1 #define GL_GLEXT_PROTOTYPES 1
17 #include "alpha_division_effect.h"
18 #include "alpha_multiplication_effect.h"
19 #include "colorspace_conversion_effect.h"
20 #include "dither_effect.h"
22 #include "effect_chain.h"
23 #include "effect_util.h"
24 #include "gamma_compression_effect.h"
25 #include "gamma_expansion_effect.h"
28 #include "resource_pool.h"
30 #include "ycbcr_conversion_effect.h"
32 using namespace Eigen;
37 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
38 : aspect_nom(aspect_nom),
39 aspect_denom(aspect_denom),
40 output_color_rgba(false),
41 output_color_ycbcr(false),
44 output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
46 resource_pool(resource_pool),
47 do_phase_timing(false) {
48 if (resource_pool == NULL) {
49 this->resource_pool = new ResourcePool();
50 owns_resource_pool = true;
52 owns_resource_pool = false;
56 EffectChain::~EffectChain()
58 for (unsigned i = 0; i < nodes.size(); ++i) {
59 delete nodes[i]->effect;
62 for (unsigned i = 0; i < phases.size(); ++i) {
63 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
66 if (owns_resource_pool) {
71 Input *EffectChain::add_input(Input *input)
74 inputs.push_back(input);
79 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
82 assert(!output_color_rgba);
83 output_format = format;
84 output_alpha_format = alpha_format;
85 output_color_rgba = true;
88 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
89 const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
92 assert(!output_color_ycbcr);
93 output_format = format;
94 output_alpha_format = alpha_format;
95 output_color_ycbcr = true;
96 output_ycbcr_format = ycbcr_format;
97 output_ycbcr_splitting = output_splitting;
99 assert(ycbcr_format.chroma_subsampling_x == 1);
100 assert(ycbcr_format.chroma_subsampling_y == 1);
103 Node *EffectChain::add_node(Effect *effect)
105 for (unsigned i = 0; i < nodes.size(); ++i) {
106 assert(nodes[i]->effect != effect);
109 Node *node = new Node;
110 node->effect = effect;
111 node->disabled = false;
112 node->output_color_space = COLORSPACE_INVALID;
113 node->output_gamma_curve = GAMMA_INVALID;
114 node->output_alpha_type = ALPHA_INVALID;
115 node->needs_mipmaps = false;
116 node->one_to_one_sampling = false;
118 nodes.push_back(node);
119 node_map[effect] = node;
120 effect->inform_added(this);
124 void EffectChain::connect_nodes(Node *sender, Node *receiver)
126 sender->outgoing_links.push_back(receiver);
127 receiver->incoming_links.push_back(sender);
130 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
132 new_receiver->incoming_links = old_receiver->incoming_links;
133 old_receiver->incoming_links.clear();
135 for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
136 Node *sender = new_receiver->incoming_links[i];
137 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
138 if (sender->outgoing_links[j] == old_receiver) {
139 sender->outgoing_links[j] = new_receiver;
145 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
147 new_sender->outgoing_links = old_sender->outgoing_links;
148 old_sender->outgoing_links.clear();
150 for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
151 Node *receiver = new_sender->outgoing_links[i];
152 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
153 if (receiver->incoming_links[j] == old_sender) {
154 receiver->incoming_links[j] = new_sender;
160 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
162 for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
163 if (sender->outgoing_links[i] == receiver) {
164 sender->outgoing_links[i] = middle;
165 middle->incoming_links.push_back(sender);
168 for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
169 if (receiver->incoming_links[i] == sender) {
170 receiver->incoming_links[i] = middle;
171 middle->outgoing_links.push_back(receiver);
175 assert(middle->incoming_links.size() == middle->effect->num_inputs());
178 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
180 assert(node->effect->needs_texture_bounce());
181 assert(input_num < node->incoming_links.size());
182 assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
183 assert(node->incoming_links[input_num]->bound_sampler_num < 8);
184 return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
187 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
189 assert(input_num < node->incoming_links.size());
190 return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
191 node->incoming_links[input_num]->bound_sampler_num < 8;
194 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
196 if (node->output_gamma_curve == GAMMA_LINEAR &&
197 node->effect->effect_type_id() != "GammaCompressionEffect") {
200 if (node->effect->num_inputs() == 0) {
201 nonlinear_inputs->push_back(node);
203 assert(node->effect->num_inputs() == node->incoming_links.size());
204 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
205 find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
210 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
213 assert(inputs.size() == effect->num_inputs());
214 Node *node = add_node(effect);
215 for (unsigned i = 0; i < inputs.size(); ++i) {
216 assert(node_map.count(inputs[i]) != 0);
217 connect_nodes(node_map[inputs[i]], node);
222 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
223 string replace_prefix(const string &text, const string &prefix)
228 while (start < text.size()) {
229 size_t pos = text.find("PREFIX(", start);
230 if (pos == string::npos) {
231 output.append(text.substr(start, string::npos));
235 output.append(text.substr(start, pos - start));
236 output.append(prefix);
239 pos += strlen("PREFIX(");
241 // Output stuff until we find the matching ), which we then eat.
243 size_t end_arg_pos = pos;
244 while (end_arg_pos < text.size()) {
245 if (text[end_arg_pos] == '(') {
247 } else if (text[end_arg_pos] == ')') {
255 output.append(text.substr(pos, end_arg_pos - pos));
266 void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
267 const string &type_specifier,
268 const string &effect_id,
269 vector<Uniform<T> > *phase_uniforms,
272 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
273 phase_uniforms->push_back(effect_uniforms[i]);
274 phase_uniforms->back().prefix = effect_id;
276 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
277 + "_" + effect_uniforms[i].name + ";\n";
282 void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
283 const string &type_specifier,
284 const string &effect_id,
285 vector<Uniform<T> > *phase_uniforms,
288 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
289 phase_uniforms->push_back(effect_uniforms[i]);
290 phase_uniforms->back().prefix = effect_id;
293 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
294 type_specifier.c_str(), effect_id.c_str(),
295 effect_uniforms[i].name.c_str(),
296 int(effect_uniforms[i].num_values));
302 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
304 for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
305 Uniform<T> &uniform = (*phase_uniforms)[i];
306 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
312 void EffectChain::compile_glsl_program(Phase *phase)
314 string frag_shader_header = read_version_dependent_file("header", "frag");
315 string frag_shader = "";
317 // Create functions and uniforms for all the texture inputs that we need.
