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,
+ size_t vao_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),
+ vao_freelist_max_length(vao_freelist_max_length),
+ texture_freelist_bytes(0)
+{
pthread_mutex_init(&lock, NULL);
}
}
assert(texture_formats.empty());
assert(texture_freelist_bytes == 0);
+
+ void *context = get_gl_context_identifier();
+ cleanup_unlinked_fbos(context);
+
+ for (map<void *, std::list<FBOFormatIterator> >::iterator context_it = fbo_freelist.begin();
+ context_it != fbo_freelist.end();
+ ++context_it) {
+ if (context_it->first != context) {
+ // If this does not hold, the client should have called clean_context() earlier.
+ assert(context_it->second.empty());
+ continue;
+ }
+ for (list<FBOFormatIterator>::const_iterator freelist_it = context_it->second.begin();
+ freelist_it != context_it->second.end();
+ ++freelist_it) {
+ FBOFormatIterator fbo_it = *freelist_it;
+ glDeleteFramebuffers(1, &fbo_it->second.fbo_num);
+ check_error();
+ fbo_formats.erase(fbo_it);
+ }
+ }
+
+ assert(fbo_formats.empty());
}
void ResourcePool::delete_program(GLuint glsl_program_num)
}
}
assert(found_program);
- glDeleteProgram(glsl_program_num);
- map<GLuint, pair<GLuint, GLuint> >::iterator shader_it =
+ map<GLuint, stack<GLuint> >::iterator instance_list_it = program_instances.find(glsl_program_num);
+ assert(instance_list_it != program_instances.end());
+
+ while (!instance_list_it->second.empty()) {
+ GLuint instance_program_num = instance_list_it->second.top();
+ instance_list_it->second.pop();
+ glDeleteProgram(instance_program_num);
+ program_masters.erase(instance_program_num);
+ }
+ program_instances.erase(instance_list_it);
+
+ map<GLuint, ShaderSpec>::iterator shader_it =
program_shaders.find(glsl_program_num);
assert(shader_it != program_shaders.end());
- glDeleteShader(shader_it->second.first);
- glDeleteShader(shader_it->second.second);
+ glDeleteShader(shader_it->second.vs_obj);
+ glDeleteShader(shader_it->second.fs_obj);
program_shaders.erase(shader_it);
}
-GLuint ResourcePool::compile_glsl_program(const string& vertex_shader, const string& fragment_shader)
+GLuint ResourcePool::compile_glsl_program(const string& vertex_shader,
+ const string& fragment_shader,
+ const vector<string>& fragment_shader_outputs)
{
GLuint glsl_program_num;
pthread_mutex_lock(&lock);
- const pair<string, string> key(vertex_shader, fragment_shader);
+
+ // Augment the fragment shader program text with the outputs, so that they become
+ // part of the key. Also potentially useful for debugging.
+ string fragment_shader_processed = fragment_shader;
+ for (unsigned output_index = 0; output_index < fragment_shader_outputs.size(); ++output_index) {
+ char buf[256];
+ snprintf(buf, sizeof(buf), "// Bound output: %s\n", fragment_shader_outputs[output_index].c_str());
+ fragment_shader_processed += buf;
+ }
+
+ const pair<string, string> key(vertex_shader, fragment_shader_processed);
if (programs.count(key)) {
// Already in the cache. Increment the refcount, or take it off the freelist
// if it's zero.
}
} else {
// Not in the cache. Compile the shaders.
- glsl_program_num = glCreateProgram();
GLuint vs_obj = compile_shader(vertex_shader, GL_VERTEX_SHADER);
- GLuint fs_obj = compile_shader(fragment_shader, GL_FRAGMENT_SHADER);
- glAttachShader(glsl_program_num, vs_obj);
- check_error();
- glAttachShader(glsl_program_num, fs_obj);
check_error();
- glLinkProgram(glsl_program_num);
+ GLuint fs_obj = compile_shader(fragment_shader_processed, GL_FRAGMENT_SHADER);
check_error();
-
- GLint success;
- glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
- if (success == GL_FALSE) {
- GLchar error_log[1024] = {0};
- glGetProgramInfoLog(glsl_program_num, 1024, NULL, error_log);
- fprintf(stderr, "Error linking program: %s\n", error_log);
- exit(1);
- }
+ glsl_program_num = link_program(vs_obj, fs_obj, fragment_shader_outputs);
if (movit_debug_level == MOVIT_DEBUG_ON) {
// Output shader to a temporary file, for easier debugging.
