+ // Special case: We could be an input and still be asked to
+ // fix our gamma; if so, we should be the only node
+ // (as node_needs_gamma_fix() would only return true in
+ // for an input in that case). That means we should insert
+ // a conversion node _after_ ourselves.
+ if (node->incoming_links.empty()) {
+ assert(node->outgoing_links.empty());
+ Node *conversion = add_node(new GammaExpansionEffect());
+ CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
+ conversion->output_gamma_curve = GAMMA_LINEAR;
+ connect_nodes(node, conversion);
+ }
+
+ // If not, go through each input that is not linear gamma,
+ // and insert a gamma conversion after it.
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ Node *input = node->incoming_links[j];
+ assert(input->output_gamma_curve != GAMMA_INVALID);
+ if (input->output_gamma_curve == GAMMA_LINEAR) {
+ continue;
+ }
+ Node *conversion = add_node(new GammaExpansionEffect());
+ CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
+ conversion->output_gamma_curve = GAMMA_LINEAR;
+ replace_sender(input, conversion);
+ connect_nodes(input, conversion);
+ }
+
+ // Re-sort topologically, and propagate the new information.
+ propagate_alpha();
+ propagate_gamma_and_color_space();
+
+ found_any = true;
+ break;
+ }
+
+ char filename[256];
+ sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
+ output_dot(filename);
+ assert(gamma_propagation_pass < 100);
+ } while (found_any);
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->output_gamma_curve != GAMMA_INVALID);
+ }
+}
+
+// Make so that the output is in the desired gamma.
+// Note that this assumes linear input gamma, so it might create the need
+// for another pass of fix_internal_gamma().
+void EffectChain::fix_output_gamma()
+{
+ Node *output = find_output_node();
+ if (output->output_gamma_curve != output_format.gamma_curve) {
+ Node *conversion = add_node(new GammaCompressionEffect());
+ CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
+ conversion->output_gamma_curve = output_format.gamma_curve;
+ connect_nodes(output, conversion);
+ }
+}
+
+// If the user has requested Y'CbCr output, we need to do this conversion
+// _after_ GammaCompressionEffect etc., but before dither (see below).
+// This is because Y'CbCr, with the exception of a special optional mode
+// in Rec. 2020 (which we currently don't support), is defined to work on
+// gamma-encoded data.
+void EffectChain::add_ycbcr_conversion_if_needed()
+{
+ assert(output_color_rgba || output_color_ycbcr);
+ if (!output_color_ycbcr) {
+ return;
+ }
+ Node *output = find_output_node();
+ Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
+ connect_nodes(output, ycbcr);
+}
+
+// If the user has requested dither, add a DitherEffect right at the end
+// (after GammaCompressionEffect etc.). This needs to be done after everything else,
+// since dither is about the only effect that can _not_ be done in linear space.
+void EffectChain::add_dither_if_needed()
+{
+ if (num_dither_bits == 0) {
+ return;
+ }
+ Node *output = find_output_node();
+ Node *dither = add_node(new DitherEffect());
+ CHECK(dither->effect->set_int("num_bits", num_dither_bits));
+ connect_nodes(output, dither);
+
+ dither_effect = dither->effect;
+}
+
+// Find the output node. This is, simply, one that has no outgoing links.
+// If there are multiple ones, the graph is malformed (we do not support
+// multiple outputs right now).
+Node *EffectChain::find_output_node()
+{
+ vector<Node *> output_nodes;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ if (node->outgoing_links.empty()) {
+ output_nodes.push_back(node);
+ }
+ }
+ assert(output_nodes.size() == 1);
+ return output_nodes[0];
+}
+
+void EffectChain::finalize()
+{
+ // Output the graph as it is before we do any conversions on it.
+ output_dot("step0-start.dot");
+
+ // Give each effect in turn a chance to rewrite its own part of the graph.
+ // Note that if more effects are added as part of this, they will be
+ // picked up as part of the same for loop, since they are added at the end.
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ nodes[i]->effect->rewrite_graph(this, nodes[i]);
+ }
+ output_dot("step1-rewritten.dot");
+
+ find_color_spaces_for_inputs();
+ output_dot("step2-input-colorspace.dot");
+
+ propagate_alpha();
+ output_dot("step3-propagated-alpha.dot");
+
+ propagate_gamma_and_color_space();
+ output_dot("step4-propagated-all.dot");
+
+ fix_internal_color_spaces();
+ fix_internal_alpha(6);
+ fix_output_color_space();
+ output_dot("step7-output-colorspacefix.dot");
+ fix_output_alpha();
+ output_dot("step8-output-alphafix.dot");
+
+ // Note that we need to fix gamma after colorspace conversion,
+ // because colorspace conversions might create needs for gamma conversions.
+ // Also, we need to run an extra pass of fix_internal_gamma() after
+ // fixing the output gamma, as we only have conversions to/from linear,
+ // and fix_internal_alpha() since GammaCompressionEffect needs
+ // postmultiplied input.
