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, GLenum intermediate_format)
36 : aspect_nom(aspect_nom),
37 aspect_denom(aspect_denom),
38 output_color_rgba(false),
39 output_color_ycbcr(false),
41 intermediate_format(intermediate_format),
43 output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
45 resource_pool(resource_pool),
46 do_phase_timing(false) {
47 if (resource_pool == NULL) {
48 this->resource_pool = new ResourcePool();
49 owns_resource_pool = true;
51 owns_resource_pool = false;
54 // Generate a VBO with some data in (shared position and texture coordinate data).
60 vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
63 EffectChain::~EffectChain()
65 for (unsigned i = 0; i < nodes.size(); ++i) {
66 delete nodes[i]->effect;
69 for (unsigned i = 0; i < phases.size(); ++i) {
70 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
73 if (owns_resource_pool) {
76 glDeleteBuffers(1, &vbo);
80 Input *EffectChain::add_input(Input *input)
83 inputs.push_back(input);
88 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
91 assert(!output_color_rgba);
92 output_format = format;
93 output_alpha_format = alpha_format;
94 output_color_rgba = true;
97 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
98 const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
101 assert(!output_color_ycbcr);
102 output_format = format;
103 output_alpha_format = alpha_format;
104 output_color_ycbcr = true;
105 output_ycbcr_format = ycbcr_format;
106 output_ycbcr_splitting = output_splitting;
108 assert(ycbcr_format.chroma_subsampling_x == 1);
109 assert(ycbcr_format.chroma_subsampling_y == 1);
112 Node *EffectChain::add_node(Effect *effect)
114 for (unsigned i = 0; i < nodes.size(); ++i) {
115 assert(nodes[i]->effect != effect);
118 Node *node = new Node;
119 node->effect = effect;
120 node->disabled = false;
121 node->output_color_space = COLORSPACE_INVALID;
122 node->output_gamma_curve = GAMMA_INVALID;
123 node->output_alpha_type = ALPHA_INVALID;
124 node->needs_mipmaps = false;
125 node->one_to_one_sampling = false;
127 nodes.push_back(node);
128 node_map[effect] = node;
129 effect->inform_added(this);
133 void EffectChain::connect_nodes(Node *sender, Node *receiver)
135 sender->outgoing_links.push_back(receiver);
136 receiver->incoming_links.push_back(sender);
139 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
141 new_receiver->incoming_links = old_receiver->incoming_links;
142 old_receiver->incoming_links.clear();
144 for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
145 Node *sender = new_receiver->incoming_links[i];
146 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
147 if (sender->outgoing_links[j] == old_receiver) {
148 sender->outgoing_links[j] = new_receiver;
154 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
156 new_sender->outgoing_links = old_sender->outgoing_links;
157 old_sender->outgoing_links.clear();
159 for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
160 Node *receiver = new_sender->outgoing_links[i];
161 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
162 if (receiver->incoming_links[j] == old_sender) {
163 receiver->incoming_links[j] = new_sender;
169 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
171 for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
172 if (sender->outgoing_links[i] == receiver) {
173 sender->outgoing_links[i] = middle;
174 middle->incoming_links.push_back(sender);
177 for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
178 if (receiver->incoming_links[i] == sender) {
179 receiver->incoming_links[i] = middle;
180 middle->outgoing_links.push_back(receiver);
184 assert(middle->incoming_links.size() == middle->effect->num_inputs());
187 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
189 assert(node->effect->needs_texture_bounce());
190 assert(input_num < node->incoming_links.size());
191 assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
192 assert(node->incoming_links[input_num]->bound_sampler_num < 8);
193 return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
196 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
198 assert(input_num < node->incoming_links.size());
199 return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
200 node->incoming_links[input_num]->bound_sampler_num < 8;
203 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
205 if (node->output_gamma_curve == GAMMA_LINEAR &&
206 node->effect->effect_type_id() != "GammaCompressionEffect") {
209 if (node->effect->num_inputs() == 0) {
210 nonlinear_inputs->push_back(node);
212 assert(node->effect->num_inputs() == node->incoming_links.size());
213 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
214 find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
219 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
222 assert(inputs.size() == effect->num_inputs());
223 Node *node = add_node(effect);
224 for (unsigned i = 0; i < inputs.size(); ++i) {
225 assert(node_map.count(inputs[i]) != 0);
226 connect_nodes(node_map[inputs[i]], node);
231 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
232 string replace_prefix(const string &text, const string &prefix)
237 while (start < text.size()) {
238 size_t pos = text.find("PREFIX(", start);
239 if (pos == string::npos) {
240 output.append(text.substr(start, string::npos));
244 output.append(text.substr(start, pos - start));
245 output.append(prefix);
248 pos += strlen("PREFIX(");
250 // Output stuff until we find the matching ), which we then eat.
252 size_t end_arg_pos = pos;
253 while (end_arg_pos < text.size()) {
254 if (text[end_arg_pos] == '(') {
256 } else if (text[end_arg_pos] == ')') {
264 output.append(text.substr(pos, end_arg_pos - pos));
275 void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
276 const string &type_specifier,
277 const string &effect_id,
278 vector<Uniform<T> > *phase_uniforms,
281 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
282 phase_uniforms->push_back(effect_uniforms[i]);
283 phase_uniforms->back().prefix = effect_id;
285 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
286 + "_" + effect_uniforms[i].name + ";\n";
291 void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
292 const string &type_specifier,
293 const string &effect_id,
294 vector<Uniform<T> > *phase_uniforms,
297 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
298 phase_uniforms->push_back(effect_uniforms[i]);
299 phase_uniforms->back().prefix = effect_id;
302 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
303 type_specifier.c_str(), effect_id.c_str(),
304 effect_uniforms[i].name.c_str(),
305 int(effect_uniforms[i].num_values));
311 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
313 for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
314 Uniform<T> &uniform = (*phase_uniforms)[i];
315 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
321 void EffectChain::compile_glsl_program(Phase *phase)
323 string frag_shader_header = read_version_dependent_file("header", "frag");
324 string frag_shader = "";
326 // Create functions and uniforms for all the texture inputs that we need.
327 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
328 Node *input = phase->inputs[i]->output_node;
330 sprintf(effect_id, "in%u", i);
331 phase->effect_ids.insert(make_pair(input, effect_id));
333 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
334 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
335 frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
336 frag_shader += "}\n";
339 Uniform<int> uniform;
340 uniform.name = effect_id;
341 uniform.value = &phase->input_samplers[i];
342 uniform.prefix = "tex";
343 uniform.num_values = 1;
344 uniform.location = -1;
345 phase->uniforms_sampler2d.push_back(uniform);
348 // Give each effect in the phase its own ID.
