mutter/cogl/cogl-attribute.c

660 lines
21 KiB
C
Raw Normal View History

/*
* Cogl
*
* An object oriented GL/GLES Abstraction/Utility Layer
*
* Copyright (C) 2010 Intel Corporation.
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2 of the License, or (at your option) any later version.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with this library. If not, see
* <http://www.gnu.org/licenses/>.
*
*
*
* Authors:
* Robert Bragg <robert@linux.intel.com>
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "cogl-util.h"
#include "cogl-context-private.h"
#include "cogl-object-private.h"
#include "cogl-journal-private.h"
#include "cogl-attribute.h"
#include "cogl-attribute-private.h"
cogl: rename CoglMaterial -> CoglPipeline This applies an API naming change that's been deliberated over for a while now which is to rename CoglMaterial to CoglPipeline. For now the new pipeline API is marked as experimental and public headers continue to talk about materials not pipelines. The CoglMaterial API is now maintained in terms of the cogl_pipeline API internally. Currently this API is targeting Cogl 2.0 so we will have time to integrate it properly with other upcoming Cogl 2.0 work. The basic reasons for the rename are: - That the term "material" implies to many people that they are constrained to fragment processing; perhaps as some kind of high-level texture abstraction. - In Clutter they get exposed by ClutterTexture actors which may be re-inforcing this misconception. - When comparing how other frameworks use the term material, a material sometimes describes a multi-pass fragment processing technique which isn't the case in Cogl. - In code, "CoglPipeline" will hopefully be a much more self documenting summary of what these objects represent; a full GPU pipeline configuration including, for example, vertex processing, fragment processing and blending. - When considering the API documentation story, at some point we need a document introducing developers to how the "GPU pipeline" works so it should become intuitive that CoglPipeline maps back to that description of the GPU pipeline. - This is consistent in terminology and concept to OpenGL 4's new pipeline object which is a container for program objects. Note: The cogl-material.[ch] files have been renamed to cogl-material-compat.[ch] because otherwise git doesn't seem to treat the change as a moving the old cogl-material.c->cogl-pipeline.c and so we loose all our git-blame history.
2010-10-27 13:54:57 -04:00
#include "cogl-pipeline.h"
#include "cogl-pipeline-private.h"
#include "cogl-pipeline-opengl-private.h"
#include "cogl-texture-private.h"
#include "cogl-framebuffer-private.h"
#include "cogl-indices-private.h"
#ifdef COGL_PIPELINE_PROGEND_GLSL
#include "cogl-pipeline-progend-glsl-private.h"
#endif
Add internal _cogl_push_source to optionally disable legacy state Some code in Cogl such as when flushing a stencil clip assumes that it can push a temporary simple pipeline to reset to a known state for internal drawing operations. However this breaks down if the application has set any legacy state because that is set globally so it will also get applied to the internal pipeline. _cogl_draw_attributes already had an internal flag to disable applying the legacy state but I think this is quite awkward to use because not all places that push a pipeline draw the attribute buffers directly so it is difficult to pass the flag down through the layers. Conceptually the legacy state is meant to be like a layer on top of the purely pipeline-based state API so I think ideally we should have an internal function to push the source without the applying the legacy state. The legacy state can't be applied as the pipeline is pushed because the global state can be modified even after it is pushed. This patch adds a _cogl_push_source() function which takes an extra boolean flag to mark whether to enable the legacy state. The value of this flag is stored alongside the pipeline in the pipeline stack. Another new internal function called _cogl_get_enable_legacy_state queries whether the top entry in the pipeline stack has legacy state enabled. cogl-primitives and the vertex array drawing code now use this to determine whether to apply the legacy state when drawing. The COGL_DRAW_SKIP_LEGACY_STATE flag is now removed. Reviewed-by: Robert Bragg <robert@linux.intel.com>
