Hi Denis,
We need a lot of upgrades in this area, but they will come with some
engineering cost.
Note that the Ductus line widening code widens curves to curves, but
many (most?) line wideners flatten everything to do widening so they
only have to deal with algorithms that widen and join lines. I haven't
applied myself to the task of creating an algorithm that can do their
"direct curve widening" technique yet so I don't have a canned solution
for that, but it would make the results of
BasicStroke.createStrokedShape() much less verbose.
Also, for a practical matter the existing down-stream rendering classes
that take the output from these widening classes only deal with lines,
but if they took the curves directly then they might be able to do a
more efficient job of flattening the segments. Until then, the only
case where it would matter much for these classes to output curves would
be for BS.createStrokedShape() and even that would probably require some
more plumbing from the code that feeds these classes.
But, I want to raise visibility of these issues here because we should
not be blindly accepting flattening as the law when we consider upgrades
to this part of the code. It is currently an "easy way out" and as we
spend time on the code we should provide resistance to accepting that
philosophy by default...
...jim
Denis Lila wrote:
That's true.
Well, if we're worried about the generated paths being verbose
and taking long to process then the problem extends beyond just
drawing round end caps. As far as I can see, whenever a path is
drawn that doesn't consist only of straight lines (i.e. an ellipse),
a flattening path iterator is being used to feed Stroker. So all
the bezier curves are still broken down into tiny straight lines,
just not by Stroker itself.
So, my question is, given a bezier curve C and a number w, is
there a way of quickly computing the control points of two bezier
curves C1, C2 such that the stuff between C1 and C2 is the widened
path?
More formally: compute the control points of C1, C2, where
C1 = {(x,y) + N(x,y)*(w/2) | (x,y) in C}
C1 = {(x,y) - N(x,y)*(w/2) | (x,y) in C}, where N(x,y) is the normal
of C at (x,y).
If we could do this easily, then we can just make a new class that
outputs bezier curves that is similar in purpose to Stroker, but that
is used only when the output can handle bezier curves. This way, the
only use left for Stroker would be when anti-aliasing, and for
every thing else we wouldn't have to use a flattening path iterator.
Thanks,
Denis.
----- "Jim Graham" <james.gra...@oracle.com> wrote:
Hi Denis,
Consider the case of using BasicStroke.createStrokedShape(). How do
you
know how many pixels the resulting path will occupy? You can't reduce
to concrete samples if you don't know the transform.
So, for rendering, then you may be correct. But for cases where the
path is being asked for then beziers are the only responsible
solution...
...jim
Denis Lila wrote:
Hello Jim.
I thought about checking the output and changing the behaviour
depending on whether the output is a PC2D or a LineSink, but I
didn't
implement it because I thought the point was to get rid of the
sampling
at this stage. However, if performance is the issue, then I guess
I'll
start working on it.
Although, I wonder whether it is really worth it. I think most lines
drawn
won't be wider than about 5 pixels, which means that the current way
will
emit about 7 lines, so that's 14 coordinates. 2 bezier quarter
circles will
require 12 coordinates. In terms of storage, there isn't much
difference, and
for lines of width 4 or smaller the current method is more
efficient.
I'm also guessing that it's harder for the rasterizer to deal with
bezier
curves than with straight lines, so is it possible that replacing
the
3.14*lineWidth/2 lines generated by the current method with 2 bezier
quarter circles isn't worth it (for small lineWidths)?
Thanks,
Denis.
----- "Jim Graham" <james.gra...@oracle.com> wrote:
Sigh - that makes sense. One issue is that the resulting paths it
generates are much more "verbose" than they need to be. This would
generally mean that it takes far more storage than it would
otherwise
need - and it means that if the result needs to be transformed then
it
would take many more computations to transform each segment than
the
bezier.
So, perhaps it would be worth having it check the type of the
output
and
do either a bezier or a bunch of lines depending on if it is a PC2D
or
a
LineSink?
