Most of the code is available under the permissive BSD license, with some optional parts under GPL. For additional details, see <ahref="#licensing">licensing</a>.
- The package comes with a command-line tool `csdr`, which lets you build DSP processing chains by shell pipes.
- The code of *libcsdr* was intended to be easy to follow.
- *libcsdr* was designed to use auto-vectorization available in *gcc*. It means that it can achieve some speedup by taking advantage of SIMD command sets available in today's CPUs (e.g. SSE on x86 and NEON on ARM).
How to compile
--------------
The project was only tested on Linux. It has the following dependencies: `libfftw3-dev`
make
sudo make install
If you compile on ARM, please edit the Makefile and tailor `PARAMS_NEON` for your CPU.
To run the examples, you will also need <ahref="http://sdr.osmocom.org/trac/wiki/rtl-sdr">rtl_sdr</a> from Osmocom, and the following packages (at least on Debian): `mplayer octave gnuplot gnuplot-x11`
Credits
-------
The library was written by Andras Retzler, HA7ILM <<randras@sdr.hu>>.
- Baseband I/Q signal is coming from an RTL-SDR USB dongle, with a center frequency of `-f 104300000` Hz, a sampling rate of `-s 240000` samples per second.
- The `rtl_sdr` tool outputs an unsigned 8-bit I/Q signal (one byte of I sample and one byte of Q coming after each other), but `libcsdr` DSP routines internally use floating point data type, so we convert the data stream of `unsigned char` to `float` by `csdr convert_u8_f`.
- We want to listen one radio station at the frequency `-f 89500000` Hz (89.5 MHz).
- No other radio station is within the sampled bandwidth, so we send the signal directly to the demodulator. (This is an easy, but not perfect solution as the anti-aliasing filter at RTL-SDR DDC is too short.)
- After FM demodulation we decimate the signal by a factor of 5 to match the rate of the audio card (240000 / 5 = 48000).
- A de-emphasis filter is used, because pre-emphasis is applied at the transmitter to compensate noise at higher frequencies. The time constant for de-emphasis for FM broadcasting in Europe is 50 microseconds (hence the `50e-6`).
- Also, `mplayer` cannot play floating point audio, so we convert our signal to a stream of 16-bit integers.
- We want to listen to one radio station, but input signal contains multiple stations, and its bandwidth is too large for sending it directly to the FM demodulator.
- We shift the signal to the center frequency of the station we want to receive: `-0.085*2400000 = -204000`, so basically we will listen to the radio station centered at 89504000 Hz.
- We decimate the signal by a factor of 10. The transition bandwidth of the FIR filter used for decimation will be 10% of total bandwidth (as of parameter 0.05 is 10% of 0.5). Hamming window will be used for windowed FIR filter design.
I/Q source ------------> FIR decimation ------------> FM demod -> frac. decimation ---------> deemphasis -> sound card
*Note:* there is an example shell script that does this for you (without the unnecessary shift operation). If you just want to listen to FM radio, type:
csdr-fm 89.5 20
The first parameter is the frequency in MHz, and the second optional parameter is the RTL-SDR tuner gain in dB.
- It will design a filter that lets only the positive frequencies pass (low cut is 0, high cut is 0.5 - these are relative to the sampling rate).
- If `--octave` and everything that follows is removed from the command, you get only the taps. E. g. the raw output of `firdes_lowpass_f` can be easily copied to C code.
- It is a modified Weaver-demodulator. The complex FIR filter removes the lower sideband and lets only the upper pass (USB). If you want to demodulate LSB, change `bandpass_fir_fft_cc 0 0.05` to `bandpass_fir_fft_cc -0.05 0`.
- We calculate the Fast Fourier Transform by `csdr fft_cc` on the first 1024 samples of every block of 1200000 complex samples coming after each other. (We calculate FFT from 1024 samples and then skip 1200000-1024=1198976 samples. This way we will calculate FFT two times every second.)
- The window used for FFT is the Hamming window, and the output consists of commands that can be directly interpreted by GNU Octave which plots us the spectrum.
Usage
-----
Some basic concepts on using *libcsdr*:
### Data types
Function name endings found in *libcsdr* mean the input and output data types of the particular function. (This is similar to GNU Radio naming conventions).
Data types are noted as it follows:
-`f` is `float` (single percision)
-`c` is `complexf` (two single precision floating point values in a struct)
-`u8` is `unsigned char` of 1 byte/8 bits (e. g. the output of `rtl_sdr` is of `u8`)
`csdr` should be considered as a reference implementation on using `libcsdr`. For additional details on how to use the library, check `csdr.c` and `libcsdr.c`.
Regarding `csdr`, the first command-line parameter is the name of a function, others are the parameters for the given function. Compulsory parameters are noted as `<parameter>`, optional parameters are noted as `[parameter]`.
