This documents some aspects of the internals of GNU LilyPond. Some of this stuff comes from e-mail I wrote, some from e-mail others wrote, some are large comments taken away from the headers. This page may be a little incoherent. Unfortunately, it is also quite outdated. A more thorough and understandable document is in the works.
You should use doc++
to take a peek at the sources.
GNU LilyPond is a "multi-pass" system. The different passes have been created so that they do not depend on each other. In a later stage some parts may be moved into libraries, or seperate programs, or they might be integrated in larger systems.
No difficult algorithms. The .ly file is read, and converted to a list
of Scores
, which each contain Music
and paper/midi-definitions.
The music is walked through in time-order. The iterators which do the walking report Music to Translators which use this information to create elements, either MIDI or "visual" elements. The translators form a hierarchy; the ones for paper output are Engravers, for MIDI Performers.
The translators swallow Music (mostly atomic gobs called Requests), create elements, broadcast them to other translators on higher or same level in the hierarchy:
The stem of a voice A is broadcast to the staff which contains A, but not to the stems, beams and noteheads of a different voice (say B) or a different staff. The stem and noteheads of A are coupled, because the the Note_heads_engraver broadcasts its heads, and the Stem_engraver catches these.
The engraver which agrees to handle a request decides whether to to honor the request, ignore it, or merge it with other requests. Merging of requests is preferably done with other requests done by members of the same voicegroups (beams, brackets, stems). In this way you can put the voices of 2 instruments in a conductor's score so they make chords (the Beam requests of both instruments will be merged).
Breakable stuff (eg. clefs and bars) are copied into pre and postbreaks.
Some dependencies are resolved, such as the direction of stems, beams, and "horizontal" placement issues (the order of clefs, keys etc, placement of chords in multi-voice music),
The lines and horizontal positions of the columns are determined.
Through some magical interactions with Line_of_score and Super_elem (check out the source) the "lines" are produced.
All other spanners can figure across which lines they are spread. If applicable, they break themselves into pieces. After this, each piece (or, if there are no pieces, the original spanner itself) throws out any dependencies which are in the wrong line.
Some items and all spanners need computation after the Paper_column positions are determined. Examples: slurs, vertical positions of staffs.
Most information is stored in the form of a request. In music typesetting, the user might want to cram a lot more symbols on the paper than actually fits. To reflect this idea (the user asks more than we can do), the container for this data is called Request.
In a lot of other formats this would be called an 'Event'
Barcheck_req
Note_req
Rest_req
Span_req
Spanner
Beam_req
Dynamic
Dynamic
request carries a time, measured
from the start of its note.
In the previous section the idea of Request has been explained, but this only solves one half of the problem. The other half is deciding which requests should be honored, which should merged with other requests, and which should be ignored. Consider this input
\type Staff < % chord { \meter 2/4; [c8 c8] } {\meter 2/4; [e8 e8] } >
Both the cs and es are part of a staff (they are in the same Voice_group), so they should share meters, but the two [ ] pairs should be merged.
The judge in this "allocation" problem a set of brokers: the requests
are transmitted to so-called engravers which respond if they want to
accept a request eg, the Notehead_engraver
will accept
Note_req
s, and turn down Slur_req
s. If the Music_iterator
cannot find a engraver that wants the request, it is junked (with a
warning message).
After all requests have been either assigned, or junked, the Engraver
will process the requests (which usually means creating an Item
or Spanner
). If a Request_engraver
creates something, it
tells the enclosing context. If all items/spanners have been created,
then each Engraver is notified of any created Score_element, via a
broadcasting system.
c4
produces:
Note_request (duration 1/4) Stem_request (duration 1/4)
Note_request will be taken by a Notehead_engraver
, stem_request
will be taken by a Stem_beam_engraver
. Notehead_engraver
creates a Notehead
, Stem_beam_engraver
creates a
Stem
. Both announce this to the Staff_engraver. Staff_engraver
will tell Stem_beam_engraver
about the Notehead
, which
will add the Notehead
to the Stem
it just created.
