libadawata is a C++17/20 library that implements a real-time compatible wavetable oscillator with Antiderivative Anti-Aliasing based on a Infinite Impulse Response filter (ADAA-IIR for short).

Introduction

This is based on the work of Leonardo Gabrielli, Stefano D'Angelo, Pier Paolo La Pastina and Stefano Squartini :
Antiderivative Antialiasing for Arbitrary Waveform Generation - August 2022

This implementation can provide really strong anti-aliasing, with filters of orders up to 16, for a complexity similar to an oversampling x8 (see benchmarks).

It was first published at the Audio Developer Conference 2023, you can find the slides here

What changed since ADC23

libadawata now uses a time-based cross-fading between mipmap tables (~= octaves)
aarch64 implementations was fixed
clang-14 support was added
ffast-math is no longer invasive (made PRIVATE)

Requirements

libadawata requires a C++17 or C++20 compliant compiler. It has been tested with the following setup

Arch	Compiler	Config	Status
x84_64	g++-11	SSE2	OK
x84_64	g++-11	SSE2 + FMA3	OK
x84_64	g++-10	SSE2	OK
x84_64	g++-10	SSE2 + FMA3	OK
x84_64	clang-14	SSE2 + FMA3	OK
aarch64	g++ 10	neon64	OK

It depends on the following libraries :

xsimd
libsamplerate
libsndfile (only for tests)

See Licensing for licensing details

How to use

Everything is under the adwt:: namespace. There are two classes to use :

adwt::Oscillator : the main oscillator class
adwt::WavetableData : the class used to load wavetable into the oscillator

Here is a basic example on how to create an oscillator, and load a waveform :

#include <adwt/adwt.hpp>
#include <iostream>
#include <vector>

const auto num_waveforms = 128;
const auto waveform_len = 2048;
auto my_wavetable = std::vector<float>(num_waveforms * waveform_len);
// TODO: Fill the vector with your wavetable

// Load the wavetable
auto wavetable_data = adwt::WavetableData::build(my_wavetable, num_waveforms, 44100);
if (wavetable_data == nullptr) {
    std::cerr << "Failed to load the wavetable" << std::endl;
    std::abort();
}

// Create your oscillator and pick a filter
auto osc = adwt::Oscillator<adwt::FilterType::kType5>{};

// Init your oscillator
if (osc.init(std::move(wavetable_data)) != 0) {
  std::cerr << "Failed to init oscillator" << std::endl;
  std::abort();
}

You can then safely generate audio on your audio thread :

constexpr auto kBlockSize = 256;
const auto phases_input = std::array<float, kBlockSize>{};
auto audio_output = std::array<float, kBlockSize>{};

osc.process(phases_input, audio_output);

In order to change the wavetable loaded in the oscillator, the workflow is :

Asynchronously compute a new wavetable by calling adwt::WavetableData::build()
Synchronously swap the wavetable using adwt::Oscillator::swapWavetable()

You can select the waveform to use in the wavetable with adwt::Oscillator::setWaveform()

Span implementation

When compiled without c++20 support, this library comes with Microsoft's GSL span implementation. The adwt::Span alias either alias gsl::span or std::span.

Limitations

Due to the nature of the algorithm (see the slides), it has limitations on how fast the frequency of the signal can vary without an exponential increase of the complexity.

Note that the cross fading strategy changed since the ADC23 presentation. libadawata now uses a more classic time-based cross-fading between the different mipmap tables.

When creating a WavetableData object, the caller can precise how long the cross-fading takes (default to 5ms). As long as you limit variations to no more than a single octave change every 5ms, no artifacts above -90dB will appear.

Setting a new frequency

To avoid such problems when you just want to play at a new arbitrary frequency, you can use the adwt::Oscillator::resetInternals() method to prepare the recursive state for your new frequency.

The cross-fading / amplitude envelope to transition from the previous frequency to the new one is left to the caller.

Pick a filter

This code uses SIMD optimisation to compute several orders of the IIR filter at the same time. You can check the slides for more details, but I suggest using an filter of order 8, which should perform well on any platform with at least 128bit of SIMD register.

Check filter_type.hpp for more details on which filter are available, and it should be easy to add support for new filter.

Build

You can build libadawata with cmake >= 3.16

Testing

Automated testing is really hard for this kind of tasks, as robust and unbiased measurements are limited and not always reliable.

libadawata is shipped with two sets of tests

Unit Tests

These tests only test the basic interface of libadawata. They don't check the generated audio content.

You can enable them using -DBUILD_TESTING=ON. You can run them with CTest or manually :

./build/tests/libadawata_tests

Quality Tests

These check provide GUI checks for the user. You can enable them using -DENABLE_QUALITY_TESTS=ON

You will also need some additional python modules, listed in the requirements.txt file, which you can install using :

pip install -r requirements.txt

Sweep test

This will print the spectrogram of a logarithmic sweep test. This is useful for identifying cross-fading issues.

You can run it with cmake :

cmake -B build -DBUILD_TESTING=ON -DENABLE_QUALITY_TESTS=ON .
cmake --build build --target sweep_test

SNR test

This will compute the SNR of static tones over a logarithmic range of frequencies. This is usefull to identify degrading quality issue of the whole algorithm.

You can run it with cmake :

cmake -B build -DBUILD_TESTING=ON -DENABLE_QUALITY_TESTS=ON .
cmake --build build --target snr_test

I strongly encourage you to run this test on Release mode

Licensing

This work is licensed under the MIT License.

It depends on the following libraries :

xsimd licensed under the BSD3-clause license
libsamplerate licensed under the BSD2-clause license
libsndfile (only for tests) licensed under the LGPL v2.1

The benchmarks are licensed under GPL3, see details here

Contributions

Contributions are very welcome. I do plan to improve this library a bit (see milestones) but keep in mind this is a hobby project.

Algorithm improvement milestones

Experiment with filter design to find the best one for each order
Move to a more classic time-based cross-fading

QoL Milestones

R&D

This code was developed based on the python experimentations I made, which you can find here

Contact

If you want to discuss about it you can open an issue or you can find me on :

maxmarsc/libadawata