Re: Approximates cross-staff slurs in VerticalAxisGroup vertical-skylines. (issue 6498077)

[mtsolo]

On 2012/09/03 13:46:33, mike7 wrote:
> On 3 sept. 2012, at 07:07, mailto:address@hidden wrote:

> > On 2012/09/02 20:38:28, Keith wrote:
> >> On 2012/09/02 06:25:58, MikeSol wrote:
> >
> >>> It's not a copy of the original slur because it is using
> >>> pure heights and offsets.
> >
> >> I saw you interrogating SlurStub regarding its purity, but did not
> > notice that
> >> SlurStub took any different shape based on 'pure' estimates.
> >
> >> The SlurStubs in the regtest were shaped just like the printed
Slurs,
> > but now I
> >> see the difference in the Chopin prelude, with
> >
> >> \layout { \context { \Score
> >>   \override SlurStub #'color = #green
> >>   \override SlurStub #'transparent = ##f
> >>   \override Slur #'color = #darkgreen
> >>   \override PhrasingSlur #'color = #darkgreen }}
> >
> >> The SlurStubs are not sufficiently accurate to help, and they cost
me
> > time.
> >
> >> Chopin op45:   patch           'skylines'      2.16
> >>    real        0m21.666s       0m16.245s       0m12.862s
> >>    user        0m20.814s       0m15.573s       0m12.232s
> >>    sys         0m0.240s        0m0.254s        0m0.217s
> >
> > I'm working on a new version of the patch.  There is no way to
improve
> > accuracy of the curve, but the Y-offset is often wrong and that can
be
> > improved.
> >
> > The time hike seems awfully expensive - there must be a way to
optimize
> > it.  I'll post something that works when I have it and then we can
> > figure out how to optimize it.
> >
> > http://codereview.appspot.com/6498077/
> >
> > _______________________________________________
> > lilypond-devel mailing list
> > mailto:address@hidden
> > https://lists.gnu.org/mailman/listinfo/lilypond-devel

> New version posted.  I'm also getting a 25% increase on the Op. 45
test case.
> I am getting severe hikes across the board, though, even in scores
without cross
> staff slurs.  What's odd is that the hikes are happening in the
"Drawing
> systems" phase (I think), although I'm positive via prints to the
command line
> that the SlurStub's control points and vertical skylines are never
being
> calculated.

> Lemme know if you see something slipping through the cracks - gprof
shows a hike
> in stencil integral and skyline code for my patch, so it must be
generating
> extra skylines somewhere.

> Cheers,
> MS

After doing some digging, it looks like that even if not a single new
grob is created (meaning if I comment out all SlurStub stuff in the
engraver), the time increase is still there.  It is tough to see where
this is coming from using profiling tools, but it seems like the hike is
in the Drawing Systems phase.  I'm guessing that it may have to do with
all of the maybe_pure stuff.  If anyone has time to do profiling I'd
appreciate help on this patch - it leads to an appreciable improvement
in several scores I've checked out w/ cross-staff slurs, but it
obviously can't be put on a countdown until the performance hit gets
taken care of.

http://codereview.appspot.com/6498077/