2011-09-14 07:17:09 -04:00
#include "cogl-private.h"
#include <string.h>
#include <stdio.h>
#include <stdlib.h>
/* This isn't defined in the GLES headers */
#ifndef GL_UNSIGNED_INT
#define GL_UNSIGNED_INT 0x1405
#endif
static void _cogl_attribute_free (CoglAttribute *attribute);
COGL_OBJECT_DEFINE (Attribute, attribute);
static CoglBool
validate_cogl_attribute_name (const char *name,
char **real_attribute_name,
CoglAttributeNameID *name_id,
CoglBool *normalized,
int *layer_number)
{
name = name + 5; /* skip "cogl_" */
*normalized = FALSE;
*layer_number = 0;
if (strcmp (name, "position_in") == 0)
*name_id = COGL_ATTRIBUTE_NAME_ID_POSITION_ARRAY;
else if (strcmp (name, "color_in") == 0)
{
*name_id = COGL_ATTRIBUTE_NAME_ID_COLOR_ARRAY;
*normalized = TRUE;
}
else if (strcmp (name, "tex_coord_in") == 0)
{
*real_attribute_name = "cogl_tex_coord0_in";
*name_id = COGL_ATTRIBUTE_NAME_ID_TEXTURE_COORD_ARRAY;
}
else if (strncmp (name, "tex_coord", strlen ("tex_coord")) == 0)
{
char *endptr;
*layer_number = strtoul (name + 9, &endptr, 10);
if (strcmp (endptr, "_in") != 0)
{
g_warning ("Texture coordinate attributes should either be named "
"\"cogl_tex_coord_in\" or named with a texture unit index "
"like \"cogl_tex_coord2_in\"\n");
return FALSE;
}
*name_id = COGL_ATTRIBUTE_NAME_ID_TEXTURE_COORD_ARRAY;
}
else if (strcmp (name, "normal_in") == 0)
{
*name_id = COGL_ATTRIBUTE_NAME_ID_NORMAL_ARRAY;
*normalized = TRUE;
}
else
{
g_warning ("Unknown cogl_* attribute name cogl_%s\n", name);
return FALSE;
}
return TRUE;
}
CoglAttributeNameState *
_cogl_attribute_register_attribute_name (CoglContext *context,
const char *name)
{
CoglAttributeNameState *name_state = g_new (CoglAttributeNameState, 1);
int name_index = context->n_attribute_names++;
char *name_copy = g_strdup (name);
name_state->name = NULL;
name_state->name_index = name_index;
if (strncmp (name, "cogl_", 5) == 0)
{
if (!validate_cogl_attribute_name (name,
&name_state->name,
&name_state->name_id,
&name_state->normalized_default,
&name_state->layer_number))
goto error;
}
else
{
name_state->name_id = COGL_ATTRIBUTE_NAME_ID_CUSTOM_ARRAY;
name_state->normalized_default = FALSE;
name_state->layer_number = 0;
}
if (name_state->name == NULL)
name_state->name = name_copy;
g_hash_table_insert (context->attribute_name_states_hash,
name_copy, name_state);
if (G_UNLIKELY (context->attribute_name_index_map == NULL))
context->attribute_name_index_map =
g_array_new (FALSE, FALSE, sizeof (void *));
g_array_set_size (context->attribute_name_index_map, name_index + 1);
g_array_index (context->attribute_name_index_map,
CoglAttributeNameState *, name_index) = name_state;
return name_state;
error:
g_free (name_state);
return NULL;
}
static CoglBool
validate_n_components (const CoglAttributeNameState *name_state,
int n_components)
{
switch (name_state->name_id)
{
case COGL_ATTRIBUTE_NAME_ID_POSITION_ARRAY:
if (G_UNLIKELY (n_components == 1))
{
g_critical ("glVertexPointer doesn't allow 1 component vertex "
"positions so we currently only support \"cogl_vertex\" "
"attributes where n_components == 2, 3 or 4");
return FALSE;
}
break;
case COGL_ATTRIBUTE_NAME_ID_COLOR_ARRAY:
if (G_UNLIKELY (n_components != 3 && n_components != 4))
{
g_critical ("glColorPointer expects 3 or 4 component colors so we "
"currently only support \"cogl_color\" attributes where "
"n_components == 3 or 4");
return FALSE;
}
break;
case COGL_ATTRIBUTE_NAME_ID_TEXTURE_COORD_ARRAY:
break;
case COGL_ATTRIBUTE_NAME_ID_NORMAL_ARRAY:
if (G_UNLIKELY (n_components != 3))
{
g_critical ("glNormalPointer expects 3 component normals so we "
"currently only support \"cogl_normal\" attributes "
"where n_components == 3");
return FALSE;
}
break;
case COGL_ATTRIBUTE_NAME_ID_CUSTOM_ARRAY:
return TRUE;
}
return TRUE;
}
CoglAttribute *
cogl_attribute_new (CoglAttributeBuffer *attribute_buffer,
const char *name,
size_t stride,
size_t offset,
int n_components,
CoglAttributeType type)
{
CoglAttribute *attribute = g_slice_new (CoglAttribute);
CoglBuffer *buffer = COGL_BUFFER (attribute_buffer);
CoglContext *ctx = buffer->context;
attribute->is_buffered = TRUE;
attribute->name_state =
g_hash_table_lookup (ctx->attribute_name_states_hash, name);
if (!attribute->name_state)
{