349 for (unsigned i = 0; i < phase->effects.size(); ++i) {
350 Node *node = phase->effects[i];
352 sprintf(effect_id, "eff%u", i);
353 phase->effect_ids.insert(make_pair(node, effect_id));
356 for (unsigned i = 0; i < phase->effects.size(); ++i) {
357 Node *node = phase->effects[i];
358 const string effect_id = phase->effect_ids[node];
359 if (node->incoming_links.size() == 1) {
360 frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
362 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
364 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
370 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
371 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
372 frag_shader += "#undef PREFIX\n";
373 frag_shader += "#undef FUNCNAME\n";
374 if (node->incoming_links.size() == 1) {
375 frag_shader += "#undef INPUT\n";
377 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
379 sprintf(buf, "#undef INPUT%d\n", j + 1);
385 frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
387 // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
388 vector<string> frag_shader_outputs; // In order.
389 if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) {
390 switch (output_ycbcr_splitting) {
391 case YCBCR_OUTPUT_INTERLEAVED:
393 frag_shader_outputs.push_back("FragColor");
395 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
396 frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
397 frag_shader_outputs.push_back("Y");
398 frag_shader_outputs.push_back("Chroma");
400 case YCBCR_OUTPUT_PLANAR:
401 frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
402 frag_shader_outputs.push_back("Y");
403 frag_shader_outputs.push_back("Cb");
404 frag_shader_outputs.push_back("Cr");
410 if (output_color_rgba) {
411 // Note: Needs to come in the header, because not only the
412 // output needs to see it (YCbCrConversionEffect and DitherEffect
414 frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
415 frag_shader_outputs.push_back("RGBA");
418 frag_shader.append(read_file("footer.frag"));
420 // Collect uniforms from all effects and output them. Note that this needs
421 // to happen after output_fragment_shader(), even though the uniforms come
422 // before in the output source, since output_fragment_shader() is allowed
423 // to register new uniforms (e.g. arrays that are of unknown length until
424 // finalization time).
425 // TODO: Make a uniform block for platforms that support it.
426 string frag_shader_uniforms = "";
427 for (unsigned i = 0; i < phase->effects.size(); ++i) {
428 Node *node = phase->effects[i];
429 Effect *effect = node->effect;
430 const string effect_id = phase->effect_ids[node];
431 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
432 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
433 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
434 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
435 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
436 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
437 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
438 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
439 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
440 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
441 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
442 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
445 frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
447 string vert_shader = read_version_dependent_file("vs", "vert");
449 // If we're the last phase and need to flip the picture to compensate for
450 // the origin, tell the vertex shader so.
451 if (phase->output_node->outgoing_links.empty() && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
452 const string needle = "#define FLIP_ORIGIN 0";
453 size_t pos = vert_shader.find(needle);
454 assert(pos != string::npos);
456 vert_shader[pos + needle.size() - 1] = '1';
459 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
460 GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
461 GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
462 if (position_attribute_index != -1) {
463 phase->attribute_indexes.insert(position_attribute_index);
465 if (texcoord_attribute_index != -1) {
466 phase->attribute_indexes.insert(texcoord_attribute_index);
469 // Collect the resulting location numbers for each uniform.
470 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
471 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
472 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
473 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
474 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
475 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
476 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
477 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
480 // Construct GLSL programs, starting at the given effect and following
481 // the chain from there. We end a program every time we come to an effect
482 // marked as "needs texture bounce", one that is used by multiple other
483 // effects, every time we need to bounce due to output size change
484 // (not all size changes require ending), and of course at the end.
486 // We follow a quite simple depth-first search from the output, although
487 // without recursing explicitly within each phase.
488 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
490 if (completed_effects->count(output)) {
491 return (*completed_effects)[output];
494 Phase *phase = new Phase;
495 phase->output_node = output;
497 // If the output effect has one-to-one sampling, we try to trace this
498 // status down through the dependency chain. This is important in case
499 // we hit an effect that changes output size (and not sets a virtual
500 // output size); if we have one-to-one sampling, we don't have to break
502 output->one_to_one_sampling = output->effect->one_to_one_sampling();
504 // Effects that we have yet to calculate, but that we know should
505 // be in the current phase.
506 stack<Node *> effects_todo_this_phase;
507 effects_todo_this_phase.push(output);
509 while (!effects_todo_this_phase.empty()) {
510 Node *node = effects_todo_this_phase.top();
511 effects_todo_this_phase.pop();
513 if (node->effect->needs_mipmaps()) {
514 node->needs_mipmaps = true;
517 // This should currently only happen for effects that are inputs
518 // (either true inputs or phase outputs). We special-case inputs,
519 // and then deduplicate phase outputs below.
520 if (node->effect->num_inputs() == 0) {
521 if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
525 assert(completed_effects->count(node) == 0);
528 phase->effects.push_back(node);
530 // Find all the dependencies of this effect, and add them to the stack.
531 vector<Node *> deps = node->incoming_links;
532 assert(node->effect->num_inputs() == deps.size());
533 for (unsigned i = 0; i < deps.size(); ++i) {
534 bool start_new_phase = false;
536 if (node->effect->needs_texture_bounce() &&
537 !deps[i]->effect->is_single_texture() &&
538 !deps[i]->effect->override_disable_bounce()) {
539 start_new_phase = true;
542 // Propagate information about needing mipmaps down the chain,
543 // breaking the phase if we notice an incompatibility.
545 // Note that we cannot do this propagation as a normal pass,
546 // because it needs information about where the phases end
547 // (we should not propagate the flag across phases).
548 if (node->needs_mipmaps) {
549 if (deps[i]->effect->num_inputs() == 0) {
550 Input *input = static_cast<Input *>(deps[i]->effect);
551 start_new_phase |= !input->can_supply_mipmaps();
553 deps[i]->needs_mipmaps = true;
557 if (deps[i]->outgoing_links.size() > 1) {
558 if (!deps[i]->effect->is_single_texture()) {
559 // More than one effect uses this as the input,
560 // and it is not a texture itself.
561 // The easiest thing to do (and probably also the safest
562 // performance-wise in most cases) is to bounce it to a texture
563 // and then let the next passes read from that.
564 start_new_phase = true;
566 assert(deps[i]->effect->num_inputs() == 0);
568 // For textures, we try to be slightly more clever;
569 // if none of our outputs need a bounce, we don't bounce
570 // but instead simply use the effect many times.