318 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
319 Node *input = phase->inputs[i]->output_node;
321 sprintf(effect_id, "in%u", i);
322 phase->effect_ids.insert(make_pair(input, effect_id));
324 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
325 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
326 frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
327 frag_shader += "}\n";
330 Uniform<int> uniform;
331 uniform.name = effect_id;
332 uniform.value = &phase->input_samplers[i];
333 uniform.prefix = "tex";
334 uniform.num_values = 1;
335 uniform.location = -1;
336 phase->uniforms_sampler2d.push_back(uniform);
339 // Give each effect in the phase its own ID.
340 for (unsigned i = 0; i < phase->effects.size(); ++i) {
341 Node *node = phase->effects[i];
343 sprintf(effect_id, "eff%u", i);
344 phase->effect_ids.insert(make_pair(node, effect_id));
347 for (unsigned i = 0; i < phase->effects.size(); ++i) {
348 Node *node = phase->effects[i];
349 const string effect_id = phase->effect_ids[node];
350 if (node->incoming_links.size() == 1) {
351 frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
353 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
355 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
361 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
362 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
363 frag_shader += "#undef PREFIX\n";
364 frag_shader += "#undef FUNCNAME\n";
365 if (node->incoming_links.size() == 1) {
366 frag_shader += "#undef INPUT\n";
368 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
370 sprintf(buf, "#undef INPUT%d\n", j + 1);
376 frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
378 // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
379 if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) {
380 switch (output_ycbcr_splitting) {
381 case YCBCR_OUTPUT_INTERLEAVED:
384 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
385 frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
387 case YCBCR_OUTPUT_PLANAR:
388 frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
394 if (output_color_rgba) {
395 // Note: Needs to come in the header, because not only the
396 // output needs to see it (YCbCrConversionEffect and DitherEffect
398 frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
401 frag_shader.append(read_file("footer.frag"));
403 // Collect uniforms from all effects and output them. Note that this needs
404 // to happen after output_fragment_shader(), even though the uniforms come
405 // before in the output source, since output_fragment_shader() is allowed
406 // to register new uniforms (e.g. arrays that are of unknown length until
407 // finalization time).
408 // TODO: Make a uniform block for platforms that support it.
409 string frag_shader_uniforms = "";
410 for (unsigned i = 0; i < phase->effects.size(); ++i) {
411 Node *node = phase->effects[i];
412 Effect *effect = node->effect;
413 const string effect_id = phase->effect_ids[node];
414 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
415 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
416 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
417 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
418 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
419 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
420 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
421 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
422 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
423 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
426 frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
428 string vert_shader = read_version_dependent_file("vs", "vert");
430 // If we're the last phase and need to flip the picture to compensate for
431 // the origin, tell the vertex shader so.
432 if (phase->output_node->outgoing_links.empty() && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
433 const string needle = "#define FLIP_ORIGIN 0";
434 size_t pos = vert_shader.find(needle);
435 assert(pos != string::npos);
437 vert_shader[pos + needle.size() - 1] = '1';
440 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader);
442 // Collect the resulting location numbers for each uniform.
443 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
444 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
445 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
446 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
447 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
448 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
449 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
450 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
453 // Construct GLSL programs, starting at the given effect and following
454 // the chain from there. We end a program every time we come to an effect
455 // marked as "needs texture bounce", one that is used by multiple other
456 // effects, every time we need to bounce due to output size change
457 // (not all size changes require ending), and of course at the end.
459 // We follow a quite simple depth-first search from the output, although
460 // without recursing explicitly within each phase.
461 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
463 if (completed_effects->count(output)) {
464 return (*completed_effects)[output];
467 Phase *phase = new Phase;
468 phase->output_node = output;
470 // If the output effect has one-to-one sampling, we try to trace this
471 // status down through the dependency chain. This is important in case
472 // we hit an effect that changes output size (and not sets a virtual
473 // output size); if we have one-to-one sampling, we don't have to break
475 output->one_to_one_sampling = output->effect->one_to_one_sampling();
477 // Effects that we have yet to calculate, but that we know should
478 // be in the current phase.
479 stack<Node *> effects_todo_this_phase;
480 effects_todo_this_phase.push(output);
482 while (!effects_todo_this_phase.empty()) {
483 Node *node = effects_todo_this_phase.top();
484 effects_todo_this_phase.pop();
486 if (node->effect->needs_mipmaps()) {
487 node->needs_mipmaps = true;
490 // This should currently only happen for effects that are inputs
491 // (either true inputs or phase outputs). We special-case inputs,
492 // and then deduplicate phase outputs below.
493 if (node->effect->num_inputs() == 0) {
494 if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
498 assert(completed_effects->count(node) == 0);
501 phase->effects.push_back(node);
503 // Find all the dependencies of this effect, and add them to the stack.
504 vector<Node *> deps = node->incoming_links;
505 assert(node->effect->num_inputs() == deps.size());
506 for (unsigned i = 0; i < deps.size(); ++i) {
507 bool start_new_phase = false;
509 if (node->effect->needs_texture_bounce() &&
510 !deps[i]->effect->is_single_texture() &&
511 !deps[i]->effect->override_disable_bounce()) {
512 start_new_phase = true;
515 // Propagate information about needing mipmaps down the chain,
516 // breaking the phase if we notice an incompatibility.
518 // Note that we cannot do this propagation as a normal pass,
519 // because it needs information about where the phases end
520 // (we should not propagate the flag across phases).
521 if (node->needs_mipmaps) {
522 if (deps[i]->effect->num_inputs() == 0) {
523 Input *input = static_cast<Input *>(deps[i]->effect);
524 start_new_phase |= !input->can_supply_mipmaps();
526 deps[i]->needs_mipmaps = true;
530 if (deps[i]->outgoing_links.size() > 1) {
531 if (!deps[i]->effect->is_single_texture()) {
532 // More than one effect uses this as the input,
533 // and it is not a texture itself.
534 // The easiest thing to do (and probably also the safest
535 // performance-wise in most cases) is to bounce it to a texture
536 // and then let the next passes read from that.
