#define GL_GLEXT_PROTOTYPES 1
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <assert.h>
#include <locale.h>
#include <math.h>
if (owns_resource_pool) {
delete resource_pool;
}
- for (map<void *, GLuint>::const_iterator fbo_it = fbos.begin();
- fbo_it != fbos.end(); ++fbo_it) {
- glDeleteFramebuffers(1, &fbo_it->second);
- check_error();
- }
}
Input *EffectChain::add_input(Input *input)
void EffectChain::compile_glsl_program(Phase *phase)
{
- string frag_shader = read_file("header.frag");
+ string frag_shader = read_version_dependent_file("header", "frag");
// Create functions for all the texture inputs that we need.
for (unsigned i = 0; i < phase->inputs.size(); ++i) {
frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
- frag_shader += "\treturn texture2D(tex_" + string(effect_id) + ", tc);\n";
+ frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
frag_shader += "}\n";
frag_shader += "\n";
}
frag_shader += "\n";
}
frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
- frag_shader.append(read_file("footer.frag"));
+ frag_shader.append(read_version_dependent_file("footer", "frag"));
- phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader);
+ string vert_shader = read_version_dependent_file("vs", "vert");
+ phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader);
// Prepare the geometry for the fullscreen quad used in this phase.
// (We have separate VAOs per shader, since the bindings can in theory
glDepthMask(GL_FALSE);
check_error();
- if (phases.size() > 1) {
- if (fbos.count(context) == 0) {
- glGenFramebuffers(1, &fbo);
- check_error();
- fbos.insert(make_pair(context, fbo));
- } else {
- fbo = fbos[context];
- }
- glBindFramebuffer(GL_FRAMEBUFFER, fbo);
- check_error();
- }
-
set<Phase *> generated_mipmaps;
// We choose the simplest option of having one texture per output,
// since otherwise this turns into an (albeit simple) register allocation problem.
map<Phase *, GLuint> output_textures;
- for (unsigned phase = 0; phase < phases.size(); ++phase) {
+ for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
+ Phase *phase = phases[phase_num];
+
// Find a texture for this phase.
- inform_input_sizes(phases[phase]);
- if (phase != phases.size() - 1) {
- find_output_size(phases[phase]);
+ inform_input_sizes(phase);
+ if (phase_num != phases.size() - 1) {
+ find_output_size(phase);
- GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height);
- output_textures.insert(make_pair(phases[phase], tex_num));
+ GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
+ output_textures.insert(make_pair(phase, tex_num));
}
- const GLuint glsl_program_num = phases[phase]->glsl_program_num;
+ const GLuint glsl_program_num = phase->glsl_program_num;
check_error();
glUseProgram(glsl_program_num);
check_error();
// Set up RTT inputs for this phase.
- for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
+ for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
glActiveTexture(GL_TEXTURE0 + sampler);
- Phase *input = phases[phase]->inputs[sampler];
+ Phase *input = phase->inputs[sampler];
input->output_node->bound_sampler_num = sampler;
glBindTexture(GL_TEXTURE_2D, output_textures[input]);
check_error();
- if (phases[phase]->input_needs_mipmaps) {
+ if (phase->input_needs_mipmaps) {
if (generated_mipmaps.count(input) == 0) {
glGenerateMipmap(GL_TEXTURE_2D);
check_error();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
check_error();
- string texture_name = string("tex_") + phases[phase]->effect_ids[input->output_node];
+ string texture_name = string("tex_") + phase->effect_ids[input->output_node];
glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler);
check_error();
}
// And now the output.
- if (phase == phases.size() - 1) {
+ if (phase_num == phases.size() - 1) {
// Last phase goes to the output the user specified.
glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
check_error();
CHECK(dither_effect->set_int("output_height", height));
}
} else {
- glFramebufferTexture2D(
- GL_FRAMEBUFFER,
- GL_COLOR_ATTACHMENT0,
- GL_TEXTURE_2D,
- output_textures[phases[phase]],
- 0);
- check_error();
- GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
- assert(status == GL_FRAMEBUFFER_COMPLETE);
- glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
+ fbo = resource_pool->create_fbo(context, output_textures[phase]);
+ glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+ glViewport(0, 0, phase->output_width, phase->output_height);
}
// Give the required parameters to all the effects.