572 // Strictly speaking, we could bounce it for some outputs
573 // and use it directly for others, but the processing becomes
574 // somewhat simpler if the effect is only used in one such way.
575 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
576 Node *rdep = deps[i]->outgoing_links[j];
577 start_new_phase |= rdep->effect->needs_texture_bounce();
582 if (deps[i]->effect->sets_virtual_output_size()) {
583 assert(deps[i]->effect->changes_output_size());
584 // If the next effect sets a virtual size to rely on OpenGL's
585 // bilinear sampling, we'll really need to break the phase here.
586 start_new_phase = true;
587 } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
588 // If the next effect changes size and we don't have one-to-one sampling,
589 // we also need to break here.
590 start_new_phase = true;
593 if (start_new_phase) {
594 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
596 effects_todo_this_phase.push(deps[i]);
598 // Propagate the one-to-one status down through the dependency.
599 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
600 deps[i]->effect->one_to_one_sampling();
605 // No more effects to do this phase. Take all the ones we have,
606 // and create a GLSL program for it.
607 assert(!phase->effects.empty());
609 // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
610 // that would be nondeterministic and thus reduce cacheability.
611 // TODO: Make this even more deterministic.
612 vector<Phase *> dedup_inputs;
613 set<Phase *> seen_inputs;
614 for (size_t i = 0; i < phase->inputs.size(); ++i) {
615 if (seen_inputs.insert(phase->inputs[i]).second) {
616 dedup_inputs.push_back(phase->inputs[i]);
619 swap(phase->inputs, dedup_inputs);
621 // Allocate samplers for each input.
622 phase->input_samplers.resize(phase->inputs.size());
624 // We added the effects from the output and back, but we need to output
625 // them in topological sort order in the shader.
626 phase->effects = topological_sort(phase->effects);
628 // Figure out if we need mipmaps or not, and if so, tell the inputs that.
629 phase->input_needs_mipmaps = false;
630 for (unsigned i = 0; i < phase->effects.size(); ++i) {
631 Node *node = phase->effects[i];
632 phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
634 for (unsigned i = 0; i < phase->effects.size(); ++i) {
635 Node *node = phase->effects[i];
636 if (node->effect->num_inputs() == 0) {
637 Input *input = static_cast<Input *>(node->effect);
638 assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
639 CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
643 // Tell each node which phase it ended up in, so that the unit test
644 // can check that the phases were split in the right place.
645 // Note that this ignores that effects may be part of multiple phases;
646 // if the unit tests need to test such cases, we'll reconsider.
647 for (unsigned i = 0; i < phase->effects.size(); ++i) {
648 phase->effects[i]->containing_phase = phase;
651 // Actually make the shader for this phase.
652 compile_glsl_program(phase);
654 // Initialize timers.
655 if (movit_timer_queries_supported) {
656 phase->time_elapsed_ns = 0;
657 phase->num_measured_iterations = 0;
660 assert(completed_effects->count(output) == 0);
661 completed_effects->insert(make_pair(output, phase));
662 phases.push_back(phase);
666 void EffectChain::output_dot(const char *filename)
668 if (movit_debug_level != MOVIT_DEBUG_ON) {
672 FILE *fp = fopen(filename, "w");
678 fprintf(fp, "digraph G {\n");
679 fprintf(fp, " output [shape=box label=\"(output)\"];\n");
680 for (unsigned i = 0; i < nodes.size(); ++i) {
681 // Find out which phase this event belongs to.
682 vector<int> in_phases;
683 for (unsigned j = 0; j < phases.size(); ++j) {
684 const Phase* p = phases[j];
685 if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
686 in_phases.push_back(j);
690 if (in_phases.empty()) {
691 fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
692 } else if (in_phases.size() == 1) {
693 fprintf(fp, " n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
694 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
695 (in_phases[0] % 8) + 1);
697 // If we had new enough Graphviz, style="wedged" would probably be ideal here.
699 fprintf(fp, " n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
700 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
701 (in_phases[0] % 8) + 1);
704 char from_node_id[256];
705 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
707 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
708 char to_node_id[256];
709 snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
711 vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
712 output_dot_edge(fp, from_node_id, to_node_id, labels);
715 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
717 vector<string> labels = get_labels_for_edge(nodes[i], NULL);
718 output_dot_edge(fp, from_node_id, "output", labels);
726 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
728 vector<string> labels;
730 if (to != NULL && to->effect->needs_texture_bounce()) {
731 labels.push_back("needs_bounce");
733 if (from->effect->changes_output_size()) {
734 labels.push_back("resize");
737 switch (from->output_color_space) {
738 case COLORSPACE_INVALID:
739 labels.push_back("spc[invalid]");
741 case COLORSPACE_REC_601_525:
742 labels.push_back("spc[rec601-525]");
744 case COLORSPACE_REC_601_625:
745 labels.push_back("spc[rec601-625]");
751 switch (from->output_gamma_curve) {
753 labels.push_back("gamma[invalid]");
756 labels.push_back("gamma[sRGB]");
758 case GAMMA_REC_601: // and GAMMA_REC_709
759 labels.push_back("gamma[rec601/709]");
765 switch (from->output_alpha_type) {
767 labels.push_back("alpha[invalid]");
770 labels.push_back("alpha[blank]");
772 case ALPHA_POSTMULTIPLIED:
773 labels.push_back("alpha[postmult]");
782 void EffectChain::output_dot_edge(FILE *fp,
783 const string &from_node_id,
784 const string &to_node_id,
785 const vector<string> &labels)
787 if (labels.empty()) {
788 fprintf(fp, " %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
790 string label = labels[0];
791 for (unsigned k = 1; k < labels.size(); ++k) {
792 label += ", " + labels[k];
794 fprintf(fp, " %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
798 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
800 unsigned scaled_width, scaled_height;
802 if (float(width) * aspect_denom >= float(height) * aspect_nom) {
803 // Same aspect, or W/H > aspect (image is wider than the frame).
804 // In either case, keep width, and adjust height.
805 scaled_width = width;
806 scaled_height = lrintf(width * aspect_denom / aspect_nom);
808 // W/H < aspect (image is taller than the frame), so keep height,
810 scaled_width = lrintf(height * aspect_nom / aspect_denom);
811 scaled_height = height;
814 // We should be consistently larger or smaller then the existing choice,
815 // since we have the same aspect.