537 start_new_phase = true;
539 assert(deps[i]->effect->num_inputs() == 0);
541 // For textures, we try to be slightly more clever;
542 // if none of our outputs need a bounce, we don't bounce
543 // but instead simply use the effect many times.
545 // Strictly speaking, we could bounce it for some outputs
546 // and use it directly for others, but the processing becomes
547 // somewhat simpler if the effect is only used in one such way.
548 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
549 Node *rdep = deps[i]->outgoing_links[j];
550 start_new_phase |= rdep->effect->needs_texture_bounce();
555 if (deps[i]->effect->sets_virtual_output_size()) {
556 assert(deps[i]->effect->changes_output_size());
557 // If the next effect sets a virtual size to rely on OpenGL's
558 // bilinear sampling, we'll really need to break the phase here.
559 start_new_phase = true;
560 } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
561 // If the next effect changes size and we don't have one-to-one sampling,
562 // we also need to break here.
563 start_new_phase = true;
566 if (start_new_phase) {
567 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
569 effects_todo_this_phase.push(deps[i]);
571 // Propagate the one-to-one status down through the dependency.
572 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
573 deps[i]->effect->one_to_one_sampling();
578 // No more effects to do this phase. Take all the ones we have,
579 // and create a GLSL program for it.
580 assert(!phase->effects.empty());
582 // Deduplicate the inputs.
583 sort(phase->inputs.begin(), phase->inputs.end());
584 phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end());
586 // Allocate samplers for each input.
587 phase->input_samplers.resize(phase->inputs.size());
589 // We added the effects from the output and back, but we need to output
590 // them in topological sort order in the shader.
591 phase->effects = topological_sort(phase->effects);
593 // Figure out if we need mipmaps or not, and if so, tell the inputs that.
594 phase->input_needs_mipmaps = false;
595 for (unsigned i = 0; i < phase->effects.size(); ++i) {
596 Node *node = phase->effects[i];
597 phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
599 for (unsigned i = 0; i < phase->effects.size(); ++i) {
600 Node *node = phase->effects[i];
601 if (node->effect->num_inputs() == 0) {
602 Input *input = static_cast<Input *>(node->effect);
603 assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
604 CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
608 // Tell each node which phase it ended up in, so that the unit test
609 // can check that the phases were split in the right place.
610 // Note that this ignores that effects may be part of multiple phases;
611 // if the unit tests need to test such cases, we'll reconsider.
612 for (unsigned i = 0; i < phase->effects.size(); ++i) {
613 phase->effects[i]->containing_phase = phase;
616 // Actually make the shader for this phase.
617 compile_glsl_program(phase);
619 // Initialize timer objects.
620 if (movit_timer_queries_supported) {
621 glGenQueries(1, &phase->timer_query_object);
622 phase->time_elapsed_ns = 0;
623 phase->num_measured_iterations = 0;
626 assert(completed_effects->count(output) == 0);
627 completed_effects->insert(make_pair(output, phase));
628 phases.push_back(phase);
632 void EffectChain::output_dot(const char *filename)
634 if (movit_debug_level != MOVIT_DEBUG_ON) {
638 FILE *fp = fopen(filename, "w");
644 fprintf(fp, "digraph G {\n");
645 fprintf(fp, " output [shape=box label=\"(output)\"];\n");
646 for (unsigned i = 0; i < nodes.size(); ++i) {
647 // Find out which phase this event belongs to.
648 vector<int> in_phases;
649 for (unsigned j = 0; j < phases.size(); ++j) {
650 const Phase* p = phases[j];
651 if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
652 in_phases.push_back(j);
656 if (in_phases.empty()) {
657 fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
658 } else if (in_phases.size() == 1) {
659 fprintf(fp, " n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
660 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
661 (in_phases[0] % 8) + 1);
663 // If we had new enough Graphviz, style="wedged" would probably be ideal here.
665 fprintf(fp, " n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
666 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
667 (in_phases[0] % 8) + 1);
670 char from_node_id[256];
671 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
673 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
674 char to_node_id[256];
675 snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
677 vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
678 output_dot_edge(fp, from_node_id, to_node_id, labels);
681 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
683 vector<string> labels = get_labels_for_edge(nodes[i], NULL);
684 output_dot_edge(fp, from_node_id, "output", labels);
692 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
694 vector<string> labels;
696 if (to != NULL && to->effect->needs_texture_bounce()) {
697 labels.push_back("needs_bounce");
699 if (from->effect->changes_output_size()) {
700 labels.push_back("resize");
703 switch (from->output_color_space) {
704 case COLORSPACE_INVALID:
705 labels.push_back("spc[invalid]");
707 case COLORSPACE_REC_601_525:
708 labels.push_back("spc[rec601-525]");
710 case COLORSPACE_REC_601_625:
711 labels.push_back("spc[rec601-625]");
717 switch (from->output_gamma_curve) {
719 labels.push_back("gamma[invalid]");
722 labels.push_back("gamma[sRGB]");
724 case GAMMA_REC_601: // and GAMMA_REC_709
725 labels.push_back("gamma[rec601/709]");
731 switch (from->output_alpha_type) {
733 labels.push_back("alpha[invalid]");
736 labels.push_back("alpha[blank]");
738 case ALPHA_POSTMULTIPLIED:
739 labels.push_back("alpha[postmult]");
748 void EffectChain::output_dot_edge(FILE *fp,
749 const string &from_node_id,
750 const string &to_node_id,
751 const vector<string> &labels)
753 if (labels.empty()) {
754 fprintf(fp, " %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
756 string label = labels[0];
757 for (unsigned k = 1; k < labels.size(); ++k) {
758 label += ", " + labels[k];
760 fprintf(fp, " %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
764 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
766 unsigned scaled_width, scaled_height;
768 if (float(width) * aspect_denom >= float(height) * aspect_nom) {
769 // Same aspect, or W/H > aspect (image is wider than the frame).
770 // In either case, keep width, and adjust height.
771 scaled_width = width;
772 scaled_height = lrintf(width * aspect_denom / aspect_nom);
774 // W/H < aspect (image is taller than the frame), so keep height,
776 scaled_width = lrintf(height * aspect_nom / aspect_denom);
777 scaled_height = height;
780 // We should be consistently larger or smaller then the existing choice,
781 // since we have the same aspect.