- unsigned sampler_num = phases[phase]->inputs.size();
- for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
- Node *node = phases[phase]->effects[i];
+ unsigned sampler_num = phase->inputs.size();
+ for (unsigned i = 0; i < phase->effects.size(); ++i) {
+ Node *node = phase->effects[i];
unsigned old_sampler_num = sampler_num;
- node->effect->set_gl_state(glsl_program_num, phases[phase]->effect_ids[node], &sampler_num);
+ node->effect->set_gl_state(glsl_program_num, phase->effect_ids[node], &sampler_num);
check_error();
if (node->effect->is_single_texture()) {
}
}
- glBindVertexArray(phases[phase]->vao);
+ glBindVertexArray(phase->vao);
check_error();
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
check_error();
- for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
- Node *node = phases[phase]->effects[i];
+ for (unsigned i = 0; i < phase->effects.size(); ++i) {
+ Node *node = phase->effects[i];
node->effect->clear_gl_state();
}
+ if (phase_num != phases.size() - 1) {
+ resource_pool->release_fbo(fbo);
+ }
}
for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
// the EffectChain holds textures and other OpenGL objects that are tied to the
// context.
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <stdio.h>
#include <map>
#include <set>
std::map<Effect *, Node *> node_map;
Effect *dither_effect;
- std::map<void *, GLuint> fbos; // One for each OpenGL context.
std::vector<Input *> inputs; // Also contained in nodes.
std::vector<Phase *> phases;
//
// Note that this also contains the tests for some of the simpler effects.
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <assert.h>
#include "effect.h"
TEST(EffectChainTest, RewritingWorksAndTexturesAreAskedForsRGB) {
unsigned char data[] = {
- 0, 64,
- 128, 255,
+ 0, 0, 0, 255,
+ 64, 64, 64, 255,
+ 128, 128, 128, 255,
+ 255, 255, 255, 255,
};
- float expected_data[4] = {
- 1.0f, 0.9771f,
- 0.8983f, 0.0f,
+ float expected_data[] = {
+ 1.0000f, 1.0000f, 1.0000f, 1.0000f,
+ 0.9771f, 0.9771f, 0.9771f, 1.0000f,
+ 0.8983f, 0.8983f, 0.8983f, 1.0000f,
+ 0.0000f, 0.0000f, 0.0000f, 1.0000f
};
- float out_data[4];
- EffectChainTester tester(NULL, 2, 2);
- tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_sRGB);
+ float out_data[4 * 4];
+ EffectChainTester tester(NULL, 1, 4);
+ tester.add_input(data, FORMAT_RGBA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_sRGB);
RewritingEffect<InvertEffect> *effect = new RewritingEffect<InvertEffect>();
tester.get_chain()->add_effect(effect);
- tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_sRGB);
+ tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_sRGB);
Node *node = effect->replaced_node;
ASSERT_EQ(1, node->incoming_links.size());
EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
EXPECT_EQ("GammaCompressionEffect", node->outgoing_links[0]->effect->effect_type_id());
- expect_equal(expected_data, out_data, 2, 2);
+ expect_equal(expected_data, out_data, 4, 4);
}
TEST(EffectChainTest, RewritingWorksAndColorspaceConversionsAreInserted) {
GammaCurve overridden_gamma_curve;
};
- TEST(EffectChainTester, HandlesInputChangingColorspace) {
+ TEST(EffectChainTest, HandlesInputChangingColorspace) {
const int size = 4;
float data[size] = {
#include <map>
#include <string>
#include <utility>
+#include <epoxy/gl.h>
-#include "glew.h"
#include "init.h"
#include "resource_pool.h"
#include "util.h"
namespace movit {
ResourcePool::ResourcePool(size_t program_freelist_max_length,
- size_t texture_freelist_max_bytes)
+ size_t texture_freelist_max_bytes,
+ size_t fbo_freelist_max_length)
: program_freelist_max_length(program_freelist_max_length),
texture_freelist_max_bytes(texture_freelist_max_bytes),
- texture_freelist_bytes(0) {
+ fbo_freelist_max_length(fbo_freelist_max_length),
+ texture_freelist_bytes(0)
+ {
pthread_mutex_init(&lock, NULL);
}
}
assert(texture_formats.empty());
assert(texture_freelist_bytes == 0);
+
+ for (list<GLuint>::const_iterator freelist_it = fbo_freelist.begin();
+ freelist_it != fbo_freelist.end();
+ ++freelist_it) {
+ GLuint free_fbo_num = *freelist_it;
+ assert(fbo_formats.count(free_fbo_num) != 0);
+ fbo_formats.erase(free_fbo_num);
+ glDeleteFramebuffers(1, &free_fbo_num);
+ check_error();
+ }
+ assert(fbo_formats.empty());
}
void ResourcePool::delete_program(GLuint glsl_program_num)
case GL_SRGB8_ALPHA8:
format = GL_RGBA;
break;
+ case GL_RGB32F:
+ case GL_RGB16F:
+ case GL_RGB8:
+ format = GL_RGB;
+ break;
case GL_RG32F:
case GL_RG16F:
+ case GL_RG8:
format = GL_RG;
break;
+ case GL_R32F:
+ case GL_R16F:
case GL_R8:
format = GL_RED;
break;
assert(false);
}
+ // Same with type; GLES is stricter than desktop OpenGL here.