816 assert(!(scaled_width < *output_width && scaled_height > *output_height));
817 assert(!(scaled_height < *output_height && scaled_width > *output_width));
819 if (scaled_width >= *output_width && scaled_height >= *output_height) {
820 *output_width = scaled_width;
821 *output_height = scaled_height;
825 // Propagate input texture sizes throughout, and inform effects downstream.
826 // (Like a lot of other code, we depend on effects being in topological order.)
827 void EffectChain::inform_input_sizes(Phase *phase)
829 // All effects that have a defined size (inputs and RTT inputs)
830 // get that. Reset all others.
831 for (unsigned i = 0; i < phase->effects.size(); ++i) {
832 Node *node = phase->effects[i];
833 if (node->effect->num_inputs() == 0) {
834 Input *input = static_cast<Input *>(node->effect);
835 node->output_width = input->get_width();
836 node->output_height = input->get_height();
837 assert(node->output_width != 0);
838 assert(node->output_height != 0);
840 node->output_width = node->output_height = 0;
843 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
844 Phase *input = phase->inputs[i];
845 input->output_node->output_width = input->virtual_output_width;
846 input->output_node->output_height = input->virtual_output_height;
847 assert(input->output_node->output_width != 0);
848 assert(input->output_node->output_height != 0);
851 // Now propagate from the inputs towards the end, and inform as we go.
852 // The rules are simple:
854 // 1. Don't touch effects that already have given sizes (ie., inputs
855 // or effects that change the output size).
856 // 2. If all of your inputs have the same size, that will be your output size.
857 // 3. Otherwise, your output size is 0x0.
858 for (unsigned i = 0; i < phase->effects.size(); ++i) {
859 Node *node = phase->effects[i];
860 if (node->effect->num_inputs() == 0) {
863 unsigned this_output_width = 0;
864 unsigned this_output_height = 0;
865 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
866 Node *input = node->incoming_links[j];
867 node->effect->inform_input_size(j, input->output_width, input->output_height);
869 this_output_width = input->output_width;
870 this_output_height = input->output_height;
871 } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
873 this_output_width = 0;
874 this_output_height = 0;
877 if (node->effect->changes_output_size()) {
878 // We cannot call get_output_size() before we've done inform_input_size()
880 unsigned real_width, real_height;
881 node->effect->get_output_size(&real_width, &real_height,
882 &node->output_width, &node->output_height);
883 assert(node->effect->sets_virtual_output_size() ||
884 (real_width == node->output_width &&
885 real_height == node->output_height));
887 node->output_width = this_output_width;
888 node->output_height = this_output_height;
893 // Note: You should call inform_input_sizes() before this, as the last effect's
894 // desired output size might change based on the inputs.
895 void EffectChain::find_output_size(Phase *phase)
897 Node *output_node = phase->effects.back();
899 // If the last effect explicitly sets an output size, use that.
900 if (output_node->effect->changes_output_size()) {
901 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
902 &phase->virtual_output_width, &phase->virtual_output_height);
903 assert(output_node->effect->sets_virtual_output_size() ||
904 (phase->output_width == phase->virtual_output_width &&
905 phase->output_height == phase->virtual_output_height));
909 // If all effects have the same size, use that.
910 unsigned output_width = 0, output_height = 0;
911 bool all_inputs_same_size = true;
913 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
914 Phase *input = phase->inputs[i];
915 assert(input->output_width != 0);
916 assert(input->output_height != 0);
917 if (output_width == 0 && output_height == 0) {
918 output_width = input->virtual_output_width;
919 output_height = input->virtual_output_height;
920 } else if (output_width != input->virtual_output_width ||
921 output_height != input->virtual_output_height) {
922 all_inputs_same_size = false;
925 for (unsigned i = 0; i < phase->effects.size(); ++i) {
926 Effect *effect = phase->effects[i]->effect;
927 if (effect->num_inputs() != 0) {
931 Input *input = static_cast<Input *>(effect);
932 if (output_width == 0 && output_height == 0) {
933 output_width = input->get_width();
934 output_height = input->get_height();
935 } else if (output_width != input->get_width() ||
936 output_height != input->get_height()) {
937 all_inputs_same_size = false;
941 if (all_inputs_same_size) {
942 assert(output_width != 0);
943 assert(output_height != 0);
944 phase->virtual_output_width = phase->output_width = output_width;
945 phase->virtual_output_height = phase->output_height = output_height;
949 // If not, fit all the inputs into the current aspect, and select the largest one.
952 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
953 Phase *input = phase->inputs[i];
954 assert(input->output_width != 0);
955 assert(input->output_height != 0);
956 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
958 for (unsigned i = 0; i < phase->effects.size(); ++i) {
959 Effect *effect = phase->effects[i]->effect;
960 if (effect->num_inputs() != 0) {
964 Input *input = static_cast<Input *>(effect);
965 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
967 assert(output_width != 0);
968 assert(output_height != 0);
969 phase->virtual_output_width = phase->output_width = output_width;
970 phase->virtual_output_height = phase->output_height = output_height;
973 void EffectChain::sort_all_nodes_topologically()
975 nodes = topological_sort(nodes);
978 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
980 set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
981 vector<Node *> sorted_list;
982 for (unsigned i = 0; i < nodes.size(); ++i) {
983 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
985 reverse(sorted_list.begin(), sorted_list.end());
989 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
991 if (nodes_left_to_visit->count(node) == 0) {
994 nodes_left_to_visit->erase(node);
995 for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
996 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
998 sorted_list->push_back(node);
1001 void EffectChain::find_color_spaces_for_inputs()
1003 for (unsigned i = 0; i < nodes.size(); ++i) {
1004 Node *node = nodes[i];
1005 if (node->disabled) {
1008 if (node->incoming_links.size() == 0) {
1009 Input *input = static_cast<Input *>(node->effect);
1010 node->output_color_space = input->get_color_space();
1011 node->output_gamma_curve = input->get_gamma_curve();
1013 Effect::AlphaHandling alpha_handling = input->alpha_handling();
1014 switch (alpha_handling) {
1015 case Effect::OUTPUT_BLANK_ALPHA:
1016 node->output_alpha_type = ALPHA_BLANK;
1018 case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1019 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1021 case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1022 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1024 case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1025 case Effect::DONT_CARE_ALPHA_TYPE:
1030 if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1031 assert(node->output_gamma_curve == GAMMA_LINEAR);
1037 // Propagate gamma and color space information as far as we can in the graph.
1038 // The rules are simple: Anything where all the inputs agree, get that as
1039 // output as well. Anything else keeps having *_INVALID.
1040 void EffectChain::propagate_gamma_and_color_space()
1042 // We depend on going through the nodes in order.
