Previously when a pipeline is added to the cache it would never be
removed. If the application is generating a lot of unique pipelines
this can end up effectively leaking a large number of resources
including the GL program objects. Arguably this isn't really a problem
because if the application is generating that many unique pipelines
then it is doing something wrong anyway. It also implies that it will
be recompiling shaders very often so the cache leaking will likely be
the least of the problems.
This patch makes it keep track of which pipelines in the cache are in
use. The cache now returns a struct representing the entry instead of
directly returning the pipeline. This entry contains a usage counter
which the pipeline backends can use to mark when there is a pipeline
alive that is using the cache entry. When the hash table decides that
it's a good time to prune some entries, it will make a list of all of
the pipelines that are not in use and then remove the least recently
used half of the pipelines. That way it is less likely to remove
pipelines that the application is actually regenerating often even if
they aren't in use all of the time.
When the cache is pruned the hash table makes a note of how small the
cache could be if it removed all of the unused pipelines. The hash
table starts pruning when there are more entries than twice this
minimum expected size. The idea is that if that case it hit then the
hash table is more than half full of useless pipelines so the
application is generating lots of redundant pipelines and it is a good
time to remove them.
Reviewed-by: Robert Bragg <robert@linux.intel.com>
(cherry picked from commit c21aac22992bb7fef5a8d0913130b8245e67f2eb)
Conflicts:
cogl/driver/gl/cogl-pipeline-fragend-glsl.c
cogl/driver/gl/cogl-pipeline-progend-glsl.c
cogl/driver/gl/cogl-pipeline-vertend-glsl.c
cogl/driver/gl/gl/cogl-pipeline-fragend-arbfp.c
Previously on GLES2 where there is no builtin point size uniform then
we would always add a line to the vertex shader to write to the
builtin point size output because when generating the shader it is not
possible to determine if the pipeline will be used to draw points or
not. This patch changes it so that the default point size is 0.0f
which is documented to have undefined results when drawing points.
That way we can avoid adding the point size code to the shader in that
case. The assumption is that any application that is drawing points
will probably have explicitly set the point size on the pipeline
anyway so it is not a big deal to change the default size from 1.0f.
This adds a new pipeline state flag to track whether the point size is
non-zero. This needs to be its own state because altering it needs to
cause a different shader to be added to the pipeline cache. The state
flags that affect the vertex shader have been changed from a constant
to a runtime function because they will be different depending on
whether there is a builtin point size uniform.
There is also a unit test to ensure that changing the point size does
or doesn't generate a new shader depending on the values.
Reviewed-by: Robert Bragg <robert@linux.intel.com>
(cherry picked from commit b2eba06e16b587acbf5c57944a70ceccecb4f175)
Conflicts:
cogl/cogl-pipeline-private.h
cogl/cogl-pipeline-state-private.h
cogl/cogl-pipeline-state.c
cogl/cogl-pipeline.c
The pipeline cache contains three separate hash tables, one for the
state affecting the vertex shaders, one for the fragment shaders and
one for the resulting combined program. Previously these hash tables
had a fair bit of duplicated code to calculate the hashes, check for
equality and copy the pipeline when it is added. This patch moves the
common bits of code to a new type called CoglPipelineHashTable which
just wraps a GHashTable with a given set of state flags to use for
hashing and checking for equality.
Reviewed-by: Robert Bragg <robert@linux.intel.com>
(cherry picked from commit 402796430c839038339e531363b8c2463f9b2a9e)
Conflicts:
cogl/Makefile.am
The coding style has for a long time said to avoid using redundant glib
data types such as gint or gchar etc because we feel that they make the
code look unnecessarily foreign to developers coming from outside of the
Gnome developer community.
Note: When we tried to find the historical rationale for the types we
just found that they were apparently only added for consistent syntax
highlighting which didn't seem that compelling.
