Using cogl_rectangle (and thus the journal) in
_cogl_add_path_to_stencil_buffer means we have to consider all the state
that the journal may change in case it may interfer with the direct GL calls
used. This has proven to be error prone and in this case the journal is an
unnecissary overhead. We now simply call glRectf instead of using
cogl_rectangle.
For small runs of text like icon labels, we can get better performance
going through the Cogl journal since text may then be batched together
with other geometry.
For larger runs of text though we still use VBOs since the cost of logging
the quads becomes too expensive, including the software transform which
isn't at all optimized at this point. VBOs also have the further advantage
of avoiding repeated validation of vertices by the driver and repeated
mapping of data into the GPU so long as the text doesn't change.
Currently the threshold is 100 vertices/25 quads. This number was plucked
out of thin air and should be tuned later.
With this change I see ~180% fps improvment for test-text. (x61s + i965 +
Mesa 7.6-devel)
We were missing the simplest test of all: are the two CoglHandles equal and
are the flush option flags for each material equal? This should improve
batching for some common cases.
Whenever we modify a material we call _cogl_material_pre_change_notify which
checks to see if the material is referenced by the journal and if so flushes
if before we modify the material.
Since the journal logs material colors directly into a vertex array (to
avoid us repeatedly calling glColor) then we know we never need to flush
the journal when material colors change.
Since most Clutter actors aren't much more than textured quads; flushing the
journal typically involves lots of 'change modelview; draw quad' sequences.
The amount of overhead involved in uploading a new modelview and queuing
that primitive is huge in comparison to simply transforming 4 vertices by
the current modelview when logging quads. (Note if your GPU supports HW
vertex transform, then it still does the projective and viewport transforms)
At the same time a --cogl-debug=disable-software-transform option has been
added for comparison and debugging.
This change allows typical pick scenes to be batched into a single draw call
and I'm seeing test-pick run over 200% faster with this. (i965 + Mesa
7.6-devel)
Enabling this option makes Cogl trace how the journal is managing to batch
your rectangles. The journal staggers how it emmits state to the GL driver
and the batches will normally get smaller for each stage, but ideally you
don't want to be in a situation where Cogl is only able to draw one quad per
modelview change and draw call.
E.g. this is a fairly ideal example:
BATCHING: journal len = 101
BATCHING: vbo offset batch len = 101
BATCHING: material batch len = 101
BATCHING: modelview batch len = 101
This isn't:
BATCHING: journal len = 1
BATCHING: vbo offset batch len = 1
BATCHING: material batch len = 1
BATCHING: modelview batch len = 1
BATCHING: journal len = 1
BATCHING: vbo offset batch len = 1
BATCHING: material batch len = 1
BATCHING: modelview batch len = 1
<repeat>
When this option is used Cogl will print a trace of all quads that get
logged into the journal, and a trace of quads as they get flushed.
If you are seeing a bug with the geometry being drawn by Cogl this may give
some clues by letting you sanity check the numbers being logged vs the
numbers being emitted.
For testing the VBO fallback paths it helps to be able to disable the
COGL_FEATURE_VBOS feature flag. When VBOs aren't available Cogl should use
client side malloc()'d buffers instead.
Previously we only used the Cogl matrix stack API for indirect contexts, but
it's too costly to keep on requesting modelview matrices from GL (for
logging in the journal) even for direct rendering.
I also experimented with a patch for mesa to improve performance and
discussed this with upstream, but we agreed to consider the GL matrix API
essentially deprecated. (For reference the GLES 2 and GL 3 specs have
removed the matrix APIs)
CoglColors shouldn't be compared using memcmp since they may contain
uninitialized padding bytes.
The prototype is also suitable for passing to g_hash_table_new as the
key_equal_func.
_cogl_pango_display_list_add_texture now uses this instead of memcmp.
We now put the color of materials into the vertex array used by the journal
instead of calling glColor() but the number of requests for the material
color were quite expensive so we have changed the material color to
internally be byte components instead of floats to avoid repeat conversions
and added _cogl_material_get_colorubv as a fast-path for the journal to
copy data into the vertex array.
To improve batching of geometry in the Cogl journal we need to avoid modifying
materials midscene.
Currently cogl_set_source_color and cogl_set_source_texture simply modify a
single shared material. In the future we can improve this so they use a pool
of materials that gets recycled as the journal is flushed, but for now we
give all ClutterRectangles their own private materials for painting with.
Currently cogl_set_source_color uses a single shared material which means
each actor that uses it causes the journal to flush if the color changes.
Until we improve cogl_set_source_color to use a pool of materials that can
be recycled as the journal is flushed we avoid mid-scene material changes by
giving all actors a private material instead.
The number of material layers enabled when logging a quad in the journal
determines the stride of the corresponding vertex data (since we need a set
of texture coordinates for each layer.) By padding data in the case where we
have only one layer we can avoid a change in stride if we are mixing single
and double layer primitives in a scene (e.g. relevent for a composite
manager that may use 2 layers for all shaped windows) Avoiding stride
changes means we can minimize calls to gl{Vertex,Color}Pointer when flushing
the journal.
Since we need to update the texcoord pointers when the actual number of
layers changes, this adds another batch_and_call() stage to deal with
glTexCoordPointer and enabling/disabling the client arrays.
Previously the journal was always flushed at the end of
_cogl_rectangles_with_multitexture_coords, (i.e. the end of any
cogl_rectangle* calls) but now we have broadened the potential for batching
geometry. In ideal circumstances we will only flush once per scene.
