Adaptive filter algorithm examples and DSO build helper for ATFA

ATFA is the Ambiente de Testes para Filtros Adaptativos (Testing environment for adaptive filters).

This repository

• provides examples of adaptive filtering algorithms implemented in C++ and C, and exposing the API that ATFA expects;
• explains how to understand the examples, how to modify them, how to write your own algorithms, and how to compile them;
• provides scripts to compile the algorithms into the file format that ATFA expects, namely dynamic shared objects, or DSOs (*.so).

Quick How-To

The algorithms must be compiled and linked into an *.so file. This can be done manually, or by using the helper script build.py. Aditionally, the script build-all.sh can also be used to compile all of the examples at once.

If you just want to compile the examples to use from ATFA, type:

$ cd atfa-examples
$ bash build-all.sh

See the section on using the build.py script for more information.

How to write algorithms

ATFA accepts custom adaptive filtering algorithms in the form of dynamic shared object (*.so) files. Each adaptive filter DSO must export the following symbols:

• adapf_init
• adapf_restart
• adapf_close
• adapf_run
• adapf_getw
• adapf_title
• adapf_listing

The last two (adapf_title and adapf_listing) are generated automatically by the build.py script. Below, each of the functions to be exported is explained:

The adapf_* functions

void *adapf_init()

The adapf_init function initializes (e.g., acquires memory for) the data structures that the algorithm will use. This data structure might include, for example, the impulse response vector (vector of filter coefficients), the input vector, and any other kind of parameter that the algorithm needs to keep track of.

This function must return a pointer to the initialized data.

void *adapf_restart(void *data)

This function resets the information contained in data (e.g., zeroes out the impulse response that has been learned so far), and returns a pointer to the zeroed data. This function might free the memory pointed to by the old data and return a pointer newly allocated-and-initialized memory.

int adapf_close(void *data)

This function releases the resources acquired in adapf_init. ATFA will call it before closing the access to the DSO, in order to prevent memory leaking. The pointer passed to adapf_close is the one returned by adapf_init, and the value returned from adapf_close should be zero if it fails and nonzero otherwise.

float adapf_run(void *data, float sample, float y)

This is the function which implements the algorithm itself. Given an audio sample from the input of the filter (sample) and an audio sample from the desired signal (y), adapf_run should calculate the filter output, say y_hat, and then return the error y - y_hat.

This function can read from and write to *data in order to access and/or update values such as the filter coefficients, the input vector (with memory), etc.

adapf_run is the only function that's called from within the audio run-time loop. The other functions are all allowed to request resources to the OS, and perform slow operations in general, but this one isn't.

void adapf_getw(void *data, float **begin, unsigned *n)

This function is called by ATFA in order to introspect the filter coefficients. It must place at *begin a pointer to the first element of the array of filter coefficients, and at *n the number of elements that can be read from that array (that is, the number of coefficients, which is N+1, where N is the filter order).

const char *adapf_title()

This function should just return a static string containing a short "name" for the adaptive filter algorithm. For example, if the algorithm is an implementation of the NLMS algorithm with mu=0.7 and Delta=1.5, then the returned string could be "NLMS: 0.7, 1.5" or maybe just "NLMS".

const char *adapf_listing()

This function must return a static string, just like adapf_title. The string returned from adapf_listing is an algorithmic description of the adaptive filter, which is shown to the users when they click the "Show filter code" button in ATFA.

Examples

Inside the LMS/, NLMS/, and AP/ directories, you will find C++ source files implementing these three adaptive filtering algorithms.

The no-op/ direcory contains a file for a dummy algorithm, which does nothing at all. The source file inside static-w/ implements a static (instead of an adaptive) filter (the filter coefficients are never updated). Both of these can be used for test purposes.

All of the above are implemented in C++. See LMS-C/ for an example using plain C.

How to compile an algorithm

To compile an algorithm into a DSO file to be used with ATFA, you can do so manually—just generate an *.so that exports all of the functions mentioned above—, or you can use the build.py helper script.

To use this script, open a shell inside the directory containing your source file(s), and then run the script. For example, to compile the LMS/ algorithm, type:

$ cd atfa-examples
$ cd LMS
$ python3 ../build.py

By default, build.py will compile and link all of the *.c and *.cpp files in the current directory. To use a custom set of sources, just pass them as arguments:

$ python3 ../build.py source1.cpp source2.cpp

The full set of source files specified on the command-line (or present in the directory, if none is specified) must contain the definitions of all of the functions mentioned above, except atfa_title and atfa_listing. These will be automatically generated.

By default, the title is made up based on the command line, and the algorithm listing is just a transcription of the source files. If you want to use a custom title, use the -t/--title flag. If you want to specify a custom file containing the listing, use the -l/--listing flag:

$ python3 ../build.py --title "My Algorithm" --listing my_algorithm.txt

If you want to provide the atfa_title or atfa_listing functions yourself, use the +t or +l flags:

$ python3 ../build.py main_source.cpp title_and_listing.cpp +tl

For more options, type:

$ python3 ../build.py --help

Using build-all.sh

You can also use the convenience script build-all.sh. This script will just call python3 ../build.py from inside each subdirectory of the current directory. It will also pass the name of the directory in the --title flag.