This is a complete overhaul of the data structures used to manage
CoglMaterial state.
We have these requirements that were aiming to meet:
(Note: the references to "renderlists" correspond to the effort to
support scenegraph level shuffling of Clutter actor primitives so we can
minimize GPU state changes)
Sparse State:
We wanted a design that allows sparse descriptions of state so it scales
well as we make CoglMaterial responsible for more and more state. It
needs to scale well in terms of memory usage and the cost of operations
we need to apply to materials such as comparing, copying and flushing
their state. I.e. we would rather have these things scale by the number
of real changes a material represents not by how much overall state
CoglMaterial becomes responsible for.
Cheap Copies:
As we add support for renderlists in Clutter we will need to be able to
get an immutable handle for a given material's current state so that we
can retain a record of a primitive with its associated material without
worrying that changes to the original material will invalidate that
record.
No more flush override options:
We want to get rid of the flush overrides mechanism we currently use to
deal with texture fallbacks, wrap mode changes and to handle the use of
highlevel CoglTextures that need to be resolved into lowlevel textures
before flushing the material state.
The flush options structure has been expanding in size and the structure
is logged with every journal entry so it is not an approach that scales
well at all. It also makes flushing material state that much more
complex.
Weak Materials:
Again for renderlists we need a way to create materials derived from
other materials but without the strict requirement that modifications to
the original material wont affect the derived ("weak") material. The
only requirement is that its possible to later check if the original
material has been changed.
A summary of the new design:
A CoglMaterial now basically represents a diff against its parent.
Each material has a single parent and a mask of state that it changes.
Each group of state (such as the blending state) has an "authority"
which is found by walking up from a given material through its ancestors
checking the difference mask until a match for that group is found.
There is only one root node to the graph of all materials, which is the
default material first created when Cogl is being initialized.
All the groups of state are divided into two types, such that
infrequently changed state belongs in a separate "BigState" structure
that is only allocated and attached to a material when necessary.
CoglMaterialLayers are another sparse structure. Like CoglMaterials they
represent a diff against their parent and all the layers are part of
another graph with the "default_layer_0" layer being the root node that
Cogl creates during initialization.
Copying a material is now basically just a case of slice allocating a
CoglMaterial, setting the parent to be the source being copied and
zeroing the mask of changes.
Flush overrides should now be handled by simply relying on the cheapness
of copying a material and making changes to it. (This will be done in a
follow on commit)
Weak material support will be added in a follow on commit.
This adds a _cogl_bind_gl_texture_transient function that should be used
instead of glBindTexture so we can have a consistent cache of the
textures bound to each texture unit so we can avoid some redundant
binding.
_cogl_atlas_texture_blit_begin binds a texture to use as the
destination and it expects it to stay bound until
_cogl_atlas_texture_end_blit is called. However there was a call to
_cogl_journal_flush directly after setting up the blit state which
could cause the wrong texture to be bound. This just moves the flush
to before the call to _cogl_atlas_texture_blit_begin.
This was breaking test-cogl-sub-texture.
1) Always flush when migrating textures out of an atlas because although
it's true that the original texture data will remain valid in the
original texture we can't assume that journal entries have resolved the
GL texture that will be used. This is only true if a layer0_override has
been used.
2) Don't flush at the point of creating a new atlas simply flush
immediately before reorganizing an atlas. This means we are now assuming
that we will never see recursion due to atlas textures being modified
during a journal flush. This means it's the responsibility of the
primitives code to _ensure_mipmaps for example not the responsibility of
_cogl_material_flush_gl_state.
The CoglAtlasTexture struct was not being freed in
_cogl_atlas_texture_free so there would be a small leak whenever a
texture was destroyed.
Thanks to Robert Bragg for spotting this.
GL supports setting different wrap modes for the s, t and r
coordinates so we should design the backend interface to support that
also. The r coordinate is not currently used by any of the backends
but we might as well have it to make life easier if we ever add
support for 3D textures.
http://bugzilla.openedhand.com/show_bug.cgi?id=2063
Add a return result from CoglTexture.transform_quad_coords_to_gl(),
so that we can properly determine the nature of repeats in
the face of GL_TEXTURE_RECTANGLE_ARB, where the returned
coordinates are not normalized.
