Commodore-Amiga Midi Driver
CAMD is an Amiga shared library which provides a general device driver
for MIDI data, so that applications can share MIDI data with each other in
real-time, and interface to MIDI hardware in a device-independent way.
The goals of CAMD are:
1. To encourage development of music software and synchronized
multimedia applications by providing a working driver to the public.
2. To enable music and multimedia applications to operate concurrently
and exchange MIDI data in real-time.
3. To improve on existing "freeware" drivers by enhancing performance,
by providing hooks for getting raw input, and by simplifying the interface
as much as possible.
Operating System Requirements
The CAMD library functions identically under 2.0 or 1.3. However, many
of the various utilities that come with CAMD require GadTools or other 2.0
features. One of the most important of these is the "MIDI Ports"
preferences editor, which is used to tell CAMD which hardware ports to
use. Note that if the preferences file is missing, CAMD will always fall
back on opening the Amiga's internal serial port. What this means is that
applications that run under 1.3 can use CAMD, but until a preferences
editor that works under 1.3 has been created the functionality will be
The MIDI System
The MIDI distribution system is based on the idea of "linkages"
(called MidiLinks) between applications. Each application or hardware
driver can establish linkages to other applications or hardware drivers.
Multiple links can be established to a single source, so that more than
one application can see the MIDI stream coming out of a hardware port or
application output. Similarly, more than one application can send a MIDI
stream to a single hardware port or application input. The ability to have
one application send data to another allows "pipelining" of the MIDI
stream, for example connecting an interactive composing program to a
sequencer and running both concurrently.
Note that there is no requirement that the data sent actually be valid
musical data -- it is possible for a pair of applications to set up a private
linkage, and communicate anything they want, as long as it follows the
syntactic rules of MIDI. However, it is suggested that such linkages be
hidden from the user (using a special bit which makes a linkage private),
since the eventually there will be a "patch editor" which will allow the
user to manipulate linkages without the application being aware of it.
Each MIDI application must create a MidiNode. This structure is used
as a central dispatch point for incoming and outgoing messages, and holds
all of the application-specific information, including:
-- location and size of input buffers.
-- the name of the application
-- the icon to be used for the application in the patch editor.
-- the address of the task to signal when messages are received, and
the signal bit to use.
Each MIDI message sent or received is contained in a MidiMsg
structure. This 8-byte structure contains a timestamp, the actual MIDI
bytes (up to 3) and a link number (so that applications which have several
input links can determine which one received the message). Note that since
the message is so small, the entire message is copied when MIDI data is
transferred, rather than passing pointers around.
How MIDI Data is received
MIDI applications can be either task-based or callback based. A
task-based application uses a signal to wait for incoming MIDI data. Once
the signal is received, the application can call GetMidi() to actually
look at what was received. All incoming messages are queued, and there is
a seperate queue for system exclusive messages (which can be quite long).
Each incoming MIDI event is both timestamped and marked with the linkage
number (settable by the application) from the link that it came in on.
Some people have questioned whether a task can respond fast enough to
incoming MIDI data to meet professional standards of timing accuracy. Our
experimentation has determined that a high-priority task (say, 30 or so),
can meet these reqiuirements, even when the disk drive is running.
However, if the application's handling of MIDI is very fast, it may be
better to use a callback. The callback occurs in the context of the
sender, so it is best to be quick so as not to slow down the sending task.
(Note that the sender will always be a task, and not an interrupt, since
the actual hardware drivers are serviced via a task) The callback is
invoked through a standard Hook structure. Using a callback avoids the
overhead of task switching, and can allow improved overall performance.
How MIDI Data is Sent
Sending MIDI data is very simple, mainly a matter of filling out a
MidiMsg structure and calling PutMidi(). Note that if the receive buffer
is full, then the function will fail, rather than waiting for the receive
buffer to empty.
For those of you not familier with MIDI, system exclusive messages
(called SysEx for short) are a kind of escape hatch in the MIDI spec which
allows developers to define their own messages. Unlike other MIDI events
which are limited to 3 bytes or less, SysEx messages can be any length. In
CAMD, SysEx messages are handled by placing the header of the message (the
first three bytes) in the regular receive queue as a MidiMsg, and placing
the full message in a seperate buffer. The receiver can look at the first
three bytes, and decide whether they want to read the rest by calling
GetSysEx() or throw it away by calling SkipSysEx();
Sending SysEx is done by calling the function PutSysEx().
