Differences

This shows you the differences between two versions of the page.

--- project:sdr:fcl [2016/04/17 20:35] – jenda
+++ project:sdr:fcl [2016/11/25 07:29] (current) – ↷ Links adapted because of a move operation ruza
@@ Line 1: / Line 1: @@
+====== FCL, a fast spectrum channelizer ======
+FCL splits wideband spectrum from your SDR into narrowband channels. This is useful for networks that have lots of transmitters side-by-side, like for example [[http://jenda.hrach.eu/gitweb/?p=fmrec;a=summary|FM broadcast]], [[:project:sdr:tetra|TETRA]], [[project:tetrapol:start|Tetrapol]] and [[project:gsm:start|GSM]].
+It supports everything you would expect from a decent channelizer:
+  * Channelization 8-)
+  * Oversampling and overlapping channels.
+  * On-the-fly frequency correction and signal power detection.
+  * Multiple input formats (float, unsigned and signed integer).
+  * Multithreading.
+We use a straightforward implementation that is cache- and vector-friendly and we have a sane pthreads scheduler. At least on our computers, FCL is 2 to 5 times faster than [[http://gnuradio.org/doc/doxygen/classgr_1_1filter_1_1pfb__channelizer__ccf.html|the GnuRadio implementation]]. Hence //fcl// stands for //Fastest Channelizer in [[https://en.wikipedia.org/wiki/Litom%C4%9B%C5%99ice|Litoměřice]]//.
+You can find description of the mathematical principles of the channelizer in Chapter 2 of [[https://jenda.hrach.eu/thesis.pdf|my thesis]].
+The following chart compares FCL with GnuRadio 3.7.10. FCL was tuned using the hints below on this page, GnuRadio was profiled using volk_profile. Frequency correction in FCL was disabled.
+{{ :project:sdr:fcl.png?nolink |}}
+**Practical results:** Raspberry Pi 3 + rtl-sdr is able to receive 13 Tetra or Tetrapol channels in real time. As it maxes the usage of CPU and memory, proper cooling is necessary!
+===== Downloads =====
+https://jenda.hrach.eu/gitweb/?p=fcl;a=summary
+===== Usage =====
+Build:
+<code>
+make
+</code>
+<note tip>On Raspberry Pi, adding "-march=native -mthumb-interwork -mfloat-abi=hard -mfpu=neon -mtune=native" to CFLAGS makes things even faster. Make sure [[:user:jenda:rpi-gcc|you have correct GCC and core libraries!]]</note>
+<note>It builds .so modules for FIR of lengths 8-24 (where compiler can unroll loops) and one generic. These are loaded dynamically runtime.</note>
+Example:
+<code>
+mkfifo /tmp/myout.ch
+rtl_sdr -f 90e6 -s 2e6 -g 49 -p 44 - | ./fcl -n 20 -s 13 -f "./fir.py 2e6 55e3 11e3" -c 3,8 -i U8 -o /tmp/myout.ch
+</code>
+i.e. read data as uint8_t, channelize to 20 channels with 20/13 oversample, select channels 3 and 8 (they are interleaved in the output stream, use the GnuRadio Deinterleave block to get them), output to /tmp/myout.ch
+The tool listens on localhost:3333 (can be changed with -b and -p) and supports these commands:
+  * setrot - set frequency correction in //milliradians per sample//
+  * getrot - print current frequency correction
+  * setchannels - set channels you want to output, format same as with -c option, e.g. "1,8,17"
+  * getpwr - print relative power of your channels. You can use this to e.g. detect BTSs
+    * getpwn N - increase number of FFT bins N-times (i.e. get much more detailed view of the spectrum); 1 <= N <= 32
+  * gethisto - print histogram (evenly spaced at [0, 1] absolute values of input floats) of samples
+There is a convenient wrapper around gnuradio firdes called fir.py.
+<code>
+usage: ./fir.py samplerate passband transition [rcos]
+passband/samplerate ... how much of spectrum to pass
+transition          ... either transition width in Hz (for non-rcos filter), lower value means higher signal quality and higher computational complexity, or filter length in taps for rcos filter
+rcos                ... use root raised cosine filter instead of Hamming window.
+</code>
+There is and example of how to use FCL to receive TETRA in examples/tetra.py. Other examples may follow as we port other brmlab tools to FCL.
+===== Performance hints =====
+  * The most expensive part is evaluating FIR filters. Computational demands scale almost linearly with filter length. So try to design your filters as short as possible.
+  * On fast computers with many cores, you can try increasing BUFSIZE and RETURNWORK in finetune.h. You can gain ~10 %.
+===== Roadmap =====
+  * write unit tests