software/libmapper/readme

This library is a system for representing input and output signals on a network and allowing arbitrary "mappings" to be dynamically created between them.

A "mapping" consists of an Open Sound Control stream being established between a source and a destination signal--the source is translated to the destination's expected format, with some mathematical expression used to condition the transmitted values as desired. This can be used for example to connect a set of sensors to a synthesizer's input parameters.

To get started quickly with libmapper, be sure to read the tutorial, found in the "doc" folder in this distribution.

History of the mapper project

This project began life as a tool for helping composers and performers to more easily play with custom-built musical instruments, such that a program that reads data from the instrument (e.g. from an Arduino embedded microcontroller) can be connected to the inputs of a sound synthesizer. The first version of this software was written entirely in Cycling 74's Max/MSP, but in order to make it more universally useful and cross-platform, we decided to re-implement the protocol in C; hence, libmapper.

We were already using Open Sound Control for this purpose, but needed a way to dynamically change the mappings--not only to modify the destinations of OSC messages, but also to change scaling, or to introduce derivatives, integration, frequency filtering, smoothing, etc. Eventually this grew into the use of a multicast bus to publish signal metadata and administrate peer-to-peer connections between machines on a local subnet, with the ability to specify arbitrary mathematical expressions that can be used to perform signal conditioning.

It has been designed with the idea of decentralization in mind. We specifically wanted to avoid keeping a "master" node which is responsible for arbitrating connections or republishing data, and we also wanted to avoid needing to keep a central database of serial numbers or other unique identifiers for devices produced in the lab. Instead, common communication is performed on a multicast UDP port, that all nodes listen on, and this is used to implement a collision-handling protocol that can assign a unique ID to each device that appears. Actual signal data is sent directly from a sender device to the receiver's IP and UDP port.

Advantages of libmapper

This library is, on one level, our response to the widely acknowledged lack of a "query" protocol for Open Sound Control. We are aware for example of current work on using the ZeroConf protocol for publishing OSC services, and various efforts to standardize a certain set of OSC address conventions.

However, libmapper covers considerably more ground than simply the ability to publish namespaces, and handles a direct need that we experience in our daily work. We evaluated the use of ZeroConf early on in this work, but we eventually decided that sending OSC messages over multicast was just as easy and required the integration of fewer dependencies. With the addition of being able to control message routing and to dynamically specify signal conditioning equations, we feel this work represents a significant enough effort that it is worth making available to the general public.

It can also be seen as an open source alternative to some commercial products such as Camille Troillard's Osculator and STEIM's Junxion, albeit certainly more "barebones" for the moment.

A major difference in libmapper's approach to handling devices is the idea that each "driver" can be a separate process, on the same or different machines in the network. By creating a C library with a basic interface and by providing SWIG bindings, we hope to enable a very wide variety of programming language bindings, to allow any choice of development strategy.^[At this time, the SWIG bindings only work for Python. Support for more languages, particularly Java, is in the works.] Another advantage of a C library is portability: we have demonstrated libmapper working on a Gumstix device, an ethernet-enabled ARM-based microcomputer running Linux that can be easily embedded directly into an instrument.

We also provide an "external" Max/MSP and PureData object, to help integrate libmapper support into programs created in these popular dataflow languages.

Known limitations

The "devices and signals" metaphor encapsulated by libmapper is of course not the be-all and end-all of mapping. We assume homogeneous vectors of floating-point or integer numbers, while in reality more complex structured data might be useful to transmit.

Particularly if libmapper is to be used for visual art, transmission of matrixes or even image formats might be interesting, although there is no standard way to represent this in Open Sound Control other than as a binary "blob". It's possible that in this case a protocol such as HTTP that supports MIME type headers is simply a more appropriate choice.