Optional parameters have safe defaults, for more info look at the code.
Along with copying its input samples to the output, it prints a warning message to *stderr* if it finds any IEEE floating point NaN values among the samples.
floatdump_f
It prints any floating point input samples.
The format string used is `"%g "`.
flowcontrol <data_rate><reads_per_second>
It limits the data rate of a stream to a given `data_rate` number of bytes per second.
It copies `data_rate / reads_per_second` bytes from the input to the output, doing it `reads_per_second` times every second.
`rate` is a floating point number between -0.5 and 0.5.
`rate` is relative to the sampling rate.
Internally, a sine and cosine wave is generated to perform this function, and this function uses `math.h` for this purpose, which is quite accurate, but not always very fast.
Internally, this function uses trigonometric addition formulas to generate sine and cosine, which is a bit faster. (About 4 times on the machine I have tested it on.)
Internally, this function uses a look-up table (LUT) to recall the values of the sine function (for the first quadrant).
The higher the table size is, the smaller the phase error is.
decimating_shift_addition_cc <rate> [decimation]
It shifts the input signal in the frequency domain, and also decimates it, without filtering. It will be useful as a part of the FFT channelizer implementation (to be done).
It cannot be used as a channelizer by itself, use `fir_decimate_cc` instead.
It is an AM demodulator that uses `sqrt`. On some architectures `sqrt` can be directly calculated by dedicated CPU instructions, but on others it may be slower.
Low-pass FIR filter design function to output real taps, with a `cutoff_rate` proportional to the sampling frequency, using the windowed sinc filter design method.
`cutoff_rate` can be between 0 and 0.5.
`length` is the number of filter taps to output, and should be odd.
The longer the filter kernel is, the shorter the transition bandwidth is, but the more CPU time it takes to process the filter.
The transition bandwidth (proportional to the sampling rate) can be calculated as: `transition_bw = 4 / length`.
Some functions (below) require the `transition_bw` to be given instead of filter `length`. Try to find the best compromise between speed and accuracy by changing this parameter.
`window` is the window function used to compensate finite filter length. Its typical values are: `HAMMING`, `BLACKMAN`, `BOXCAR`. For the actual list of values, run: `cpp libcsdr.c | grep window\ ==`
The `--octave` parameter lets you directly view the filter response in `octave`. For more information, look at the [Usage by example] section.
-`hang_time` is the number of samples to wait before strating to increase the gain after a peak.
-`reference` is the reference level for the AGC. It tries to keep the amplitude of the output signal close to that.
-`attack_rate` is the rate of decreasing the signal level if it gets higher than it used to be before.
-`decay_rate` is the rate of increasing the signal level if it gets lower than it used to be before.
- AGC won't increase the gain over `max_gain`.
-`attack_wait` is the number of sampels to wait before starting to decrease the gain, because sometimes very short peaks happen, and we don't want them to spoil the reception by substantially decreasing the gain of the AGC.
-`filter_alpha` is the parameter of the loop filter.
Its default parameters work best for an audio signal sampled at 48000 Hz.
It is a faster AGC that linearly changes the gain, taking the highest amplitude peak in the buffer into consideration. Its output will never exceed `-reference ... reference`.
It performs an FFT on the first `fft_size` samples out of `out_of_every_n_samples`, thus skipping `out_of_every_n_samples - fft_size` samples in the input.
It can draw the spectrum by using `--octave`, for more information, look at the [Usage by example] section.
FFTW can be faster if we let it optimalize a while before starting the first transform, hence the `--benchmark` switch.
Encodes the audio stream to IMA ADPCM, which decreases the size to 25% of the original.
decode_ima_adpcm_u8_i16
Decodes the audio stream from IMA ADPCM.
compress_fft_adpcm_f_u8 <fft_size>
Encodes the FFT output vectors of `fft_size`. It should be used on the data output from `logpower_cf`.
It resets the ADPCM encoder at the beginning of every vector, and to compensate it, `COMPRESS_FFT_PAD_N` samples are added at beginning (these equal to the first relevant sample).
The actual number of padding samples can be determined by running `cat csdr.c | grep "define COMPRESS_FFT_PAD_N"`.
fft_exchange_sides_ff <fft_size>
It exchanges the first and second part of the FFT vector, to prepare it for the waterfall/spectrum display. It should operate on the data output from `logpower_cf`.
It converts a real signal to a double sideband complex signal centered around DC.
It does so by generating a complex signal:
* the real part of which is the input real signal,
* the imaginary part of which is `q_value` (0 by default).