Also, it isn't really that difficult to for Renderer to include
its
own
Cubic/Quadratic flattening code, but it might involve more
calculations
than the round-cap code since it would have to be written for
arbitrary
beziers whereas if you know it is a quarter circle then it is
easier
to
know how far to subdivide... :-(
...jim
Denis Lila wrote:
So, I have been thinking about this, and I can't see a good
way to do it that wouldn't involve heavy changes to Pisces.
In order for Stroker to generate Bezier quarter circles, it would
have to implement a curveTo method, which means Stroker should
start implementing PathConsumer2D and instead of using a LineSink
output it would have to use a PathConsumer2D output (either that,
or
LineSink should include a curveTo method, but then there won't
really
be any difference between a LineSink and a PathConsumer2D. By the
way,
LineSink doesn't have any implemented methods, so why is it an
abstract
class as opposed to an interface?)
Stroker is used in 3 ways:
1. As an implementation of BasicStroke's createStrokedShape
method.
This
uses a Path2D object as output.
2. As a way of feeding a PathConsumer2D without calling
createStrokedShape
to generate an intermediate Shape. This uses a PathConsumer2D
output.
3. As a way of feeding lines to a Renderer object, which
generates
alpha
tiles used for anti-aliasing that are fed to a cache and
extracted
as needed
by an AATileGenerator. Obviously, Stroker's output here is a
Renderer.
1 and 2 aren't problems, because the underlying output objects
support
Bezier curves. 3, however, doesn't, and it seems like implementing
a
curveTo method for Renderer would be very difficult because the
way
it
generates alpha tiles is by scanning the drawn edges with
horizontal
scan lines, and for each scan line finding the x-intersections of
the scan
lines and the edges. Then it determines the alpha values (I'm not
too sure
how it does this).
In order to implement Bezier curves in Renderer, we would have to
have
a quick way of computing, for each scan line, all its
intersections
with
however many Bezier curves are being drawn.
I haven't given much thought to how this could be done, as I am
not
very
familiar with Bezier curves, but it doesn't seem easy enough to
justify
fixing such a small bug.
----- Original Message -----
From: "Jim Graham" <james.gra...@oracle.com>
To: "Denis Lila" <dl...@redhat.com>
Cc: 2d-dev@openjdk.java.net
Sent: Wednesday, June 9, 2010 7:42:33 PM GMT -05:00 US/Canada
Eastern
Subject: Re: [OpenJDK 2D-Dev] Fix for drawing round endcaps on
scaled lines.
I don't understand - why do we generate sample points based on
the
size
of the cap? Why not generate a pair of bezier quarter-circles
and
let
the rasterizer deal with sampling?
...jim
Denis Lila wrote:
Hello.
I think I have a fix for this bug:
http://icedtea.classpath.org/bugzilla/show_bug.cgi?id=506
Basically, the problem is that if there is a magnifying affine
transformation set on the graphics object and one tries to draw a
line
with small thickness and round end caps, the end caps appear
jagged.
This is because the computation of the length of the array that
contains the points on the "pen" with which the decoration is
drawn
does not take into account the size of the pen after the
magnification
of the affine transformation. So, for example, if the line length
was
set to 1, and the transformation was a scaling by 10, the
resulting
pen would have a diameter of 10, but only 3 pen points would be
computed (pi*untransformedLineWidth), so the end cap looks like a
triangle.
My fix computes an approximation of the circumference of the
transformed pen (which is an ellipse) and uses that as the number
of
points on the pen. The approximation is crude, but it is simple,
faster than alternatives
(http://en.wikipedia.org/wiki/Ellipse#Circumference), and I can
say
from observations that it works fairly well.
There is also icing on the cake, in the form of slight
improvements
in performance when the scaling is a zooming out. Example: if the
original line width was 100, but g2d.scale(0.1,0.1) was set, then
the
resulting line would have a width of 10, so only ~31 points are
necessary for the decoration to look like a circle, but without
this
patch, about 314 points are computed (and a line is emitted to
each
one of them).
I appreciate any feedback.
Regards,
Denis Lila.
