15 #include "alpha_division_effect.h"
16 #include "alpha_multiplication_effect.h"
17 #include "colorspace_conversion_effect.h"
18 #include "dither_effect.h"
20 #include "effect_chain.h"
21 #include "effect_util.h"
22 #include "gamma_compression_effect.h"
23 #include "gamma_expansion_effect.h"
26 #include "resource_pool.h"
28 #include "ycbcr_conversion_effect.h"
30 using namespace Eigen;
35 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
36 : aspect_nom(aspect_nom),
37 aspect_denom(aspect_denom),
38 output_color_rgba(false),
39 output_color_ycbcr(false),
42 output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
44 resource_pool(resource_pool),
45 do_phase_timing(false) {
46 if (resource_pool == NULL) {
47 this->resource_pool = new ResourcePool();
48 owns_resource_pool = true;
50 owns_resource_pool = false;
53 // Generate a VBO with some data in (shared position and texture coordinate data).
59 vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
62 EffectChain::~EffectChain()
64 for (unsigned i = 0; i < nodes.size(); ++i) {
65 delete nodes[i]->effect;
68 for (unsigned i = 0; i < phases.size(); ++i) {
69 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
72 if (owns_resource_pool) {
75 glDeleteBuffers(1, &vbo);
79 Input *EffectChain::add_input(Input *input)
82 inputs.push_back(input);
87 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
90 assert(!output_color_rgba);
91 output_format = format;
92 output_alpha_format = alpha_format;
93 output_color_rgba = true;
96 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
97 const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
100 assert(!output_color_ycbcr);
101 output_format = format;
102 output_alpha_format = alpha_format;
103 output_color_ycbcr = true;
104 output_ycbcr_format = ycbcr_format;
105 output_ycbcr_splitting = output_splitting;
107 assert(ycbcr_format.chroma_subsampling_x == 1);
108 assert(ycbcr_format.chroma_subsampling_y == 1);
111 Node *EffectChain::add_node(Effect *effect)
113 for (unsigned i = 0; i < nodes.size(); ++i) {
114 assert(nodes[i]->effect != effect);
117 Node *node = new Node;
118 node->effect = effect;
119 node->disabled = false;
120 node->output_color_space = COLORSPACE_INVALID;
121 node->output_gamma_curve = GAMMA_INVALID;
122 node->output_alpha_type = ALPHA_INVALID;
123 node->needs_mipmaps = false;
124 node->one_to_one_sampling = false;
126 nodes.push_back(node);
127 node_map[effect] = node;
128 effect->inform_added(this);
132 void EffectChain::connect_nodes(Node *sender, Node *receiver)
134 sender->outgoing_links.push_back(receiver);
135 receiver->incoming_links.push_back(sender);
138 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
140 new_receiver->incoming_links = old_receiver->incoming_links;
141 old_receiver->incoming_links.clear();
143 for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
144 Node *sender = new_receiver->incoming_links[i];
145 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
146 if (sender->outgoing_links[j] == old_receiver) {
147 sender->outgoing_links[j] = new_receiver;
153 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
155 new_sender->outgoing_links = old_sender->outgoing_links;
156 old_sender->outgoing_links.clear();
158 for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
159 Node *receiver = new_sender->outgoing_links[i];
160 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
161 if (receiver->incoming_links[j] == old_sender) {
162 receiver->incoming_links[j] = new_sender;
168 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
170 for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
171 if (sender->outgoing_links[i] == receiver) {
172 sender->outgoing_links[i] = middle;
173 middle->incoming_links.push_back(sender);
176 for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
177 if (receiver->incoming_links[i] == sender) {
178 receiver->incoming_links[i] = middle;
179 middle->outgoing_links.push_back(receiver);
183 assert(middle->incoming_links.size() == middle->effect->num_inputs());
186 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
188 assert(node->effect->needs_texture_bounce());
189 assert(input_num < node->incoming_links.size());
190 assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
191 assert(node->incoming_links[input_num]->bound_sampler_num < 8);
192 return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
195 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
197 assert(input_num < node->incoming_links.size());
198 return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
199 node->incoming_links[input_num]->bound_sampler_num < 8;
202 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
204 if (node->output_gamma_curve == GAMMA_LINEAR &&
205 node->effect->effect_type_id() != "GammaCompressionEffect") {
208 if (node->effect->num_inputs() == 0) {
209 nonlinear_inputs->push_back(node);
211 assert(node->effect->num_inputs() == node->incoming_links.size());
212 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
213 find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
218 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
221 assert(inputs.size() == effect->num_inputs());
222 Node *node = add_node(effect);
223 for (unsigned i = 0; i < inputs.size(); ++i) {
224 assert(node_map.count(inputs[i]) != 0);
225 connect_nodes(node_map[inputs[i]], node);
230 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
231 string replace_prefix(const string &text, const string &prefix)
236 while (start < text.size()) {
237 size_t pos = text.find("PREFIX(", start);
238 if (pos == string::npos) {
239 output.append(text.substr(start, string::npos));
243 output.append(text.substr(start, pos - start));
244 output.append(prefix);
247 pos += strlen("PREFIX(");
249 // Output stuff until we find the matching ), which we then eat.
251 size_t end_arg_pos = pos;
252 while (end_arg_pos < text.size()) {
253 if (text[end_arg_pos] == '(') {
255 } else if (text[end_arg_pos] == ')') {
263 output.append(text.substr(pos, end_arg_pos - pos));
274 void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
275 const string &type_specifier,
276 const string &effect_id,
277 vector<Uniform<T> > *phase_uniforms,
280 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
281 phase_uniforms->push_back(effect_uniforms[i]);
282 phase_uniforms->back().prefix = effect_id;
284 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
285 + "_" + effect_uniforms[i].name + ";\n";
290 void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
291 const string &type_specifier,
292 const string &effect_id,
293 vector<Uniform<T> > *phase_uniforms,
296 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
297 phase_uniforms->push_back(effect_uniforms[i]);
298 phase_uniforms->back().prefix = effect_id;
301 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
302 type_specifier.c_str(), effect_id.c_str(),
303 effect_uniforms[i].name.c_str(),
304 int(effect_uniforms[i].num_values));
310 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
312 for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
313 Uniform<T> &uniform = (*phase_uniforms)[i];
314 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
320 void EffectChain::compile_glsl_program(Phase *phase)
322 string frag_shader_header = read_version_dependent_file("header", "frag");
323 string frag_shader = "";
325 // Create functions and uniforms for all the texture inputs that we need.