Up until now we have been continuing to use some of the platform
specific type such as gint{8,16,32,64} and gsize but this patch switches
us over to using the standard c99 equivalents instead so we can further
ensure that our code looks familiar to the widest range of C developers
who might potentially contribute to Cogl.
So instead of using the gint{8,16,32,64} and guint{8,16,32,64} types this
switches all Cogl code to instead use the int{8,16,32,64}_t and
uint{8,16,32,64}_t c99 types instead.
Instead of gsize we now use size_t
For now we are not going to use the c99 _Bool type and instead we have
introduced a new CoglBool type to use instead of gboolean.
Reviewed-by: Neil Roberts <neil@linux.intel.com>
(cherry picked from commit 5967dad2400d32ca6319cef6cb572e81bf2c15f0)
This adds two new public experimental functions for attaching
CoglSnippets to two hook points on a CoglPipeline:
void cogl_pipeline_add_vertex_hook (CoglPipeline *, CoglSnippet *)
void cogl_pipeline_add_fragment_hook (CoglPipeline *, CoglSnippet *)
The hooks are intended to be around the entire vertex or fragment
processing. That means the pre string in the snippet will be inserted
at the very top of the main function and the post function will be
inserted at the very end. The declarations get inserted in the global
scope.
The snippets are stored in two separate linked lists with a structure
containing an enum representing the hook point and a pointer to the
snippet. The lists are meant to be for hooks that affect the vertex
shader and fragment shader respectively. Although there are currently
only two hooks and the names match these two lists, the intention is
*not* that each new hook will be in a separate list. The separation of
the lists is just to make it easier to determine which shader needs to
be regenerated when a new snippet is added.
When a pipeline becomes the authority for either the vertex or
fragment snipper state, it simply copies the entire list from the
previous authority (although of course the shader snippet objects are
referenced instead of copied so it doesn't duplicate the source
strings).
Each string is inserted into its own block in the shader. This means
that each string has its own scope so it doesn't need to worry about
name collisions with variables in other snippets. However it does mean
that the pre and post strings can't share variables. It could be
possible to wrap both parts in one block and then wrap the actual
inner hook code in another block, however this would mean that any
further snippets within the outer snippet would be able to see those
variables. Perhaps something to consider would be to put each snippet
into its own function which calls another function between the pre and
post strings to do further processing.
The pipeline cache for generated programs was previously shared with
the fragment shader cache because the state that affects vertex
shaders was a subset of the state that affects fragment shaders. This
is no longer the case because there is a separate state mask for
vertex snippets so the program cache now has its own hash table.
Reviewed-by: Robert Bragg <robert@linux.intel.com>
It seems that cogl-context-private.h needs to be included before including
any of the pipeline-related stuff to avoid build errors on C89 compilers.
This is due to the recent cogl-pipeline decoupling, seems like.
Reviewed-by: Robert Bragg <robert@linux.intel.com>
The CoglPipelineCache is now extended to store templates for state
affecting vertex shaders and combined programs. The GLSL fragend,
vertend and progend now uses this to get cached shaders and a program.
When a new pipeline is created it will now get hashed three times if
the GLSL backends are in use (once for the fragend, once for the
vertend and once for the progend). Ideally we should add some way for
the progend to check its cache before the fragends and vertends are
checked so that it can bypass them entirely if it can find a cached
combined program.
The pipeline cache is now handled in CoglPipelineCache instead of
directly in the ARBfp fragend. The flags needed to hash a pipeline
should be exactly the same for the ARBfp and GLSL fragends so it's
convenient to share the code. The hash table now stores the actual
pipeline as the value instead of the private data so that the two
fragends can attach their data to it. That way it's possible to use
the same pipeline key with ancestors that are using different
fragends.
The hash table is created with g_hash_table_new_full to set a
destructor for the key and value and there is a destructor for
CoglPipelineCache that gets called when the CoglContext is
destroyed. That way we no longer leak the pipelines and shader state
when the context is desroyed.
