I'm about to commit the attached (quite large) patch. It's all about
accelerating alpha compositing of images with and without the RENDER
extension, though it has random bug fixes scattered in it.
How fast? It's pretty hard to quantify because it depends a lot on:
- Whether RENDER is present or not
- Whether RENDER is hardware accelerated
- Whether the target drawable is in video ram or not
- System details
Pre-patch timings on my machine were (PIII 1.1ghz, no hw accel for
RENDER):
Dest in video RAM: 2.19 megapixels/s
Dest in system RAM: 2.15 megapixels/s
Current timings on my machine are:
Dest in video RAM (RENDER): 1.15 megapixels/s
Dest in system RAM (RENDER): 8.67 megapixels/s
Dest in video RAM (RENDER disabled): 3.01 megapixels/s
Dest in system RAM (RENDER disabled): 18.33 megapixels/s
Plans are under way to get the RENDER software speeds up.
Unfortunately this patch will break linux-fb and win32. The necessary
fixes are:
- Implement _gdk_windowing_get_bits_for_depth()
- Implement _gdk_image_new_for_depth()
- Implement the _copy_to_image() virtual function
for windows/pixmaps.
I want to make sure we have win32 fixed before we do 1.3.13 - I looked
at doing it myself, but there were a lot of GDI objects flying around,
and without a setup to compile and test, it didn't look very
attractive.
Regards,
Owen
Thu Jan 3 22:18:15 2002 Owen Taylor <otaylor redhat com>
* gdk/x11/gdkdrawable-x11.c gdk/x11/gdkprivate-x11.h
(_gdk_x11_have_render): Private function to tell if we have RENDER
extension.
* gdk/x11/gdkgc-x11.c (_gdk_x11_gc_get_fg_picture): Return
None if we don't have RENDER extension.
* gdk/x11/gdkpango-x11.c (gdk_pango_context_get): Don't
use Xft unless we have render extension.
* gdk/x11/gdkdrawable-x11.c (gdk_x11_drawable_get_picture):
Handle missing render extension.
* gdk/gdkdraw.c gdk/gdkdrawable.h gdk/gdkpixmap.c gdk/gdkwindow.c
gdk/gdkinternals.h: Add a private copy_to_image() virtual function
to the GdkDrawable vtable that extends get_image() to allow
copying onto existing images. Make the default implementation of
get_image() use this so that backends don't have to implement
both. Add private wrapper _gdk_drawable_copy_to_image().
* gdk/x11/gdkimage-x11.c gdk/x11/gdkprivate-x11.c
gdk/x11/gdkdrawable-x11.c (_gdk_x11_copy_to_image): Implement
copy_to_image() semantics, speed up by using ShmPixmaps and
XCopyArea when possible, XFlush() after ungrabbing the server,
generally redo the logic once again.
* gdk/gdkinternals.h gdk/x11/gdkimage-x11.c
_gdk_windowing_bits_per_depth(): Function to convert from depth to
bits-per-pixel. (We assume only one bpp per depth - X requires
this.)
* gdk/gdkinternals.h gdk/gdkrgb.c gdk/gdkimage.c: Move the GdkRGB
scratch image code into a generic _gdk_image_get_scratch() chunk
of code that we can use other places we need scratch images.
* gdk/gdkimage.c gdk/x11/gdkimage.c gdk/gdkinternals.h:
Add _gdk_image_new_for_depth() as the backend
to _gdk_image_new() to allowing creating images with
a depth and no visual.
* gdk/gdkpixbuf-drawable.c: Fix so that getting
parts of images not at 0,0 actually works.
* gdk/gdkdrawable.h gdk/gdkinternals.h gdk/gdkdraw.c
gdk/gdkwindow.c gdk/gdkpixmap.c gdk/gdkpixbuf-render.c:
- Add a new GdkDrawableClass vfunc _draw_pixbuf, and
_gdk_draw_pixbuf() [ will be made public later ], to allow
backends to accelerate drawing pixbufs.
- Move the implementation of gdk_pixbuf_render_to_drawable_alpha()
to be the default implementation.
- Update docs for gdk_pixbuf_render_to_drawable_alpha().
- Optimize the default implementation by using
_gdk_image_copy_to_pixmap() and scratch shared images, and
special casing the compositing.
* gdk/x11/gdkdrawable-x11.c: Accelerate _gdk_draw_pixbuf()
with alpha using the RENDER extension.
* gdk/gdkpixbuf-drawable.c (gdk_pixbuf_get_from_drawable):
Optimize by _gdk_image_copy_to_pixmap() and scratch images.
* tests/testrgb.c: Add test for speed of alpha composition,
reduce the number of iterations since alpha composition
can be a bit slow.
* gdk/x11/gdkimage-x11.c gdk/gdkprivate-x11.h (_gdk_x11_image_get_shm_pixmap):
Private function to get a ShmPixmap for an image, if possible.
Attachment:
gdk-draw.diff.gz
Description: Composite speedup patch