1043 sort_all_nodes_topologically();
1045 for (unsigned i = 0; i < nodes.size(); ++i) {
1046 Node *node = nodes[i];
1047 if (node->disabled) {
1050 assert(node->incoming_links.size() == node->effect->num_inputs());
1051 if (node->incoming_links.size() == 0) {
1052 assert(node->output_color_space != COLORSPACE_INVALID);
1053 assert(node->output_gamma_curve != GAMMA_INVALID);
1057 Colorspace color_space = node->incoming_links[0]->output_color_space;
1058 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1059 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1060 if (node->incoming_links[j]->output_color_space != color_space) {
1061 color_space = COLORSPACE_INVALID;
1063 if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1064 gamma_curve = GAMMA_INVALID;
1068 // The conversion effects already have their outputs set correctly,
1069 // so leave them alone.
1070 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1071 node->output_color_space = color_space;
1073 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1074 node->effect->effect_type_id() != "GammaExpansionEffect") {
1075 node->output_gamma_curve = gamma_curve;
1080 // Propagate alpha information as far as we can in the graph.
1081 // Similar to propagate_gamma_and_color_space().
1082 void EffectChain::propagate_alpha()
1084 // We depend on going through the nodes in order.
1085 sort_all_nodes_topologically();
1087 for (unsigned i = 0; i < nodes.size(); ++i) {
1088 Node *node = nodes[i];
1089 if (node->disabled) {
1092 assert(node->incoming_links.size() == node->effect->num_inputs());
1093 if (node->incoming_links.size() == 0) {
1094 assert(node->output_alpha_type != ALPHA_INVALID);
1098 // The alpha multiplication/division effects are special cases.
1099 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1100 assert(node->incoming_links.size() == 1);
1101 assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1102 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1105 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1106 assert(node->incoming_links.size() == 1);
1107 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1108 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1112 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1113 // because they are the only one that _need_ postmultiplied alpha.
1114 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1115 node->effect->effect_type_id() == "GammaExpansionEffect") {
1116 assert(node->incoming_links.size() == 1);
1117 if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1118 node->output_alpha_type = ALPHA_BLANK;
1119 } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1120 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1122 node->output_alpha_type = ALPHA_INVALID;
1127 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1128 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1129 // taken care of above. Rationale: Even if you could imagine
1130 // e.g. an effect that took in an image and set alpha=1.0
1131 // unconditionally, it wouldn't make any sense to have it as
1132 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1133 // got its input pre- or postmultiplied, so it wouldn't know
1134 // whether to divide away the old alpha or not.
1135 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1136 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1137 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1138 alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1140 // If the node has multiple inputs, check that they are all valid and
1142 bool any_invalid = false;
1143 bool any_premultiplied = false;
1144 bool any_postmultiplied = false;
1146 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1147 switch (node->incoming_links[j]->output_alpha_type) {
1152 // Blank is good as both pre- and postmultiplied alpha,
1153 // so just ignore it.
1155 case ALPHA_PREMULTIPLIED:
1156 any_premultiplied = true;
1158 case ALPHA_POSTMULTIPLIED:
1159 any_postmultiplied = true;
1167 node->output_alpha_type = ALPHA_INVALID;
1171 // Inputs must be of the same type.
1172 if (any_premultiplied && any_postmultiplied) {
1173 node->output_alpha_type = ALPHA_INVALID;
1177 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1178 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1179 // If the effect has asked for premultiplied alpha, check that it has got it.
1180 if (any_postmultiplied) {
1181 node->output_alpha_type = ALPHA_INVALID;
1182 } else if (!any_premultiplied &&
1183 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1184 // Blank input alpha, and the effect preserves blank alpha.
1185 node->output_alpha_type = ALPHA_BLANK;
1187 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1190 // OK, all inputs are the same, and this effect is not going
1192 assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1193 if (any_premultiplied) {
1194 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1195 } else if (any_postmultiplied) {
1196 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1198 node->output_alpha_type = ALPHA_BLANK;
1204 bool EffectChain::node_needs_colorspace_fix(Node *node)
1206 if (node->disabled) {
1209 if (node->effect->num_inputs() == 0) {
1213 // propagate_gamma_and_color_space() has already set our output
1214 // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1215 if (node->output_color_space == COLORSPACE_INVALID) {
1218 return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1221 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1222 // the graph. Our strategy is not always optimal, but quite simple:
1223 // Find an effect that's as early as possible where the inputs are of
1224 // unacceptable colorspaces (that is, either different, or, if the effect only
1225 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1226 // propagate the information anew, and repeat until there are no more such
1228 void EffectChain::fix_internal_color_spaces()
1230 unsigned colorspace_propagation_pass = 0;
1234 for (unsigned i = 0; i < nodes.size(); ++i) {
1235 Node *node = nodes[i];
1236 if (!node_needs_colorspace_fix(node)) {
1240 // Go through each input that is not sRGB, and insert
1241 // a colorspace conversion after it.
1242 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1243 Node *input = node->incoming_links[j];
1244 assert(input->output_color_space != COLORSPACE_INVALID);
1245 if (input->output_color_space == COLORSPACE_sRGB) {
1248 Node *conversion = add_node(new ColorspaceConversionEffect());
1249 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1250 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1251 conversion->output_color_space = COLORSPACE_sRGB;
1252 replace_sender(input, conversion);
1253 connect_nodes(input, conversion);
1256 // Re-sort topologically, and propagate the new information.
1257 propagate_gamma_and_color_space();
1264 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1265 output_dot(filename);
1266 assert(colorspace_propagation_pass < 100);
1267 } while (found_any);
1269 for (unsigned i = 0; i < nodes.size(); ++i) {
1270 Node *node = nodes[i];
1271 if (node->disabled) {
1274 assert(node->output_color_space != COLORSPACE_INVALID);
1278 bool EffectChain::node_needs_alpha_fix(Node *node)
1280 if (node->disabled) {
1284 // propagate_alpha() has already set our output to ALPHA_INVALID if the
1285 // inputs differ or we are otherwise in mismatch, so we can rely on that.
1286 return (node->output_alpha_type == ALPHA_INVALID);
1289 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1290 // the graph. Similar to fix_internal_color_spaces().
1291 void EffectChain::fix_internal_alpha(unsigned step)
1293 unsigned alpha_propagation_pass = 0;
1297 for (unsigned i = 0; i < nodes.size(); ++i) {
1298 Node *node = nodes[i];
1299 if (!node_needs_alpha_fix(node)) {
1303 // If we need to fix up GammaExpansionEffect, then clearly something
1304 // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1306 assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1308 AlphaType desired_type = ALPHA_PREMULTIPLIED;
1310 // GammaCompressionEffect is special; it needs postmultiplied alpha.