+ GLenum type;
+ switch (internal_format) {
+ case GL_RGBA32F_ARB:
+ case GL_RGBA16F_ARB:
+ case GL_RGB32F:
+ case GL_RGB16F:
+ case GL_RG32F:
+ case GL_RG16F:
+ case GL_R32F:
+ case GL_R16F:
+ type = GL_FLOAT;
+ break;
+ case GL_SRGB8_ALPHA8:
+ case GL_RGBA8:
+ case GL_RGB8:
+ case GL_RG8:
+ case GL_R8:
+ type = GL_UNSIGNED_BYTE;
+ break;
+ default:
+ // TODO: Add more here as needed.
+ assert(false);
+ }
+
+
GLuint texture_num;
glGenTextures(1, &texture_num);
check_error();
glBindTexture(GL_TEXTURE_2D, texture_num);
check_error();
- glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, GL_UNSIGNED_BYTE, NULL);
+ glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, type, NULL);
check_error();
glBindTexture(GL_TEXTURE_2D, 0);
check_error();
texture_formats.erase(free_texture_num);
glDeleteTextures(1, &free_texture_num);
check_error();
+
+ // Delete any FBO related to this texture.
+ for (list<GLuint>::iterator fbo_freelist_it = fbo_freelist.begin();
+ fbo_freelist_it != fbo_freelist.end(); ) {
+ GLuint fbo_num = *fbo_freelist_it;
+ map<GLuint, FBO>::const_iterator format_it = fbo_formats.find(fbo_num);
+ assert(format_it != fbo_formats.end());
+ if (format_it->second.texture_num == free_texture_num) {
+ glDeleteFramebuffers(1, &fbo_num);
+ fbo_freelist.erase(fbo_freelist_it++);
+ } else {
+ ++fbo_freelist_it;
+ }
+ }
+ }
+ pthread_mutex_unlock(&lock);
+ }
+
+ GLuint ResourcePool::create_fbo(void *context, GLuint texture_num)
+ {
+ pthread_mutex_lock(&lock);
+ // See if there's an FBO on the freelist we can use.
+ for (list<GLuint>::iterator freelist_it = fbo_freelist.begin();
+ freelist_it != fbo_freelist.end();
+ ++freelist_it) {
+ GLuint fbo_num = *freelist_it;
+ map<GLuint, FBO>::const_iterator format_it = fbo_formats.find(fbo_num);
+ assert(format_it != fbo_formats.end());
+ if (format_it->second.context == context &&
+ format_it->second.texture_num == texture_num) {
+ fbo_freelist.erase(freelist_it);
+ pthread_mutex_unlock(&lock);
+ return fbo_num;
+ }
+ }
+
+ // Create a new one.