CoglAttributeNameState *name_state =
_cogl_attribute_register_attribute_name (ctx, name);
if (!name_state)
goto error;
attribute->name_state = name_state;
}
attribute->d.buffered.attribute_buffer = cogl_object_ref (attribute_buffer);
attribute->d.buffered.stride = stride;
attribute->d.buffered.offset = offset;
attribute->d.buffered.n_components = n_components;
attribute->d.buffered.type = type;
attribute->immutable_ref = 0;
if (attribute->name_state->name_id != COGL_ATTRIBUTE_NAME_ID_CUSTOM_ARRAY)
{
if (!validate_n_components (attribute->name_state, n_components))
return NULL;
attribute->normalized =
attribute->name_state->normalized_default;
}
else
attribute->normalized = FALSE;
return _cogl_attribute_object_new (attribute);
error:
_cogl_attribute_free (attribute);
return NULL;
}
static CoglAttribute *
_cogl_attribute_new_const (CoglContext *context,
const char *name,
int n_components,
int n_columns,
CoglBool transpose,
const float *value)
{
CoglAttribute *attribute = g_slice_new (CoglAttribute);
attribute->name_state =
g_hash_table_lookup (context->attribute_name_states_hash, name);
if (!attribute->name_state)
{
CoglAttributeNameState *name_state =
_cogl_attribute_register_attribute_name (context, name);
if (!name_state)
goto error;
attribute->name_state = name_state;
}
if (!validate_n_components (attribute->name_state, n_components))
goto error;
attribute->is_buffered = FALSE;
attribute->normalized = FALSE;
attribute->d.constant.context = cogl_object_ref (context);
attribute->d.constant.boxed.v.array = NULL;
if (n_columns == 1)
{
_cogl_boxed_value_set_float (&attribute->d.constant.boxed,
n_components,
1,
value);
}
else
{
/* FIXME: Up until GL[ES] 3 only square matrices were supported
* and we don't currently expose non-square matrices in Cogl.
*/
_COGL_RETURN_VAL_IF_FAIL (n_columns == n_components, NULL);
_cogl_boxed_value_set_matrix (&attribute->d.constant.boxed,
n_columns,
1,
transpose,
value);
}
return _cogl_attribute_object_new (attribute);
error:
_cogl_attribute_free (attribute);
return NULL;
}
CoglAttribute *
cogl_attribute_new_const_1f (CoglContext *context,
const char *name,
float value)
{
return _cogl_attribute_new_const (context,
name,
1, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
&value);
}
CoglAttribute *
cogl_attribute_new_const_2fv (CoglContext *context,
const char *name,
const float *value)
{
return _cogl_attribute_new_const (context,
name,
2, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
value);
}
CoglAttribute *
cogl_attribute_new_const_3fv (CoglContext *context,
const char *name,
const float *value)
{
return _cogl_attribute_new_const (context,
name,
3, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
value);
}
CoglAttribute *
cogl_attribute_new_const_4fv (CoglContext *context,
const char *name,
const float *value)
{
return _cogl_attribute_new_const (context,
name,
4, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
value);
}
CoglAttribute *
cogl_attribute_new_const_2f (CoglContext *context,
const char *name,
float component0,
float component1)
{
float vec2[2] = { component0, component1 };
return _cogl_attribute_new_const (context,
name,
2, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
vec2);
}
CoglAttribute *
cogl_attribute_new_const_3f (CoglContext *context,
const char *name,
float component0,
float component1,
float component2)
{
float vec3[3] = { component0, component1, component2 };
return _cogl_attribute_new_const (context,
name,
3, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
vec3);
}
CoglAttribute *
cogl_attribute_new_const_4f (CoglContext *context,
const char *name,
float component0,
float component1,
float component2,
float component3)
{
float vec4[4] = { component0, component1, component2, component3 };
return _cogl_attribute_new_const (context,
name,
4, /* n_components */
1, /* 1 column vector */
FALSE, /* no transpose */
vec4);
}
CoglAttribute *
cogl_attribute_new_const_2x2fv (CoglContext *context,
const char *name,
const float *matrix2x2,
CoglBool transpose)
{
return _cogl_attribute_new_const (context,
name,
2, /* n_components */
2, /* 2 column vector */
FALSE, /* no transpose */
matrix2x2);
}
CoglAttribute *
cogl_attribute_new_const_3x3fv (CoglContext *context,
const char *name,
const float *matrix3x3,
CoglBool transpose)
{
return _cogl_attribute_new_const (context,
name,