perror(filename);
exit(1);
}
- fprintf(fp, "%s\n", fragment_shader.c_str());
+ fprintf(fp, "%s\n", fragment_shader_processed.c_str());
fclose(fp);
}
programs.insert(make_pair(key, glsl_program_num));
program_refcount.insert(make_pair(glsl_program_num, 1));
- program_shaders.insert(make_pair(glsl_program_num, make_pair(vs_obj, fs_obj)));
+ ShaderSpec spec;
+ spec.vs_obj = vs_obj;
+ spec.fs_obj = fs_obj;
+ spec.fragment_shader_outputs = fragment_shader_outputs;
+ program_shaders.insert(make_pair(glsl_program_num, spec));
+ stack<GLuint> instances;
+ instances.push(glsl_program_num);
+ program_instances.insert(make_pair(glsl_program_num, instances));
+ program_masters.insert(make_pair(glsl_program_num, glsl_program_num));
}
pthread_mutex_unlock(&lock);
return glsl_program_num;
}
+GLuint ResourcePool::link_program(GLuint vs_obj,
+ GLuint fs_obj,
+ const vector<string>& fragment_shader_outputs)
+{
+ GLuint glsl_program_num = glCreateProgram();
+ check_error();
+ glAttachShader(glsl_program_num, vs_obj);
+ check_error();
+ glAttachShader(glsl_program_num, fs_obj);
+ check_error();
+
+ // Bind the outputs, if we have multiple ones.
+ if (fragment_shader_outputs.size() > 1) {
+ for (unsigned output_index = 0; output_index < fragment_shader_outputs.size(); ++output_index) {
+ glBindFragDataLocation(glsl_program_num, output_index,
+ fragment_shader_outputs[output_index].c_str());
+ }
+ }
+
+ glLinkProgram(glsl_program_num);
+ check_error();
+
+ GLint success;
+ glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
+ if (success == GL_FALSE) {
+ GLchar error_log[1024] = {0};
+ glGetProgramInfoLog(glsl_program_num, 1024, NULL, error_log);
+ fprintf(stderr, "Error linking program: %s\n", error_log);
+ exit(1);
+ }
+
+ return glsl_program_num;
+}
+
void ResourcePool::release_glsl_program(GLuint glsl_program_num)
{
pthread_mutex_lock(&lock);
pthread_mutex_unlock(&lock);
}
+GLuint ResourcePool::use_glsl_program(GLuint glsl_program_num)
+{
+ pthread_mutex_lock(&lock);
+ assert(program_instances.count(glsl_program_num));
+ stack<GLuint> &instances = program_instances[glsl_program_num];
+
+ GLuint instance_program_num;
+ if (!instances.empty()) {
+ // There's an unused instance of this program; just return it.
+ instance_program_num = instances.top();
+ instances.pop();
+ } else {
+ // We need to clone this program. (unuse_glsl_program()
+ // will later put it onto the list.)
+ map<GLuint, ShaderSpec>::iterator shader_it =
+ program_shaders.find(glsl_program_num);
+ assert(shader_it != program_shaders.end());
+
+ instance_program_num = link_program(
+ shader_it->second.vs_obj,
+ shader_it->second.fs_obj,
+ shader_it->second.fragment_shader_outputs);
+ program_masters.insert(make_pair(instance_program_num, glsl_program_num));
+ }
+ pthread_mutex_unlock(&lock);
+
+ glUseProgram(instance_program_num);
+ return instance_program_num;
+}
+
+void ResourcePool::unuse_glsl_program(GLuint instance_program_num)
+{
+ pthread_mutex_lock(&lock);
+
+ map<GLuint, GLuint>::const_iterator master_it = program_masters.find(instance_program_num);
+ assert(master_it != program_masters.end());
+
+ assert(program_instances.count(master_it->second));
+ stack<GLuint> &instances = program_instances[master_it->second];
+
+ instances.push(instance_program_num);
+
+ pthread_mutex_unlock(&lock);
+}
+
GLuint ResourcePool::create_2d_texture(GLint internal_format, GLsizei width, GLsizei height)
{
+ assert(width > 0);
+ assert(height > 0);
+
pthread_mutex_lock(&lock);
// See if there's a texture on the freelist we can use.
