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;
55 // Generate a VBO with some data in (shared position and texture coordinate data).
61 vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
64 EffectChain::~EffectChain()
66 for (unsigned i = 0; i < nodes.size(); ++i) {
67 delete nodes[i]->effect;
70 for (unsigned i = 0; i < phases.size(); ++i) {
71 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
74 if (owns_resource_pool) {
77 glDeleteBuffers(1, &vbo);
81 Input *EffectChain::add_input(Input *input)
84 inputs.push_back(input);
89 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
92 assert(!output_color_rgba);
93 output_format = format;
94 output_alpha_format = alpha_format;
95 output_color_rgba = true;
98 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
99 const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
102 assert(!output_color_ycbcr);
103 output_format = format;
104 output_alpha_format = alpha_format;
105 output_color_ycbcr = true;
106 output_ycbcr_format = ycbcr_format;
107 output_ycbcr_splitting = output_splitting;
109 assert(ycbcr_format.chroma_subsampling_x == 1);
110 assert(ycbcr_format.chroma_subsampling_y == 1);
113 Node *EffectChain::add_node(Effect *effect)
115 for (unsigned i = 0; i < nodes.size(); ++i) {
116 assert(nodes[i]->effect != effect);
119 Node *node = new Node;
120 node->effect = effect;
121 node->disabled = false;
122 node->output_color_space = COLORSPACE_INVALID;
123 node->output_gamma_curve = GAMMA_INVALID;
124 node->output_alpha_type = ALPHA_INVALID;
125 node->needs_mipmaps = false;
126 node->one_to_one_sampling = false;
128 nodes.push_back(node);
129 node_map[effect] = node;
130 effect->inform_added(this);
134 void EffectChain::connect_nodes(Node *sender, Node *receiver)
136 sender->outgoing_links.push_back(receiver);
137 receiver->incoming_links.push_back(sender);
140 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
142 new_receiver->incoming_links = old_receiver->incoming_links;
143 old_receiver->incoming_links.clear();
145 for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
146 Node *sender = new_receiver->incoming_links[i];
147 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
148 if (sender->outgoing_links[j] == old_receiver) {
149 sender->outgoing_links[j] = new_receiver;
155 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
157 new_sender->outgoing_links = old_sender->outgoing_links;
158 old_sender->outgoing_links.clear();
160 for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
161 Node *receiver = new_sender->outgoing_links[i];
162 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
163 if (receiver->incoming_links[j] == old_sender) {
164 receiver->incoming_links[j] = new_sender;
170 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
172 for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
173 if (sender->outgoing_links[i] == receiver) {
174 sender->outgoing_links[i] = middle;
175 middle->incoming_links.push_back(sender);
178 for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
179 if (receiver->incoming_links[i] == sender) {
180 receiver->incoming_links[i] = middle;
181 middle->outgoing_links.push_back(receiver);
185 assert(middle->incoming_links.size() == middle->effect->num_inputs());
188 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
190 assert(node->effect->needs_texture_bounce());
191 assert(input_num < node->incoming_links.size());
192 assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
193 assert(node->incoming_links[input_num]->bound_sampler_num < 8);
194 return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
197 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
199 assert(input_num < node->incoming_links.size());
200 return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
201 node->incoming_links[input_num]->bound_sampler_num < 8;
204 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
206 if (node->output_gamma_curve == GAMMA_LINEAR &&
207 node->effect->effect_type_id() != "GammaCompressionEffect") {
210 if (node->effect->num_inputs() == 0) {
211 nonlinear_inputs->push_back(node);
213 assert(node->effect->num_inputs() == node->incoming_links.size());
214 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
215 find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
220 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
223 assert(inputs.size() == effect->num_inputs());
224 Node *node = add_node(effect);
225 for (unsigned i = 0; i < inputs.size(); ++i) {
226 assert(node_map.count(inputs[i]) != 0);
227 connect_nodes(node_map[inputs[i]], node);
232 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
233 string replace_prefix(const string &text, const string &prefix)
238 while (start < text.size()) {
239 size_t pos = text.find("PREFIX(", start);
240 if (pos == string::npos) {
241 output.append(text.substr(start, string::npos));
245 output.append(text.substr(start, pos - start));
246 output.append(prefix);
249 pos += strlen("PREFIX(");
251 // Output stuff until we find the matching ), which we then eat.
253 size_t end_arg_pos = pos;
254 while (end_arg_pos < text.size()) {
255 if (text[end_arg_pos] == '(') {
257 } else if (text[end_arg_pos] == ')') {
265 output.append(text.substr(pos, end_arg_pos - pos));
276 void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
277 const string &type_specifier,
278 const string &effect_id,
279 vector<Uniform<T> > *phase_uniforms,
282 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
283 phase_uniforms->push_back(effect_uniforms[i]);
284 phase_uniforms->back().prefix = effect_id;
286 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
287 + "_" + effect_uniforms[i].name + ";\n";
292 void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
293 const string &type_specifier,
294 const string &effect_id,
295 vector<Uniform<T> > *phase_uniforms,
298 for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
299 phase_uniforms->push_back(effect_uniforms[i]);
300 phase_uniforms->back().prefix = effect_id;
303 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
304 type_specifier.c_str(), effect_id.c_str(),
305 effect_uniforms[i].name.c_str(),
306 int(effect_uniforms[i].num_values));
312 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
314 for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
315 Uniform<T> &uniform = (*phase_uniforms)[i];
316 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
322 void EffectChain::compile_glsl_program(Phase *phase)
324 string frag_shader_header = read_version_dependent_file("header", "frag");
325 string frag_shader = "";
327 // Create functions and uniforms for all the texture inputs that we need.