+ GLuint fbo_num;
+ glGenFramebuffers(1, &fbo_num);
+ check_error();
+ glBindFramebuffer(GL_FRAMEBUFFER, fbo_num);
+ check_error();
+ glFramebufferTexture2D(
+ GL_FRAMEBUFFER,
+ GL_COLOR_ATTACHMENT0,
+ GL_TEXTURE_2D,
+ texture_num,
+ 0);
+ check_error();
+ GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+ assert(status == GL_FRAMEBUFFER_COMPLETE);
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ check_error();
+
+ FBO fbo_format;
+ fbo_format.context = context;
+ fbo_format.texture_num = texture_num;
+ assert(fbo_formats.count(fbo_num) == 0);
+ fbo_formats.insert(make_pair(fbo_num, fbo_format));
+
+ pthread_mutex_unlock(&lock);
+ return fbo_num;
+ }
+
+ void ResourcePool::release_fbo(GLuint fbo_num)
+ {
+ pthread_mutex_lock(&lock);
+ fbo_freelist.push_front(fbo_num);
+ assert(fbo_formats.count(fbo_num) != 0);
+
+ while (fbo_freelist.size() > fbo_freelist_max_length) {
+ GLuint free_fbo_num = fbo_freelist.front();
+ fbo_freelist.pop_front();
+ assert(fbo_formats.count(free_fbo_num) != 0);
+ fbo_formats.erase(free_fbo_num);
+ glDeleteFramebuffers(1, &free_fbo_num);
+ check_error();
}
pthread_mutex_unlock(&lock);
}
case GL_RGBA16F_ARB:
bytes_per_pixel = 8;
break;
+ case GL_RGB32F_ARB:
+ bytes_per_pixel = 12;
+ break;
+ case GL_RGB16F_ARB:
+ bytes_per_pixel = 6;
+ break;
case GL_RGBA8:
case GL_SRGB8_ALPHA8:
bytes_per_pixel = 4;
break;
+ case GL_RGB8:
+ case GL_SRGB8:
+ bytes_per_pixel = 3;
+ break;
case GL_RG32F:
bytes_per_pixel = 8;
break;
case GL_RG16F:
bytes_per_pixel = 4;
break;
+ case GL_R32F:
+ bytes_per_pixel = 4;
+ break;
+ case GL_R16F:
+ bytes_per_pixel = 2;
+ break;
case GL_R8:
bytes_per_pixel = 1;
break;
// safely called from multiple threads at the same time, provided they have
// separate (but sharing) OpenGL contexts.
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <pthread.h>
#include <stddef.h>
#include <list>
// This means you should be prepared for actual memory usage of the freelist being
// twice this estimate or more.
ResourcePool(size_t program_freelist_max_length = 100,
- size_t texture_freelist_max_bytes = 100 << 20); // 100 MB.
+ size_t texture_freelist_max_bytes = 100 << 20, // 100 MB.
+ size_t fbo_freelist_max_length = 100);
~ResourcePool();
// All remaining functions are intended for calls from EffectChain only.
// or fetch a previous used if possible. Unbinds GL_TEXTURE_2D afterwards.
// Keeps ownership of the texture; you must call release_2d_texture() instead
// of deleting it when you no longer want it.
- //
- // Note: Currently we do not actually have a freelist, but this will change soon.
GLuint create_2d_texture(GLint internal_format, GLsizei width, GLsizei height);
void release_2d_texture(GLuint texture_num);
+ // Allocate an FBO with the the given texture bound as a framebuffer attachment,
+ // or fetch a previous used if possible. Unbinds GL_FRAMEBUFFER afterwards.
+ // Keeps ownership of the FBO; you must call release_fbo() of deleting
+ // it when you no longer want it. You can get an appropriate context
+ // pointer from get_gl_context_identifier().
+ //
+ // NOTE: In principle, the FBO doesn't have a resolution or pixel format;
+ // you can bind almost whatever texture you want to it. However, changing
+ // textures can have an adverse effect on performance due to validation,
+ // in particular on NVidia cards. Also, keep in mind that FBOs are not
+ // shareable across contexts.
+ GLuint create_fbo(void *context, GLuint texture_num);
+ void release_fbo(GLuint fbo_num);
+
private:
// Delete the given program and both its shaders.
void delete_program(GLuint program_num);
// Protects all the other elements in the class.
pthread_mutex_t lock;
- size_t program_freelist_max_length, texture_freelist_max_bytes;
+ size_t program_freelist_max_length, texture_freelist_max_bytes, fbo_freelist_max_length;
// A mapping from vertex/fragment shader source strings to compiled program number.
std::map<std::pair<std::string, std::string>, GLuint> programs;
std::list<GLuint> texture_freelist;
size_t texture_freelist_bytes;
+ struct FBO {
+ void *context;
+ GLuint texture_num;
+ };
+
+ // A mapping from FBO number to format details. This is filled if the
+ // FBO is given out to a client or on the freelist, but not if it is
+ // deleted from the freelist.
+ std::map<GLuint, FBO> fbo_formats;
+
+ // A list of all FBOs that are release but not freed (most recently freed
+ // first). Once this reaches <fbo_freelist_max_length>, the last element
+ // will be deleted.
+ std::list<GLuint> fbo_freelist;
+
// See the caveats at the constructor.
static size_t estimate_texture_size(const Texture2D &texture_format);
};