3, /* n_components */
3, /* 3 column vector */
FALSE, /* no transpose */
matrix3x3);
}
CoglAttribute *
cogl_attribute_new_const_4x4fv (CoglContext *context,
const char *name,
const float *matrix4x4,
CoglBool transpose)
{
return _cogl_attribute_new_const (context,
name,
4, /* n_components */
4, /* 4 column vector */
FALSE, /* no transpose */
matrix4x4);
}
CoglBool
cogl_attribute_get_normalized (CoglAttribute *attribute)
{
_COGL_RETURN_VAL_IF_FAIL (cogl_is_attribute (attribute), FALSE);
return attribute->normalized;
}
static void
warn_about_midscene_changes (void)
{
static CoglBool seen = FALSE;
if (!seen)
{
g_warning ("Mid-scene modification of attributes has "
"undefined results\n");
seen = TRUE;
}
}
void
cogl_attribute_set_normalized (CoglAttribute *attribute,
CoglBool normalized)
{
_COGL_RETURN_IF_FAIL (cogl_is_attribute (attribute));
if (G_UNLIKELY (attribute->immutable_ref))
warn_about_midscene_changes ();
attribute->normalized = normalized;
}
CoglAttributeBuffer *
cogl_attribute_get_buffer (CoglAttribute *attribute)
{
_COGL_RETURN_VAL_IF_FAIL (cogl_is_attribute (attribute), NULL);
_COGL_RETURN_VAL_IF_FAIL (attribute->is_buffered, NULL);
return attribute->d.buffered.attribute_buffer;
}
void
cogl_attribute_set_buffer (CoglAttribute *attribute,
CoglAttributeBuffer *attribute_buffer)
{
_COGL_RETURN_IF_FAIL (cogl_is_attribute (attribute));
_COGL_RETURN_IF_FAIL (attribute->is_buffered);
if (G_UNLIKELY (attribute->immutable_ref))
warn_about_midscene_changes ();
cogl_object_ref (attribute_buffer);
cogl_object_unref (attribute->d.buffered.attribute_buffer);
attribute->d.buffered.attribute_buffer = attribute_buffer;
}
CoglAttribute *
_cogl_attribute_immutable_ref (CoglAttribute *attribute)
{
CoglBuffer *buffer = COGL_BUFFER (attribute->d.buffered.attribute_buffer);
_COGL_RETURN_VAL_IF_FAIL (cogl_is_attribute (attribute), NULL);
attribute->immutable_ref++;
_cogl_buffer_immutable_ref (buffer);
return attribute;
}
void
_cogl_attribute_immutable_unref (CoglAttribute *attribute)
{
CoglBuffer *buffer = COGL_BUFFER (attribute->d.buffered.attribute_buffer);
_COGL_RETURN_IF_FAIL (cogl_is_attribute (attribute));
_COGL_RETURN_IF_FAIL (attribute->immutable_ref > 0);
attribute->immutable_ref--;
_cogl_buffer_immutable_unref (buffer);
}
static void
_cogl_attribute_free (CoglAttribute *attribute)
{
if (attribute->is_buffered)
cogl_object_unref (attribute->d.buffered.attribute_buffer);
else
_cogl_boxed_value_destroy (&attribute->d.constant.boxed);
g_slice_free (CoglAttribute, attribute);
}
static CoglBool
cogl: rename CoglMaterial -> CoglPipeline This applies an API naming change that's been deliberated over for a while now which is to rename CoglMaterial to CoglPipeline. For now the new pipeline API is marked as experimental and public headers continue to talk about materials not pipelines. The CoglMaterial API is now maintained in terms of the cogl_pipeline API internally. Currently this API is targeting Cogl 2.0 so we will have time to integrate it properly with other upcoming Cogl 2.0 work. The basic reasons for the rename are: - That the term "material" implies to many people that they are constrained to fragment processing; perhaps as some kind of high-level texture abstraction. - In Clutter they get exposed by ClutterTexture actors which may be re-inforcing this misconception. - When comparing how other frameworks use the term material, a material sometimes describes a multi-pass fragment processing technique which isn't the case in Cogl. - In code, "CoglPipeline" will hopefully be a much more self documenting summary of what these objects represent; a full GPU pipeline configuration including, for example, vertex processing, fragment processing and blending. - When considering the API documentation story, at some point we need a document introducing developers to how the "GPU pipeline" works so it should become intuitive that CoglPipeline maps back to that description of the GPU pipeline. - This is consistent in terminology and concept to OpenGL 4's new pipeline object which is a container for program objects. Note: The cogl-material.[ch] files have been renamed to cogl-material-compat.[ch] because otherwise git doesn't seem to treat the change as a moving the old cogl-material.c->cogl-pipeline.c and so we loose all our git-blame history.
