The R Project SVN R

\documentclass[a4paper]{article}
%\VignetteIndexEntry{Display Lists in grid}
%\VignettePackage{grid}
%\VignetteDepends{gridBase}
\newcommand{\grid}{{\tt grid}}
\newcommand{\gridBase}{{\tt gridBase}}
\newcommand{\lattice}{{\tt lattice}}
\newcommand{\R}{{\tt R}}
\setlength{\parindent}{0in}
\setlength{\parskip}{.1in}
\setlength{\textwidth}{140mm}
\setlength{\oddsidemargin}{10mm}

\title{Display Lists in \grid}
\author{Paul Murrell}

\begin{document}
\maketitle

<<echo=FALSE, results=hide>>=
library(graphics)
library(grid)
ps.options(pointsize=12)
options(width=60)

@
A display list is a record of drawing operations.  It is used
to redraw graphics output when a graphics window is resized,
when graphics output is copied from one device to another, and
when graphics output is edited (via {\tt grid.edit}).

There are two display lists that can be used when working with 
\grid{}.  \R{}'s graphics engine maintains a display list and
\grid{} maintains its own display list.  The former is maintained at
the C code level and records both base graphics output and 
\grid{} graphics output.  The latter is maintained at the 
R code level and only records \grid{} output.

In standard usage,
the graphics engine's display list is used to redraw when a window
is resized and when copying between devices.  \grid{}'s display list
is used for redrawing when editing \grid{} output.

There are two main problems with this standard usage:
\begin{enumerate}
\item The graphics engine display list only records graphics output;
none of the calculations leading up to producing the output are
recorded.  This particularly impacts on plots which perform calculations
based on the physical dimensions of the device -- an example is
the legend function which performs calculations in order to
arrange the elements of the legend.  The effect can be seen from 
any example which uses the legend function.  Try running
{\tt example(legend)} then resize the device (make it quite
tall and thin or quite wide and fat);  the legend will start 
to look pretty sick.

{\bf NOTE:} that this is a problem with the graphics engine
display list -- it is not specific to \grid{}.  In fact, much
of \grid{}'s behaviour is protected from this problem because
things like \grid{} units are ``declarative'' and will be reevaluated 
on each redraw.  However, there are situations where \grid{} output
can be afflicted, in particular, whenever the {\tt convertUnit()}
function (or one of its variants) is used (the help file
for {\tt convertUnit()} gives an example).  

A situation where this problem becomes very relevant for 
\grid{} output is when the \gridBase{} package is used.
This is a situation where lots of calculations are performed
in order to align base and grid output, but these calculations
are not recorded on the graphics engine display list, so
if the device is resized the output will become very yukky.

\item \grid{}'s display list does not record base graphics 
output\footnote{This is not quite true;  it is possible
to include base graphics output on the \grid{} display list
as we will see later.} so if both base and \grid{} output appear
on the same device then the result of editing will not redraw the 
base output.  The following code provides a simple example:

<<guts1, eval=FALSE>>=
plot(1:10)
par(new=TRUE)
grid.rect(width=0.5, height=0.5, gp=gpar(lwd=3), name="gr")

<<ex1, echo=FALSE, fig=TRUE, width=4, height=3, include=FALSE>>=
<<guts1>>
grid.rect(w=0.99, h=0.99, gp=gpar(lty="dashed"))

@
\includegraphics[width=4in, height=3in]{displaylist-ex1}

<<ex2, echo=FALSE, fig=TRUE, width=4, height=3, include=FALSE>>=
grid.rect(width=0.5, height=0.5, gp=gpar(col="red", lwd=3))
grid.rect(w=0.99, h=0.99, gp=gpar(lty="dashed"))

<<eval=FALSE>>=
grid.edit("gr", gp=gpar(col="red", lwd=3))

@
\includegraphics[width=4in, height=3in]{displaylist-ex2}

After the {\tt grid.edit}, the rectangle has been redrawn, but the
base plot has not.
\end{enumerate}

\section*{Avoiding the graphics engine display list}

The problems described above can all be avoided if \grid{} is able 
to control all redrawing of output.  There are several parts to this 
solution:
\begin{itemize}
\item It must be possible to avoid using the graphics engine 
display list.  This facility is now provided by the 
{\tt engine.display.list()} function.  The following code
will stop \grid{} output from being recorded on the 
graphics engine display list:

<<results=hide>>=
engine.display.list(FALSE)

@
\item \grid{} must be alerted whenever a device has been resized or
output is being
copied between devices.  This is handled in C code;  if the 
graphics engine display list is not being used then \grid{}
automatically redraws its own display list whenever a resize
or copy occurs.

\item Calculations must be recorded on \grid{}'s display list.
If graphics output is based on calculations which will not be consistent
when the device is resized (e.g., a call to {\tt convertUnit()},
or calls to the \gridBase{} functions),
then a device resize will produce inconsistent output unless
the calculations are recorded somehow on \grid{}'s display list.
This appears to be exactly the same problem as we had with
the graphics engine display list; the difference is that there {\it is}
a way to record calculations on \grid{}'s display list.

When a \grid{} grob is drawn the {\tt drawDetails} method
for that grob will be called;  if calculations are put within
a {\tt drawDetails} method, then the calculations will be
performed every time the grob is drawn.
\end{itemize}

All of 
this means that it is possible, for example, to use {\tt convertUnit()}
and have the result consistent across device resizese or 
copies\footnote{In 
each of the examples that follow, you should execute the example
code, resize the device to see any inconsistency, then close the
device before trying the next example.}.
This next piece of code is an example where the output becomes
inconsistent when the device is resized.  We specify a width for 
the rectangle in inches, but convert it (gratuitously) to
NPC coordinates -- when the device is resized, the NPC coordinates
will no longer correspond to 1''.