328 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
329 Node *input = phase->inputs[i]->output_node;
331 sprintf(effect_id, "in%u", i);
332 phase->effect_ids.insert(make_pair(input, effect_id));
334 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
335 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
336 frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
337 frag_shader += "}\n";
340 Uniform<int> uniform;
341 uniform.name = effect_id;
342 uniform.value = &phase->input_samplers[i];
343 uniform.prefix = "tex";
344 uniform.num_values = 1;
345 uniform.location = -1;
346 phase->uniforms_sampler2d.push_back(uniform);
349 // Give each effect in the phase its own ID.
350 for (unsigned i = 0; i < phase->effects.size(); ++i) {
351 Node *node = phase->effects[i];
353 sprintf(effect_id, "eff%u", i);
354 phase->effect_ids.insert(make_pair(node, effect_id));
357 for (unsigned i = 0; i < phase->effects.size(); ++i) {
358 Node *node = phase->effects[i];
359 const string effect_id = phase->effect_ids[node];
360 if (node->incoming_links.size() == 1) {
361 frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
363 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
365 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
371 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
372 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
373 frag_shader += "#undef PREFIX\n";
374 frag_shader += "#undef FUNCNAME\n";
375 if (node->incoming_links.size() == 1) {
376 frag_shader += "#undef INPUT\n";
378 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
380 sprintf(buf, "#undef INPUT%d\n", j + 1);
386 frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
388 // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
389 vector<string> frag_shader_outputs; // In order.
390 if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) {
391 switch (output_ycbcr_splitting) {
392 case YCBCR_OUTPUT_INTERLEAVED:
394 frag_shader_outputs.push_back("FragColor");
396 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
397 frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
398 frag_shader_outputs.push_back("Y");
399 frag_shader_outputs.push_back("Chroma");
401 case YCBCR_OUTPUT_PLANAR:
402 frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
403 frag_shader_outputs.push_back("Y");
404 frag_shader_outputs.push_back("Cb");
405 frag_shader_outputs.push_back("Cr");
411 if (output_color_rgba) {
412 // Note: Needs to come in the header, because not only the
413 // output needs to see it (YCbCrConversionEffect and DitherEffect
415 frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
416 frag_shader_outputs.push_back("RGBA");
419 frag_shader.append(read_file("footer.frag"));
421 // Collect uniforms from all effects and output them. Note that this needs
422 // to happen after output_fragment_shader(), even though the uniforms come
423 // before in the output source, since output_fragment_shader() is allowed
424 // to register new uniforms (e.g. arrays that are of unknown length until
425 // finalization time).
426 // TODO: Make a uniform block for platforms that support it.
427 string frag_shader_uniforms = "";
428 for (unsigned i = 0; i < phase->effects.size(); ++i) {
429 Node *node = phase->effects[i];
430 Effect *effect = node->effect;
431 const string effect_id = phase->effect_ids[node];
432 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
433 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
434 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
435 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
436 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
437 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
438 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
439 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
440 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
441 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
442 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
443 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
446 frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
448 string vert_shader = read_version_dependent_file("vs", "vert");
450 // If we're the last phase and need to flip the picture to compensate for
451 // the origin, tell the vertex shader so.
452 if (phase->output_node->outgoing_links.empty() && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
453 const string needle = "#define FLIP_ORIGIN 0";
454 size_t pos = vert_shader.find(needle);
455 assert(pos != string::npos);
457 vert_shader[pos + needle.size() - 1] = '1';
460 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
461 GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
462 GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
463 if (position_attribute_index != -1) {
464 phase->attribute_indexes.insert(position_attribute_index);
466 if (texcoord_attribute_index != -1) {
467 phase->attribute_indexes.insert(texcoord_attribute_index);
470 // Collect the resulting location numbers for each uniform.
471 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
472 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
473 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
474 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
475 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
476 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
477 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
478 collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
481 // Construct GLSL programs, starting at the given effect and following
482 // the chain from there. We end a program every time we come to an effect
483 // marked as "needs texture bounce", one that is used by multiple other
484 // effects, every time we need to bounce due to output size change
485 // (not all size changes require ending), and of course at the end.
487 // We follow a quite simple depth-first search from the output, although
488 // without recursing explicitly within each phase.
489 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
491 if (completed_effects->count(output)) {
492 return (*completed_effects)[output];
495 Phase *phase = new Phase;
496 phase->output_node = output;
498 // If the output effect has one-to-one sampling, we try to trace this
499 // status down through the dependency chain. This is important in case
500 // we hit an effect that changes output size (and not sets a virtual
501 // output size); if we have one-to-one sampling, we don't have to break
503 output->one_to_one_sampling = output->effect->one_to_one_sampling();
505 // Effects that we have yet to calculate, but that we know should
506 // be in the current phase.
507 stack<Node *> effects_todo_this_phase;
508 effects_todo_this_phase.push(output);
510 while (!effects_todo_this_phase.empty()) {
511 Node *node = effects_todo_this_phase.top();
512 effects_todo_this_phase.pop();
514 if (node->effect->needs_mipmaps()) {
515 node->needs_mipmaps = true;
518 // This should currently only happen for effects that are inputs
519 // (either true inputs or phase outputs). We special-case inputs,
520 // and then deduplicate phase outputs below.
521 if (node->effect->num_inputs() == 0) {
522 if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
526 assert(completed_effects->count(node) == 0);
529 phase->effects.push_back(node);
531 // Find all the dependencies of this effect, and add them to the stack.
532 vector<Node *> deps = node->incoming_links;
533 assert(node->effect->num_inputs() == deps.size());
534 for (unsigned i = 0; i < deps.size(); ++i) {
535 bool start_new_phase = false;
537 if (node->effect->needs_texture_bounce() &&
538 !deps[i]->effect->is_single_texture() &&
539 !deps[i]->effect->override_disable_bounce()) {
540 start_new_phase = true;
543 // Propagate information about needing mipmaps down the chain,
544 // breaking the phase if we notice an incompatibility.