That aside, mapping techniques based on table-lookup or other data-oriented processes don't fit directly into the "connection" paradigm used here. Of course, many of these techniques can easily be implemented as "intermediate" devices that sit between a sender and receiver. Another category of mappings that can currently only be solved this way include many-to-one mappings--libmapper currently has no special handling of multiple devices sending to one receiver, and the values are simply interleaved, whereas what is intended is probably some combining function such as addition, multiplication, or thresholding. It will be possible to solve this latter issue by using some form of receiver-side routers which can handle expressions containing multiple input variables (See future plans below). Many-to-one mapping functionality was not an immediate priority, since we feel that in practise it is actually quite rare to have a case where such a combining function should be arbitrarily modifiable by a user. In the real world, dependencies between signals often have a semantic significance which would be better handled internally to the sender or the receiver rather than in the mapper layer. Nonetheless, we hope to tackle this problem eventually.

One impact of peer-to-peer messaging is that it may suffer from redundancy. In general it may be more efficient to send all data once to a central node, and then out once more to nodes that request it, at the expense of weighing down the bandwidth of that particular central node. In libmapper's case, if 50 nodes subscribe to a particular signal, it will be repeated that many times by the originating node. Dealing with centralized-vs-decentralized efficiency issues by automatically optimizing decisions on how messages are distributed and where routing takes place is not impossible, but represents non-trivial work--for example, in libmapper the concept of a "router" is actually an independent node that a device sends messages to, which then translates and rebroadcasts; it just happens that the router is embedded in the sender because that was the most efficient place to have it in our scenario. In any case, given that optimal message passing efficiency is not a near-term goal, it goes without saying that your choice of using libmapper should depend on what type of network you envision: if you need a large number of receivers for the same signal you might consider one of the many alternatives, such as the SenseWorld DataNetwork, a SuperCollider-based mapping framework using a central-server approach.

Future plans

We chose to use Open Sound Control for this because of its compatibility with a wide variety of audio and media software. It is a practical way to transmit arbitrary data in a well-defined binary format. In practice, since the actual connections are peer-to-peer, they could technically be implemented using a variety of protocols. High-frequency data for example might be better transmitted using the JACK Audio Connection Kit or something similar. In addition, on embedded devices, it might make sense to "transmit" data between sensors and synthesizers through shared memory, while still allowing the use of the libmapper admin protocol and GUIs to dynamically experiment with mapping.

Although we provide a basic cross-platform GUI using Qt, the most advanced user interface is still the Max/MSP implementation. Although this works extremely well, it is limited to platforms supported by Max/MSP. (In other words, not Linux.) There are currently plans to create a web-based front-end that will be cross-platform. Better ways of visualizing and interacting with the network are also an active research topic.

For implementing intermediate devices that "learn" mappings automatically, a useful feature would be to request snapshots of the states of all inputs on a destination device. This has already been implemented in the previous Max/MSP version, and we have examples of neural network and SVM intermediates, but this feature has not yet made it into the C version of libmapper.

Currently the only supported data types are 32-bit integers and 32-bit floating point. This covers most of our use cases, but we hope to expand this catalogue eventually.

In the syntax for mathematical expressions, we include a method for indexing previous values to allow basic filter construction. This is done by index, but we also implemented a syntax for accessing previous state according to time in seconds. This feature is not yet useable, but in the future interpolation will be performed to allow referencing of time-accurate values.

Currently any "filter" implemented by an expression may suffer somewhat from jitter due to irregular timing in the sender, network delay, etc. We plan to tackle this problem by allowing timestamped signals using OSC bundles, which will likely require some changes to liblo, the OSC implementation used by libmapper.

As mentioned above, the ability to design many-to-one mapping connections will also be explored in future development of libmapper.

Getting libmapper

The latest version of libmapper source code and binaries can be downloaded from the IDMIL website at:

http://idmil.org/software/libmapper

License

This software was written in the Input Devices and Music Interaction Laboratory at McGill University in Montreal, and is copyright those found in the AUTHORS file. It is licensed under the GNU Lesser General Public License version 2.1 or later. Please see COPYING for details.

In accordance with the LGPL, you are allowed to use it in commercial products provided it remains dynamically linked, such that libmapper always remains free to modify. If you'd like to use it in an outstanding context, please contact the AUTHORS to seek an agreement.

Dependencies of libmapper are:

Optional dependencies for the Python bindings:

  • SWIG, GPL3, LGPL-compatible for generated code

  • Python, Python license, LGPL-compatible

Used only in the examples:

  • RtAudio, MIT license, LGPL-compatible