+ fix_internal_gamma_by_asking_inputs(9);
+ fix_internal_gamma_by_inserting_nodes(10);
+ fix_output_gamma();
+ output_dot("step11-output-gammafix.dot");
+ propagate_alpha();
+ output_dot("step12-output-alpha-propagated.dot");
+ fix_internal_alpha(13);
+ output_dot("step14-output-alpha-fixed.dot");
+ fix_internal_gamma_by_asking_inputs(15);
+ fix_internal_gamma_by_inserting_nodes(16);
+
+ output_dot("step17-before-ycbcr.dot");
+ add_ycbcr_conversion_if_needed();
+
+ output_dot("step18-before-dither.dot");
+ add_dither_if_needed();
+
+ output_dot("step19-final.dot");
+
+ // Construct all needed GLSL programs, starting at the output.
+ // We need to keep track of which effects have already been computed,
+ // as an effect with multiple users could otherwise be calculated
+ // multiple times.
+ map<Node *, Phase *> completed_effects;
+ construct_phase(find_output_node(), &completed_effects);
+
+ output_dot("step20-split-to-phases.dot");
+
+ assert(phases[0]->inputs.empty());
+
+ finalized = true;
+}
+
+void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
+{
+ assert(finalized);
+
+ // This needs to be set anew, in case we are coming from a different context
+ // from when we initialized.
+ glDisable(GL_DITHER);
+
+ // Save original viewport.
+ GLuint x = 0, y = 0;
+
+ if (width == 0 && height == 0) {
+ GLint viewport[4];
+ glGetIntegerv(GL_VIEWPORT, viewport);
+ x = viewport[0];
+ y = viewport[1];
+ width = viewport[2];
+ height = viewport[3];
+ }
+
+ // Basic state.
+ glDisable(GL_BLEND);
+ check_error();
+ glDisable(GL_DEPTH_TEST);
+ check_error();
+ glDepthMask(GL_FALSE);
+ check_error();
+
+ // Generate a VAO. All the phases should have exactly the same vertex attributes,
+ // so it's safe to reuse this.
+ float vertices[] = {
+ 0.0f, 2.0f,
+ 0.0f, 0.0f,
+ 2.0f, 0.0f
+ };
+
+ GLuint vao;
+ glGenVertexArrays(1, &vao);
+ check_error();
+ glBindVertexArray(vao);
+ check_error();
+
+ GLuint position_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
+ GLuint texcoord_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
+
+ 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_num = 0; phase_num < phases.size(); ++phase_num) {
+ Phase *phase = phases[phase_num];
+
+ if (do_phase_timing) {
+ glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object);
+ }
+ if (phase_num == phases.size() - 1) {
+ // Last phase goes to the output the user specified.
+ glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
+ check_error();
+ GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+ assert(status == GL_FRAMEBUFFER_COMPLETE);
+ glViewport(x, y, width, height);
+ if (dither_effect != NULL) {
+ CHECK(dither_effect->set_int("output_width", width));
+ CHECK(dither_effect->set_int("output_height", height));
+ }
+ }
+ execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
+ if (do_phase_timing) {
+ glEndQuery(GL_TIME_ELAPSED);
+ }
+ }
+
+ for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
+ texture_it != output_textures.end();
+ ++texture_it) {
+ resource_pool->release_2d_texture(texture_it->second);
+ }
+
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ check_error();
+ glUseProgram(0);
+ check_error();
+
+ cleanup_vertex_attribute(phases[0]->glsl_program_num, "position", position_vbo);
+ cleanup_vertex_attribute(phases[0]->glsl_program_num, "texcoord", texcoord_vbo);
+
+ glDeleteVertexArrays(1, &vao);
+ check_error();
+
+ if (do_phase_timing) {
+ // Get back the timer queries.
+ for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
+ Phase *phase = phases[phase_num];
+ GLint available = 0;
+ while (!available) {
+ glGetQueryObjectiv(phase->timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
+ }
+ GLuint64 time_elapsed;
+ glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
+ phase->time_elapsed_ns += time_elapsed;
+ ++phase->num_measured_iterations;
+ }
+ }
+}
+
+void EffectChain::enable_phase_timing(bool enable)
+{
+ if (enable) {
+ assert(movit_timer_queries_supported);
+ }
+ this->do_phase_timing = enable;
+}
+
+void EffectChain::reset_phase_timing()
+{
+ for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
+ Phase *phase = phases[phase_num];
+ phase->time_elapsed_ns = 0;
+ phase->num_measured_iterations = 0;
+ }
+}
+
+void EffectChain::print_phase_timing()
+{
+ double total_time_ms = 0.0;
+ for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
+ Phase *phase = phases[phase_num];
+ double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
+ printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms);
+ for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
+ if (effect_num != 0) {
+ printf(", ");
+ }
+ printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
+ }
+ printf("]\n");
+ total_time_ms += avg_time_ms;
+ }
+ printf("Total: %5.1f ms\n", total_time_ms);
+}
+
+void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLuint> *output_textures, set<Phase *> *generated_mipmaps)
+{
+ GLuint fbo = 0;
+
+ // Find a texture for this phase.
+ inform_input_sizes(phase);
+ if (!last_phase) {
+ find_output_size(phase);
+
+ 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 = phase->glsl_program_num;
+ check_error();
+ glUseProgram(glsl_program_num);
+ check_error();
+
+ // Set up RTT inputs for this phase.
+ for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
+ glActiveTexture(GL_TEXTURE0 + sampler);
+ Phase *input = phase->inputs[sampler];
+ input->output_node->bound_sampler_num = sampler;
+ glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);