326 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
327 Node *input = phase->inputs[i]->output_node;
329 sprintf(effect_id, "in%u", i);
330 phase->effect_ids.insert(make_pair(input, effect_id));
332 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
333 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
334 frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
335 frag_shader += "}\n";
338 Uniform<int> uniform;
339 uniform.name = effect_id;
340 uniform.value = &phase->input_samplers[i];
341 uniform.prefix = "tex";
342 uniform.num_values = 1;
343 uniform.location = -1;
344 phase->uniforms_sampler2d.push_back(uniform);
347 // Give each effect in the phase its own ID.
348 for (unsigned i = 0; i < phase->effects.size(); ++i) {
349 Node *node = phase->effects[i];
351 sprintf(effect_id, "eff%u", i);
352 phase->effect_ids.insert(make_pair(node, effect_id));
355 for (unsigned i = 0; i < phase->effects.size(); ++i) {
356 Node *node = phase->effects[i];
357 const string effect_id = phase->effect_ids[node];
358 if (node->incoming_links.size() == 1) {
359 frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
361 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
363 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
369 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
370 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
371 frag_shader += "#undef PREFIX\n";
372 frag_shader += "#undef FUNCNAME\n";
373 if (node->incoming_links.size() == 1) {
374 frag_shader += "#undef INPUT\n";
376 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
378 sprintf(buf, "#undef INPUT%d\n", j + 1);
384 frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
386 // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
387 vector<string> frag_shader_outputs; // In order.
388 if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) {
389 switch (output_ycbcr_splitting) {
390 case YCBCR_OUTPUT_INTERLEAVED:
392 frag_shader_outputs.push_back("FragColor");
394 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
395 frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
396 frag_shader_outputs.push_back("Y");
397 frag_shader_outputs.push_back("Chroma");
399 case YCBCR_OUTPUT_PLANAR:
400 frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
401 frag_shader_outputs.push_back("Y");
402 frag_shader_outputs.push_back("Cb");
403 frag_shader_outputs.push_back("Cr");
409 if (output_color_rgba) {
410 // Note: Needs to come in the header, because not only the
411 // output needs to see it (YCbCrConversionEffect and DitherEffect
413 frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
414 frag_shader_outputs.push_back("RGBA");
417 frag_shader.append(read_file("footer.frag"));
419 // Collect uniforms from all effects and output them. Note that this needs
420 // to happen after output_fragment_shader(), even though the uniforms come
421 // before in the output source, since output_fragment_shader() is allowed
422 // to register new uniforms (e.g. arrays that are of unknown length until
423 // finalization time).
424 // TODO: Make a uniform block for platforms that support it.
425 string frag_shader_uniforms = "";
426 for (unsigned i = 0; i < phase->effects.size(); ++i) {
427 Node *node = phase->effects[i];
428 Effect *effect = node->effect;
429 const string effect_id = phase->effect_ids[node];
430 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
431 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
432 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
433 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
434 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
435 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
436 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
437 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
438 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
439 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
440 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
441 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
444 frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
446 string vert_shader = read_version_dependent_file("vs", "vert");
448 // If we're the last phase and need to flip the picture to compensate for
449 // the origin, tell the vertex shader so.
450 if (phase->output_node->outgoing_links.empty() && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
451 const string needle = "#define FLIP_ORIGIN 0";
452 size_t pos = vert_shader.find(needle);
453 assert(pos != string::npos);
455 vert_shader[pos + needle.size() - 1] = '1';
458 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
459 GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
460 GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
461 if (position_attribute_index != -1) {
462 phase->attribute_indexes.insert(position_attribute_index);
464 if (texcoord_attribute_index != -1) {
465 phase->attribute_indexes.insert(texcoord_attribute_index);
468 // Collect the resulting location numbers for each uniform.
469 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
470 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
471 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
472 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
473 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
474 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
475 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
476 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
479 // Construct GLSL programs, starting at the given effect and following
480 // the chain from there. We end a program every time we come to an effect
481 // marked as "needs texture bounce", one that is used by multiple other
482 // effects, every time we need to bounce due to output size change
483 // (not all size changes require ending), and of course at the end.
485 // We follow a quite simple depth-first search from the output, although
486 // without recursing explicitly within each phase.
487 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
489 if (completed_effects->count(output)) {
490 return (*completed_effects)[output];
493 Phase *phase = new Phase;
494 phase->output_node = output;
496 // If the output effect has one-to-one sampling, we try to trace this
497 // status down through the dependency chain. This is important in case
498 // we hit an effect that changes output size (and not sets a virtual
499 // output size); if we have one-to-one sampling, we don't have to break
501 output->one_to_one_sampling = output->effect->one_to_one_sampling();
503 // Effects that we have yet to calculate, but that we know should
504 // be in the current phase.
505 stack<Node *> effects_todo_this_phase;
506 effects_todo_this_phase.push(output);
508 while (!effects_todo_this_phase.empty()) {
509 Node *node = effects_todo_this_phase.top();
510 effects_todo_this_phase.pop();
512 if (node->effect->needs_mipmaps()) {
513 node->needs_mipmaps = true;
516 // This should currently only happen for effects that are inputs
517 // (either true inputs or phase outputs). We special-case inputs,
518 // and then deduplicate phase outputs below.
519 if (node->effect->num_inputs() == 0) {
520 if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
524 assert(completed_effects->count(node) == 0);
527 phase->effects.push_back(node);
529 // Find all the dependencies of this effect, and add them to the stack.
530 vector<Node *> deps = node->incoming_links;
531 assert(node->effect->num_inputs() == deps.size());
532 for (unsigned i = 0; i < deps.size(); ++i) {
533 bool start_new_phase = false;
535 if (node->effect->needs_texture_bounce() &&
536 !deps[i]->effect->is_single_texture() &&
537 !deps[i]->effect->override_disable_bounce()) {
538 start_new_phase = true;
541 // Propagate information about needing mipmaps down the chain,
542 // breaking the phase if we notice an incompatibility.
544 // Note that we cannot do this propagation as a normal pass,
545 // because it needs information about where the phases end
546 // (we should not propagate the flag across phases).
547 if (node->needs_mipmaps) {
548 if (deps[i]->effect->num_inputs() == 0) {
549 Input *input = static_cast<Input *>(deps[i]->effect);
550 start_new_phase |= !input->can_supply_mipmaps();
552 deps[i]->needs_mipmaps = true;
556 if (deps[i]->outgoing_links.size() > 1) {
557 if (!deps[i]->effect->is_single_texture()) {
558 // More than one effect uses this as the input,
559 // and it is not a texture itself.