The comment "We also work out whether any of the texture
coordinates are outside the range [0.0,1.0]. We need to do
this after calling transform_coords_to_gl in case the texture
backend is munging the coordinates (such as in the sub texture
backend)." is disregarded and removed, since it's actually
the virtual coordinates that determine whether we repeat,
not the GL coordinates.
Warnings about disregarded layers are used in all cases where
applicable, including for subtextures.
http://bugzilla.openedhand.com/show_bug.cgi?id=2016
Signed-off-by: Neil Roberts <neil@linux.intel.com>
Since using addresses that might change is something that finally
the FSF acknowledge as a plausible scenario (after changing address
twice), the license blurb in the source files should use the URI
for getting the license in case the library did not come with it.
Not that URIs cannot possibly change, but at least it's easier to
set up a redirection at the same place.
As a side note: this commit closes the oldes bug in Clutter's bug
report tool.
http://bugzilla.openedhand.com/show_bug.cgi?id=521
If FBOs aren't supported then it will end up very slow to reorganize
the atlas. Also currently the CoglTexture2D backend will refuse to
create any textures anyway so the full atlas texture won't be created.
cogl_texture_2d_new may fail in certain circumstances so
cogl_atlas_texture_reserve_space should detect this and also
fail. This will cause cogl_texture_new to fallback to a sliced
texture.
Thanks to Vladimir Ivakin for reporting this problem.
When the atlas is reorganised we could potentially be moving around
textures that are already referenced in the journal. We therefore need
to flush the journal otherwise they will be rendered with incorrect
texture coordinates. We also need to flush the journal even if we are
not reorganizing so that we can rely on the old texture contents
remaining in the atlas after migrating a texture out.
We've had complaints that our Cogl code/headers are a bit "special" so
this is a first pass at tidying things up by giving them some
consistency. These changes are all consistent with how new code in Cogl
is being written, but the style isn't consistently applied across all
code yet.
There are two parts to this patch; but since each one required a large
amount of effort to maintain tidy indenting it made sense to combine the
changes to reduce the time spent re indenting the same lines.
The first change is to use a consistent style for declaring function
prototypes in headers. Cogl headers now consistently use this style for
prototypes:
return_type
cogl_function_name (CoglType arg0,
CoglType arg1);
Not everyone likes this style, but it seems that most of the currently
active Cogl developers agree on it.
The second change is to constrain the use of redundant glib data types
in Cogl. Uses of gint, guint, gfloat, glong, gulong and gchar have all
been replaced with int, unsigned int, float, long, unsigned long and char
respectively. When talking about pixel data; use of guchar has been
replaced with guint8, otherwise unsigned char can be used.
The glib types that we continue to use for portability are gboolean,
gint{8,16,32,64}, guint{8,16,32,64} and gsize.
The general intention is that Cogl should look palatable to the widest
range of C programmers including those outside the Gnome community so
- especially for the public API - we want to minimize the number of
foreign looking typedefs.
The Cogl atlas code was using _cogl_texture_prepare_for_upload with a
NULL pointer for the dst_bmp to determine the internal format of the
texture without converting the bitmap. It needs to do this to decide
whether the texture will go in the atlas before wasting time on the
conversion. This use of the function is a little confusing so that
part of it has been split out into a new function called
_cogl_texture_determine_internal_format. The code to decide whether a
premult conversion is needed has also been split out.
Cogl accepts a pixel format for both the data in memory and the
internal format to be used for the texture. If they do not match then
it would convert them using the CoglBitmap functions before uploading
the data. However, GL also lets you specify both formats so it makes
more sense to let GL do the conversion. The driver may need the
texture in a specific format so it may end up being converted anyway.
The cogl_texture_upload_data functions have been removed and replaced
with a single function to prepare the bitmap. This will only do the
premultiplication conversion because that is the only part that GL
can't do directly.
The premult part of _cogl_convert_premult has now been split out as
_cogl_convert_premult_status. _cogl_convert_premult has been renamed
to _cogl_convert_format to make it less confusing. The premult
conversion is now done in-place instead of copying the
buffer. Previously it was copying the buffer once for the format
conversion and then copying it again for the premult conversion. The
premult conversion never changes the size of the buffer so it's quite
easy to do in place. We can also use the separated out function
independently.