for (list<GLuint>::iterator freelist_it = texture_freelist.begin();
switch (internal_format) {
case GL_RGBA32F_ARB:
case GL_RGBA16F_ARB:
+ case GL_RGBA16:
case GL_RGBA8:
+ case GL_RGB10_A2:
case GL_SRGB8_ALPHA8:
format = GL_RGBA;
break;
+ case GL_RGB32F:
+ case GL_RGB16F:
+ case GL_RGB16:
+ case GL_R11F_G11F_B10F:
+ case GL_RGB8:
+ case GL_RGB10:
+ case GL_SRGB8:
+ case GL_RGB565:
+ case GL_RGB9_E5:
+ format = GL_RGB;
+ break;
case GL_RG32F:
case GL_RG16F:
+ case GL_RG16:
+ case GL_RG8:
format = GL_RG;
break;
case GL_R32F:
case GL_R16F:
+ case GL_R16:
case GL_R8:
format = GL_RED;
break;
switch (internal_format) {
case GL_RGBA32F_ARB:
case GL_RGBA16F_ARB:
+ case GL_RGB32F:
+ case GL_RGB16F:
+ case GL_R11F_G11F_B10F:
+ case GL_RGB9_E5:
case GL_RG32F:
case GL_RG16F:
case GL_R32F:
case GL_R16F:
type = GL_FLOAT;
break;
- case GL_RGBA8:
+ case GL_RGBA16:
+ case GL_RGB16:
+ case GL_RG16:
+ case GL_R16:
+ type = GL_UNSIGNED_SHORT;
+ break;
case GL_SRGB8_ALPHA8:
+ case GL_SRGB8:
+ case GL_RGBA8:
+ case GL_RGB8:
+ case GL_RGB10_A2:
+ case GL_RGB10:
+ case GL_RG8:
case GL_R8:
type = GL_UNSIGNED_BYTE;
break;
+ case GL_RGB565:
+ type = GL_UNSIGNED_SHORT_5_6_5;
+ break;
default:
// TODO: Add more here as needed.
assert(false);
texture_freelist_bytes += estimate_texture_size(texture_formats[texture_num]);
while (texture_freelist_bytes > texture_freelist_max_bytes) {
- GLuint free_texture_num = texture_freelist.front();
- texture_freelist.pop_front();
+ GLuint free_texture_num = texture_freelist.back();
+ texture_freelist.pop_back();
assert(texture_formats.count(free_texture_num) != 0);
texture_freelist_bytes -= estimate_texture_size(texture_formats[free_texture_num]);
texture_formats.erase(free_texture_num);
glDeleteTextures(1, &free_texture_num);
check_error();
+
+ // Unlink any lingering FBO related to this texture. We might
+ // not be in the right context, so don't delete it right away;
+ // the cleanup in release_fbo() (which calls cleanup_unlinked_fbos())
+ // will take care of actually doing that later.
+ for (map<pair<void *, GLuint>, FBO>::iterator format_it = fbo_formats.begin();
+ format_it != fbo_formats.end();
+ ++format_it) {
+ for (unsigned i = 0; i < num_fbo_attachments; ++i) {
+ if (format_it->second.texture_num[i] == free_texture_num) {
+ format_it->second.texture_num[i] = GL_INVALID_INDEX;
+ }
+ }
+ }
+ }
+ pthread_mutex_unlock(&lock);
+}
+
+GLuint ResourcePool::create_fbo(GLuint texture0_num, GLuint texture1_num, GLuint texture2_num, GLuint texture3_num)
+{
+ void *context = get_gl_context_identifier();
+
+ // Make sure we are filled from the bottom.
+ assert(texture0_num != 0);
+ if (texture1_num == 0) {
+ assert(texture2_num == 0);
+ }
+ if (texture2_num == 0) {
+ assert(texture3_num == 0);
+ }
+
+ pthread_mutex_lock(&lock);
+ if (fbo_freelist.count(context) != 0) {
+ // See if there's an FBO on the freelist we can use.