2010-10-27 13:54:57 -04:00
validate_layer_cb (CoglPipeline *pipeline,
int layer_index,
void *user_data)
{
Add a strong CoglTexture type to replace CoglHandle As part of the on going, incremental effort to purge the non type safe CoglHandle type from the Cogl API this patch tackles most of the CoglHandle uses relating to textures. We'd postponed making this change for quite a while because we wanted to have a clearer understanding of how we wanted to evolve the texture APIs towards Cogl 2.0 before exposing type safety here which would be difficult to change later since it would imply breaking APIs. The basic idea that we are steering towards now is that CoglTexture can be considered to be the most primitive interface we have for any object representing a texture. The texture interface would provide roughly these methods: cogl_texture_get_width cogl_texture_get_height cogl_texture_can_repeat cogl_texture_can_mipmap cogl_texture_generate_mipmap; cogl_texture_get_format cogl_texture_set_region cogl_texture_get_region Besides the texture interface we will then start to expose types corresponding to specific texture types: CoglTexture2D, CoglTexture3D, CoglTexture2DSliced, CoglSubTexture, CoglAtlasTexture and CoglTexturePixmapX11. We will then also expose an interface for the high-level texture types we have (such as CoglTexture2DSlice, CoglSubTexture and CoglAtlasTexture) called CoglMetaTexture. CoglMetaTexture is an additional interface that lets you iterate a virtual region of a meta texture and get mappings of primitive textures to sub-regions of that virtual region. Internally we already have this kind of abstraction for dealing with sliced texture, sub-textures and atlas textures in a consistent way, so this will just make that abstraction public. The aim here is to clarify that there is a difference between primitive textures (CoglTexture2D/3D) and some of the other high-level textures, and also enable developers to implement primitives that can support meta textures since they can only be used with the cogl_rectangle API currently. The thing that's not so clean-cut with this are the texture constructors we have currently; such as cogl_texture_new_from_file which no longer make sense when CoglTexture is considered to be an interface. These will basically just become convenient factory functions and it's just a bit unusual that they are within the cogl_texture namespace. It's worth noting here that all the texture type APIs will also have their own type specific constructors so these functions will only be used for the convenience of being able to create a texture without really wanting to know the details of what type of texture you need. Longer term for 2.0 we may come up with replacement names for these factory functions or the other thing we are considering is designing some asynchronous factory functions instead since it's so often detrimental to application performance to be blocked waiting for a texture to be uploaded to the GPU. Reviewed-by: Neil Roberts <neil@linux.intel.com>
2011-08-24 16:30:34 -04:00
CoglTexture *texture =
cogl_pipeline_get_layer_texture (pipeline, layer_index);
CoglFlushLayerState *state = user_data;
CoglBool status = TRUE;
/* invalid textures will be handled correctly in
cogl: rename CoglMaterial -> CoglPipeline This applies an API naming change that's been deliberated over for a while now which is to rename CoglMaterial to CoglPipeline. For now the new pipeline API is marked as experimental and public headers continue to talk about materials not pipelines. The CoglMaterial API is now maintained in terms of the cogl_pipeline API internally. Currently this API is targeting Cogl 2.0 so we will have time to integrate it properly with other upcoming Cogl 2.0 work. The basic reasons for the rename are: - That the term "material" implies to many people that they are constrained to fragment processing; perhaps as some kind of high-level texture abstraction. - In Clutter they get exposed by ClutterTexture actors which may be re-inforcing this misconception. - When comparing how other frameworks use the term material, a material sometimes describes a multi-pass fragment processing technique which isn't the case in Cogl. - In code, "CoglPipeline" will hopefully be a much more self documenting summary of what these objects represent; a full GPU pipeline configuration including, for example, vertex processing, fragment processing and blending. - When considering the API documentation story, at some point we need a document introducing developers to how the "GPU pipeline" works so it should become intuitive that CoglPipeline maps back to that description of the GPU pipeline. - This is consistent in terminology and concept to OpenGL 4's new pipeline object which is a container for program objects. Note: The cogl-material.[ch] files have been renamed to cogl-material-compat.[ch] because otherwise git doesn't seem to treat the change as a moving the old cogl-material.c->cogl-pipeline.c and so we loose all our git-blame history.