In summary the journal works like this:
When you use any of the cogl_rectangle* APIs then nothing is emitted to the
GPU at this point, we just log one or more quads into the journal. A
journal entry consists of the quad coordinates, an associated material
reference, and a modelview matrix. Ideally the journal only gets flushed
once at the end of a scene, but in fact there are things to consider that
may cause unwanted flushing, including:
- modifying materials mid-scene
This is because each quad in the journal has an associated material
reference (i.e. not copy), so if you try and modify a material that is
already referenced in the journal we force a flush first)
NOTE: For now this means you should avoid using cogl_set_source_color()
since that currently uses a single shared material. Later we
should change it to use a pool of materials that is recycled
when the journal is flushed.
- modifying any state that isn't currently logged, such as depth, fog and
backface culling enables.
The first thing that happens when flushing, is to upload all the vertex data
associated with the journal into a single VBO.
We then go through a process of splitting up the journal into batches that
have compatible state so they can be emitted to the GPU together. This is
currently broken up into 3 levels so we can stagger the state changes:
1) we break the journal up according to changes in the number of material layers
associated with logged quads. The number of layers in a material determines
the stride of the associated vertices, so we have to update our vertex
array offsets at this level. (i.e. calling gl{Vertex,Color},Pointer etc)
2) we further split batches up according to material compatability. (e.g.
materials with different textures) We flush material state at this level.
3) Finally we split batches up according to modelview changes. At this level
we update the modelview matrix and actually emit the actual draw command.
This commit is largely about putting the initial design in-place; this will be
followed by other changes that take advantage of the extended batching.
Removed the G_PARAM_CONSTRUCT from the property registration of
"load-async" and "load-data-async". It made it impossible to use only
load-data-async, as the async loading state would be unset when
load-async got set it's default FALSE value.
Use signed integers while combining window space clip rectangles, so we avoid
arithmatic errors later resulting in glScissor getting negative width and
height arguments.
Previously this was RGBA_8888. It souldn't really make a difference but for
consistency we expect almost all textures in use to have an internaly
premultiplied pixel format.
_cogl_texture_download_from_gl needs to create transient CoglBitmaps when
downloading sliced textures from GL, and then copies these as subregions
into the final target_bitmap. _cogl_texture_download_from_gl also supports
target_bitmaps with a different format to the source CoglTexture being
downloaded.
The problem was that in the case of slice textures we were always looking
at the format of the CoglTexture, not of the target_bitmap when setting
up the transient slice bitmap.
To allow for flushing of batched geometry within Cogl we can't support users
directly calling glReadPixels. glReadPixels is also awkward, not least
because it returns upside down image data.
All the unit tests have been swithed over and clutter_stage_read_pixels now
sits on top of this too.
I've found this is something I do quite often when debugging rendering
problems since its a simple way to wipe out lots of geometry and removes a
lot of unpredictable noise when logging geometry passing through the Cogl
journal.
We were calculating our vertex stride and allocating our vertex array
differently depending on whether the user passed TRUE for use_color or not.
The problem was that we were always writting color data to the array
regardless of use_color.
There was also a bug with _cogl_texture_sliced_polygon in that it was
writing byte color components but we were expecting float components. We
now use byte components in _cogl_multitexture_unsliced_polygon too and pass
GL_UNSIGNED_BYTE to glColorPointer.
test-cogl-tex-getset was assuming it was dealing with
COGL_PIXEL_FORMAT_RGBA_8888 but since merging the premultiplcation branch
the pixel format is actually COGL_PIXEL_FORMAT_RGBA_8888_PRE
Cogl already add similar defines but with the CLUTTER namespace
(CLUTTER_COGL_HAS_GL and CLUTTER_COGL_HAS_GLES). Let's just add two
similar defines with the COGL namespace. Removing the CLUTTER_COGL ones
could break applications silently for no real good reason.
While grepping through the public headers looking for invalid use of
private HAVE_* defines, I stumbled upon two out of sync comments. Yes
it's a very minor trivial change.
The clutter_actor_pick() function just emits the ::pick signal
on the actor. Nobody should be using it, since the paint() method
is already context sensitive and will result in a ::pick emission
by itself. The clutter_actor_pick() is just confusing things.
The Clutter API reference should have a section on how to port
applications from older version of Clutter to the new API.
The first guide deals on how to port animations created with
ClutterEffect to clutter_actor_animate().
Texture data is now in premultiplied format and the shader should
output a premultiplied color if the default blend mode is being
used. Shaders that directly manipulate the rgb values now
unpremultiply and premultiply again afterwards.
In the int-to-float switch for actor properties, the ::size-change signal
was moved to use floats instead of integers. Sub-pixel precision for image
size is meaningless, though, so we should revert it back to ints.
Fixes bug:
http://bugzilla.openedhand.com/show_bug.cgi?id=1659
Instead of using a specific function to check whether the X
server supports the XInput extension we can use the generic
Xlib function XQueryExtension(). This cuts down the extra
checks inside the configure.ac and simplifies the code inside
clutter_x11_register_xinput().
Currently, XInput support requires a function call. In order to
make it easier for people to test it, we can also add a command
line switch that moves the pointer device detection and handling
to XInput. This should ensure that, at least for people building
Clutter with --enable-xinput, applications can be easily migrated
and regressions can be caught.
Now that we can safely check for an uninitialized Clutter we
don't have side effects in calling one of the functions like
clutter_x11_enable_xinput(), which require to be called before
any other Clutter function.
The StageManager singleton instance is already kept around
by the clutter_stage_manager_get_default() function; there is
no need to have it inside the main Clutter context as well.
Instead of using _clutter_context_get_default() and checking the
is_initialized flag, we should use the newly added private function
that does not cause side effects, especially for functions that have
to be called before any other Clutter function.