1311 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1312 assert(node->incoming_links.size() == 1);
1313 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1314 desired_type = ALPHA_POSTMULTIPLIED;
1317 // Go through each input that is not premultiplied alpha, and insert
1318 // a conversion before it.
1319 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1320 Node *input = node->incoming_links[j];
1321 assert(input->output_alpha_type != ALPHA_INVALID);
1322 if (input->output_alpha_type == desired_type ||
1323 input->output_alpha_type == ALPHA_BLANK) {
1327 if (desired_type == ALPHA_PREMULTIPLIED) {
1328 conversion = add_node(new AlphaMultiplicationEffect());
1330 conversion = add_node(new AlphaDivisionEffect());
1332 conversion->output_alpha_type = desired_type;
1333 replace_sender(input, conversion);
1334 connect_nodes(input, conversion);
1337 // Re-sort topologically, and propagate the new information.
1338 propagate_gamma_and_color_space();
1346 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1347 output_dot(filename);
1348 assert(alpha_propagation_pass < 100);
1349 } while (found_any);
1351 for (unsigned i = 0; i < nodes.size(); ++i) {
1352 Node *node = nodes[i];
1353 if (node->disabled) {
1356 assert(node->output_alpha_type != ALPHA_INVALID);
1360 // Make so that the output is in the desired color space.
1361 void EffectChain::fix_output_color_space()
1363 Node *output = find_output_node();
1364 if (output->output_color_space != output_format.color_space) {
1365 Node *conversion = add_node(new ColorspaceConversionEffect());
1366 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1367 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1368 conversion->output_color_space = output_format.color_space;
1369 connect_nodes(output, conversion);
1371 propagate_gamma_and_color_space();
1375 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1376 void EffectChain::fix_output_alpha()
1378 Node *output = find_output_node();
1379 assert(output->output_alpha_type != ALPHA_INVALID);
1380 if (output->output_alpha_type == ALPHA_BLANK) {
1381 // No alpha output, so we don't care.
1384 if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1385 output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1386 Node *conversion = add_node(new AlphaDivisionEffect());
1387 connect_nodes(output, conversion);
1389 propagate_gamma_and_color_space();
1391 if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1392 output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1393 Node *conversion = add_node(new AlphaMultiplicationEffect());
1394 connect_nodes(output, conversion);
1396 propagate_gamma_and_color_space();
1400 bool EffectChain::node_needs_gamma_fix(Node *node)
1402 if (node->disabled) {
1406 // Small hack since the output is not an explicit node:
1407 // If we are the last node and our output is in the wrong
1408 // space compared to EffectChain's output, we need to fix it.
1409 // This will only take us to linear, but fix_output_gamma()
1410 // will come and take us to the desired output gamma
1413 // This needs to be before everything else, since it could
1414 // even apply to inputs (if they are the only effect).
1415 if (node->outgoing_links.empty() &&
1416 node->output_gamma_curve != output_format.gamma_curve &&
1417 node->output_gamma_curve != GAMMA_LINEAR) {
1421 if (node->effect->num_inputs() == 0) {
1425 // propagate_gamma_and_color_space() has already set our output
1426 // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1427 // except for GammaCompressionEffect.
1428 if (node->output_gamma_curve == GAMMA_INVALID) {
1431 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1432 assert(node->incoming_links.size() == 1);
1433 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1436 return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1439 // Very similar to fix_internal_color_spaces(), but for gamma.
1440 // There is one difference, though; before we start adding conversion nodes,
1441 // we see if we can get anything out of asking the sources to deliver
1442 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1443 // does that part, while fix_internal_gamma_by_inserting_nodes()
1444 // inserts nodes as needed afterwards.
1445 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1447 unsigned gamma_propagation_pass = 0;
1451 for (unsigned i = 0; i < nodes.size(); ++i) {
1452 Node *node = nodes[i];
1453 if (!node_needs_gamma_fix(node)) {
1457 // See if all inputs can give us linear gamma. If not, leave it.
1458 vector<Node *> nonlinear_inputs;
1459 find_all_nonlinear_inputs(node, &nonlinear_inputs);
1460 assert(!nonlinear_inputs.empty());
1463 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1464 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1465 all_ok &= input->can_output_linear_gamma();
1472 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1473 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1474 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1477 // Re-sort topologically, and propagate the new information.
1478 propagate_gamma_and_color_space();
1485 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1486 output_dot(filename);
1487 assert(gamma_propagation_pass < 100);
1488 } while (found_any);
1491 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1493 unsigned gamma_propagation_pass = 0;
1497 for (unsigned i = 0; i < nodes.size(); ++i) {
1498 Node *node = nodes[i];
1499 if (!node_needs_gamma_fix(node)) {
1503 // Special case: We could be an input and still be asked to
1504 // fix our gamma; if so, we should be the only node
1505 // (as node_needs_gamma_fix() would only return true in
1506 // for an input in that case). That means we should insert
1507 // a conversion node _after_ ourselves.
1508 if (node->incoming_links.empty()) {
1509 assert(node->outgoing_links.empty());
1510 Node *conversion = add_node(new GammaExpansionEffect());
1511 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1512 conversion->output_gamma_curve = GAMMA_LINEAR;
1513 connect_nodes(node, conversion);
1516 // If not, go through each input that is not linear gamma,
1517 // and insert a gamma conversion after it.
1518 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1519 Node *input = node->incoming_links[j];
1520 assert(input->output_gamma_curve != GAMMA_INVALID);
1521 if (input->output_gamma_curve == GAMMA_LINEAR) {
1524 Node *conversion = add_node(new GammaExpansionEffect());
1525 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1526 conversion->output_gamma_curve = GAMMA_LINEAR;
1527 replace_sender(input, conversion);
1528 connect_nodes(input, conversion);
1531 // Re-sort topologically, and propagate the new information.
1533 propagate_gamma_and_color_space();
1540 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1541 output_dot(filename);
1542 assert(gamma_propagation_pass < 100);
1543 } while (found_any);
1545 for (unsigned i = 0; i < nodes.size(); ++i) {
1546 Node *node = nodes[i];
1547 if (node->disabled) {
1550 assert(node->output_gamma_curve != GAMMA_INVALID);
1554 // Make so that the output is in the desired gamma.
1555 // Note that this assumes linear input gamma, so it might create the need
1556 // for another pass of fix_internal_gamma().
