Archive for May, 2012

Handling Errors in ABC

May 24, 2012

When I started thinking I should write yesterday’s post and started testing the Ed Reavy ABC file, I quickly ran into one of the module’s limitations that I thought I should fix before claiming ABC was a useful program.

When I started trying the Reavy file, if the ABC file ran into an error in the ABC file, it would frequently just hang. Now, it was relatively easy to track down the erroneous ABC at that point, because the ABC Actions class prints all the title fields it finds as it finds them. You’d watch the list of tune names going by, and if it stopped the list before the tune file was done and just sat there, you knew something was wrong with that tune.

That’s okay for a toy program that only I play with, but it’s unacceptable for a serious program other people can use. I had a theory about what caused it: when the grammar hit something it didn’t understand, it started backtracking to try to find a correct way of interpreting the file. Since there was no correct way, that resulted in a painfully long attempt to repeatedly re-interpret the bad data.

So I asked how to stop backtracking grammar regexes on #perl6. They helpfully suggested changing the ABC grammar’s regex methods to token, because token suppresses backtracking. I tried just changing one method, and that didn’t help. I tried changing them all, and that failed some tests. I changed one method back, and suddenly all the tests passed, and the grammar no longer hung on ABC file errors!

That was a huge improvement, but it was still only marginally helpful at figuring out what the problem in ABC file was. Once I was done with yesterday’s post, I tried to figure out how to get better information about where the parsing problem was. I hit upon a simple method. I added a $.current-tune attribute to the ABC::Actions class. When the action for a tune title fires, it clears this attribute and sets it to the tune title:

    method header_field($/) {
        if $<header_field_name> eq "T" {
            $*ERR.say: "Parsing " ~ $<header_field_data>;
            $.current-tune = $<header_field_data> ~ "\n";
        make ~$<header_field_name> => ~$<header_field_data>;

When the action method for the bar token fires, it adds the string of the bar to $.current-tune. When the action method for line_of_music fires, it adds a newline. And when a fatal error occurs parsing an ABC file, it prints out this attribute, resulting in error messages that look like this:

Unhandled exception: Did not match ABC grammar: last tune understood:
 Silent The Lonely Glen
g/2f/2|ed/2B/2 AG/2A/2|B/2d^c/2 dg/2f/2|ed/2B/2 AG/2A/2|

While this isn’t an exact reproduction of what’s in the tune, it’s close enough to easily show where the problem is:

g/2f/2|ed/2B/2 AG/2A/2|B/2d^c/2 dg/2f/2|ed/2B/2 AG/2A/2|
B/2>A/2G/2/F/2 eg/2f/2|ed/2B/2 ba/2g/2|g/2f/2e/2^d/2|

You just look at where the grammar stopped understanding the file, and the error pops right out if you know ABC: B/2>A/2G/2/F/2. Delete that duration not attached to a note, and the grammar will zip right through the tune.

I’m sure there are more sophisticated ways to handle errors in parsing available in Perl 6, because the STD grammar uses them to give you great error messages. But this is a simple technique which is pretty handy.


About ABC

May 23, 2012

I’ve referenced the ABC module a number of times over the years on this blog, but I don’t think I’ve ever properly explained it. I think it’s particularly relevant with the recurring discussions going on about Perl 6’s “production readiness”, because at this point the script in the module is a productive tool I am using on a regular basis. One that would have been harder to write in almost any other programming language, I should add.

ABC format is a designed to make it very easy to enter single-line music. It’s designed to be a terse ASCII representation of the music, and it’s absolutely terrific for notating jigs and reels and things like that. Here’s an example ABC hornpipe:

T:Two Weeks To Wait
C:Solomon Foster
GF|:E2GB AGFA|(3Bcd ed gedB|A2GB ABGA|BgdB A2BG|
E2GB AGFA|(3Bcd ed gedB|A2GB ABGF|E2ED E2GF:|
|:E2GB edef|g2fg edBG|A2GB ABGA|BgdB e2ef|
g2fg edBG|~A3B AGEG|~A3B AGFG|E2ED E2GF:|

A tune considers of a header section with information about the tune and the basic glue to define the tune, followed by a list of notes, barlines, repeat signs, and other symbols.