2010-10-27 13:54:57 -04:00
* _cogl_pipeline_flush_layers_gl_state */
Add a strong CoglTexture type to replace CoglHandle As part of the on going, incremental effort to purge the non type safe CoglHandle type from the Cogl API this patch tackles most of the CoglHandle uses relating to textures. We'd postponed making this change for quite a while because we wanted to have a clearer understanding of how we wanted to evolve the texture APIs towards Cogl 2.0 before exposing type safety here which would be difficult to change later since it would imply breaking APIs. The basic idea that we are steering towards now is that CoglTexture can be considered to be the most primitive interface we have for any object representing a texture. The texture interface would provide roughly these methods: cogl_texture_get_width cogl_texture_get_height cogl_texture_can_repeat cogl_texture_can_mipmap cogl_texture_generate_mipmap; cogl_texture_get_format cogl_texture_set_region cogl_texture_get_region Besides the texture interface we will then start to expose types corresponding to specific texture types: CoglTexture2D, CoglTexture3D, CoglTexture2DSliced, CoglSubTexture, CoglAtlasTexture and CoglTexturePixmapX11. We will then also expose an interface for the high-level texture types we have (such as CoglTexture2DSlice, CoglSubTexture and CoglAtlasTexture) called CoglMetaTexture. CoglMetaTexture is an additional interface that lets you iterate a virtual region of a meta texture and get mappings of primitive textures to sub-regions of that virtual region. Internally we already have this kind of abstraction for dealing with sliced texture, sub-textures and atlas textures in a consistent way, so this will just make that abstraction public. The aim here is to clarify that there is a difference between primitive textures (CoglTexture2D/3D) and some of the other high-level textures, and also enable developers to implement primitives that can support meta textures since they can only be used with the cogl_rectangle API currently. The thing that's not so clean-cut with this are the texture constructors we have currently; such as cogl_texture_new_from_file which no longer make sense when CoglTexture is considered to be an interface. These will basically just become convenient factory functions and it's just a bit unusual that they are within the cogl_texture namespace. It's worth noting here that all the texture type APIs will also have their own type specific constructors so these functions will only be used for the convenience of being able to create a texture without really wanting to know the details of what type of texture you need. Longer term for 2.0 we may come up with replacement names for these factory functions or the other thing we are considering is designing some asynchronous factory functions instead since it's so often detrimental to application performance to be blocked waiting for a texture to be uploaded to the GPU. Reviewed-by: Neil Roberts <neil@linux.intel.com>
2011-08-24 16:30:34 -04:00
if (texture == NULL)
goto validated;
_cogl_texture_flush_journal_rendering (texture);
/* Give the texture a chance to know that we're rendering
non-quad shaped primitives. If the texture is in an atlas it
will be migrated */
_cogl_texture_ensure_non_quad_rendering (texture);
/* We need to ensure the mipmaps are ready before deciding
* anything else about the texture because the texture storate
* could completely change if it needs to be migrated out of the
* atlas and will affect how we validate the layer.