782 assert(!(scaled_width < *output_width && scaled_height > *output_height));
783 assert(!(scaled_height < *output_height && scaled_width > *output_width));
785 if (scaled_width >= *output_width && scaled_height >= *output_height) {
786 *output_width = scaled_width;
787 *output_height = scaled_height;
791 // Propagate input texture sizes throughout, and inform effects downstream.
792 // (Like a lot of other code, we depend on effects being in topological order.)
793 void EffectChain::inform_input_sizes(Phase *phase)
795 // All effects that have a defined size (inputs and RTT inputs)
796 // get that. Reset all others.
797 for (unsigned i = 0; i < phase->effects.size(); ++i) {
798 Node *node = phase->effects[i];
799 if (node->effect->num_inputs() == 0) {
800 Input *input = static_cast<Input *>(node->effect);
801 node->output_width = input->get_width();
802 node->output_height = input->get_height();
803 assert(node->output_width != 0);
804 assert(node->output_height != 0);
806 node->output_width = node->output_height = 0;
809 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
810 Phase *input = phase->inputs[i];
811 input->output_node->output_width = input->virtual_output_width;
812 input->output_node->output_height = input->virtual_output_height;
813 assert(input->output_node->output_width != 0);
814 assert(input->output_node->output_height != 0);
817 // Now propagate from the inputs towards the end, and inform as we go.
818 // The rules are simple:
820 // 1. Don't touch effects that already have given sizes (ie., inputs
821 // or effects that change the output size).
822 // 2. If all of your inputs have the same size, that will be your output size.
823 // 3. Otherwise, your output size is 0x0.
824 for (unsigned i = 0; i < phase->effects.size(); ++i) {
825 Node *node = phase->effects[i];
826 if (node->effect->num_inputs() == 0) {
829 unsigned this_output_width = 0;
830 unsigned this_output_height = 0;
831 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
832 Node *input = node->incoming_links[j];
833 node->effect->inform_input_size(j, input->output_width, input->output_height);
835 this_output_width = input->output_width;
836 this_output_height = input->output_height;
837 } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
839 this_output_width = 0;
840 this_output_height = 0;
843 if (node->effect->changes_output_size()) {
844 // We cannot call get_output_size() before we've done inform_input_size()
846 unsigned real_width, real_height;
847 node->effect->get_output_size(&real_width, &real_height,
848 &node->output_width, &node->output_height);
849 assert(node->effect->sets_virtual_output_size() ||
850 (real_width == node->output_width &&
851 real_height == node->output_height));
853 node->output_width = this_output_width;
854 node->output_height = this_output_height;
859 // Note: You should call inform_input_sizes() before this, as the last effect's
860 // desired output size might change based on the inputs.
861 void EffectChain::find_output_size(Phase *phase)
863 Node *output_node = phase->effects.back();
865 // If the last effect explicitly sets an output size, use that.
866 if (output_node->effect->changes_output_size()) {
867 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
868 &phase->virtual_output_width, &phase->virtual_output_height);
869 assert(output_node->effect->sets_virtual_output_size() ||
870 (phase->output_width == phase->virtual_output_width &&
871 phase->output_height == phase->virtual_output_height));
875 // If all effects have the same size, use that.
876 unsigned output_width = 0, output_height = 0;
877 bool all_inputs_same_size = true;
879 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
880 Phase *input = phase->inputs[i];
881 assert(input->output_width != 0);
882 assert(input->output_height != 0);
883 if (output_width == 0 && output_height == 0) {
884 output_width = input->virtual_output_width;
885 output_height = input->virtual_output_height;
886 } else if (output_width != input->virtual_output_width ||
887 output_height != input->virtual_output_height) {
888 all_inputs_same_size = false;
891 for (unsigned i = 0; i < phase->effects.size(); ++i) {
892 Effect *effect = phase->effects[i]->effect;
893 if (effect->num_inputs() != 0) {
897 Input *input = static_cast<Input *>(effect);
898 if (output_width == 0 && output_height == 0) {
899 output_width = input->get_width();
900 output_height = input->get_height();
901 } else if (output_width != input->get_width() ||
902 output_height != input->get_height()) {
903 all_inputs_same_size = false;
907 if (all_inputs_same_size) {
908 assert(output_width != 0);
909 assert(output_height != 0);
910 phase->virtual_output_width = phase->output_width = output_width;
911 phase->virtual_output_height = phase->output_height = output_height;
915 // If not, fit all the inputs into the current aspect, and select the largest one.
918 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
919 Phase *input = phase->inputs[i];
920 assert(input->output_width != 0);
921 assert(input->output_height != 0);
922 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
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 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
933 assert(output_width != 0);
934 assert(output_height != 0);
935 phase->virtual_output_width = phase->output_width = output_width;
936 phase->virtual_output_height = phase->output_height = output_height;
939 void EffectChain::sort_all_nodes_topologically()
941 nodes = topological_sort(nodes);
944 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
946 set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
947 vector<Node *> sorted_list;
948 for (unsigned i = 0; i < nodes.size(); ++i) {
949 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
951 reverse(sorted_list.begin(), sorted_list.end());
955 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
957 if (nodes_left_to_visit->count(node) == 0) {
960 nodes_left_to_visit->erase(node);
961 for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
962 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
964 sorted_list->push_back(node);
967 void EffectChain::find_color_spaces_for_inputs()
969 for (unsigned i = 0; i < nodes.size(); ++i) {
970 Node *node = nodes[i];
971 if (node->disabled) {
974 if (node->incoming_links.size() == 0) {
975 Input *input = static_cast<Input *>(node->effect);
976 node->output_color_space = input->get_color_space();
977 node->output_gamma_curve = input->get_gamma_curve();
979 Effect::AlphaHandling alpha_handling = input->alpha_handling();
980 switch (alpha_handling) {
981 case Effect::OUTPUT_BLANK_ALPHA:
982 node->output_alpha_type = ALPHA_BLANK;
984 case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
985 node->output_alpha_type = ALPHA_PREMULTIPLIED;
987 case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
988 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
990 case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
991 case Effect::DONT_CARE_ALPHA_TYPE:
996 if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
997 assert(node->output_gamma_curve == GAMMA_LINEAR);
1003 // Propagate gamma and color space information as far as we can in the graph.