With `q_value = 0` it is an AM-DSB/SC modulator. If you want to get an AM-DSB signal, you will have to add a carrier to it.
add_dcoffset_cc
It adds a DC offset to the complex signal: `i_output = 0.5 + i_input / 2, q_output = q_input / 2`
convert_f_samplerf <wait_for_this_sample>
It converts a real signal to the `-mRF` input format of [https://github.com/F5OEO/rpitx](rpitx), so it allows you to generate frequency modulation. The input signal will be the modulating signal. The `<wait_for_this_sample>` parameter is the value for `rpitx` indicating the time to wait between samples. For a sampling rate of 48 ksps, this is 20833.
fmmod_fc
It generates a complex FM modulated output from a real input signal.
fixed_amplitude_cc <new_amplitude>
It changes the amplitude of every complex input sample to a fixed value. It does not change the phase information of the samples.
Some parameters can be changed while the `csdr` process is running. To achieve this, some `csdr` functions have special parameters. You have to supply a fifo previously created by the `mkfifo` command. Processing will only start after the first control command has been received by `csdr` over the FIFO.
`csdr` has three modes of determining the buffer sizes, which can be chosen by the appropriate environment variables:
* default: 16k or 1k buffer is chosen based on function,
* dynamic buffer size determination: input buffer size is recommended by the previous process, output buffer size is determined by the process,
* fixed buffer sizes.
`csdr` uses two buffer sizes by **default**.
* For operations handling the full-bandwidth I/Q data from the receiver, a buffer size of 16384 samples is used.
* For operations handling only a selected channel, a buffer size of 1024 samples is used (see `env_csdr_fixed_bufsize` in the code).
`csdr` now has an experimental feature called **dynamic buffer size determination**, which is switched on by issuing `export CSDR_DYNAMIC_BUFSIZE_ON=1` in the shell before running `csdr`. If it is enabled:
* All `csdr` processes in a DSP chain acquire their recommended input buffer size from the previous `csdr` process. This information is in the first 8 bytes of the input stream.
* Each process can decide whether to use this or choose another input buffer size (if that's more practical).
* Every process sends out its output buffer size to the next process. Then it startss processing data.
* The DSP chain should start with a `csdr setbuf <buffer_size>` process, which only copies data from the input to the output, but also sends out the given buffer size information to the next process.
* The 8 information included in the beginning of the stream is:
* a preamble of the bytes 'C','S','D','R' (4 bytes),
* the buffer size stored as `int` (4 bytes).
* This size always counts as samples, as we expect that the user takes care of connecting the functions with right data types to each other.
> I added this feature while researching how to decrease the latency of a DSP chain consisting of several multirate algorithms.
> For example, a `csdr fir_decimate_cc 10` would use an input buffer of 10240, and an output buffer of 1024. The next process in the chain, `csdr bandpass_fir_fft_cc` would automatically adjust to it, using a buffer of 1024 for both input and output.
> In contrast to original expectations, using dynamic buffer sizes didn't decrease the latency much.
If dynamic buffer size determination is disabled, you can still set a **fixed buffer size** with `export CSDR_FIXED_BUFSIZE=<buffer_size>`.
For debug purposes, buffer sizes of all processes can be printed using `export CSDR_PRINT_BUFSIZES=1`.
If you add your own functions to `csdr`, you have to initialize the buffers before doing the processing. Buffer size will be stored in the global variable `the_bufsize`.
Example of initialization if the process generates N output samples for N input samples:
if(!sendbufsize(initialize_buffers())) return -2;
Example of initalization if the process generates N/D output samples for N input samples:
if(!initialize_buffers()) return -2;
sendbufsize(the_bufsize/D);
Example of initialization if the process allocates memory for itself, and it doesn't want to use the global buffers:
getbufsize(); //dummy
sendbufsize(my_own_bufsize);
Example of initialization if the process always works with a fixed output size, regardless of the input:
`csdr` was tested with GNU Radio Companion flowgraphs. These flowgraphs are available under the directory `grc_tests`, and they require the <ahref="https://github.com/simonyiszk/gr-ha5kfu">gr-ha5kfu</a> set of blocks for GNU Radio.
*sdr.js* is *libcsdr* compiled to JavaScript code with *Emscripten*. Nowadays JavaScript runs quite fast in browsers, as all major browser vendors included JavaScript JIT machines into their product. You can find a <ahref="https://kripken.github.io/mloc_emscripten_talk/cppcon.html">great introductory slideshow here</a> on the concept behind *Emscripten* and *asm.js*.
The purpose of *sdr.js* is to make SDR DSP processing available in the web browser. However, it is not easy to use in production yet. By now, only those functions have wrappers that the front-end of OpenWebRX uses.
To compile *sdr.js*, first get <ahref="http://emscripten.org/">Emscripten</a>. (It turns out that there is an *emscripten* package in Ubuntu repositories.)
