Forward taps (1,) and Feed-back taps (1,0)!
Where does in a running sum does a division take place?

> (a) Why am I getting the initial zero samples from the file block in
> gnuradio_companion and non-zero values when using a vector_source in
> unit tests?

If these zeros are not in the file you're reading, your block has a
bug!


> (b) What can I do about it (here specifically as a fix to the
> situation, and a general guideline to always remember)?

good question, but we'd need to know your code, your motivation for a
circular buffer, and why you're implementing a running sum yourself!

Best regards,
Marcus

On Fri, 2018-03-30 at 23:19 +0530, Anshul Thakur wrote:
> Hi,
>
> I used a circular buffer of finite size to keep the past 'N' power
> values of the sample stream in my block as a part of creating a
> running sum. This buffer is initialized to 0 in the constructor.
> The running sum of powers is used to compute the average power used
> in computing signal correlation.
>
> I have a capture stream (cfile) to test the operation of the block.
> The test case uses a vector_source_c block to read the contents of
> the file into memory. The unit tests pass without error.
>
> However, when I use the block in a flowgraph in that reads the same
> file from a file source block gnuradio_companion, I am getting the
> first few sample values as 0 which cause a divide by zero
> problem. This messes up the rest of the running sum. I don't want to
> put an 'if' block that checks for the zero condition as it is not
> expected that a device/stream would ever spit out zero values.
>
> (a) Why am I getting the initial zero samples from the file block in
> gnuradio_companion and non-zero values when using a vector_source in
> unit tests?
>
> (b) What can I do about it (here specifically as a fix to the
> situation, and a general guideline to always remember)?
>
> I am using GNURadio version 3.7.12.
>
> Regards,
> Anshul

> _______________________________________________
> Discuss-gnuradio mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/discuss-gnuradio

_______________________________________________
Discuss-gnuradio mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio

_______________________________________________
Discuss-gnuradio mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio