README for gtk+/perf
--------------------
This is a framework for testing performance in GTK+. For GTK+, being
performant does not only mean "paint widgets fast". It also means
things like the time needed to set up widgets, to map and draw a
window for the first time, and emitting/propagating signals.
The following is accurate as of 2006/Jun/14.
Background
----------
A widget's lifetime looks more or less like this:
1. Instantiation
2. Size request
3. Size allocate
5. Realize
4. Map
5. Expose
6. Destroy
Some of these stages are particularly interesting:
- Instantiation means creating the widget. This may be as simple as a
few malloc()s and setting some fields. It could also be a
complicated operation if the widget needs to contact an external
server to create itself, or if it needs to read data files.
- Size requisition is when GTK+ asks the widget, "how big do you want
to be on the screen"? This can be an expensive operation. The
widget has to measure its text, measure its icons (and thus load its
icons), and generally run through its internal layout code.
- Realization is when the widget creates its GDK resources, like its
GdkWindow and graphics contexts it may need. This could be
expensive if the widget needs to load data files for cursors or
backgrounds.
- Expose is when the widget gets repainted. This will happen many
times throughout the lifetime of the widget: every time you drag a
window on top of it, every time its data changes and it needs to
redraw, every time it gets resized.
GtkWidgetProfiler is a mechanism to let you get individual timings for
each of the stages in the lifetime of a widget. It also lets you run
some stages many times in a sequence, so that you can run a real
profiler and get an adequate number of samples. For example,
GtkWidgetProfiler lets you say "repaint this widget 1000 times".
Why is this not as simple as doing
start_timer ();
for (i = 0; i < 1000; i++) {
gtk_widget_queue_draw (widget);
while (gtk_events_pending ())
gtk_main_iteration ();
}
stop_timer ();
Huh?
Because X is an asynchronous window system. So, when you send the
"paint" commands, your program will regain control but it will take
some time for the X server to actually process those commands.
GtkWidgetProfiler has special code to wait for the X server and give
you accurate timings.
Using the framework
-------------------
Right now the framework is very simple; it just has utility functions
to time widget creation, mapping, exposure, and destruction. To run
such a test, you use the GtkWidgetProfiler object in
gtkwidgetprofiler.h.
The gtk_widget_profiler_profile_boot() function will emit the
"create-widget" signal so that you can create your widget for
testing. It will then take timings for the widget, and emit the
"report" signal as appropriate.
The "create-widget" signal:
The handler has this form:
GtkWidget *create_widget_callback (GtkWidgetProfiler *profiler,
gpointer user_data);
You need to create a widget in your handler, and return it. Do not
show the widget; the profiler will do that by itself at the right
time, and will actually complain if you show the widget.
The "report" signal:
This function will get called when the profiler wants to report that
it finished timing an important stage in the lifecycle of your
widget. The handler has this form:
void report_callback (GtkWidgetProfiler *profiler,
GtkWidgetProfilerReport report,
GtkWidget *widget,
gdouble elapsed,
gpointer user_data);
The "report" argument tells you what happened to your widget:
GTK_WIDGET_PROFILER_REPORT_CREATE. A timer gets started right
before the profiler emits the "create-widget" signal,, and it gets
stopped when your callback returns with the new widget. This
measures the time it takes to set up your widget, but not show it.
GTK_WIDGET_PROFILER_REPORT_MAP. A timer gets started right before
the profiler calls gtk_widget_show_all() on your widget, and it
gets stopped when the the widget has been mapped.
GTK_WIDGET_PROFILER_REPORT_EXPOSE. A timer gets started right before
the profiler starts waiting for GTK+ and the X server to finish
painting your widget, and it gets stopped when the widget is fully
painted to the screen.
GTK_WIDGET_PROFILER_REPORT_DESTROY. A timer gets started right
before the profiler calls gtk_widget_destroy() on your widget, and
it gets stopped when gtk_widget_destroy() returns.
As a very basic example of using GtkWidgetProfiler is this:
----------------------------------------------------------------------
#include <stdio.h>
#include <gtk/gtk.h>
#include "gtkwidgetprofiler.h"
static GtkWidget *
create_widget_cb (GtkWidgetProfiler *profiler, gpointer data)
{
GtkWidget *window;
window = gtk_window_new (GTK_WINDOW_TOPLEVEL);
/* ... fill the window with widgets, and don't show them ... */
return window;
}
static void
report_cb (GtkWidgetProfiler *profiler, GtkWidgetProfilerReport report, GtkWidget *widget, gdouble elapsed, gpointer data)
{
const char *type;
switch (report) {
case GTK_WIDGET_PROFILER_REPORT_CREATE:
type = "widget creation";
break;
case GTK_WIDGET_PROFILER_REPORT_MAP:
type = "widget map";
break;
case GTK_WIDGET_PROFILER_REPORT_EXPOSE:
type = "widget expose";
break;
case GTK_WIDGET_PROFILER_REPORT_DESTROY:
type = "widget destruction";
break;
default:
g_assert_not_reached ();
type = NULL;
}
fprintf (stderr, "%s: %g sec\n", type, elapsed);
if (report == GTK_WIDGET_PROFILER_REPORT_DESTROY)
fputs ("\n", stderr);
}
int
main (int argc, char **argv)
{
GtkWidgetProfiler *profiler;
gtk_init (&argc, &argv);
profiler = gtk_widget_profiler_new ();
g_signal_connect (profiler, "create-widget",
G_CALLBACK (create_widget_cb), NULL);
g_signal_connect (profiler, "report",
G_CALLBACK (report_cb), NULL);
gtk_widget_profiler_set_num_iterations (profiler, 100);
gtk_widget_profiler_profile_boot (profiler);
gtk_widget_profiler_profile_expose (profiler);
return 0;
}
----------------------------------------------------------------------
Getting meaningful results
--------------------------
Getting times for widget creation/mapping/exposing/destruction is
interesting, but how do you actually find the places that need
optimizing?
Why, you run the tests under a profiler, of course.
FIXME: document how to do this.
Feedback
--------
Please mail your feedback to Federico Mena-Quintero <federico@novell.com>.
This performance framework is a work in progress.