<<results=hide>>=
grid.rect(width=convertWidth(unit(1, "inches"), "npc"))

@
This second piece of code demonstrates that it is not enough simply
to use \grid{}'s display list;  the calculations are still not being
recorded so the inconsistency will still occur.

<<results=hide>>=
engine.display.list(FALSE)
grid.rect(width=convertWidth(unit(1, "inches"), "npc"))

@
Finally, the next piece of code demonstrates that, if we only
use \grid{}'s display list {\em and} we
place the calculations within a {\tt drawDetails} method, then
the output remains consistent across device resizes and 
copies\footnote{In order to avoid drawing the rectangle 
multiple times, we have draw the rectangle
 but not record it on the display list ({\tt recording=FALSE}; only
the {\tt "myrect"} grob is recorded on the display list).}.

<<results=hide>>=
engine.display.list(FALSE)
drawDetails.myrect <- function(x, x.wrapped, recording) {
  gr <- rectGrob(width=convertWidth(unit(1, "inches"), "npc"))
  grid.draw(gr, recording=FALSE)
}
grid.draw(grob(cl="myrect"))
 

@
A more impressive use of this facility can be obtained from an example
using 
\gridBase{} to combine base and \grid{} graphics output.
Here I replicate the last example from the ``gridBase'' vignette --
a set of base pie charts within \grid{} viewports within a base plot.
In this case, I can produce all of the grobs required in the normal 
manner -- their locations and sizes are not based on special 
calculations\footnote{The example is wrapped inside a check for
whether the {\tt gridBase} package is installed so that the code
will still ``run'' on systems without {\tt gridBase}.}.

<<results=hide>>=
if ("gridBase" %in% .packages(all.available=TRUE)) {
library(gridBase)

x <- c(0.88, 1.00, 0.67, 0.34)      
y <- c(0.87, 0.43, 0.04, 0.94)
z <- matrix(runif(4*2), ncol=2)

maxpiesize <- unit(1, "inches")
totals <- apply(z, 1, sum)
sizemult <- totals/max(totals)

gs <- segmentsGrob(x0=unit(c(rep(0, 4), x),
                        rep(c("npc", "native"), each=4)),
                x1=unit(c(x, x), rep("native", 8)),
                y0=unit(c(y, rep(0, 4)),
                        rep(c("native", "npc"), each=4)),
                y1=unit(c(y, y), rep("native", 8)),
                gp=gpar(lty="dashed", col="grey"))
gr <- rectGrob(gp=gpar(col="grey", fill="white", lty="dashed"))
}

@
What is important is that I place the calls to the \gridBase{} functions
within the  {\tt drawDetails} method so that they are performed
every time the grob is drawn {\em and} the calls to the base graphics
functions are in here too so that they are called for every redraw.

<<results=hide>>=
drawDetails.pieplot <- function(grob, grob.wrapped, recording) {
  plot(x, y, xlim=c(-0.2, 1.2), ylim=c(-0.2, 1.2), type="n")
  vps <- baseViewports()
  par(new=TRUE)
  pushViewport(vps$inner, vps$figure, vps$plot, recording=FALSE)
  grid.draw(grob$gs, recording=FALSE)
  for (i in 1:4) { 
    pushViewport(viewport(x=unit(x[i], "native"),
                           y=unit(y[i], "native"),
                           width=sizemult[i]*maxpiesize,
                           height=sizemult[i]*maxpiesize),
                 recording=FALSE)
    grid.draw(grob$gr, recording=FALSE)
    par(plt=gridPLT(), new=TRUE)
    pie(z[i,], radius=1, labels=rep("", 2))
    popViewport(recording=FALSE)
  }
  popViewport(3, recording=FALSE)
}

@
The ``pieplot'' is created by assembling the component grobs
into a collective grob of the appropriate class; the {\tt drawDetails}
method takes care of actually producing the output.
An important additional step here is that I turn off the graphics engine
display list for the base graphics also (otherwise I would
get multiple copies of the base graphics output).  

<<results=hide>>=
if ("gridBase" %in% .packages()) {
engine.display.list(FALSE)
dev.control(displaylist="inhibit")

grid.draw(grob(gs=gs, gr=gr, cl="pieplot"))
}

@

The output from this example can be resized safely;  \grid{} handles
all of the redrawing, 
 and performs all of the actions within the 
{\tt drawDetails} method for each redraw,
including redrawing the base graphics output!

As a final example, we will harness the \grid{} display list
purely to achieve consistency in base graphics output.
The following reproduces the last example from the 
{\tt legend()} help page, but produces output which can be
resized without the legend going crazy.

<<results=hide>>=
drawDetails.mylegend <- function(grob, grob.wrapped, recording) {
   x <- 0:64/64
   y <- sin(3*pi*x)
   plot(x, y, type="l", col="blue", main = "points with bg \& legend(*, pt.bg)")
   points(x, y, pch=21, bg="white")
   legend(.4,1, "sin(c x)", pch=21, pt.bg="white", lty=1, col = "blue")
}
engine.display.list(FALSE)
dev.control(displaylist="inhibit")
grid.draw(grob(cl="mylegend"))


@

\section*{Disadvantages to using \grid{} for all redraws}

The main problem here is that redrawing \grid{}'s display list
is slower than redrawing the graphics engine display list.
This is just the price that must be paid in order to obtain the
advantages of using \grid{}'s display list.

@
\end{document}