The internal format of the atlas texture is still set to the
appropriate format so Cogl will disable blending for textures that are
intended to be RGB. This should end up ignoring the alpha channel from
the texture in the atlas. This makes the code slightly easier to
maintain and should also improve the chances of batching.
Previously the atlas textures were being created with whatever format
the first sub texture is in. Only three formats are supported so this
only matters if the first texture is a premultiplied alpha
texture. Instead it now masks out the premultiplied bit so that the
textures are always either RGB_888 or RGBA_8888.
When uploading texture data it was just calling cogl_texture_set_data
on the large texture. This would attempt to convert the data to the
format of the large texture. All of the textures with alpha channels
are stored together regardless of whether they are premultiplied so
this was causing premultiplied textures to be unpremultiplied
again. It now just uploads the data ignoring the premult bit of the
format so that it only gets converted once.
The sub texture backend doesn't work well as a completely general
texture backend because for example when rendering with cogl_polygon
it needs to be able to tranform arbitrary texture coordinates without
reference to the other coordintes. This can't be done when the texture
coordinates are a multiple of one because sometimes the coordinate
should represent the left or top edge and sometimes it should
represent the bottom or top edge. For example if the s coordinates are
0 and 1 then 1 represents the right edge but if they are 1 and 2 then
1 represents the left edge.
Instead the sub-textures are now documented not to support coordinates
outside the range [0,1]. The coordinates for the sub-region are now
represented as integers as this helps avoid rounding issues. The
region can no longer be a super-region of the texture as this
simplifies the code quite a lot.
There are two new texture virtual functions:
transform_quad_coords_to_gl - This transforms two pairs of coordinates
representing a quad. It will return FALSE if the coordinates can
not be transformed. The sub texture backend uses this to detect
coordinates that require repeating which causes cogl-primitives
to use manual repeating.
ensure_non_quad_rendering - This is used in cogl_polygon and
cogl_vertex_buffer to inform the texture backend that
transform_quad_to_gl is going to be used. The atlas backend
migrates the texture out of the atlas when it hits this.
Mipmaps don't work very well in the current atlas because there is not
enough padding between the textures. If ensure_mipmaps is called it
will now create a new texture and migrate the atlased texture to
it. It will use the same blit mechanism as when migrating so it will
try to use an FBO for a fast blit. However if this is not possible it
will end up downloading the data for the entire atlas which is not
ideal.
When reorganizing the textures, we can avoid downloading the entire
texture data if we bind the source texture in a framebuffer object and
copy the destination using glCopyTexSubImage2D. This is also
implemented using a much faster path in Mesa.
Currently it is calling the GL framebuffer API directly but ideally it
would use the Cogl offscreen API. However there is no way to tell Cogl
not to create a stencil renderbuffer which seems like a waste in this
situation.
If FBOs are not available it will fallback to reading back the entire
texture data as before.
This adds an 'atlas' category to the COGL_DEBUG environment
variable. When enabled Cogl will display messages when textures are
added to the atlas and when the atlas is reorganized.
When space can't be found in the atlas for a new texture it will now
try to reorganize the atlas to make space. A new CoglAtlas is created
and all of the textures are readded in decreasing size order. If the
textures still don't fit then the size of the atlas is doubled until
either we find a space or we reach the texture size limits. If we
successfully find an organization that fits then all of the textures
will be migrated to a new texture. This involves copying the texture
data into CPU memory and then uploading it again. Potentially it could
eventually use a PBO or an FBO to transfer the image without going
through the CPU.
The algorithm for laying out the textures works a lot better if the
rectangles are added in order so we might eventually want some API for
creating multiple textures in one go to avoid reorganizing the atlas
as far as possible.
This adds a CoglAtlas type which is a data structure that keeps track
of unused sub rectangles of a larger rectangle. There is a new atlased
texture backend which uses this to put multiple textures into a single
larger texture.
Currently the atlas is always sized 256x256 and the textures are never
moved once they are put in. Eventually it needs to be able to
reorganise the atlas and grow it if necessary. It also needs to
migrate the textures out of the atlas if mipmaps are required.