1557 void EffectChain::fix_output_gamma()
1559 Node *output = find_output_node();
1560 if (output->output_gamma_curve != output_format.gamma_curve) {
1561 Node *conversion = add_node(new GammaCompressionEffect());
1562 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1563 conversion->output_gamma_curve = output_format.gamma_curve;
1564 connect_nodes(output, conversion);
1568 // If the user has requested Y'CbCr output, we need to do this conversion
1569 // _after_ GammaCompressionEffect etc., but before dither (see below).
1570 // This is because Y'CbCr, with the exception of a special optional mode
1571 // in Rec. 2020 (which we currently don't support), is defined to work on
1572 // gamma-encoded data.
1573 void EffectChain::add_ycbcr_conversion_if_needed()
1575 assert(output_color_rgba || output_color_ycbcr);
1576 if (!output_color_ycbcr) {
1579 Node *output = find_output_node();
1580 Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
1581 connect_nodes(output, ycbcr);
1584 // If the user has requested dither, add a DitherEffect right at the end
1585 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1586 // since dither is about the only effect that can _not_ be done in linear space.
1587 void EffectChain::add_dither_if_needed()
1589 if (num_dither_bits == 0) {
1592 Node *output = find_output_node();
1593 Node *dither = add_node(new DitherEffect());
1594 CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1595 connect_nodes(output, dither);
1597 dither_effect = dither->effect;
1600 // Find the output node. This is, simply, one that has no outgoing links.
1601 // If there are multiple ones, the graph is malformed (we do not support
1602 // multiple outputs right now).
1603 Node *EffectChain::find_output_node()
1605 vector<Node *> output_nodes;
1606 for (unsigned i = 0; i < nodes.size(); ++i) {
1607 Node *node = nodes[i];
1608 if (node->disabled) {
1611 if (node->outgoing_links.empty()) {
1612 output_nodes.push_back(node);
1615 assert(output_nodes.size() == 1);
1616 return output_nodes[0];
1619 void EffectChain::finalize()
1621 // Output the graph as it is before we do any conversions on it.
1622 output_dot("step0-start.dot");
1624 // Give each effect in turn a chance to rewrite its own part of the graph.
1625 // Note that if more effects are added as part of this, they will be
1626 // picked up as part of the same for loop, since they are added at the end.
1627 for (unsigned i = 0; i < nodes.size(); ++i) {
1628 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1630 output_dot("step1-rewritten.dot");
1632 find_color_spaces_for_inputs();
1633 output_dot("step2-input-colorspace.dot");
1636 output_dot("step3-propagated-alpha.dot");
1638 propagate_gamma_and_color_space();
1639 output_dot("step4-propagated-all.dot");
1641 fix_internal_color_spaces();
1642 fix_internal_alpha(6);
1643 fix_output_color_space();
1644 output_dot("step7-output-colorspacefix.dot");
1646 output_dot("step8-output-alphafix.dot");
1648 // Note that we need to fix gamma after colorspace conversion,
1649 // because colorspace conversions might create needs for gamma conversions.
1650 // Also, we need to run an extra pass of fix_internal_gamma() after
1651 // fixing the output gamma, as we only have conversions to/from linear,
1652 // and fix_internal_alpha() since GammaCompressionEffect needs
1653 // postmultiplied input.
1654 fix_internal_gamma_by_asking_inputs(9);
1655 fix_internal_gamma_by_inserting_nodes(10);
1657 output_dot("step11-output-gammafix.dot");
1659 output_dot("step12-output-alpha-propagated.dot");
1660 fix_internal_alpha(13);
1661 output_dot("step14-output-alpha-fixed.dot");
1662 fix_internal_gamma_by_asking_inputs(15);
1663 fix_internal_gamma_by_inserting_nodes(16);
1665 output_dot("step17-before-ycbcr.dot");
1666 add_ycbcr_conversion_if_needed();
1668 output_dot("step18-before-dither.dot");
1669 add_dither_if_needed();
1671 output_dot("step19-final.dot");
1673 // Construct all needed GLSL programs, starting at the output.
1674 // We need to keep track of which effects have already been computed,
1675 // as an effect with multiple users could otherwise be calculated
1677 map<Node *, Phase *> completed_effects;
1678 construct_phase(find_output_node(), &completed_effects);
1680 output_dot("step20-split-to-phases.dot");
1682 assert(phases[0]->inputs.empty());
1687 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1691 // This needs to be set anew, in case we are coming from a different context
1692 // from when we initialized.
1694 glDisable(GL_DITHER);
1696 glEnable(GL_FRAMEBUFFER_SRGB);
1699 // Save original viewport.
1700 GLuint x = 0, y = 0;
1702 if (width == 0 && height == 0) {
1704 glGetIntegerv(GL_VIEWPORT, viewport);
1707 width = viewport[2];
1708 height = viewport[3];
1713 glDisable(GL_BLEND);
1715 glDisable(GL_DEPTH_TEST);
1717 glDepthMask(GL_FALSE);
1720 // Generate a VAO that will be used during the entire execution,
1721 // and bind the VBO, since it contains all the data.
1723 glGenVertexArrays(1, &vao);
1725 glBindVertexArray(vao);
1727 glBindBuffer(GL_ARRAY_BUFFER, vbo);
1729 set<GLint> bound_attribute_indices;
1731 set<Phase *> generated_mipmaps;
1733 // We choose the simplest option of having one texture per output,
1734 // since otherwise this turns into an (albeit simple) register allocation problem.
1735 map<Phase *, GLuint> output_textures;
1737 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1738 Phase *phase = phases[phase_num];
1740 if (do_phase_timing) {
1741 GLuint timer_query_object;
1742 if (phase->timer_query_objects_free.empty()) {
1743 glGenQueries(1, &timer_query_object);
1745 timer_query_object = phase->timer_query_objects_free.front();
1746 phase->timer_query_objects_free.pop_front();
1748 glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
1749 phase->timer_query_objects_running.push_back(timer_query_object);
1751 if (phase_num == phases.size() - 1) {
1752 // Last phase goes to the output the user specified.
1753 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1755 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1756 assert(status == GL_FRAMEBUFFER_COMPLETE);
1757 glViewport(x, y, width, height);
1758 if (dither_effect != NULL) {
1759 CHECK(dither_effect->set_int("output_width", width));
1760 CHECK(dither_effect->set_int("output_height", height));
1763 execute_phase(phase, phase_num == phases.size() - 1, &bound_attribute_indices, &output_textures, &generated_mipmaps);
1764 if (do_phase_timing) {
1765 glEndQuery(GL_TIME_ELAPSED);
1769 for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1770 texture_it != output_textures.end();
1772 resource_pool->release_2d_texture(texture_it->second);
1775 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1780 glBindBuffer(GL_ARRAY_BUFFER, 0);
1782 glBindVertexArray(0);
1784 glDeleteVertexArrays(1, &vao);
1787 if (do_phase_timing) {
1788 // Get back the timer queries.