546 // Note that we cannot do this propagation as a normal pass,
547 // because it needs information about where the phases end
548 // (we should not propagate the flag across phases).
549 if (node->needs_mipmaps) {
550 if (deps[i]->effect->num_inputs() == 0) {
551 Input *input = static_cast<Input *>(deps[i]->effect);
552 start_new_phase |= !input->can_supply_mipmaps();
554 deps[i]->needs_mipmaps = true;
558 if (deps[i]->outgoing_links.size() > 1) {
559 if (!deps[i]->effect->is_single_texture()) {
560 // More than one effect uses this as the input,
561 // and it is not a texture itself.
562 // The easiest thing to do (and probably also the safest
563 // performance-wise in most cases) is to bounce it to a texture
564 // and then let the next passes read from that.
565 start_new_phase = true;
567 assert(deps[i]->effect->num_inputs() == 0);
569 // For textures, we try to be slightly more clever;
570 // if none of our outputs need a bounce, we don't bounce
571 // but instead simply use the effect many times.
573 // Strictly speaking, we could bounce it for some outputs
574 // and use it directly for others, but the processing becomes
575 // somewhat simpler if the effect is only used in one such way.
576 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
577 Node *rdep = deps[i]->outgoing_links[j];
578 start_new_phase |= rdep->effect->needs_texture_bounce();
583 if (deps[i]->effect->sets_virtual_output_size()) {
584 assert(deps[i]->effect->changes_output_size());
585 // If the next effect sets a virtual size to rely on OpenGL's
586 // bilinear sampling, we'll really need to break the phase here.
587 start_new_phase = true;
588 } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
589 // If the next effect changes size and we don't have one-to-one sampling,
590 // we also need to break here.
591 start_new_phase = true;
594 if (start_new_phase) {
595 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
597 effects_todo_this_phase.push(deps[i]);
599 // Propagate the one-to-one status down through the dependency.
600 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
601 deps[i]->effect->one_to_one_sampling();
606 // No more effects to do this phase. Take all the ones we have,
607 // and create a GLSL program for it.
608 assert(!phase->effects.empty());
610 // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
611 // that would be nondeterministic and thus reduce cacheability.
612 // TODO: Make this even more deterministic.
613 vector<Phase *> dedup_inputs;
614 set<Phase *> seen_inputs;
615 for (size_t i = 0; i < phase->inputs.size(); ++i) {
616 if (seen_inputs.insert(phase->inputs[i]).second) {
617 dedup_inputs.push_back(phase->inputs[i]);
620 swap(phase->inputs, dedup_inputs);
622 // Allocate samplers for each input.
623 phase->input_samplers.resize(phase->inputs.size());
625 // We added the effects from the output and back, but we need to output
626 // them in topological sort order in the shader.
627 phase->effects = topological_sort(phase->effects);
629 // Figure out if we need mipmaps or not, and if so, tell the inputs that.
630 phase->input_needs_mipmaps = false;
631 for (unsigned i = 0; i < phase->effects.size(); ++i) {
632 Node *node = phase->effects[i];
633 phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
635 for (unsigned i = 0; i < phase->effects.size(); ++i) {
636 Node *node = phase->effects[i];
637 if (node->effect->num_inputs() == 0) {
638 Input *input = static_cast<Input *>(node->effect);
639 assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
640 CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
644 // Tell each node which phase it ended up in, so that the unit test
645 // can check that the phases were split in the right place.
646 // Note that this ignores that effects may be part of multiple phases;
647 // if the unit tests need to test such cases, we'll reconsider.
648 for (unsigned i = 0; i < phase->effects.size(); ++i) {
649 phase->effects[i]->containing_phase = phase;
652 // Actually make the shader for this phase.
653 compile_glsl_program(phase);
655 // Initialize timer objects.
656 if (movit_timer_queries_supported) {
657 glGenQueries(1, &phase->timer_query_object);
658 phase->time_elapsed_ns = 0;
659 phase->num_measured_iterations = 0;
662 assert(completed_effects->count(output) == 0);
663 completed_effects->insert(make_pair(output, phase));
664 phases.push_back(phase);
668 void EffectChain::output_dot(const char *filename)
670 if (movit_debug_level != MOVIT_DEBUG_ON) {
674 FILE *fp = fopen(filename, "w");
680 fprintf(fp, "digraph G {\n");
681 fprintf(fp, " output [shape=box label=\"(output)\"];\n");
682 for (unsigned i = 0; i < nodes.size(); ++i) {
683 // Find out which phase this event belongs to.
684 vector<int> in_phases;
685 for (unsigned j = 0; j < phases.size(); ++j) {
686 const Phase* p = phases[j];
687 if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
688 in_phases.push_back(j);
692 if (in_phases.empty()) {
693 fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
694 } else if (in_phases.size() == 1) {
695 fprintf(fp, " n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
696 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
697 (in_phases[0] % 8) + 1);
699 // If we had new enough Graphviz, style="wedged" would probably be ideal here.