*/
cogl: rename CoglMaterial -> CoglPipeline This applies an API naming change that's been deliberated over for a while now which is to rename CoglMaterial to CoglPipeline. For now the new pipeline API is marked as experimental and public headers continue to talk about materials not pipelines. The CoglMaterial API is now maintained in terms of the cogl_pipeline API internally. Currently this API is targeting Cogl 2.0 so we will have time to integrate it properly with other upcoming Cogl 2.0 work. The basic reasons for the rename are: - That the term "material" implies to many people that they are constrained to fragment processing; perhaps as some kind of high-level texture abstraction. - In Clutter they get exposed by ClutterTexture actors which may be re-inforcing this misconception. - When comparing how other frameworks use the term material, a material sometimes describes a multi-pass fragment processing technique which isn't the case in Cogl. - In code, "CoglPipeline" will hopefully be a much more self documenting summary of what these objects represent; a full GPU pipeline configuration including, for example, vertex processing, fragment processing and blending. - When considering the API documentation story, at some point we need a document introducing developers to how the "GPU pipeline" works so it should become intuitive that CoglPipeline maps back to that description of the GPU pipeline. - This is consistent in terminology and concept to OpenGL 4's new pipeline object which is a container for program objects. Note: The cogl-material.[ch] files have been renamed to cogl-material-compat.[ch] because otherwise git doesn't seem to treat the change as a moving the old cogl-material.c->cogl-pipeline.c and so we loose all our git-blame history.
2010-10-27 13:54:57 -04:00
_cogl_pipeline_pre_paint_for_layer (pipeline, layer_index);
if (!_cogl_texture_can_hardware_repeat (texture))
{
g_warning ("Disabling layer %d of the current source material, "
"because texturing with the vertex buffer API is not "
"currently supported using sliced textures, or textures "
"with waste\n", layer_index);
/* XXX: maybe we can add a mechanism for users to forcibly use
* textures with waste where it would be their responsability to use
* texture coords in the range [0,1] such that sampling outside isn't
* required. We can then use a texture matrix (or a modification of
* the users own matrix) to map 1 to the edge of the texture data.
*
* Potentially, given the same guarantee as above we could also
* support a single sliced layer too. We would have to redraw the
* vertices once for each layer, each time with a fiddled texture
* matrix.
*/
state->fallback_layers |= (1 << state->unit);
cogl: rename CoglMaterial -> CoglPipeline This applies an API naming change that's been deliberated over for a while now which is to rename CoglMaterial to CoglPipeline. For now the new pipeline API is marked as experimental and public headers continue to talk about materials not pipelines. The CoglMaterial API is now maintained in terms of the cogl_pipeline API internally. Currently this API is targeting Cogl 2.0 so we will have time to integrate it properly with other upcoming Cogl 2.0 work. The basic reasons for the rename are: - That the term "material" implies to many people that they are constrained to fragment processing; perhaps as some kind of high-level texture abstraction. - In Clutter they get exposed by ClutterTexture actors which may be re-inforcing this misconception. - When comparing how other frameworks use the term material, a material sometimes describes a multi-pass fragment processing technique which isn't the case in Cogl. - In code, "CoglPipeline" will hopefully be a much more self documenting summary of what these objects represent; a full GPU pipeline configuration including, for example, vertex processing, fragment processing and blending. - When considering the API documentation story, at some point we need a document introducing developers to how the "GPU pipeline" works so it should become intuitive that CoglPipeline maps back to that description of the GPU pipeline. - This is consistent in terminology and concept to OpenGL 4's new pipeline object which is a container for program objects. Note: The cogl-material.[ch] files have been renamed to cogl-material-compat.[ch] because otherwise git doesn't seem to treat the change as a moving the old cogl-material.c->cogl-pipeline.c and so we loose all our git-blame history.