560 // The easiest thing to do (and probably also the safest
561 // performance-wise in most cases) is to bounce it to a texture
562 // and then let the next passes read from that.
563 start_new_phase = true;
565 assert(deps[i]->effect->num_inputs() == 0);
567 // For textures, we try to be slightly more clever;
568 // if none of our outputs need a bounce, we don't bounce
569 // but instead simply use the effect many times.
571 // Strictly speaking, we could bounce it for some outputs
572 // and use it directly for others, but the processing becomes
573 // somewhat simpler if the effect is only used in one such way.
574 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
575 Node *rdep = deps[i]->outgoing_links[j];
576 start_new_phase |= rdep->effect->needs_texture_bounce();
581 if (deps[i]->effect->sets_virtual_output_size()) {
582 assert(deps[i]->effect->changes_output_size());
583 // If the next effect sets a virtual size to rely on OpenGL's
584 // bilinear sampling, we'll really need to break the phase here.
585 start_new_phase = true;
586 } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
587 // If the next effect changes size and we don't have one-to-one sampling,
588 // we also need to break here.
589 start_new_phase = true;
592 if (start_new_phase) {
593 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
595 effects_todo_this_phase.push(deps[i]);
597 // Propagate the one-to-one status down through the dependency.
598 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
599 deps[i]->effect->one_to_one_sampling();
604 // No more effects to do this phase. Take all the ones we have,
605 // and create a GLSL program for it.
606 assert(!phase->effects.empty());
608 // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
609 // that would be nondeterministic and thus reduce cacheability.
610 // TODO: Make this even more deterministic.
611 vector<Phase *> dedup_inputs;
612 set<Phase *> seen_inputs;
613 for (size_t i = 0; i < phase->inputs.size(); ++i) {
614 if (seen_inputs.insert(phase->inputs[i]).second) {
615 dedup_inputs.push_back(phase->inputs[i]);
618 swap(phase->inputs, dedup_inputs);
620 // Allocate samplers for each input.
621 phase->input_samplers.resize(phase->inputs.size());
623 // We added the effects from the output and back, but we need to output
624 // them in topological sort order in the shader.
625 phase->effects = topological_sort(phase->effects);
627 // Figure out if we need mipmaps or not, and if so, tell the inputs that.
628 phase->input_needs_mipmaps = false;
629 for (unsigned i = 0; i < phase->effects.size(); ++i) {
630 Node *node = phase->effects[i];
631 phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
633 for (unsigned i = 0; i < phase->effects.size(); ++i) {
634 Node *node = phase->effects[i];
635 if (node->effect->num_inputs() == 0) {
636 Input *input = static_cast<Input *>(node->effect);
637 assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
638 CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
642 // Tell each node which phase it ended up in, so that the unit test
643 // can check that the phases were split in the right place.
644 // Note that this ignores that effects may be part of multiple phases;
645 // if the unit tests need to test such cases, we'll reconsider.
646 for (unsigned i = 0; i < phase->effects.size(); ++i) {
647 phase->effects[i]->containing_phase = phase;
650 // Actually make the shader for this phase.
651 compile_glsl_program(phase);
653 // Initialize timers.
654 if (movit_timer_queries_supported) {
655 phase->time_elapsed_ns = 0;
656 phase->num_measured_iterations = 0;
659 assert(completed_effects->count(output) == 0);
660 completed_effects->insert(make_pair(output, phase));
661 phases.push_back(phase);
665 void EffectChain::output_dot(const char *filename)
667 if (movit_debug_level != MOVIT_DEBUG_ON) {
671 FILE *fp = fopen(filename, "w");
677 fprintf(fp, "digraph G {\n");
678 fprintf(fp, " output [shape=box label=\"(output)\"];\n");
679 for (unsigned i = 0; i < nodes.size(); ++i) {
680 // Find out which phase this event belongs to.
681 vector<int> in_phases;
682 for (unsigned j = 0; j < phases.size(); ++j) {
683 const Phase* p = phases[j];
684 if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
685 in_phases.push_back(j);
689 if (in_phases.empty()) {
690 fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
691 } else if (in_phases.size() == 1) {
692 fprintf(fp, " n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
693 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
694 (in_phases[0] % 8) + 1);
696 // If we had new enough Graphviz, style="wedged" would probably be ideal here.
698 fprintf(fp, " n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
699 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
700 (in_phases[0] % 8) + 1);
703 char from_node_id[256];
704 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
706 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
707 char to_node_id[256];
708 snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
710 vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
711 output_dot_edge(fp, from_node_id, to_node_id, labels);
714 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
716 vector<string> labels = get_labels_for_edge(nodes[i], NULL);
717 output_dot_edge(fp, from_node_id, "output", labels);
725 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
727 vector<string> labels;
729 if (to != NULL && to->effect->needs_texture_bounce()) {
730 labels.push_back("needs_bounce");
732 if (from->effect->changes_output_size()) {
733 labels.push_back("resize");
736 switch (from->output_color_space) {
737 case COLORSPACE_INVALID:
738 labels.push_back("spc[invalid]");
740 case COLORSPACE_REC_601_525:
741 labels.push_back("spc[rec601-525]");
743 case COLORSPACE_REC_601_625:
744 labels.push_back("spc[rec601-625]");
750 switch (from->output_gamma_curve) {
752 labels.push_back("gamma[invalid]");
755 labels.push_back("gamma[sRGB]");
757 case GAMMA_REC_601: // and GAMMA_REC_709
758 labels.push_back("gamma[rec601/709]");
764 switch (from->output_alpha_type) {
766 labels.push_back("alpha[invalid]");
769 labels.push_back("alpha[blank]");
771 case ALPHA_POSTMULTIPLIED:
772 labels.push_back("alpha[postmult]");
781 void EffectChain::output_dot_edge(FILE *fp,
782 const string &from_node_id,
783 const string &to_node_id,
784 const vector<string> &labels)
786 if (labels.empty()) {
787 fprintf(fp, " %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
789 string label = labels[0];
790 for (unsigned k = 1; k < labels.size(); ++k) {
791 label += ", " + labels[k];
793 fprintf(fp, " %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
797 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
799 unsigned scaled_width, scaled_height;
801 if (float(width) * aspect_denom >= float(height) * aspect_nom) {
802 // Same aspect, or W/H > aspect (image is wider than the frame).
803 // In either case, keep width, and adjust height.
804 scaled_width = width;
805 scaled_height = lrintf(width * aspect_denom / aspect_nom);
807 // W/H < aspect (image is taller than the frame), so keep height,
809 scaled_width = lrintf(height * aspect_nom / aspect_denom);
810 scaled_height = height;
813 // We should be consistently larger or smaller then the existing choice,
814 // since we have the same aspect.