701 fprintf(fp, " n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
702 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
703 (in_phases[0] % 8) + 1);
706 char from_node_id[256];
707 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
709 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
710 char to_node_id[256];
711 snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
713 vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
714 output_dot_edge(fp, from_node_id, to_node_id, labels);
717 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
719 vector<string> labels = get_labels_for_edge(nodes[i], NULL);
720 output_dot_edge(fp, from_node_id, "output", labels);
728 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
730 vector<string> labels;
732 if (to != NULL && to->effect->needs_texture_bounce()) {
733 labels.push_back("needs_bounce");
735 if (from->effect->changes_output_size()) {
736 labels.push_back("resize");
739 switch (from->output_color_space) {
740 case COLORSPACE_INVALID:
741 labels.push_back("spc[invalid]");
743 case COLORSPACE_REC_601_525:
744 labels.push_back("spc[rec601-525]");
746 case COLORSPACE_REC_601_625:
747 labels.push_back("spc[rec601-625]");
753 switch (from->output_gamma_curve) {
755 labels.push_back("gamma[invalid]");
758 labels.push_back("gamma[sRGB]");
760 case GAMMA_REC_601: // and GAMMA_REC_709
761 labels.push_back("gamma[rec601/709]");
767 switch (from->output_alpha_type) {
769 labels.push_back("alpha[invalid]");
772 labels.push_back("alpha[blank]");
774 case ALPHA_POSTMULTIPLIED:
775 labels.push_back("alpha[postmult]");
784 void EffectChain::output_dot_edge(FILE *fp,
785 const string &from_node_id,
786 const string &to_node_id,
787 const vector<string> &labels)
789 if (labels.empty()) {
790 fprintf(fp, " %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
792 string label = labels[0];
793 for (unsigned k = 1; k < labels.size(); ++k) {
794 label += ", " + labels[k];
796 fprintf(fp, " %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
800 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
802 unsigned scaled_width, scaled_height;
804 if (float(width) * aspect_denom >= float(height) * aspect_nom) {
805 // Same aspect, or W/H > aspect (image is wider than the frame).
806 // In either case, keep width, and adjust height.
807 scaled_width = width;
808 scaled_height = lrintf(width * aspect_denom / aspect_nom);
810 // W/H < aspect (image is taller than the frame), so keep height,
812 scaled_width = lrintf(height * aspect_nom / aspect_denom);
813 scaled_height = height;
816 // We should be consistently larger or smaller then the existing choice,
817 // since we have the same aspect.
818 assert(!(scaled_width < *output_width && scaled_height > *output_height));
819 assert(!(scaled_height < *output_height && scaled_width > *output_width));
821 if (scaled_width >= *output_width && scaled_height >= *output_height) {
822 *output_width = scaled_width;
823 *output_height = scaled_height;
827 // Propagate input texture sizes throughout, and inform effects downstream.
828 // (Like a lot of other code, we depend on effects being in topological order.)
829 void EffectChain::inform_input_sizes(Phase *phase)
831 // All effects that have a defined size (inputs and RTT inputs)
832 // get that. Reset all others.
833 for (unsigned i = 0; i < phase->effects.size(); ++i) {
834 Node *node = phase->effects[i];
835 if (node->effect->num_inputs() == 0) {
836 Input *input = static_cast<Input *>(node->effect);
837 node->output_width = input->get_width();
838 node->output_height = input->get_height();
839 assert(node->output_width != 0);
840 assert(node->output_height != 0);
842 node->output_width = node->output_height = 0;
845 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
846 Phase *input = phase->inputs[i];
847 input->output_node->output_width = input->virtual_output_width;
848 input->output_node->output_height = input->virtual_output_height;
849 assert(input->output_node->output_width != 0);
850 assert(input->output_node->output_height != 0);
853 // Now propagate from the inputs towards the end, and inform as we go.
854 // The rules are simple:
856 // 1. Don't touch effects that already have given sizes (ie., inputs
857 // or effects that change the output size).
858 // 2. If all of your inputs have the same size, that will be your output size.
859 // 3. Otherwise, your output size is 0x0.
860 for (unsigned i = 0; i < phase->effects.size(); ++i) {
861 Node *node = phase->effects[i];
862 if (node->effect->num_inputs() == 0) {
865 unsigned this_output_width = 0;
866 unsigned this_output_height = 0;
867 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
868 Node *input = node->incoming_links[j];
869 node->effect->inform_input_size(j, input->output_width, input->output_height);
871 this_output_width = input->output_width;
872 this_output_height = input->output_height;
873 } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
875 this_output_width = 0;
876 this_output_height = 0;
879 if (node->effect->changes_output_size()) {
880 // We cannot call get_output_size() before we've done inform_input_size()
882 unsigned real_width, real_height;
883 node->effect->get_output_size(&real_width, &real_height,
884 &node->output_width, &node->output_height);
885 assert(node->effect->sets_virtual_output_size() ||
886 (real_width == node->output_width &&
887 real_height == node->output_height));
889 node->output_width = this_output_width;
890 node->output_height = this_output_height;
895 // Note: You should call inform_input_sizes() before this, as the last effect's
896 // desired output size might change based on the inputs.
897 void EffectChain::find_output_size(Phase *phase)
899 Node *output_node = phase->effects.back();
901 // If the last effect explicitly sets an output size, use that.
902 if (output_node->effect->changes_output_size()) {
903 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
904 &phase->virtual_output_width, &phase->virtual_output_height);
905 assert(output_node->effect->sets_virtual_output_size() ||
906 (phase->output_width == phase->virtual_output_width &&
907 phase->output_height == phase->virtual_output_height));
911 // If all effects have the same size, use that.
912 unsigned output_width = 0, output_height = 0;
913 bool all_inputs_same_size = true;
915 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
916 Phase *input = phase->inputs[i];
917 assert(input->output_width != 0);
918 assert(input->output_height != 0);
919 if (output_width == 0 && output_height == 0) {
920 output_width = input->virtual_output_width;
921 output_height = input->virtual_output_height;
922 } else if (output_width != input->virtual_output_width ||
923 output_height != input->virtual_output_height) {
924 all_inputs_same_size = false;
927 for (unsigned i = 0; i < phase->effects.size(); ++i) {
928 Effect *effect = phase->effects[i]->effect;
929 if (effect->num_inputs() != 0) {
933 Input *input = static_cast<Input *>(effect);
934 if (output_width == 0 && output_height == 0) {
935 output_width = input->get_width();
936 output_height = input->get_height();
937 } else if (output_width != input->get_width() ||
938 output_height != input->get_height()) {
939 all_inputs_same_size = false;
943 if (all_inputs_same_size) {
944 assert(output_width != 0);
945 assert(output_height != 0);
946 phase->virtual_output_width = phase->output_width = output_width;
947 phase->virtual_output_height = phase->output_height = output_height;
951 // If not, fit all the inputs into the current aspect, and select the largest one.