The ABC module’s script uses the ABC module’s ABC library to read the file, and outputs the music in the Lilypond format. Lilypond produces very nice output, but the file format is a pain. For instance, the (3Bcd triplet in the above file is translated to this: \times 2/3 { b'8[ cis''8 d''8] }. But the PDF Lilypond outputs from it is beautiful.

I’ve been using this tool to handle my sheet music creation needs for about a year now, but in the last six weeks I’ve really stepped things up. I’ve added multi-note stems, chord names, text annotations, ties, slurs, and many other markings I needed, as well as fixing a number of bugs. As a test I threw the Ed Reavy ABC collection at it today. (After deleting the comment lines in the beginning, because I haven’t added them to the parser yet.) In the process I found four mistakes in the ABC file which threw the parser. Once they were fixed, it processed all 127 tunes in about 22 seconds. I haven’t examined the 44 pages of sheet music output by Lilypond closely yet, but it looks like all the problems I see with the output are actually errors (that parse) in the ABC file.

In conclusion, while it doesn’t implement the entire ABC standard yet, this is a perfectly usable ABC to Lilypond translator, capable of handling fairly complex single-staff ABC files quickly and accurately. It’s definitely reached the point where it’s more than reasonable for other people to start using it.


May 21, 2012

Shortly after I wrote my last blog post, I started actively working on the ABC module again.  First I had to enhance a few features so it could handle some whistle parts for a little musical I was asked to play for.  Next I needed some enhancements to handle whistle / fiddle parts for the standard wedding music pieces for an upcoming gig.  Finally getting obsessed with ABC updates, I decided to throw the code at a collection of 159 tunes I put together a few years ago.

Suddenly I was running into hard limitations of the ABC module.  It was always slow, but taking 40 seconds to process my file of three tunes wasn’t really a problem for me.   I did add some progress notification messages, and was shocked to learn it wasn’t the parsing of the ABC files that was slow, it was processing the parsed result.  But when I got things working well enough it was actually trying to process all 159 tunes, it would run for 15 or 20 minutes and then crash.  (This is under Niecza, BTW.)

I spent Friday and Saturday trying to optimize bits of code in there that I knew were terrible, with no noticeable improvement.  Finally I did a profile run under mono.  I found the results didn’t really mean anything to me, but I sent it to sorear++ to see what he thought.  And he said, “Why do you need to use .eval in”

Well, I quickly checked, and found all the uses were intended to convert strings like "3/4" to Rats.  You may say “Why were you doing that?  Shouldn’t .Numeric work just as well?”  And the answer is yes, it does.  (In my defense, I’m pretty sure that didn’t work when I started working on the ABC module.)

But now here is the tricky bit. Here’s a simple script to demonstrate the issue:

my $a = 0;
for 1..400 {
    $a += "1/$_";
say $a;

On my machine, that takes 1.3 seconds to run under Niecza. Now let’s make a simple change:

my $a = 0;
for 1..400 {
    $a += "1/$_".eval;
say $a;

This version, with just the .eval added, takes 39.0 seconds to run. Huge difference, eh?

It turned out that even though the evals were happening per measure of music and not per note, they were so slow they were completely swamping the performance of my script. With the evals, processing the file took about 28 seconds. Without them, 3.7 seconds. And remember the big ABC file that would run for 15 minutes and then croak? Without the evals, the entire file processes in 31.8 seconds. That’s easily 30 times faster.

Lesson? Don’t use eval unless you really, really need to.

This advice holds for Rakudo as well: the above fraction-summing scripts took .3 and 8.5 seconds to run, respectively. (Yes, Rakudo is doing a lot better performance-wise on those scripts than Niecza did!)