To reduce the load on the system, MIDI data can be filtered so that
only useful data shows up in the application's input buffer. Each MidiLink
has a set of filter bits, can allow incoming messages to be ignored (not
placed in the receive queue). The first set of filter bits correspond to
the 16 MIDI channels. (For those of you unfamilier with MIDI, the low
nybble of the first MIDI byte contains the channel number). If the
incoming MIDI messages is on a channel which does not correspond to one of
the bits set in the filter word. the message is skipped. A second filter
is based on the type of the event, of which CAMD breaks up into 14
-- note on/off
-- program change
-- pitch bend
-- controller change MSB
-- controller change LSB
-- controller change boolean switch
-- controller change single byte
-- controller parameter change
-- undefined controllers
-- mode change messages
-- channel after touch
-- polyphonic after touch
-- system real-time messages (MIDI clock, MTC Quarter Frame)
-- system common messages (Start, Stop, etc)
-- system exclusive messages
In addition, there is a special filtering system for SysEx messages
which allows them to be filtered based on the first byte or the first
three bytes after the SysEx header byte. If the first byte only is used,
then three different filters can be specified. If the first three bytes
are used, then only one filter can be specified.
One of the nice thing about MidiLinks is that they can be established
even if the other application or hardware driver hasn't been loaded yet.
This is because a MidiLink does not connect directly to the other
application, but rather it connects to a "meeting place" or "rendevous
point" for MidiLinks called a Cluster. Each cluster is referred to by
name. For example, if I establish an output link to the cluster "foo", and
someone else establishes an input link to that same cluster, then any data
that my application sends to that link will be received by that other
application. If a third application creates an input link to "foo", then
it will also receive the data, whereas if another application creates an
output link to "foo" then it's MIDI data will be merged with mine, and
distributed to all the input links.
So the rules of clusters are as follows:
-- The first attempt to link to a cluster creates the cluster, and
the last link to leave deletes it.
-- Each sender link to a cluster is merged with all the other senders.
-- Each receiver link to a cluster gets a copy of what all the other
In addition, there are some other properties of clusters:
Participants: The library function MidiLinkConnected() can be used to
check a cluster to see if there are any linkages of the opposite type. For
example, a sender could check to see if anybody is listening or if they
are just talking to vacuum. Similarly, a receiver could check to see if
there are any senders. In addition, you can request to be notified (via
signal) whenever the participants in a cluster change. This feature is
primarily used by the hardware interface in the library itself -- it
allows a driver to be shut down (and freeing the hardware resources) when
there are no applications using it.
Cluster Comments: For purposes of building user interface to select
clusters, each link to a cluster can specify a "comment", up to 34
characters long, which describes what this cluster actually is. However,
since there can only be one comment for a cluster, the comment from the
first link is the one used.
One of the advantages of the cluster model is that applications can be
started up in any order and still work. The following are suggestions as
to how applications should handle linkages:
1. An application should allow the user to see a list of existing
clusters (which CAMD can provide), or allow the user to type in a new
cluster name. [Issue: Should the word "cluster" be used in UI?]
2. The application should save the current linkages either in the
applications "settings", or embedded in the document or performance file
(perhaps using an IFF chunk). When the application is restarted (or that
performance loaded or whatever) the application should then automatically
establish the links specified.
If every application does this, then it will be easy for the user to set
up the same configuration of linkages as they did last time, even if they
launch their applications in a different order. Even if the applications
are invoked via a script, the network of applications can come into
(Eventually we will want to have a "performance manager" which can
"snapshot" a whole network of apps, allowing the user to go back to that
"state" later. This might be done using ARexx).
Each incoming MIDI message may be timestamped, however, you must
supply a source of timing information (the recommended method is to use
RealTime.library, however many other timestamp sources are possible).
You can tell CAMD to use a particular timig source. The MidiNode
contains a pointer (of type LONG *), which may be pointed to the source of
timestamps. Whenever a MidiMsg is received, the longword that is pointed
to by this pointer is used as the current time, and copied into the
MidiMsg. Normally, what you would want to do is point this pointer at a
longword that was contually being updated. Note that in this fashion, your
application can have timestamps in any format it wants, since CAMD never
looks at the timestamp field once it is set.
One important point is that the timestamp is set at the time the
message is placed into the receiver's buffer. It would have been nice to
timestamp the messages at the interrupt time of the first MIDI status
byte, however this would have made the cluster model of distribution
impossible. Application which desire ultimate accuracy should probably
adjust the timestamp to compensate for the length of the MidiMsg.
Interfacing to Hardware
CAMD maintains a list of hardware drivers which are used to access
serial ports, DSP boards, or any other hardware interface. A preferences
file (ENV:sys/midi.prefs) is used to indicate which hardware drivers
should be loaded and which ports should be activated. Note that using file
notification, CAMD is capable of changing this setup at any time, even
while applications are running.
Hardware drivers live in the directory DEVS:midi. They can be created
by third-party developers, and are fairly simple.
CAMD maintains a task for each input MIDI stream. This task is
responsible for reading bytes from the hardware, parsing them into
MidiMsgs, and sending them to a cluster. The cluster name is specified in
the preferences file, on a per-port basis.
CAMD has a long and convoluted history. It was originally created at
Carnegie-Mellon university by Roger B. Dannenberg and Jean-Christophe
Dhellemmes. After that is was worked on by Bill Barton, and later by
Darius Taghavy, followed by Carolyn Scheppner. The final form of the
design was conceived by David Joiner (a.k.a. Talin) and implemented by Joe