954 for (unsigned i = 0; i < phase->inputs.size(); ++i) {
955 Phase *input = phase->inputs[i];
956 assert(input->output_width != 0);
957 assert(input->output_height != 0);
958 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
960 for (unsigned i = 0; i < phase->effects.size(); ++i) {
961 Effect *effect = phase->effects[i]->effect;
962 if (effect->num_inputs() != 0) {
966 Input *input = static_cast<Input *>(effect);
967 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
969 assert(output_width != 0);
970 assert(output_height != 0);
971 phase->virtual_output_width = phase->output_width = output_width;
972 phase->virtual_output_height = phase->output_height = output_height;
975 void EffectChain::sort_all_nodes_topologically()
977 nodes = topological_sort(nodes);
980 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
982 set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
983 vector<Node *> sorted_list;
984 for (unsigned i = 0; i < nodes.size(); ++i) {
985 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
987 reverse(sorted_list.begin(), sorted_list.end());
991 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
993 if (nodes_left_to_visit->count(node) == 0) {
996 nodes_left_to_visit->erase(node);
997 for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
998 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
1000 sorted_list->push_back(node);
1003 void EffectChain::find_color_spaces_for_inputs()
1005 for (unsigned i = 0; i < nodes.size(); ++i) {
1006 Node *node = nodes[i];
1007 if (node->disabled) {
1010 if (node->incoming_links.size() == 0) {
1011 Input *input = static_cast<Input *>(node->effect);
1012 node->output_color_space = input->get_color_space();
1013 node->output_gamma_curve = input->get_gamma_curve();
1015 Effect::AlphaHandling alpha_handling = input->alpha_handling();
1016 switch (alpha_handling) {
1017 case Effect::OUTPUT_BLANK_ALPHA:
1018 node->output_alpha_type = ALPHA_BLANK;
1020 case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1021 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1023 case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1024 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1026 case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1027 case Effect::DONT_CARE_ALPHA_TYPE:
1032 if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1033 assert(node->output_gamma_curve == GAMMA_LINEAR);
1039 // Propagate gamma and color space information as far as we can in the graph.
1040 // The rules are simple: Anything where all the inputs agree, get that as
1041 // output as well. Anything else keeps having *_INVALID.
1042 void EffectChain::propagate_gamma_and_color_space()
1044 // We depend on going through the nodes in order.
1045 sort_all_nodes_topologically();
1047 for (unsigned i = 0; i < nodes.size(); ++i) {
1048 Node *node = nodes[i];
1049 if (node->disabled) {
1052 assert(node->incoming_links.size() == node->effect->num_inputs());
1053 if (node->incoming_links.size() == 0) {
1054 assert(node->output_color_space != COLORSPACE_INVALID);
1055 assert(node->output_gamma_curve != GAMMA_INVALID);
1059 Colorspace color_space = node->incoming_links[0]->output_color_space;
1060 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1061 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1062 if (node->incoming_links[j]->output_color_space != color_space) {
1063 color_space = COLORSPACE_INVALID;
1065 if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1066 gamma_curve = GAMMA_INVALID;
1070 // The conversion effects already have their outputs set correctly,
1071 // so leave them alone.
1072 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1073 node->output_color_space = color_space;
1075 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1076 node->effect->effect_type_id() != "GammaExpansionEffect") {
1077 node->output_gamma_curve = gamma_curve;
1082 // Propagate alpha information as far as we can in the graph.
1083 // Similar to propagate_gamma_and_color_space().
1084 void EffectChain::propagate_alpha()
1086 // We depend on going through the nodes in order.
1087 sort_all_nodes_topologically();
1089 for (unsigned i = 0; i < nodes.size(); ++i) {
1090 Node *node = nodes[i];
1091 if (node->disabled) {
1094 assert(node->incoming_links.size() == node->effect->num_inputs());
1095 if (node->incoming_links.size() == 0) {
1096 assert(node->output_alpha_type != ALPHA_INVALID);
1100 // The alpha multiplication/division effects are special cases.
1101 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1102 assert(node->incoming_links.size() == 1);
1103 assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1104 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1107 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1108 assert(node->incoming_links.size() == 1);
1109 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1110 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1114 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1115 // because they are the only one that _need_ postmultiplied alpha.
1116 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1117 node->effect->effect_type_id() == "GammaExpansionEffect") {
1118 assert(node->incoming_links.size() == 1);
1119 if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1120 node->output_alpha_type = ALPHA_BLANK;
1121 } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1122 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1124 node->output_alpha_type = ALPHA_INVALID;
1129 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1130 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1131 // taken care of above. Rationale: Even if you could imagine
1132 // e.g. an effect that took in an image and set alpha=1.0
1133 // unconditionally, it wouldn't make any sense to have it as
1134 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1135 // got its input pre- or postmultiplied, so it wouldn't know
1136 // whether to divide away the old alpha or not.
1137 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1138 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1139 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1140 alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1142 // If the node has multiple inputs, check that they are all valid and
1144 bool any_invalid = false;
1145 bool any_premultiplied = false;
1146 bool any_postmultiplied = false;
1148 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1149 switch (node->incoming_links[j]->output_alpha_type) {
1154 // Blank is good as both pre- and postmultiplied alpha,
1155 // so just ignore it.
1157 case ALPHA_PREMULTIPLIED:
1158 any_premultiplied = true;
1160 case ALPHA_POSTMULTIPLIED:
1161 any_postmultiplied = true;
1169 node->output_alpha_type = ALPHA_INVALID;
1173 // Inputs must be of the same type.