815 assert(!(scaled_width < *output_width && scaled_height > *output_height));
816 assert(!(scaled_height < *output_height && scaled_width > *output_width));
818 if (scaled_width >= *output_width && scaled_height >= *output_height) {
819 *output_width = scaled_width;
820 *output_height = scaled_height;
824 // Propagate input texture sizes throughout, and inform effects downstream.
825 // (Like a lot of other code, we depend on effects being in topological order.)
826 void EffectChain::inform_input_sizes(Phase *phase)
828 // All effects that have a defined size (inputs and RTT inputs)
829 // get that. Reset all others.
830 for (unsigned i = 0; i < phase->effects.size(); ++i) {
831 Node *node = phase->effects[i];
832 if (node->effect->num_inputs() == 0) {
833 Input *input = static_cast<Input *>(node->effect);
834 node->output_width = input->get_width();
835 node->output_height = input->get_height();
836 assert(node->output_width != 0);
837 assert(node->output_height != 0);
839 node->output_width = node->output_height = 0;
842 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
843 Phase *input = phase->inputs[i];
844 input->output_node->output_width = input->virtual_output_width;
845 input->output_node->output_height = input->virtual_output_height;
846 assert(input->output_node->output_width != 0);
847 assert(input->output_node->output_height != 0);
850 // Now propagate from the inputs towards the end, and inform as we go.
851 // The rules are simple:
853 // 1. Don't touch effects that already have given sizes (ie., inputs
854 // or effects that change the output size).
855 // 2. If all of your inputs have the same size, that will be your output size.
856 // 3. Otherwise, your output size is 0x0.
857 for (unsigned i = 0; i < phase->effects.size(); ++i) {
858 Node *node = phase->effects[i];
859 if (node->effect->num_inputs() == 0) {
862 unsigned this_output_width = 0;
863 unsigned this_output_height = 0;
864 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
865 Node *input = node->incoming_links[j];
866 node->effect->inform_input_size(j, input->output_width, input->output_height);
868 this_output_width = input->output_width;
869 this_output_height = input->output_height;
870 } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
872 this_output_width = 0;
873 this_output_height = 0;
876 if (node->effect->changes_output_size()) {
877 // We cannot call get_output_size() before we've done inform_input_size()
879 unsigned real_width, real_height;
880 node->effect->get_output_size(&real_width, &real_height,
881 &node->output_width, &node->output_height);
882 assert(node->effect->sets_virtual_output_size() ||
883 (real_width == node->output_width &&
884 real_height == node->output_height));
886 node->output_width = this_output_width;
887 node->output_height = this_output_height;
892 // Note: You should call inform_input_sizes() before this, as the last effect's
893 // desired output size might change based on the inputs.
894 void EffectChain::find_output_size(Phase *phase)
896 Node *output_node = phase->effects.back();
898 // If the last effect explicitly sets an output size, use that.
899 if (output_node->effect->changes_output_size()) {
900 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
901 &phase->virtual_output_width, &phase->virtual_output_height);
902 assert(output_node->effect->sets_virtual_output_size() ||
903 (phase->output_width == phase->virtual_output_width &&
904 phase->output_height == phase->virtual_output_height));
908 // If all effects have the same size, use that.
909 unsigned output_width = 0, output_height = 0;
910 bool all_inputs_same_size = true;
912 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
913 Phase *input = phase->inputs[i];
914 assert(input->output_width != 0);
915 assert(input->output_height != 0);
916 if (output_width == 0 && output_height == 0) {
917 output_width = input->virtual_output_width;
918 output_height = input->virtual_output_height;
919 } else if (output_width != input->virtual_output_width ||
920 output_height != input->virtual_output_height) {
921 all_inputs_same_size = false;
924 for (unsigned i = 0; i < phase->effects.size(); ++i) {
925 Effect *effect = phase->effects[i]->effect;
926 if (effect->num_inputs() != 0) {
930 Input *input = static_cast<Input *>(effect);
931 if (output_width == 0 && output_height == 0) {
932 output_width = input->get_width();
933 output_height = input->get_height();
934 } else if (output_width != input->get_width() ||
935 output_height != input->get_height()) {
936 all_inputs_same_size = false;
940 if (all_inputs_same_size) {
941 assert(output_width != 0);
942 assert(output_height != 0);
943 phase->virtual_output_width = phase->output_width = output_width;
944 phase->virtual_output_height = phase->output_height = output_height;
948 // If not, fit all the inputs into the current aspect, and select the largest one.
951 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
952 Phase *input = phase->inputs[i];
953 assert(input->output_width != 0);
954 assert(input->output_height != 0);
955 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
957 for (unsigned i = 0; i < phase->effects.size(); ++i) {
958 Effect *effect = phase->effects[i]->effect;
959 if (effect->num_inputs() != 0) {
963 Input *input = static_cast<Input *>(effect);
964 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
966 assert(output_width != 0);
967 assert(output_height != 0);
968 phase->virtual_output_width = phase->output_width = output_width;
969 phase->virtual_output_height = phase->output_height = output_height;
972 void EffectChain::sort_all_nodes_topologically()
974 nodes = topological_sort(nodes);
977 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
979 set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
980 vector<Node *> sorted_list;
981 for (unsigned i = 0; i < nodes.size(); ++i) {
982 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
984 reverse(sorted_list.begin(), sorted_list.end());
988 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
990 if (nodes_left_to_visit->count(node) == 0) {
993 nodes_left_to_visit->erase(node);
994 for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
995 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
997 sorted_list->push_back(node);
1000 void EffectChain::find_color_spaces_for_inputs()
1002 for (unsigned i = 0; i < nodes.size(); ++i) {
1003 Node *node = nodes[i];
1004 if (node->disabled) {
1007 if (node->incoming_links.size() == 0) {
1008 Input *input = static_cast<Input *>(node->effect);
1009 node->output_color_space = input->get_color_space();
1010 node->output_gamma_curve = input->get_gamma_curve();
1012 Effect::AlphaHandling alpha_handling = input->alpha_handling();
1013 switch (alpha_handling) {
1014 case Effect::OUTPUT_BLANK_ALPHA:
1015 node->output_alpha_type = ALPHA_BLANK;
1017 case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1018 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1020 case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1021 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1023 case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1024 case Effect::DONT_CARE_ALPHA_TYPE:
1029 if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1030 assert(node->output_gamma_curve == GAMMA_LINEAR);
1036 // Propagate gamma and color space information as far as we can in the graph.
