Attached is a C implementation of a block I wish was in Gnu Radio.
Anyone done anything similar-enough as a block
in Gnu Radio that would make a decent starting point?
It's basically a block that takes a fixed-size vector input, and a list
of "descriptors" that describe output channels that need to be carved out of
the input vector (a list of start:end which are simply offsets into
the vector). It's intended to be used after a FFT+complex-to-MAG block to
process the resulting vector and select sub-bands and turn them into
power-estimate channels. I see nothing in Gnu Radio that does this,
other than things like the FFT channelizer, PFB channelizer, etc.
Neither of those is efficient in this case, because the channels are of
non-uniform distribution and width, and if the target is simply power
estimation, that part has already basically been done by the
FFT+complex-to-MAG stage.
--
Marcus Leech
Principal Investigator
Shirleys Bay Radio Astronomy Consortium
http://www.sbrac.org
/*
* A specialzed vector de-mux for estimating power within defined sub-bands of
a vector that
* is the output of an FFT+complex-to-MAG
*
* It takes in a fixed-size float vector, and a list of descriptors describing
the sub-bands that are to
* be added together to form an output channel. The output is a number of
channels (one per descriptor).
*
* Consider the following example:
*
* Input vector: XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
* ^^^^^^ ^^^^ ^^^^^^^
* CH 0 CH 1 CH 2 (etc)
*
* For simple multi-channel power estimation this is significantly more
efficient than many other techniques
* that might be used. Channel-width resolution is based purely on the
resolution of the FFT+complex-to-MAG
* that feeds this specialized demux. Resolutions of a few hundred Hz with
25MHz input bandwidths appear to
* be feasible with this technique. For uniformly-spaced channels with
relatively-coarse resolution, a PFB
* channel bank might be better.
*
* This "rides" on the fact that the Gnu Radio FFT implementation is very
efficient, due to its use of
* FFTW, which provides several overlapping techniques to improve performance
including:
*
* o an optimizer
* o aggressive use of SSEx/3DNow! instruction sets
* o multi-threaded FFT calculations (useful for very large FFTs, which are
required for fine resolution)
*
* In many situations, the total number of post-FFT ops you need to do is quite
modest compared to the FFT,
* since you may be ignoring the vast majority of the input vector. In the
application domain of interest
* (multi-channel power estimation for ionospheric measurements), probably
98% of the input vector isn't
* touched. But the channels of interest are:
*
* o not evenly spaced
* o not of uniform bandwidth
* o between 1 and perhaps 10 sub-band "collections" (output channels)
might be of interest
*
* invector - input vector of size 'vecsize'
* vec_begins - list of sub-channel start points (offsets from beginning of
input vector)
* vec_ends - list of sub-channel end points (offsets from beginning of
input vector)
* nchan - number of output channels (or "collections of sub-bands from
input")
* outchans - the output channels
*
*/
int
vector_power_estimator_ff (float invector[], int vecsize, int vec_begins[], int
vec_ends[], int nchan, float outchans[])
{
int i, j;
/*
* For each output channel descriptor
*/
for (i = 0; i < nchan; i++)
{
/*
* Accumulate an output channel from sub-bands in the input
*/
outchans[i] = 0.0;
for (j = vec_begins[i]; j < vec_ends[i]; j++)
{
if (j < vecsize)
{
/*
* Opportunity for volk-ifying here
*/
outchans[i] += invector[j];
}
}
}
return nchan;
}
_______________________________________________
Discuss-gnuradio mailing list
Discuss-gnuradio@gnu.org
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