1174 if (any_premultiplied && any_postmultiplied) {
1175 node->output_alpha_type = ALPHA_INVALID;
1179 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1180 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1181 // If the effect has asked for premultiplied alpha, check that it has got it.
1182 if (any_postmultiplied) {
1183 node->output_alpha_type = ALPHA_INVALID;
1184 } else if (!any_premultiplied &&
1185 alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1186 // Blank input alpha, and the effect preserves blank alpha.
1187 node->output_alpha_type = ALPHA_BLANK;
1189 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1192 // OK, all inputs are the same, and this effect is not going
1194 assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1195 if (any_premultiplied) {
1196 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1197 } else if (any_postmultiplied) {
1198 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1200 node->output_alpha_type = ALPHA_BLANK;
1206 bool EffectChain::node_needs_colorspace_fix(Node *node)
1208 if (node->disabled) {
1211 if (node->effect->num_inputs() == 0) {
1215 // propagate_gamma_and_color_space() has already set our output
1216 // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1217 if (node->output_color_space == COLORSPACE_INVALID) {
1220 return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1223 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1224 // the graph. Our strategy is not always optimal, but quite simple:
1225 // Find an effect that's as early as possible where the inputs are of
1226 // unacceptable colorspaces (that is, either different, or, if the effect only
1227 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1228 // propagate the information anew, and repeat until there are no more such
1230 void EffectChain::fix_internal_color_spaces()
1232 unsigned colorspace_propagation_pass = 0;
1236 for (unsigned i = 0; i < nodes.size(); ++i) {
1237 Node *node = nodes[i];
1238 if (!node_needs_colorspace_fix(node)) {
1242 // Go through each input that is not sRGB, and insert
1243 // a colorspace conversion after it.
1244 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1245 Node *input = node->incoming_links[j];
1246 assert(input->output_color_space != COLORSPACE_INVALID);
1247 if (input->output_color_space == COLORSPACE_sRGB) {
1250 Node *conversion = add_node(new ColorspaceConversionEffect());
1251 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1252 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1253 conversion->output_color_space = COLORSPACE_sRGB;
1254 replace_sender(input, conversion);
1255 connect_nodes(input, conversion);
1258 // Re-sort topologically, and propagate the new information.
1259 propagate_gamma_and_color_space();
1266 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1267 output_dot(filename);
1268 assert(colorspace_propagation_pass < 100);
1269 } while (found_any);
1271 for (unsigned i = 0; i < nodes.size(); ++i) {
1272 Node *node = nodes[i];
1273 if (node->disabled) {
1276 assert(node->output_color_space != COLORSPACE_INVALID);
1280 bool EffectChain::node_needs_alpha_fix(Node *node)
1282 if (node->disabled) {
1286 // propagate_alpha() has already set our output to ALPHA_INVALID if the
1287 // inputs differ or we are otherwise in mismatch, so we can rely on that.
1288 return (node->output_alpha_type == ALPHA_INVALID);
1291 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1292 // the graph. Similar to fix_internal_color_spaces().
1293 void EffectChain::fix_internal_alpha(unsigned step)
1295 unsigned alpha_propagation_pass = 0;
1299 for (unsigned i = 0; i < nodes.size(); ++i) {
1300 Node *node = nodes[i];
1301 if (!node_needs_alpha_fix(node)) {
1305 // If we need to fix up GammaExpansionEffect, then clearly something
1306 // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1308 assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1310 AlphaType desired_type = ALPHA_PREMULTIPLIED;
1312 // GammaCompressionEffect is special; it needs postmultiplied alpha.
1313 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1314 assert(node->incoming_links.size() == 1);
1315 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1316 desired_type = ALPHA_POSTMULTIPLIED;
1319 // Go through each input that is not premultiplied alpha, and insert
1320 // a conversion before it.
1321 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1322 Node *input = node->incoming_links[j];
1323 assert(input->output_alpha_type != ALPHA_INVALID);
1324 if (input->output_alpha_type == desired_type ||
1325 input->output_alpha_type == ALPHA_BLANK) {
1329 if (desired_type == ALPHA_PREMULTIPLIED) {
1330 conversion = add_node(new AlphaMultiplicationEffect());
1332 conversion = add_node(new AlphaDivisionEffect());
1334 conversion->output_alpha_type = desired_type;
1335 replace_sender(input, conversion);
1336 connect_nodes(input, conversion);
1339 // Re-sort topologically, and propagate the new information.
1340 propagate_gamma_and_color_space();
1348 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1349 output_dot(filename);
1350 assert(alpha_propagation_pass < 100);
1351 } while (found_any);
1353 for (unsigned i = 0; i < nodes.size(); ++i) {
1354 Node *node = nodes[i];
1355 if (node->disabled) {
1358 assert(node->output_alpha_type != ALPHA_INVALID);
1362 // Make so that the output is in the desired color space.
1363 void EffectChain::fix_output_color_space()
1365 Node *output = find_output_node();
1366 if (output->output_color_space != output_format.color_space) {
1367 Node *conversion = add_node(new ColorspaceConversionEffect());
1368 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1369 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1370 conversion->output_color_space = output_format.color_space;
1371 connect_nodes(output, conversion);
1373 propagate_gamma_and_color_space();
1377 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1378 void EffectChain::fix_output_alpha()
1380 Node *output = find_output_node();
1381 assert(output->output_alpha_type != ALPHA_INVALID);
1382 if (output->output_alpha_type == ALPHA_BLANK) {
1383 // No alpha output, so we don't care.