1037 // The rules are simple: Anything where all the inputs agree, get that as
1038 // output as well. Anything else keeps having *_INVALID.
1039 void EffectChain::propagate_gamma_and_color_space()
1041 // We depend on going through the nodes in order.
1042 sort_all_nodes_topologically();
1044 for (unsigned i = 0; i < nodes.size(); ++i) {
1045 Node *node = nodes[i];
1046 if (node->disabled) {
1049 assert(node->incoming_links.size() == node->effect->num_inputs());
1050 if (node->incoming_links.size() == 0) {
1051 assert(node->output_color_space != COLORSPACE_INVALID);
1052 assert(node->output_gamma_curve != GAMMA_INVALID);
1056 Colorspace color_space = node->incoming_links[0]->output_color_space;
1057 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1058 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1059 if (node->incoming_links[j]->output_color_space != color_space) {
1060 color_space = COLORSPACE_INVALID;
1062 if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1063 gamma_curve = GAMMA_INVALID;
1067 // The conversion effects already have their outputs set correctly,
1068 // so leave them alone.
1069 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1070 node->output_color_space = color_space;
1072 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1073 node->effect->effect_type_id() != "GammaExpansionEffect") {
1074 node->output_gamma_curve = gamma_curve;
1079 // Propagate alpha information as far as we can in the graph.
1080 // Similar to propagate_gamma_and_color_space().
1081 void EffectChain::propagate_alpha()
1083 // We depend on going through the nodes in order.
1084 sort_all_nodes_topologically();
1086 for (unsigned i = 0; i < nodes.size(); ++i) {
1087 Node *node = nodes[i];
1088 if (node->disabled) {
1091 assert(node->incoming_links.size() == node->effect->num_inputs());
1092 if (node->incoming_links.size() == 0) {
1093 assert(node->output_alpha_type != ALPHA_INVALID);
1097 // The alpha multiplication/division effects are special cases.
1098 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1099 assert(node->incoming_links.size() == 1);
1100 assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1101 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1104 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1105 assert(node->incoming_links.size() == 1);
1106 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1107 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1111 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1112 // because they are the only one that _need_ postmultiplied alpha.
1113 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1114 node->effect->effect_type_id() == "GammaExpansionEffect") {
1115 assert(node->incoming_links.size() == 1);
1116 if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1117 node->output_alpha_type = ALPHA_BLANK;
1118 } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1119 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1121 node->output_alpha_type = ALPHA_INVALID;
1126 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1127 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1128 // taken care of above. Rationale: Even if you could imagine
1129 // e.g. an effect that took in an image and set alpha=1.0
1130 // unconditionally, it wouldn't make any sense to have it as
1131 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1132 // got its input pre- or postmultiplied, so it wouldn't know
1133 // whether to divide away the old alpha or not.
1134 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1135 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1136 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1137 alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1139 // If the node has multiple inputs, check that they are all valid and
1141 bool any_invalid = false;
1142 bool any_premultiplied = false;
1143 bool any_postmultiplied = false;
1145 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1146 switch (node->incoming_links[j]->output_alpha_type) {
1151 // Blank is good as both pre- and postmultiplied alpha,
1152 // so just ignore it.
1154 case ALPHA_PREMULTIPLIED:
1155 any_premultiplied = true;
1157 case ALPHA_POSTMULTIPLIED:
1158 any_postmultiplied = true;
1166 node->output_alpha_type = ALPHA_INVALID;
1170 // Inputs must be of the same type.
1171 if (any_premultiplied && any_postmultiplied) {
1172 node->output_alpha_type = ALPHA_INVALID;
1176 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1177 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1178 // If the effect has asked for premultiplied alpha, check that it has got it.
1179 if (any_postmultiplied) {
1180 node->output_alpha_type = ALPHA_INVALID;
1181 } else if (!any_premultiplied &&
1182 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1183 // Blank input alpha, and the effect preserves blank alpha.
1184 node->output_alpha_type = ALPHA_BLANK;
1186 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1189 // OK, all inputs are the same, and this effect is not going
1191 assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1192 if (any_premultiplied) {
1193 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1194 } else if (any_postmultiplied) {
1195 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1197 node->output_alpha_type = ALPHA_BLANK;
1203 bool EffectChain::node_needs_colorspace_fix(Node *node)
1205 if (node->disabled) {
1208 if (node->effect->num_inputs() == 0) {
1212 // propagate_gamma_and_color_space() has already set our output
1213 // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1214 if (node->output_color_space == COLORSPACE_INVALID) {
1217 return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1220 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1221 // the graph. Our strategy is not always optimal, but quite simple:
1222 // Find an effect that's as early as possible where the inputs are of
1223 // unacceptable colorspaces (that is, either different, or, if the effect only
1224 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1225 // propagate the information anew, and repeat until there are no more such
1227 void EffectChain::fix_internal_color_spaces()
1229 unsigned colorspace_propagation_pass = 0;
1233 for (unsigned i = 0; i < nodes.size(); ++i) {
1234 Node *node = nodes[i];
1235 if (!node_needs_colorspace_fix(node)) {
1239 // Go through each input that is not sRGB, and insert
1240 // a colorspace conversion after it.
1241 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1242 Node *input = node->incoming_links[j];
1243 assert(input->output_color_space != COLORSPACE_INVALID);
1244 if (input->output_color_space == COLORSPACE_sRGB) {
1247 Node *conversion = add_node(new ColorspaceConversionEffect());
1248 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1249 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1250 conversion->output_color_space = COLORSPACE_sRGB;
1251 replace_sender(input, conversion);
1252 connect_nodes(input, conversion);
1255 // Re-sort topologically, and propagate the new information.
1256 propagate_gamma_and_color_space();
1263 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1264 output_dot(filename);
1265 assert(colorspace_propagation_pass < 100);
1266 } while (found_any);
1268 for (unsigned i = 0; i < nodes.size(); ++i) {
1269 Node *node = nodes[i];
1270 if (node->disabled) {
1273 assert(node->output_color_space != COLORSPACE_INVALID);
1277 bool EffectChain::node_needs_alpha_fix(Node *node)
1279 if (node->disabled) {
1283 // propagate_alpha() has already set our output to ALPHA_INVALID if the
1284 // inputs differ or we are otherwise in mismatch, so we can rely on that.
1285 return (node->output_alpha_type == ALPHA_INVALID);
1288 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1289 // the graph. Similar to fix_internal_color_spaces().