1004 // The rules are simple: Anything where all the inputs agree, get that as
1005 // output as well. Anything else keeps having *_INVALID.
1006 void EffectChain::propagate_gamma_and_color_space()
1008 // We depend on going through the nodes in order.
1009 sort_all_nodes_topologically();
1011 for (unsigned i = 0; i < nodes.size(); ++i) {
1012 Node *node = nodes[i];
1013 if (node->disabled) {
1016 assert(node->incoming_links.size() == node->effect->num_inputs());
1017 if (node->incoming_links.size() == 0) {
1018 assert(node->output_color_space != COLORSPACE_INVALID);
1019 assert(node->output_gamma_curve != GAMMA_INVALID);
1023 Colorspace color_space = node->incoming_links[0]->output_color_space;
1024 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1025 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1026 if (node->incoming_links[j]->output_color_space != color_space) {
1027 color_space = COLORSPACE_INVALID;
1029 if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1030 gamma_curve = GAMMA_INVALID;
1034 // The conversion effects already have their outputs set correctly,
1035 // so leave them alone.
1036 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1037 node->output_color_space = color_space;
1039 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1040 node->effect->effect_type_id() != "GammaExpansionEffect") {
1041 node->output_gamma_curve = gamma_curve;
1046 // Propagate alpha information as far as we can in the graph.
1047 // Similar to propagate_gamma_and_color_space().
1048 void EffectChain::propagate_alpha()
1050 // We depend on going through the nodes in order.
1051 sort_all_nodes_topologically();
1053 for (unsigned i = 0; i < nodes.size(); ++i) {
1054 Node *node = nodes[i];
1055 if (node->disabled) {
1058 assert(node->incoming_links.size() == node->effect->num_inputs());
1059 if (node->incoming_links.size() == 0) {
1060 assert(node->output_alpha_type != ALPHA_INVALID);
1064 // The alpha multiplication/division effects are special cases.
1065 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1066 assert(node->incoming_links.size() == 1);
1067 assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1068 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1071 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1072 assert(node->incoming_links.size() == 1);
1073 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1074 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1078 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1079 // because they are the only one that _need_ postmultiplied alpha.
1080 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1081 node->effect->effect_type_id() == "GammaExpansionEffect") {
1082 assert(node->incoming_links.size() == 1);
1083 if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1084 node->output_alpha_type = ALPHA_BLANK;
1085 } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1086 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1088 node->output_alpha_type = ALPHA_INVALID;
1093 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1094 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1095 // taken care of above. Rationale: Even if you could imagine
1096 // e.g. an effect that took in an image and set alpha=1.0
1097 // unconditionally, it wouldn't make any sense to have it as
1098 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1099 // got its input pre- or postmultiplied, so it wouldn't know
1100 // whether to divide away the old alpha or not.
1101 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1102 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1103 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1104 alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1106 // If the node has multiple inputs, check that they are all valid and
1108 bool any_invalid = false;
1109 bool any_premultiplied = false;
1110 bool any_postmultiplied = false;
1112 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1113 switch (node->incoming_links[j]->output_alpha_type) {
1118 // Blank is good as both pre- and postmultiplied alpha,
1119 // so just ignore it.
1121 case ALPHA_PREMULTIPLIED:
1122 any_premultiplied = true;
1124 case ALPHA_POSTMULTIPLIED:
1125 any_postmultiplied = true;
1133 node->output_alpha_type = ALPHA_INVALID;
1137 // Inputs must be of the same type.
1138 if (any_premultiplied && any_postmultiplied) {
1139 node->output_alpha_type = ALPHA_INVALID;
1143 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1144 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1145 // If the effect has asked for premultiplied alpha, check that it has got it.
1146 if (any_postmultiplied) {
1147 node->output_alpha_type = ALPHA_INVALID;
1148 } else if (!any_premultiplied &&
1149 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1150 // Blank input alpha, and the effect preserves blank alpha.
1151 node->output_alpha_type = ALPHA_BLANK;
1153 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1156 // OK, all inputs are the same, and this effect is not going
1158 assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1159 if (any_premultiplied) {
1160 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1161 } else if (any_postmultiplied) {
1162 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1164 node->output_alpha_type = ALPHA_BLANK;
1170 bool EffectChain::node_needs_colorspace_fix(Node *node)
1172 if (node->disabled) {
1175 if (node->effect->num_inputs() == 0) {
1179 // propagate_gamma_and_color_space() has already set our output
1180 // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1181 if (node->output_color_space == COLORSPACE_INVALID) {
1184 return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1187 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1188 // the graph. Our strategy is not always optimal, but quite simple:
1189 // Find an effect that's as early as possible where the inputs are of
1190 // unacceptable colorspaces (that is, either different, or, if the effect only
1191 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1192 // propagate the information anew, and repeat until there are no more such
1194 void EffectChain::fix_internal_color_spaces()
1196 unsigned colorspace_propagation_pass = 0;
1200 for (unsigned i = 0; i < nodes.size(); ++i) {
1201 Node *node = nodes[i];
1202 if (!node_needs_colorspace_fix(node)) {
1206 // Go through each input that is not sRGB, and insert
1207 // a colorspace conversion after it.
1208 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1209 Node *input = node->incoming_links[j];
1210 assert(input->output_color_space != COLORSPACE_INVALID);
1211 if (input->output_color_space == COLORSPACE_sRGB) {
1214 Node *conversion = add_node(new ColorspaceConversionEffect());
1215 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1216 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1217 conversion->output_color_space = COLORSPACE_sRGB;
1218 replace_sender(input, conversion);
1219 connect_nodes(input, conversion);
1222 // Re-sort topologically, and propagate the new information.
1223 propagate_gamma_and_color_space();
1230 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1231 output_dot(filename);
1232 assert(colorspace_propagation_pass < 100);
1233 } while (found_any);
1235 for (unsigned i = 0; i < nodes.size(); ++i) {
1236 Node *node = nodes[i];
1237 if (node->disabled) {
1240 assert(node->output_color_space != COLORSPACE_INVALID);
1244 bool EffectChain::node_needs_alpha_fix(Node *node)
1246 if (node->disabled) {
1250 // propagate_alpha() has already set our output to ALPHA_INVALID if the
1251 // inputs differ or we are otherwise in mismatch, so we can rely on that.