1386 if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1387 output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1388 Node *conversion = add_node(new AlphaDivisionEffect());
1389 connect_nodes(output, conversion);
1391 propagate_gamma_and_color_space();
1393 if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1394 output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1395 Node *conversion = add_node(new AlphaMultiplicationEffect());
1396 connect_nodes(output, conversion);
1398 propagate_gamma_and_color_space();
1402 bool EffectChain::node_needs_gamma_fix(Node *node)
1404 if (node->disabled) {
1408 // Small hack since the output is not an explicit node:
1409 // If we are the last node and our output is in the wrong
1410 // space compared to EffectChain's output, we need to fix it.
1411 // This will only take us to linear, but fix_output_gamma()
1412 // will come and take us to the desired output gamma
1415 // This needs to be before everything else, since it could
1416 // even apply to inputs (if they are the only effect).
1417 if (node->outgoing_links.empty() &&
1418 node->output_gamma_curve != output_format.gamma_curve &&
1419 node->output_gamma_curve != GAMMA_LINEAR) {
1423 if (node->effect->num_inputs() == 0) {
1427 // propagate_gamma_and_color_space() has already set our output
1428 // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1429 // except for GammaCompressionEffect.
1430 if (node->output_gamma_curve == GAMMA_INVALID) {
1433 if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1434 assert(node->incoming_links.size() == 1);
1435 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1438 return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1441 // Very similar to fix_internal_color_spaces(), but for gamma.
1442 // There is one difference, though; before we start adding conversion nodes,
1443 // we see if we can get anything out of asking the sources to deliver
1444 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1445 // does that part, while fix_internal_gamma_by_inserting_nodes()
1446 // inserts nodes as needed afterwards.
1447 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1449 unsigned gamma_propagation_pass = 0;
1453 for (unsigned i = 0; i < nodes.size(); ++i) {
1454 Node *node = nodes[i];
1455 if (!node_needs_gamma_fix(node)) {
1459 // See if all inputs can give us linear gamma. If not, leave it.
1460 vector<Node *> nonlinear_inputs;
1461 find_all_nonlinear_inputs(node, &nonlinear_inputs);
1462 assert(!nonlinear_inputs.empty());
1465 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1466 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1467 all_ok &= input->can_output_linear_gamma();
1474 for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1475 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1476 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1479 // Re-sort topologically, and propagate the new information.
1480 propagate_gamma_and_color_space();
1487 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1488 output_dot(filename);
1489 assert(gamma_propagation_pass < 100);
1490 } while (found_any);
1493 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1495 unsigned gamma_propagation_pass = 0;
1499 for (unsigned i = 0; i < nodes.size(); ++i) {
1500 Node *node = nodes[i];
1501 if (!node_needs_gamma_fix(node)) {
1505 // Special case: We could be an input and still be asked to
1506 // fix our gamma; if so, we should be the only node
1507 // (as node_needs_gamma_fix() would only return true in
1508 // for an input in that case). That means we should insert
1509 // a conversion node _after_ ourselves.
1510 if (node->incoming_links.empty()) {
1511 assert(node->outgoing_links.empty());
1512 Node *conversion = add_node(new GammaExpansionEffect());
1513 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1514 conversion->output_gamma_curve = GAMMA_LINEAR;
1515 connect_nodes(node, conversion);
1518 // If not, go through each input that is not linear gamma,
1519 // and insert a gamma conversion after it.
1520 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1521 Node *input = node->incoming_links[j];
1522 assert(input->output_gamma_curve != GAMMA_INVALID);
1523 if (input->output_gamma_curve == GAMMA_LINEAR) {
1526 Node *conversion = add_node(new GammaExpansionEffect());
1527 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1528 conversion->output_gamma_curve = GAMMA_LINEAR;
1529 replace_sender(input, conversion);
1530 connect_nodes(input, conversion);
1533 // Re-sort topologically, and propagate the new information.
1535 propagate_gamma_and_color_space();
1542 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1543 output_dot(filename);
1544 assert(gamma_propagation_pass < 100);
1545 } while (found_any);
1547 for (unsigned i = 0; i < nodes.size(); ++i) {
1548 Node *node = nodes[i];
1549 if (node->disabled) {
1552 assert(node->output_gamma_curve != GAMMA_INVALID);
1556 // Make so that the output is in the desired gamma.
1557 // Note that this assumes linear input gamma, so it might create the need
1558 // for another pass of fix_internal_gamma().
1559 void EffectChain::fix_output_gamma()
1561 Node *output = find_output_node();
1562 if (output->output_gamma_curve != output_format.gamma_curve) {
1563 Node *conversion = add_node(new GammaCompressionEffect());
1564 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1565 conversion->output_gamma_curve = output_format.gamma_curve;
1566 connect_nodes(output, conversion);
1570 // If the user has requested Y'CbCr output, we need to do this conversion
1571 // _after_ GammaCompressionEffect etc., but before dither (see below).
1572 // This is because Y'CbCr, with the exception of a special optional mode
1573 // in Rec. 2020 (which we currently don't support), is defined to work on
1574 // gamma-encoded data.
1575 void EffectChain::add_ycbcr_conversion_if_needed()
1577 assert(output_color_rgba || output_color_ycbcr);
1578 if (!output_color_ycbcr) {
1581 Node *output = find_output_node();
1582 Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
1583 connect_nodes(output, ycbcr);
1586 // If the user has requested dither, add a DitherEffect right at the end
1587 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1588 // since dither is about the only effect that can _not_ be done in linear space.
1589 void EffectChain::add_dither_if_needed()
1591 if (num_dither_bits == 0) {
1594 Node *output = find_output_node();
1595 Node *dither = add_node(new DitherEffect());
1596 CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1597 connect_nodes(output, dither);
1599 dither_effect = dither->effect;
1602 // Find the output node. This is, simply, one that has no outgoing links.
1603 // If there are multiple ones, the graph is malformed (we do not support
1604 // multiple outputs right now).