1290 void EffectChain::fix_internal_alpha(unsigned step)
1292 unsigned alpha_propagation_pass = 0;
1296 for (unsigned i = 0; i < nodes.size(); ++i) {
1297 Node *node = nodes[i];
1298 if (!node_needs_alpha_fix(node)) {
1302 // If we need to fix up GammaExpansionEffect, then clearly something
1303 // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1305 assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1307 AlphaType desired_type = ALPHA_PREMULTIPLIED;
1309 // GammaCompressionEffect is special; it needs postmultiplied alpha.
1310 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1311 assert(node->incoming_links.size() == 1);
1312 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1313 desired_type = ALPHA_POSTMULTIPLIED;
1316 // Go through each input that is not premultiplied alpha, and insert
1317 // a conversion before it.
1318 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1319 Node *input = node->incoming_links[j];
1320 assert(input->output_alpha_type != ALPHA_INVALID);
1321 if (input->output_alpha_type == desired_type ||
1322 input->output_alpha_type == ALPHA_BLANK) {
1326 if (desired_type == ALPHA_PREMULTIPLIED) {
1327 conversion = add_node(new AlphaMultiplicationEffect());
1329 conversion = add_node(new AlphaDivisionEffect());
1331 conversion->output_alpha_type = desired_type;
1332 replace_sender(input, conversion);
1333 connect_nodes(input, conversion);
1336 // Re-sort topologically, and propagate the new information.
1337 propagate_gamma_and_color_space();
1345 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1346 output_dot(filename);
1347 assert(alpha_propagation_pass < 100);
1348 } while (found_any);
1350 for (unsigned i = 0; i < nodes.size(); ++i) {
1351 Node *node = nodes[i];
1352 if (node->disabled) {
1355 assert(node->output_alpha_type != ALPHA_INVALID);
1359 // Make so that the output is in the desired color space.
1360 void EffectChain::fix_output_color_space()
1362 Node *output = find_output_node();
1363 if (output->output_color_space != output_format.color_space) {
1364 Node *conversion = add_node(new ColorspaceConversionEffect());
1365 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1366 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1367 conversion->output_color_space = output_format.color_space;
1368 connect_nodes(output, conversion);
1370 propagate_gamma_and_color_space();
1374 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1375 void EffectChain::fix_output_alpha()
1377 Node *output = find_output_node();
1378 assert(output->output_alpha_type != ALPHA_INVALID);
1379 if (output->output_alpha_type == ALPHA_BLANK) {
1380 // No alpha output, so we don't care.
1383 if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1384 output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1385 Node *conversion = add_node(new AlphaDivisionEffect());
1386 connect_nodes(output, conversion);
1388 propagate_gamma_and_color_space();
1390 if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1391 output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1392 Node *conversion = add_node(new AlphaMultiplicationEffect());
1393 connect_nodes(output, conversion);
1395 propagate_gamma_and_color_space();
1399 bool EffectChain::node_needs_gamma_fix(Node *node)
1401 if (node->disabled) {
1405 // Small hack since the output is not an explicit node:
1406 // If we are the last node and our output is in the wrong
1407 // space compared to EffectChain's output, we need to fix it.
1408 // This will only take us to linear, but fix_output_gamma()
1409 // will come and take us to the desired output gamma
1412 // This needs to be before everything else, since it could
1413 // even apply to inputs (if they are the only effect).
1414 if (node->outgoing_links.empty() &&
1415 node->output_gamma_curve != output_format.gamma_curve &&
1416 node->output_gamma_curve != GAMMA_LINEAR) {
1420 if (node->effect->num_inputs() == 0) {
1424 // propagate_gamma_and_color_space() has already set our output
1425 // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1426 // except for GammaCompressionEffect.
1427 if (node->output_gamma_curve == GAMMA_INVALID) {
1430 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1431 assert(node->incoming_links.size() == 1);
1432 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1435 return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1438 // Very similar to fix_internal_color_spaces(), but for gamma.
1439 // There is one difference, though; before we start adding conversion nodes,
1440 // we see if we can get anything out of asking the sources to deliver
1441 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1442 // does that part, while fix_internal_gamma_by_inserting_nodes()
1443 // inserts nodes as needed afterwards.
1444 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1446 unsigned gamma_propagation_pass = 0;
1450 for (unsigned i = 0; i < nodes.size(); ++i) {
1451 Node *node = nodes[i];
1452 if (!node_needs_gamma_fix(node)) {
1456 // See if all inputs can give us linear gamma. If not, leave it.
1457 vector<Node *> nonlinear_inputs;
1458 find_all_nonlinear_inputs(node, &nonlinear_inputs);
1459 assert(!nonlinear_inputs.empty());
1462 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1463 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1464 all_ok &= input->can_output_linear_gamma();
1471 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1472 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1473 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1476 // Re-sort topologically, and propagate the new information.
1477 propagate_gamma_and_color_space();
1484 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1485 output_dot(filename);
1486 assert(gamma_propagation_pass < 100);
1487 } while (found_any);
1490 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1492 unsigned gamma_propagation_pass = 0;
1496 for (unsigned i = 0; i < nodes.size(); ++i) {
1497 Node *node = nodes[i];
1498 if (!node_needs_gamma_fix(node)) {
1502 // Special case: We could be an input and still be asked to
1503 // fix our gamma; if so, we should be the only node
1504 // (as node_needs_gamma_fix() would only return true in
1505 // for an input in that case). That means we should insert
1506 // a conversion node _after_ ourselves.
1507 if (node->incoming_links.empty()) {
1508 assert(node->outgoing_links.empty());
1509 Node *conversion = add_node(new GammaExpansionEffect());
1510 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1511 conversion->output_gamma_curve = GAMMA_LINEAR;
1512 connect_nodes(node, conversion);
1515 // If not, go through each input that is not linear gamma,
1516 // and insert a gamma conversion after it.
1517 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1518 Node *input = node->incoming_links[j];
1519 assert(input->output_gamma_curve != GAMMA_INVALID);
1520 if (input->output_gamma_curve == GAMMA_LINEAR) {
1523 Node *conversion = add_node(new GammaExpansionEffect());
1524 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1525 conversion->output_gamma_curve = GAMMA_LINEAR;
1526 replace_sender(input, conversion);
1527 connect_nodes(input, conversion);
1530 // Re-sort topologically, and propagate the new information.
1532 propagate_gamma_and_color_space();
1539 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1540 output_dot(filename);
1541 assert(gamma_propagation_pass < 100);
1542 } while (found_any);
1544 for (unsigned i = 0; i < nodes.size(); ++i) {
1545 Node *node = nodes[i];
1546 if (node->disabled) {
1549 assert(node->output_gamma_curve != GAMMA_INVALID);
1553 // Make so that the output is in the desired gamma.