1252 return (node->output_alpha_type == ALPHA_INVALID);
1255 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1256 // the graph. Similar to fix_internal_color_spaces().
1257 void EffectChain::fix_internal_alpha(unsigned step)
1259 unsigned alpha_propagation_pass = 0;
1263 for (unsigned i = 0; i < nodes.size(); ++i) {
1264 Node *node = nodes[i];
1265 if (!node_needs_alpha_fix(node)) {
1269 // If we need to fix up GammaExpansionEffect, then clearly something
1270 // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1272 assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1274 AlphaType desired_type = ALPHA_PREMULTIPLIED;
1276 // GammaCompressionEffect is special; it needs postmultiplied alpha.
1277 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1278 assert(node->incoming_links.size() == 1);
1279 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1280 desired_type = ALPHA_POSTMULTIPLIED;
1283 // Go through each input that is not premultiplied alpha, and insert
1284 // a conversion before it.
1285 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1286 Node *input = node->incoming_links[j];
1287 assert(input->output_alpha_type != ALPHA_INVALID);
1288 if (input->output_alpha_type == desired_type ||
1289 input->output_alpha_type == ALPHA_BLANK) {
1293 if (desired_type == ALPHA_PREMULTIPLIED) {
1294 conversion = add_node(new AlphaMultiplicationEffect());
1296 conversion = add_node(new AlphaDivisionEffect());
1298 conversion->output_alpha_type = desired_type;
1299 replace_sender(input, conversion);
1300 connect_nodes(input, conversion);
1303 // Re-sort topologically, and propagate the new information.
1304 propagate_gamma_and_color_space();
1312 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1313 output_dot(filename);
1314 assert(alpha_propagation_pass < 100);
1315 } while (found_any);
1317 for (unsigned i = 0; i < nodes.size(); ++i) {
1318 Node *node = nodes[i];
1319 if (node->disabled) {
1322 assert(node->output_alpha_type != ALPHA_INVALID);
1326 // Make so that the output is in the desired color space.
1327 void EffectChain::fix_output_color_space()
1329 Node *output = find_output_node();
1330 if (output->output_color_space != output_format.color_space) {
1331 Node *conversion = add_node(new ColorspaceConversionEffect());
1332 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1333 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1334 conversion->output_color_space = output_format.color_space;
1335 connect_nodes(output, conversion);
1337 propagate_gamma_and_color_space();
1341 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1342 void EffectChain::fix_output_alpha()
1344 Node *output = find_output_node();
1345 assert(output->output_alpha_type != ALPHA_INVALID);
1346 if (output->output_alpha_type == ALPHA_BLANK) {
1347 // No alpha output, so we don't care.
1350 if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1351 output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1352 Node *conversion = add_node(new AlphaDivisionEffect());
1353 connect_nodes(output, conversion);
1355 propagate_gamma_and_color_space();
1357 if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1358 output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1359 Node *conversion = add_node(new AlphaMultiplicationEffect());
1360 connect_nodes(output, conversion);
1362 propagate_gamma_and_color_space();
1366 bool EffectChain::node_needs_gamma_fix(Node *node)
1368 if (node->disabled) {
1372 // Small hack since the output is not an explicit node:
1373 // If we are the last node and our output is in the wrong
1374 // space compared to EffectChain's output, we need to fix it.
1375 // This will only take us to linear, but fix_output_gamma()
1376 // will come and take us to the desired output gamma
1379 // This needs to be before everything else, since it could
1380 // even apply to inputs (if they are the only effect).
1381 if (node->outgoing_links.empty() &&
1382 node->output_gamma_curve != output_format.gamma_curve &&
1383 node->output_gamma_curve != GAMMA_LINEAR) {
1387 if (node->effect->num_inputs() == 0) {
1391 // propagate_gamma_and_color_space() has already set our output
1392 // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1393 // except for GammaCompressionEffect.
1394 if (node->output_gamma_curve == GAMMA_INVALID) {
1397 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1398 assert(node->incoming_links.size() == 1);
1399 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1402 return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1405 // Very similar to fix_internal_color_spaces(), but for gamma.
1406 // There is one difference, though; before we start adding conversion nodes,
1407 // we see if we can get anything out of asking the sources to deliver
1408 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1409 // does that part, while fix_internal_gamma_by_inserting_nodes()
1410 // inserts nodes as needed afterwards.
1411 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1413 unsigned gamma_propagation_pass = 0;
1417 for (unsigned i = 0; i < nodes.size(); ++i) {
1418 Node *node = nodes[i];
1419 if (!node_needs_gamma_fix(node)) {
1423 // See if all inputs can give us linear gamma. If not, leave it.
1424 vector<Node *> nonlinear_inputs;
1425 find_all_nonlinear_inputs(node, &nonlinear_inputs);
1426 assert(!nonlinear_inputs.empty());
1429 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1430 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1431 all_ok &= input->can_output_linear_gamma();
1438 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1439 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1440 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1443 // Re-sort topologically, and propagate the new information.
1444 propagate_gamma_and_color_space();
1451 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1452 output_dot(filename);
1453 assert(gamma_propagation_pass < 100);
1454 } while (found_any);
1457 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1459 unsigned gamma_propagation_pass = 0;
1463 for (unsigned i = 0; i < nodes.size(); ++i) {
1464 Node *node = nodes[i];
1465 if (!node_needs_gamma_fix(node)) {
1469 // Special case: We could be an input and still be asked to
1470 // fix our gamma; if so, we should be the only node
1471 // (as node_needs_gamma_fix() would only return true in
1472 // for an input in that case). That means we should insert
1473 // a conversion node _after_ ourselves.
1474 if (node->incoming_links.empty()) {
1475 assert(node->outgoing_links.empty());
1476 Node *conversion = add_node(new GammaExpansionEffect());
1477 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1478 conversion->output_gamma_curve = GAMMA_LINEAR;
1479 connect_nodes(node, conversion);
1482 // If not, go through each input that is not linear gamma,
1483 // and insert a gamma conversion after it.