To decide on merging, several engravers have been grouped. Please check init/engraver.ly.
The symbols that are printed, are generated by items and spanners (staff-elements). An item has one horizontal position, whereas a spanner spans several columns.
In music symbols depend on each other: the stems of a beam should
point in the same direction as the beam itself, so the stems of a beam
depend on the beam. In the same way do scripts depend on the direction
of the stem. To reflect this, LilyPond has the notion of dependency.
It works in the same fashion that make
uses to build programs:
before a stem is calculated, its dependencies (the beam) should be
calculated. Before a slur is calculated, its dependencies (stems,
noteheads) should be calculated.
So what is this PREBREAK and POSTBREAK stuff?
Let's take text as an example. In German some compound words change their spelling if they are broken: "backen" becomes "bak-ken". TeX has a mechanism to deal with this, you would define the spelling of "backen" in TeX in this way
\discretionary{bak-}{ken}{backen}
These 3 arguments are called "prebreak", "postbreak" and "nobreak" text.
The same problem exists when typesetting music. If a line of music is broken, the next line usually gets a clef. So in TeX terms, the clef is a postbreak. The same thing happens with meter signs: Normally the meter follows the bar. If a line is broken at that bar, the bar along with the meter stays on the "last" line, but the next line also gets a meter sign after the clef. Using the previous notation,
\discretionary{bar meter}{clef meter}{ bar meter }
In GNU Lilypond, we have the same concepts (and the same terminology). Each (nonrhythmic) symbol is typeset in a nonrhythmic column At a breakpoint, multiple symbols are printed; symbols to be printed if the line is not broken, symbols to appear on the previous line, and on the next line if it is broken.
Some terminology: I call a vertical group of symbols (notes) which start at the same time a "column". Each line of a score has notes in it, grouped in columns. The difference in starting time between those columns makes it possible to determine ideal distances between those columns.
Example:
time -----> cols: col1 col2 col3 col4 voice1 1 1 voice2 2 2 2 2 (1 is a whole note, 2 a half note.) time_difference (col1 , col2) = 0.5 wholes, time_difference (col1 , col3) = 1 wholes, time_difference (col2 , col3) = 0.5 wholes, etc.
these differences are translated into ideal distances
distance (col1,col2) = 10 pt distance (col1,col3) = 14.1 pt distance (col2,col3) = 10 pt etc.
as you can see, these distance are conflicting. So instead of satisfying all those ideals simultaneously, a compromise is sought.
This is Columbus' egg: GNU LilyPond attaches "springs" to each column-pair. each spring has an equilibrium-position which is equal to the above mentioned distance, so
spring (col1, col2) and spring (col2,col3) try to push column 1 and 3 away (to a distance of 20pt) from each other, whereas the spring between col 1 and col 3 tries to pull those two together (to a distance of 14.1 pt). The net result of this pushing and pulling is an equilibrium situation (the pushing cancels the pulling), which can be calculated as the solution of Quadratic program: it is the solution with minimum potential energy, for you physicists out there.
This algorithm for doing one line, gives a "badness" parameter for each line (the potential energy). Now one can use TeX's algorithm for making paragraphs (using this new version of "badness"): one should try to minimise the overall badness of a paragraph. GNU LilyPond also uses the concept of pre- and post-breaks.
(actually, it is a bit more complicated: each column also has a minimum distance to other columns, to prevent symbols from running into symbols of other columns.)
Please send GNU LilyPond questions and comments to gnu-music-discuss@gnu.org.
Please send comments on these web pages to
(address unknown),
send other FSF & GNU inquiries and questions to
Copyright (c) 1997, 1998, 1999 Han-Wen Nienhuys and Jan Nieuwenhuizen.
Verbatim copying and distribution of this entire article is permitted in any medium, provided this notice is preserved.