1554 // Note that this assumes linear input gamma, so it might create the need
1555 // for another pass of fix_internal_gamma().
1556 void EffectChain::fix_output_gamma()
1558 Node *output = find_output_node();
1559 if (output->output_gamma_curve != output_format.gamma_curve) {
1560 Node *conversion = add_node(new GammaCompressionEffect());
1561 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1562 conversion->output_gamma_curve = output_format.gamma_curve;
1563 connect_nodes(output, conversion);
1567 // If the user has requested Y'CbCr output, we need to do this conversion
1568 // _after_ GammaCompressionEffect etc., but before dither (see below).
1569 // This is because Y'CbCr, with the exception of a special optional mode
1570 // in Rec. 2020 (which we currently don't support), is defined to work on
1571 // gamma-encoded data.
1572 void EffectChain::add_ycbcr_conversion_if_needed()
1574 assert(output_color_rgba || output_color_ycbcr);
1575 if (!output_color_ycbcr) {
1578 Node *output = find_output_node();
1579 Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
1580 connect_nodes(output, ycbcr);
1583 // If the user has requested dither, add a DitherEffect right at the end
1584 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1585 // since dither is about the only effect that can _not_ be done in linear space.
1586 void EffectChain::add_dither_if_needed()
1588 if (num_dither_bits == 0) {
1591 Node *output = find_output_node();
1592 Node *dither = add_node(new DitherEffect());
1593 CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1594 connect_nodes(output, dither);
1596 dither_effect = dither->effect;
1599 // Find the output node. This is, simply, one that has no outgoing links.
1600 // If there are multiple ones, the graph is malformed (we do not support
1601 // multiple outputs right now).
1602 Node *EffectChain::find_output_node()
1604 vector<Node *> output_nodes;
1605 for (unsigned i = 0; i < nodes.size(); ++i) {
1606 Node *node = nodes[i];
1607 if (node->disabled) {
1610 if (node->outgoing_links.empty()) {
1611 output_nodes.push_back(node);
1614 assert(output_nodes.size() == 1);
1615 return output_nodes[0];
1618 void EffectChain::finalize()
1620 // Output the graph as it is before we do any conversions on it.
1621 output_dot("step0-start.dot");
1623 // Give each effect in turn a chance to rewrite its own part of the graph.
1624 // Note that if more effects are added as part of this, they will be
1625 // picked up as part of the same for loop, since they are added at the end.
1626 for (unsigned i = 0; i < nodes.size(); ++i) {
1627 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1629 output_dot("step1-rewritten.dot");
1631 find_color_spaces_for_inputs();
1632 output_dot("step2-input-colorspace.dot");
1635 output_dot("step3-propagated-alpha.dot");
1637 propagate_gamma_and_color_space();
1638 output_dot("step4-propagated-all.dot");
1640 fix_internal_color_spaces();
1641 fix_internal_alpha(6);
1642 fix_output_color_space();
1643 output_dot("step7-output-colorspacefix.dot");
1645 output_dot("step8-output-alphafix.dot");
1647 // Note that we need to fix gamma after colorspace conversion,
1648 // because colorspace conversions might create needs for gamma conversions.
1649 // Also, we need to run an extra pass of fix_internal_gamma() after
1650 // fixing the output gamma, as we only have conversions to/from linear,
1651 // and fix_internal_alpha() since GammaCompressionEffect needs
1652 // postmultiplied input.
1653 fix_internal_gamma_by_asking_inputs(9);
1654 fix_internal_gamma_by_inserting_nodes(10);
1656 output_dot("step11-output-gammafix.dot");
1658 output_dot("step12-output-alpha-propagated.dot");
1659 fix_internal_alpha(13);
1660 output_dot("step14-output-alpha-fixed.dot");
1661 fix_internal_gamma_by_asking_inputs(15);
1662 fix_internal_gamma_by_inserting_nodes(16);
1664 output_dot("step17-before-ycbcr.dot");
1665 add_ycbcr_conversion_if_needed();
1667 output_dot("step18-before-dither.dot");
1668 add_dither_if_needed();
1670 output_dot("step19-final.dot");
1672 // Construct all needed GLSL programs, starting at the output.
1673 // We need to keep track of which effects have already been computed,
1674 // as an effect with multiple users could otherwise be calculated
1676 map<Node *, Phase *> completed_effects;
1677 construct_phase(find_output_node(), &completed_effects);
1679 output_dot("step20-split-to-phases.dot");
1681 assert(phases[0]->inputs.empty());
1686 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1690 // This needs to be set anew, in case we are coming from a different context
1691 // from when we initialized.
1693 glDisable(GL_DITHER);
1696 // Save original viewport.
1697 GLuint x = 0, y = 0;
1699 if (width == 0 && height == 0) {
1701 glGetIntegerv(GL_VIEWPORT, viewport);
1704 width = viewport[2];
1705 height = viewport[3];
1710 glDisable(GL_BLEND);
1712 glDisable(GL_DEPTH_TEST);
1714 glDepthMask(GL_FALSE);
1717 // Generate a VAO that will be used during the entire execution,
1718 // and bind the VBO, since it contains all the data.
1720 glGenVertexArrays(1, &vao);
1722 glBindVertexArray(vao);
1724 glBindBuffer(GL_ARRAY_BUFFER, vbo);
1726 set<GLint> bound_attribute_indices;
1728 set<Phase *> generated_mipmaps;
1730 // We choose the simplest option of having one texture per output,
1731 // since otherwise this turns into an (albeit simple) register allocation problem.
1732 map<Phase *, GLuint> output_textures;
1734 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1735 Phase *phase = phases[phase_num];
1737 if (do_phase_timing) {
1738 GLuint timer_query_object;
1739 if (phase->timer_query_objects_free.empty()) {
1740 glGenQueries(1, &timer_query_object);
1742 timer_query_object = phase->timer_query_objects_free.front();
1743 phase->timer_query_objects_free.pop_front();
1745 glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
1746 phase->timer_query_objects_running.push_back(timer_query_object);
1748 if (phase_num == phases.size() - 1) {
1749 // Last phase goes to the output the user specified.