+ list<FBOFormatIterator>::iterator end = fbo_freelist[context].end();
+ for (list<FBOFormatIterator>::iterator freelist_it = fbo_freelist[context].begin();
+ freelist_it != end; ++freelist_it) {
+ FBOFormatIterator fbo_it = *freelist_it;
+ if (fbo_it->second.texture_num[0] == texture0_num &&
+ fbo_it->second.texture_num[1] == texture1_num &&
+ fbo_it->second.texture_num[2] == texture2_num &&
+ fbo_it->second.texture_num[3] == texture3_num) {
+ fbo_freelist[context].erase(freelist_it);
+ pthread_mutex_unlock(&lock);
+ return fbo_it->second.fbo_num;
+ }
+ }
+ }
+
+ // Create a new one.
+ FBO fbo_format;
+ fbo_format.texture_num[0] = texture0_num;
+ fbo_format.texture_num[1] = texture1_num;
+ fbo_format.texture_num[2] = texture2_num;
+ fbo_format.texture_num[3] = texture3_num;
+
+ glGenFramebuffers(1, &fbo_format.fbo_num);
+ check_error();
+ glBindFramebuffer(GL_FRAMEBUFFER, fbo_format.fbo_num);
+ check_error();
+
+ GLenum bufs[num_fbo_attachments];
+ unsigned num_active_attachments = 0;
+ for (unsigned i = 0; i < num_fbo_attachments; ++i, ++num_active_attachments) {
+ if (fbo_format.texture_num[i] == 0) {
+ break;
+ }
+ glFramebufferTexture2D(
+ GL_FRAMEBUFFER,
+ GL_COLOR_ATTACHMENT0 + i,
+ GL_TEXTURE_2D,
+ fbo_format.texture_num[i],
+ 0);
+ check_error();
+ bufs[i] = GL_COLOR_ATTACHMENT0 + i;
+ }
+
+ glDrawBuffers(num_active_attachments, bufs);
+ check_error();
+
+ GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+ assert(status == GL_FRAMEBUFFER_COMPLETE);
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ check_error();
+
+ pair<void *, GLuint> key(context, fbo_format.fbo_num);
+ assert(fbo_formats.count(key) == 0);
+ fbo_formats.insert(make_pair(key, fbo_format));
+
+ pthread_mutex_unlock(&lock);
+ return fbo_format.fbo_num;
+}
+
+void ResourcePool::release_fbo(GLuint fbo_num)
+{
+ void *context = get_gl_context_identifier();
+
+ pthread_mutex_lock(&lock);
+ FBOFormatIterator fbo_it = fbo_formats.find(make_pair(context, fbo_num));
+ assert(fbo_it != fbo_formats.end());
+ fbo_freelist[context].push_front(fbo_it);
+
+ // Now that we're in this context, free up any FBOs that are connected
+ // to deleted textures (in release_2d_texture).
+ cleanup_unlinked_fbos(context);
+
+ shrink_fbo_freelist(context, fbo_freelist_max_length);
+ pthread_mutex_unlock(&lock);
+}
+
+GLuint ResourcePool::create_vec2_vao(const set<GLint> &attribute_indices, GLuint vbo_num)
+{
+ void *context = get_gl_context_identifier();
+
+ pthread_mutex_lock(&lock);
+ if (vao_freelist.count(context) != 0) {
+ // See if there's a VAO the freelist we can use.
+ list<VAOFormatIterator>::iterator end = vao_freelist[context].end();
+ for (list<VAOFormatIterator>::iterator freelist_it = vao_freelist[context].begin();
+ freelist_it != end; ++freelist_it) {
+ VAOFormatIterator vao_it = *freelist_it;
+ if (vao_it->second.vbo_num == vbo_num &&
+ vao_it->second.attribute_indices == attribute_indices) {
+ vao_freelist[context].erase(freelist_it);
+ pthread_mutex_unlock(&lock);
+ return vao_it->second.vao_num;
+ }
+ }
+ }
+
+ // Create a new one.