2010-10-27 13:54:57 -04:00
state->options.flags |= COGL_PIPELINE_FLUSH_FALLBACK_MASK;
}
validated:
state->unit++;
return status;
}
void
_cogl_flush_attributes_state (CoglFramebuffer *framebuffer,
CoglPipeline *pipeline,
CoglDrawFlags flags,
CoglAttribute **attributes,
int n_attributes)
{
CoglContext *ctx = framebuffer->context;
CoglFlushLayerState layers_state;
CoglPipeline *copy = NULL;
if (!(flags & COGL_DRAW_SKIP_JOURNAL_FLUSH))
_cogl_journal_flush (framebuffer->journal);
layers_state.unit = 0;
layers_state.options.flags = 0;
layers_state.fallback_layers = 0;
if (!(flags & COGL_DRAW_SKIP_PIPELINE_VALIDATION))
cogl_pipeline_foreach_layer (pipeline,
validate_layer_cb,
&layers_state);
/* NB: _cogl_framebuffer_flush_state may disrupt various state (such
* as the pipeline state) when flushing the clip stack, so should
* always be done first when preparing to draw. We need to do this
* before setting up the array pointers because setting up the clip
* stack can cause some drawing which would change the array
* pointers. */
if (!(flags & COGL_DRAW_SKIP_FRAMEBUFFER_FLUSH))
_cogl_framebuffer_flush_state (framebuffer,
framebuffer,
COGL_FRAMEBUFFER_STATE_ALL);
cogl: Implements a software only read-pixel fast-path This adds a transparent optimization to cogl_read_pixels for when a single pixel is being read back and it happens that all the geometry of the current frame is still available in the framebuffer's associated journal. The intention is to indirectly optimize Clutter's render based picking mechanism in such a way that the 99% of cases where scenes are comprised of trivial quad primitives that can easily be intersected we can avoid the latency of kicking a GPU render and blocking for the result when we know we can calculate the result manually on the CPU probably faster than we could even kick a render. A nice property of this solution is that it maintains all the flexibility of the render based picking provided by Clutter and it can gracefully fall back to GPU rendering if actors are drawn using anything more complex than a quad for their geometry. It seems worth noting that there is a limitation to the extensibility of this approach in that it can only optimize picking a against geometry that passes through Cogl's journal which isn't something Clutter directly controls. For now though this really doesn't matter since basically all apps should end up hitting this fast-path. The current idea to address this longer term would be a pick2 vfunc for ClutterActor that can support geometry and render based input regions of actors and move this optimization up into Clutter instead. Note: currently we don't have a primitive count threshold to consider that there could be scenes with enough geometry for us to compensate for the cost of kicking a render and determine a result more efficiently by utilizing the GPU. We don't currently expect this to be common though. Note: in the future it could still be interesting to revive something like the wip/async-pbo-picking branch to provide an asynchronous read-pixels based optimization for Clutter picking in cases where more complex input regions that necessitate rendering are in use or if we do add a threshold for rendering as mentioned above.
2011-01-12 17:12:41 -05:00
/* In cogl_read_pixels we have a fast-path when reading a single
* pixel and the scene is just comprised of simple rectangles still
* in the journal. For this optimization to work we need to track
* when the framebuffer really does get drawn to. */
_cogl_framebuffer_dirty (framebuffer);
if (G_UNLIKELY (!(flags & COGL_DRAW_SKIP_LEGACY_STATE)) &&
G_UNLIKELY (ctx->legacy_state_set) &&
Add internal _cogl_push_source to optionally disable legacy state Some code in Cogl such as when flushing a stencil clip assumes that it can push a temporary simple pipeline to reset to a known state for internal drawing operations. However this breaks down if the application has set any legacy state because that is set globally so it will also get applied to the internal pipeline. _cogl_draw_attributes already had an internal flag to disable applying the legacy state but I think this is quite awkward to use because not all places that push a pipeline draw the attribute buffers directly so it is difficult to pass the flag down through the layers. Conceptually the legacy state is meant to be like a layer on top of the purely pipeline-based state API so I think ideally we should have an internal function to push the source without the applying the legacy state. The legacy state can't be applied as the pipeline is pushed because the global state can be modified even after it is pushed. This patch adds a _cogl_push_source() function which takes an extra boolean flag to mark whether to enable the legacy state. The value of this flag is stored alongside the pipeline in the pipeline stack. Another new internal function called _cogl_get_enable_legacy_state queries whether the top entry in the pipeline stack has legacy state enabled. cogl-primitives and the vertex array drawing code now use this to determine whether to apply the legacy state when drawing. The COGL_DRAW_SKIP_LEGACY_STATE flag is now removed. Reviewed-by: Robert Bragg <robert@linux.intel.com>
2011-09-14 07:17:09 -04:00
_cogl_get_enable_legacy_state ())
{
copy = cogl_pipeline_copy (pipeline);
pipeline = copy;
_cogl_pipeline_apply_legacy_state (pipeline);
}
ctx->driver_vtable->flush_attributes_state (framebuffer,
pipeline,
&layers_state,
flags,
attributes,
n_attributes);
if (copy)
cogl_object_unref (copy);
}