1750 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1752 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1753 assert(status == GL_FRAMEBUFFER_COMPLETE);
1754 glViewport(x, y, width, height);
1755 if (dither_effect != NULL) {
1756 CHECK(dither_effect->set_int("output_width", width));
1757 CHECK(dither_effect->set_int("output_height", height));
1760 execute_phase(phase, phase_num == phases.size() - 1, &bound_attribute_indices, &output_textures, &generated_mipmaps);
1761 if (do_phase_timing) {
1762 glEndQuery(GL_TIME_ELAPSED);
1766 for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1767 texture_it != output_textures.end();
1769 resource_pool->release_2d_texture(texture_it->second);
1772 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1777 glBindBuffer(GL_ARRAY_BUFFER, 0);
1779 glBindVertexArray(0);
1781 glDeleteVertexArrays(1, &vao);
1784 if (do_phase_timing) {
1785 // Get back the timer queries.
1786 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1787 Phase *phase = phases[phase_num];
1788 for (std::list<GLuint>::iterator timer_it = phase->timer_query_objects_running.begin();
1789 timer_it != phase->timer_query_objects_running.end(); ) {
1790 GLint timer_query_object = *timer_it;
1792 glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
1794 GLuint64 time_elapsed;
1795 glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
1796 phase->time_elapsed_ns += time_elapsed;
1797 ++phase->num_measured_iterations;
1798 phase->timer_query_objects_free.push_back(timer_query_object);
1799 phase->timer_query_objects_running.erase(timer_it++);
1808 void EffectChain::enable_phase_timing(bool enable)
1811 assert(movit_timer_queries_supported);
1813 this->do_phase_timing = enable;
1816 void EffectChain::reset_phase_timing()
1818 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1819 Phase *phase = phases[phase_num];
1820 phase->time_elapsed_ns = 0;
1821 phase->num_measured_iterations = 0;
1825 void EffectChain::print_phase_timing()
1827 double total_time_ms = 0.0;
1828 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1829 Phase *phase = phases[phase_num];
1830 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
1831 printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms);
1832 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
1833 if (effect_num != 0) {
1836 printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
1839 total_time_ms += avg_time_ms;
1841 printf("Total: %5.1f ms\n", total_time_ms);
1844 void EffectChain::execute_phase(Phase *phase, bool last_phase,
1845 set<GLint> *bound_attribute_indices,
1846 map<Phase *, GLuint> *output_textures,
1847 set<Phase *> *generated_mipmaps)
1851 // Find a texture for this phase.
1852 inform_input_sizes(phase);
1854 find_output_size(phase);
1856 GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
1857 output_textures->insert(make_pair(phase, tex_num));
1860 glUseProgram(phase->glsl_program_num);
1863 // Set up RTT inputs for this phase.
1864 for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
1865 glActiveTexture(GL_TEXTURE0 + sampler);
1866 Phase *input = phase->inputs[sampler];
1867 input->output_node->bound_sampler_num = sampler;
1868 glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
1870 if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
1871 glGenerateMipmap(GL_TEXTURE_2D);
1873 generated_mipmaps->insert(input);
1875 setup_rtt_sampler(sampler, phase->input_needs_mipmaps);
1876 phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform.
1879 // And now the output. (Already set up for us if it is the last phase.)
1881 fbo = resource_pool->create_fbo((*output_textures)[phase]);
1882 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1883 glViewport(0, 0, phase->output_width, phase->output_height);
1886 // Give the required parameters to all the effects.
1887 unsigned sampler_num = phase->inputs.size();
1888 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1889 Node *node = phase->effects[i];
1890 unsigned old_sampler_num = sampler_num;
1891 node->effect->set_gl_state(phase->glsl_program_num, phase->effect_ids[node], &sampler_num);
1894 if (node->effect->is_single_texture()) {
1895 assert(sampler_num - old_sampler_num == 1);
1896 node->bound_sampler_num = old_sampler_num;
1898 node->bound_sampler_num = -1;
1902 // Uniforms need to come after set_gl_state(), since they can be updated
1904 setup_uniforms(phase);
1906 // Clean up old attributes if they are no longer needed.
1907 for (set<GLint>::iterator attr_it = bound_attribute_indices->begin();
1908 attr_it != bound_attribute_indices->end(); ) {
1909 if (phase->attribute_indexes.count(*attr_it) == 0) {
1910 glDisableVertexAttribArray(*attr_it);
1912 bound_attribute_indices->erase(attr_it++);
1918 // Set up the new attributes, if needed.
1919 for (set<GLint>::iterator attr_it = phase->attribute_indexes.begin();
1920 attr_it != phase->attribute_indexes.end();
1922 if (bound_attribute_indices->count(*attr_it) == 0) {
1923 glEnableVertexAttribArray(*attr_it);
1925 glVertexAttribPointer(*attr_it, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1927 bound_attribute_indices->insert(*attr_it);
1931 glDrawArrays(GL_TRIANGLES, 0, 3);
1934 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1935 Node *node = phase->effects[i];
1936 node->effect->clear_gl_state();
1940 resource_pool->release_fbo(fbo);
1944 void EffectChain::setup_uniforms(Phase *phase)
1946 // TODO: Use UBO blocks.
1947 for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
1948 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
1949 if (uniform.location != -1) {
1950 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1953 for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
1954 const Uniform<bool> &uniform = phase->uniforms_bool[i];
1955 assert(uniform.num_values == 1);
1956 if (uniform.location != -1) {
1957 glUniform1i(uniform.location, *uniform.value);
1960 for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
1961 const Uniform<int> &uniform = phase->uniforms_int[i];
1962 if (uniform.location != -1) {
1963 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1966 for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
1967 const Uniform<float> &uniform = phase->uniforms_float[i];
1968 if (uniform.location != -1) {
1969 glUniform1fv(uniform.location, uniform.num_values, uniform.value);
1972 for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
1973 const Uniform<float> &uniform = phase->uniforms_vec2[i];
1974 if (uniform.location != -1) {
1975 glUniform2fv(uniform.location, uniform.num_values, uniform.value);
1978 for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
1979 const Uniform<float> &uniform = phase->uniforms_vec3[i];
1980 if (uniform.location != -1) {
1981 glUniform3fv(uniform.location, uniform.num_values, uniform.value);
1984 for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
1985 const Uniform<float> &uniform = phase->uniforms_vec4[i];
1986 if (uniform.location != -1) {
1987 glUniform4fv(uniform.location, uniform.num_values, uniform.value);
1990 for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
1991 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
1992 assert(uniform.num_values == 1);
1993 if (uniform.location != -1) {
1994 // Convert to float (GLSL has no double matrices).
1996 for (unsigned y = 0; y < 3; ++y) {
1997 for (unsigned x = 0; x < 3; ++x) {
1998 matrixf[y + x * 3] = (*uniform.value)(y, x);
2001 glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
2006 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
2008 glActiveTexture(GL_TEXTURE0 + sampler_num);
2011 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2014 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2017 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2019 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2023 } // namespace movit