AlgoCompSynth-One - Algorithmic Composition and Digital Sound Synthesis

M. Edward (Ed) Borasky

Un-archiving - 3 May 2024

I’m unarchiving this and changing the deployment mechanism. It will be a Distrobox container based on Arch Linux, designed for use on Fedora Atomic Desktops and Universal Blue Aurora and Bluefin.

If you’re into following along, this is happening in branch Arch-Distrobox. Watch this space!

Archiving - 31 March 2024

I am archiving the project for two reasons: 1. NVIDIA is no longer enhancing JetPack 5, and the Orin hardware is just too expensive for me at the moment, and 2. I have too many other projects to continue maintaining this one.

There may be some GPU audio projects in the future, but they will not use Jetson hardware.

What is AlgoCompSynth-One?

AlgoCompSynth-One is a collection of tools for composing music and synthesizing sound on systems with NVIDIA® GPUs. The current implementation runs on the Jetson™ hardware, running JetPack 5.1.1 or later, and Windows 11 with Ubuntu 22.04 LTS jammy running in Windows Subsystem for Linux.

I test the JetPack version on an AGX Xavier with 16 GB of RAM, but it should run on any Xavier or Orin device. It will not run on a device that only supports JetPack 4.

I test the WSL version on a laptop with a GTX 1650 and a desktop with an RTX 3090. It should run with any device supported by CUDA 11.8.

What can it do?

AlgoCompSynth-One creates a Mamba virtual environment and installs the remaining components in it. The user specifies the name of the environment when the script creates it, with a default of acs-1. This environment contains:

• Python 3 (version 3.8 on JetPack, 3.10 on Ubuntu 22.04 LTS),
• JupyterLab,
• PyTorch,
• torchaudio,
• cuSignal,
• R (this is the version packaged by Mamba, currenty version 4.2.3 on JetPack, 4.3.1 on Ubuntu 22.04 LTS),
• R package development tools,
• the reticulate R - Python interface library,
• R audio libraries as described in Sound Analysis and Synthesis with R, and
• the Quarto scientific and technical publishing system.

PyTorch and cuSignal are optimized for NVIDIA GPUs, enabling a variety of digital signal processing and artificial intelligence applications, including the emerging field of differentiable digital signal processing.

The two main textbooks I recommend if you’re just getting started with the software deployed in AlgoCompSynth-One are Sueur (2018) for the R audio packages and Müller (2021) for the Python ones.

How does it work?

AlgoCompSynth-One is a collection of installers. This has a number of advantages:

• The field is moving rapidly, so you can always get the latest tested software,
• It avoids licensing issues with distributing binaries, and
• The source code is right there for you to examine and enhance.

The downside is that the components that need to be compiled on the Jetson, torchaudio, cupy, and cuSignal, take a fair amount of time to build. I have included logfiles of my builds on a Jetson AGX Xavier so you can get an idea of what to expect for build times.

The builds only need to be done once unless you want to change versions. On a Jetson, the Python wheels downloaded or built are cached, and the install scripts will look for those first rather than doing a new build.

Who is it for?

At the moment, it’s primarily for developers who need GPU speed for digital signal processing or artificial intelligence. That’s a fairly narrow audience given the kind of power today’s CPUs have, but there are a number of audio analysis and synthesis approaches that can use GPU power.

How do I get started?

The short version is:

1. Get a Jetson Developer Kit (Xavier or later) and install Jetpack 5.1.1 or later,

   or

   get a Windows 11 machine with an NVIDIA GPU capable of supporting CUDA 11.8.

2. Clone this repository.

3. cd into JetPack5 (Jetson) or WSL-jammy (Windows Subsystem for Linux Ubuntu 22.04 LTS) and follow the instructions on the README.md there.

Roadmap

Now that the WSL port is done, I am stopping enhancements on the JetPack version. The original motivation for this project was to build powerful synthesizers on an inexpensive platform, the NVIDIA Jetson Nano. That was before COVID-19, supply chain disruptions, and NVIDIA deploying the Orin modules.

The Orin series devices are no doubt excellent, but they are intended for applications at industrial scale and budgets. A hobbyist like myself can now get better performance / dollar and payoff / effort ratios with a 40 series laptop or desktop than with a Jetson Orin or Xavier device.

I plan to track cuSignal and PyTorch releases, and ensure that my R music packages all work on the Jetson. But future enhancements will only happen on the WSL version.

Finally, a word about other Linux distros. I believe this platform will work out of the box on any x86_64 Ubuntu 22.04 LTS machine with a CUDA 11.8 compatible GPU, but I don’t have one available to test. For other distros, I don’t have the time to test them, but it should work if CUDA 11.8 and Mambaforge do, and you rewrite WSL-jammy/00linux-dependencies.sh to install the required packages from the distro.

References