1484 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1485 Node *input = node->incoming_links[j];
1486 assert(input->output_gamma_curve != GAMMA_INVALID);
1487 if (input->output_gamma_curve == GAMMA_LINEAR) {
1490 Node *conversion = add_node(new GammaExpansionEffect());
1491 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1492 conversion->output_gamma_curve = GAMMA_LINEAR;
1493 replace_sender(input, conversion);
1494 connect_nodes(input, conversion);
1497 // Re-sort topologically, and propagate the new information.
1499 propagate_gamma_and_color_space();
1506 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1507 output_dot(filename);
1508 assert(gamma_propagation_pass < 100);
1509 } while (found_any);
1511 for (unsigned i = 0; i < nodes.size(); ++i) {
1512 Node *node = nodes[i];
1513 if (node->disabled) {
1516 assert(node->output_gamma_curve != GAMMA_INVALID);
1520 // Make so that the output is in the desired gamma.
1521 // Note that this assumes linear input gamma, so it might create the need
1522 // for another pass of fix_internal_gamma().
1523 void EffectChain::fix_output_gamma()
1525 Node *output = find_output_node();
1526 if (output->output_gamma_curve != output_format.gamma_curve) {
1527 Node *conversion = add_node(new GammaCompressionEffect());
1528 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1529 conversion->output_gamma_curve = output_format.gamma_curve;
1530 connect_nodes(output, conversion);
1534 // If the user has requested Y'CbCr output, we need to do this conversion
1535 // _after_ GammaCompressionEffect etc., but before dither (see below).
1536 // This is because Y'CbCr, with the exception of a special optional mode
1537 // in Rec. 2020 (which we currently don't support), is defined to work on
1538 // gamma-encoded data.
1539 void EffectChain::add_ycbcr_conversion_if_needed()
1541 assert(output_color_rgba || output_color_ycbcr);
1542 if (!output_color_ycbcr) {
1545 Node *output = find_output_node();
1546 Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
1547 connect_nodes(output, ycbcr);
1550 // If the user has requested dither, add a DitherEffect right at the end
1551 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1552 // since dither is about the only effect that can _not_ be done in linear space.
1553 void EffectChain::add_dither_if_needed()
1555 if (num_dither_bits == 0) {
1558 Node *output = find_output_node();
1559 Node *dither = add_node(new DitherEffect());
1560 CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1561 connect_nodes(output, dither);
1563 dither_effect = dither->effect;
1566 // Find the output node. This is, simply, one that has no outgoing links.
1567 // If there are multiple ones, the graph is malformed (we do not support
1568 // multiple outputs right now).
1569 Node *EffectChain::find_output_node()
1571 vector<Node *> output_nodes;
1572 for (unsigned i = 0; i < nodes.size(); ++i) {
1573 Node *node = nodes[i];
1574 if (node->disabled) {
1577 if (node->outgoing_links.empty()) {
1578 output_nodes.push_back(node);
1581 assert(output_nodes.size() == 1);
1582 return output_nodes[0];
1585 void EffectChain::finalize()
1587 // Output the graph as it is before we do any conversions on it.
1588 output_dot("step0-start.dot");
1590 // Give each effect in turn a chance to rewrite its own part of the graph.
1591 // Note that if more effects are added as part of this, they will be
1592 // picked up as part of the same for loop, since they are added at the end.
1593 for (unsigned i = 0; i < nodes.size(); ++i) {
1594 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1596 output_dot("step1-rewritten.dot");
1598 find_color_spaces_for_inputs();
1599 output_dot("step2-input-colorspace.dot");
1602 output_dot("step3-propagated-alpha.dot");
1604 propagate_gamma_and_color_space();
1605 output_dot("step4-propagated-all.dot");
1607 fix_internal_color_spaces();
1608 fix_internal_alpha(6);
1609 fix_output_color_space();
1610 output_dot("step7-output-colorspacefix.dot");
1612 output_dot("step8-output-alphafix.dot");
1614 // Note that we need to fix gamma after colorspace conversion,
1615 // because colorspace conversions might create needs for gamma conversions.
1616 // Also, we need to run an extra pass of fix_internal_gamma() after
1617 // fixing the output gamma, as we only have conversions to/from linear,
1618 // and fix_internal_alpha() since GammaCompressionEffect needs
1619 // postmultiplied input.
1620 fix_internal_gamma_by_asking_inputs(9);
1621 fix_internal_gamma_by_inserting_nodes(10);
1623 output_dot("step11-output-gammafix.dot");
1625 output_dot("step12-output-alpha-propagated.dot");
1626 fix_internal_alpha(13);
1627 output_dot("step14-output-alpha-fixed.dot");
1628 fix_internal_gamma_by_asking_inputs(15);
1629 fix_internal_gamma_by_inserting_nodes(16);
1631 output_dot("step17-before-ycbcr.dot");
1632 add_ycbcr_conversion_if_needed();
1634 output_dot("step18-before-dither.dot");
1635 add_dither_if_needed();
1637 output_dot("step19-final.dot");
1639 // Construct all needed GLSL programs, starting at the output.
1640 // We need to keep track of which effects have already been computed,
1641 // as an effect with multiple users could otherwise be calculated
1643 map<Node *, Phase *> completed_effects;
1644 construct_phase(find_output_node(), &completed_effects);
1646 output_dot("step20-split-to-phases.dot");
1648 assert(phases[0]->inputs.empty());
1653 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1657 // This needs to be set anew, in case we are coming from a different context
1658 // from when we initialized.
1660 glDisable(GL_DITHER);
1663 // Save original viewport.
1664 GLuint x = 0, y = 0;
1666 if (width == 0 && height == 0) {
1668 glGetIntegerv(GL_VIEWPORT, viewport);
1671 width = viewport[2];
1672 height = viewport[3];
1677 glDisable(GL_BLEND);
1679 glDisable(GL_DEPTH_TEST);
1681 glDepthMask(GL_FALSE);
1684 // Generate a VAO. All the phases should have exactly the same vertex attributes,
1685 // so it's safe to reuse this.
1686 float vertices[] = {
1693 glGenVertexArrays(1, &vao);
1695 glBindVertexArray(vao);
1698 GLuint position_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
1699 GLuint texcoord_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
1701 set<Phase *> generated_mipmaps;
1703 // We choose the simplest option of having one texture per output,
1704 // since otherwise this turns into an (albeit simple) register allocation problem.