+ VAO vao_format;
+ vao_format.attribute_indices = attribute_indices;
+ vao_format.vbo_num = vbo_num;
+
+ glGenVertexArrays(1, &vao_format.vao_num);
+ check_error();
+ glBindVertexArray(vao_format.vao_num);
+ check_error();
+ glBindBuffer(GL_ARRAY_BUFFER, vbo_num);
+ check_error();
+
+ for (set<GLint>::const_iterator attr_it = attribute_indices.begin(); attr_it != attribute_indices.end(); ++attr_it) {
+ glEnableVertexAttribArray(*attr_it);
+ check_error();
+ glVertexAttribPointer(*attr_it, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+ check_error();
}
+
+ glBindVertexArray(0);
+ check_error();
+ glBindBuffer(GL_ARRAY_BUFFER, 0);
+ check_error();
+
+ pair<void *, GLuint> key(context, vao_format.vao_num);
+ assert(vao_formats.count(key) == 0);
+ vao_formats.insert(make_pair(key, vao_format));
+
+ pthread_mutex_unlock(&lock);
+ return vao_format.vao_num;
+}
+
+void ResourcePool::release_vec2_vao(GLuint vao_num)
+{
+ void *context = get_gl_context_identifier();
+
+ pthread_mutex_lock(&lock);
+ VAOFormatIterator vao_it = vao_formats.find(make_pair(context, vao_num));
+ assert(vao_it != vao_formats.end());
+ vao_freelist[context].push_front(vao_it);
+
+ shrink_vao_freelist(context, vao_freelist_max_length);
pthread_mutex_unlock(&lock);
}
+void ResourcePool::clean_context()
+{
+ void *context = get_gl_context_identifier();
+
+ // Currently, we only need to worry about FBOs and VAOs, as they are the only
+ // non-shareable resources we hold.
+ shrink_fbo_freelist(context, 0);
+ fbo_freelist.erase(context);
+
+ shrink_vao_freelist(context, 0);
+ vao_freelist.erase(context);
+}
+
+void ResourcePool::cleanup_unlinked_fbos(void *context)
+{
+ list<FBOFormatIterator>::iterator end = fbo_freelist[context].end();
+ for (list<FBOFormatIterator>::iterator freelist_it = fbo_freelist[context].begin(); freelist_it != end; ) {
+ FBOFormatIterator fbo_it = *freelist_it;
+
+ bool all_unlinked = true;
+ for (unsigned i = 0; i < num_fbo_attachments; ++i) {
+ if (fbo_it->second.texture_num[i] != 0 &&
+ fbo_it->second.texture_num[i] != GL_INVALID_INDEX) {
+ all_unlinked = false;
+ break;
+ }
+ }
+ if (all_unlinked) {
+ glDeleteFramebuffers(1, &fbo_it->second.fbo_num);
+ check_error();
+ fbo_formats.erase(fbo_it);
+ fbo_freelist[context].erase(freelist_it++);
+ } else {
+ freelist_it++;
+ }
+ }
+}
+
+void ResourcePool::shrink_fbo_freelist(void *context, size_t max_length)
+{
+ list<FBOFormatIterator> &freelist = fbo_freelist[context];
+ while (freelist.size() > max_length) {
+ FBOFormatIterator free_fbo_it = freelist.back();
+ glDeleteFramebuffers(1, &free_fbo_it->second.fbo_num);
+ check_error();
+ fbo_formats.erase(free_fbo_it);
+ freelist.pop_back();
+ }
+}
+
+void ResourcePool::shrink_vao_freelist(void *context, size_t max_length)
+{
+ list<VAOFormatIterator> &freelist = vao_freelist[context];
+ while (freelist.size() > max_length) {
+ VAOFormatIterator free_vao_it = freelist.back();
+ glDeleteVertexArrays(1, &free_vao_it->second.vao_num);
+ check_error();
+ vao_formats.erase(free_vao_it);
+ freelist.pop_back();
+ }
+}
+
size_t ResourcePool::estimate_texture_size(const Texture2D &texture_format)
{
size_t bytes_per_pixel;
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_R11F_G11F_B10F:
+ bytes_per_pixel = 4;
+ break;
+ case GL_RGB9_E5:
+ bytes_per_pixel = 4;
+ break;
case GL_RGBA8:
case GL_SRGB8_ALPHA8:
+ case GL_RGB10_A2:
+ case GL_RGB10:
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_R16F:
bytes_per_pixel = 2;
break;
+ case GL_RG8:
+ bytes_per_pixel = 2;
+ break;
case GL_R8:
bytes_per_pixel = 1;
break;
+ case GL_RGB565:
+ bytes_per_pixel = 2;
+ break;
+ case GL_RGBA16:
+ bytes_per_pixel = 8;
+ break;
+ case GL_RGB16:
+ bytes_per_pixel = 6;
+ break;
+ case GL_RG16:
+ bytes_per_pixel = 4;
+ break;
+ case GL_R16:
+ bytes_per_pixel = 2;
+ break;
default:
// TODO: Add more here as needed.
assert(false);