1789 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1790 Phase *phase = phases[phase_num];
1791 for (std::list<GLuint>::iterator timer_it = phase->timer_query_objects_running.begin();
1792 timer_it != phase->timer_query_objects_running.end(); ) {
1793 GLint timer_query_object = *timer_it;
1795 glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
1797 GLuint64 time_elapsed;
1798 glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
1799 phase->time_elapsed_ns += time_elapsed;
1800 ++phase->num_measured_iterations;
1801 phase->timer_query_objects_free.push_back(timer_query_object);
1802 phase->timer_query_objects_running.erase(timer_it++);
1811 void EffectChain::enable_phase_timing(bool enable)
1814 assert(movit_timer_queries_supported);
1816 this->do_phase_timing = enable;
1819 void EffectChain::reset_phase_timing()
1821 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1822 Phase *phase = phases[phase_num];
1823 phase->time_elapsed_ns = 0;
1824 phase->num_measured_iterations = 0;
1828 void EffectChain::print_phase_timing()
1830 double total_time_ms = 0.0;
1831 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1832 Phase *phase = phases[phase_num];
1833 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
1834 printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms);
1835 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
1836 if (effect_num != 0) {
1839 printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
1842 total_time_ms += avg_time_ms;
1844 printf("Total: %5.1f ms\n", total_time_ms);
1847 void EffectChain::execute_phase(Phase *phase, bool last_phase,
1848 set<GLint> *bound_attribute_indices,
1849 map<Phase *, GLuint> *output_textures,
1850 set<Phase *> *generated_mipmaps)
1854 // Find a texture for this phase.
1855 inform_input_sizes(phase);
1857 find_output_size(phase);
1859 GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
1860 output_textures->insert(make_pair(phase, tex_num));
1863 glUseProgram(phase->glsl_program_num);
1866 // Set up RTT inputs for this phase.
1867 for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
1868 glActiveTexture(GL_TEXTURE0 + sampler);
1869 Phase *input = phase->inputs[sampler];
1870 input->output_node->bound_sampler_num = sampler;
1871 glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
1873 if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
1874 glGenerateMipmap(GL_TEXTURE_2D);
1876 generated_mipmaps->insert(input);
1878 setup_rtt_sampler(sampler, phase->input_needs_mipmaps);
1879 phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform.
1882 // And now the output. (Already set up for us if it is the last phase.)
1884 fbo = resource_pool->create_fbo((*output_textures)[phase]);
1885 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1886 glViewport(0, 0, phase->output_width, phase->output_height);
1889 // Give the required parameters to all the effects.
1890 unsigned sampler_num = phase->inputs.size();
1891 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1892 Node *node = phase->effects[i];
1893 unsigned old_sampler_num = sampler_num;
1894 node->effect->set_gl_state(phase->glsl_program_num, phase->effect_ids[node], &sampler_num);
1897 if (node->effect->is_single_texture()) {
1898 assert(sampler_num - old_sampler_num == 1);
1899 node->bound_sampler_num = old_sampler_num;
1901 node->bound_sampler_num = -1;
1905 // Uniforms need to come after set_gl_state(), since they can be updated
1907 setup_uniforms(phase);
1909 // Clean up old attributes if they are no longer needed.
1910 for (set<GLint>::iterator attr_it = bound_attribute_indices->begin();
1911 attr_it != bound_attribute_indices->end(); ) {
1912 if (phase->attribute_indexes.count(*attr_it) == 0) {
1913 glDisableVertexAttribArray(*attr_it);
1915 bound_attribute_indices->erase(attr_it++);
1921 // Set up the new attributes, if needed.
1922 for (set<GLint>::iterator attr_it = phase->attribute_indexes.begin();
1923 attr_it != phase->attribute_indexes.end();
1925 if (bound_attribute_indices->count(*attr_it) == 0) {
1926 glEnableVertexAttribArray(*attr_it);
1928 glVertexAttribPointer(*attr_it, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1930 bound_attribute_indices->insert(*attr_it);
1934 glDrawArrays(GL_TRIANGLES, 0, 3);
1937 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1938 Node *node = phase->effects[i];
1939 node->effect->clear_gl_state();
1943 resource_pool->release_fbo(fbo);
1947 void EffectChain::setup_uniforms(Phase *phase)
1949 // TODO: Use UBO blocks.
1950 for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
1951 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
1952 if (uniform.location != -1) {
1953 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1956 for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
1957 const Uniform<bool> &uniform = phase->uniforms_bool[i];
1958 assert(uniform.num_values == 1);
1959 if (uniform.location != -1) {
1960 glUniform1i(uniform.location, *uniform.value);
1963 for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
1964 const Uniform<int> &uniform = phase->uniforms_int[i];
1965 if (uniform.location != -1) {
1966 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1969 for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
1970 const Uniform<float> &uniform = phase->uniforms_float[i];
1971 if (uniform.location != -1) {
1972 glUniform1fv(uniform.location, uniform.num_values, uniform.value);
1975 for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
1976 const Uniform<float> &uniform = phase->uniforms_vec2[i];
1977 if (uniform.location != -1) {
1978 glUniform2fv(uniform.location, uniform.num_values, uniform.value);
1981 for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
1982 const Uniform<float> &uniform = phase->uniforms_vec3[i];
1983 if (uniform.location != -1) {
1984 glUniform3fv(uniform.location, uniform.num_values, uniform.value);
1987 for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
1988 const Uniform<float> &uniform = phase->uniforms_vec4[i];
1989 if (uniform.location != -1) {
1990 glUniform4fv(uniform.location, uniform.num_values, uniform.value);
1993 for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
1994 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
1995 assert(uniform.num_values == 1);
1996 if (uniform.location != -1) {
1997 // Convert to float (GLSL has no double matrices).
1999 for (unsigned y = 0; y < 3; ++y) {
2000 for (unsigned x = 0; x < 3; ++x) {
2001 matrixf[y + x * 3] = (*uniform.value)(y, x);
2004 glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
2009 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
2011 glActiveTexture(GL_TEXTURE0 + sampler_num);
2014 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2017 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2020 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2022 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2026 } // namespace movit