1705 map<Phase *, GLuint> output_textures;
1707 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1708 Phase *phase = phases[phase_num];
1710 if (do_phase_timing) {
1711 glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object);
1713 if (phase_num == phases.size() - 1) {
1714 // Last phase goes to the output the user specified.
1715 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1717 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1718 assert(status == GL_FRAMEBUFFER_COMPLETE);
1719 glViewport(x, y, width, height);
1720 if (dither_effect != NULL) {
1721 CHECK(dither_effect->set_int("output_width", width));
1722 CHECK(dither_effect->set_int("output_height", height));
1725 execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
1726 if (do_phase_timing) {
1727 glEndQuery(GL_TIME_ELAPSED);
1731 for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1732 texture_it != output_textures.end();
1734 resource_pool->release_2d_texture(texture_it->second);
1737 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1742 cleanup_vertex_attribute(phases[0]->glsl_program_num, "position", position_vbo);
1743 cleanup_vertex_attribute(phases[0]->glsl_program_num, "texcoord", texcoord_vbo);
1745 glDeleteVertexArrays(1, &vao);
1748 if (do_phase_timing) {
1749 // Get back the timer queries.
1750 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1751 Phase *phase = phases[phase_num];
1752 GLint available = 0;
1753 while (!available) {
1754 glGetQueryObjectiv(phase->timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
1756 GLuint64 time_elapsed;
1757 glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
1758 phase->time_elapsed_ns += time_elapsed;
1759 ++phase->num_measured_iterations;
1764 void EffectChain::enable_phase_timing(bool enable)
1767 assert(movit_timer_queries_supported);
1769 this->do_phase_timing = enable;
1772 void EffectChain::reset_phase_timing()
1774 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1775 Phase *phase = phases[phase_num];
1776 phase->time_elapsed_ns = 0;
1777 phase->num_measured_iterations = 0;
1781 void EffectChain::print_phase_timing()
1783 double total_time_ms = 0.0;
1784 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1785 Phase *phase = phases[phase_num];
1786 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
1787 printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms);
1788 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
1789 if (effect_num != 0) {
1792 printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
1795 total_time_ms += avg_time_ms;
1797 printf("Total: %5.1f ms\n", total_time_ms);
1800 void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLuint> *output_textures, set<Phase *> *generated_mipmaps)
1804 // Find a texture for this phase.
1805 inform_input_sizes(phase);
1807 find_output_size(phase);
1809 GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
1810 output_textures->insert(make_pair(phase, tex_num));
1813 const GLuint glsl_program_num = phase->glsl_program_num;
1815 glUseProgram(glsl_program_num);
1818 // Set up RTT inputs for this phase.
1819 for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
1820 glActiveTexture(GL_TEXTURE0 + sampler);
1821 Phase *input = phase->inputs[sampler];
1822 input->output_node->bound_sampler_num = sampler;
1823 glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
1825 if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
1826 glGenerateMipmap(GL_TEXTURE_2D);
1828 generated_mipmaps->insert(input);
1830 setup_rtt_sampler(sampler, phase->input_needs_mipmaps);
1831 phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform.
1834 // And now the output. (Already set up for us if it is the last phase.)
1836 fbo = resource_pool->create_fbo((*output_textures)[phase]);
1837 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1838 glViewport(0, 0, phase->output_width, phase->output_height);
1841 // Give the required parameters to all the effects.
1842 unsigned sampler_num = phase->inputs.size();
1843 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1844 Node *node = phase->effects[i];
1845 unsigned old_sampler_num = sampler_num;
1846 node->effect->set_gl_state(glsl_program_num, phase->effect_ids[node], &sampler_num);
1849 if (node->effect->is_single_texture()) {
1850 assert(sampler_num - old_sampler_num == 1);
1851 node->bound_sampler_num = old_sampler_num;
1853 node->bound_sampler_num = -1;
1857 // Uniforms need to come after set_gl_state(), since they can be updated
1859 setup_uniforms(phase);
1861 glDrawArrays(GL_TRIANGLES, 0, 3);
1867 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1868 Node *node = phase->effects[i];
1869 node->effect->clear_gl_state();
1873 resource_pool->release_fbo(fbo);
1877 void EffectChain::setup_uniforms(Phase *phase)
1879 // TODO: Use UBO blocks.
1880 for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
1881 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
1882 if (uniform.location != -1) {
1883 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1886 for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
1887 const Uniform<bool> &uniform = phase->uniforms_bool[i];
1888 assert(uniform.num_values == 1);
1889 if (uniform.location != -1) {
1890 glUniform1i(uniform.location, *uniform.value);
1893 for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
1894 const Uniform<int> &uniform = phase->uniforms_int[i];
1895 if (uniform.location != -1) {
1896 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1899 for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
1900 const Uniform<float> &uniform = phase->uniforms_float[i];
1901 if (uniform.location != -1) {
1902 glUniform1fv(uniform.location, uniform.num_values, uniform.value);
1905 for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
1906 const Uniform<float> &uniform = phase->uniforms_vec2[i];
1907 if (uniform.location != -1) {
1908 glUniform2fv(uniform.location, uniform.num_values, uniform.value);
1911 for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
1912 const Uniform<float> &uniform = phase->uniforms_vec3[i];
1913 if (uniform.location != -1) {
1914 glUniform3fv(uniform.location, uniform.num_values, uniform.value);
1917 for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
1918 const Uniform<float> &uniform = phase->uniforms_vec4[i];
1919 if (uniform.location != -1) {
1920 glUniform4fv(uniform.location, uniform.num_values, uniform.value);
1923 for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
1924 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
1925 assert(uniform.num_values == 1);
1926 if (uniform.location != -1) {
1927 // Convert to float (GLSL has no double matrices).
1929 for (unsigned y = 0; y < 3; ++y) {
1930 for (unsigned x = 0; x < 3; ++x) {
1931 matrixf[y + x * 3] = (*uniform.value)(y, x);
1934 glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
1939 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
1941 glActiveTexture(GL_TEXTURE0 + sampler_num);
1944 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1947 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1950 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1952 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1956 } // namespace movit