1605 Node *EffectChain::find_output_node()
1607 vector<Node *> output_nodes;
1608 for (unsigned i = 0; i < nodes.size(); ++i) {
1609 Node *node = nodes[i];
1610 if (node->disabled) {
1613 if (node->outgoing_links.empty()) {
1614 output_nodes.push_back(node);
1617 assert(output_nodes.size() == 1);
1618 return output_nodes[0];
1621 void EffectChain::finalize()
1623 // Output the graph as it is before we do any conversions on it.
1624 output_dot("step0-start.dot");
1626 // Give each effect in turn a chance to rewrite its own part of the graph.
1627 // Note that if more effects are added as part of this, they will be
1628 // picked up as part of the same for loop, since they are added at the end.
1629 for (unsigned i = 0; i < nodes.size(); ++i) {
1630 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1632 output_dot("step1-rewritten.dot");
1634 find_color_spaces_for_inputs();
1635 output_dot("step2-input-colorspace.dot");
1638 output_dot("step3-propagated-alpha.dot");
1640 propagate_gamma_and_color_space();
1641 output_dot("step4-propagated-all.dot");
1643 fix_internal_color_spaces();
1644 fix_internal_alpha(6);
1645 fix_output_color_space();
1646 output_dot("step7-output-colorspacefix.dot");
1648 output_dot("step8-output-alphafix.dot");
1650 // Note that we need to fix gamma after colorspace conversion,
1651 // because colorspace conversions might create needs for gamma conversions.
1652 // Also, we need to run an extra pass of fix_internal_gamma() after
1653 // fixing the output gamma, as we only have conversions to/from linear,
1654 // and fix_internal_alpha() since GammaCompressionEffect needs
1655 // postmultiplied input.
1656 fix_internal_gamma_by_asking_inputs(9);
1657 fix_internal_gamma_by_inserting_nodes(10);
1659 output_dot("step11-output-gammafix.dot");
1661 output_dot("step12-output-alpha-propagated.dot");
1662 fix_internal_alpha(13);
1663 output_dot("step14-output-alpha-fixed.dot");
1664 fix_internal_gamma_by_asking_inputs(15);
1665 fix_internal_gamma_by_inserting_nodes(16);
1667 output_dot("step17-before-ycbcr.dot");
1668 add_ycbcr_conversion_if_needed();
1670 output_dot("step18-before-dither.dot");
1671 add_dither_if_needed();
1673 output_dot("step19-final.dot");
1675 // Construct all needed GLSL programs, starting at the output.
1676 // We need to keep track of which effects have already been computed,
1677 // as an effect with multiple users could otherwise be calculated
1679 map<Node *, Phase *> completed_effects;
1680 construct_phase(find_output_node(), &completed_effects);
1682 output_dot("step20-split-to-phases.dot");
1684 assert(phases[0]->inputs.empty());
1689 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1693 // This needs to be set anew, in case we are coming from a different context
1694 // from when we initialized.
1696 glDisable(GL_DITHER);
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 glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object);
1743 if (phase_num == phases.size() - 1) {
1744 // Last phase goes to the output the user specified.
1745 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1747 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1748 assert(status == GL_FRAMEBUFFER_COMPLETE);
1749 glViewport(x, y, width, height);
1750 if (dither_effect != NULL) {
1751 CHECK(dither_effect->set_int("output_width", width));
1752 CHECK(dither_effect->set_int("output_height", height));
1755 execute_phase(phase, phase_num == phases.size() - 1, &bound_attribute_indices, &output_textures, &generated_mipmaps);
1756 if (do_phase_timing) {
1757 glEndQuery(GL_TIME_ELAPSED);
1761 for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1762 texture_it != output_textures.end();
1764 resource_pool->release_2d_texture(texture_it->second);
1767 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1772 glBindBuffer(GL_ARRAY_BUFFER, 0);
1774 glBindVertexArray(0);
1776 glDeleteVertexArrays(1, &vao);
1779 if (do_phase_timing) {
1780 // Get back the timer queries.
1781 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1782 Phase *phase = phases[phase_num];
1783 GLint available = 0;
1784 while (!available) {
1785 glGetQueryObjectiv(phase->timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
1787 GLuint64 time_elapsed;
1788 glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
1789 phase->time_elapsed_ns += time_elapsed;
1790 ++phase->num_measured_iterations;
1795 void EffectChain::enable_phase_timing(bool enable)
1798 assert(movit_timer_queries_supported);
1800 this->do_phase_timing = enable;
1803 void EffectChain::reset_phase_timing()
1805 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1806 Phase *phase = phases[phase_num];
1807 phase->time_elapsed_ns = 0;
1808 phase->num_measured_iterations = 0;
1812 void EffectChain::print_phase_timing()
1814 double total_time_ms = 0.0;
1815 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1816 Phase *phase = phases[phase_num];
1817 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
1818 printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms);
1819 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
1820 if (effect_num != 0) {
1823 printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
1826 total_time_ms += avg_time_ms;
1828 printf("Total: %5.1f ms\n", total_time_ms);
1831 void EffectChain::execute_phase(Phase *phase, bool last_phase,
1832 set<GLint> *bound_attribute_indices,
1833 map<Phase *, GLuint> *output_textures,
1834 set<Phase *> *generated_mipmaps)
1838 // Find a texture for this phase.
1839 inform_input_sizes(phase);
1841 find_output_size(phase);
1843 GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
1844 output_textures->insert(make_pair(phase, tex_num));
1847 glUseProgram(phase->glsl_program_num);
1850 // Set up RTT inputs for this phase.
1851 for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
1852 glActiveTexture(GL_TEXTURE0 + sampler);
1853 Phase *input = phase->inputs[sampler];
1854 input->output_node->bound_sampler_num = sampler;
1855 glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
1857 if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
1858 glGenerateMipmap(GL_TEXTURE_2D);
1860 generated_mipmaps->insert(input);
1862 setup_rtt_sampler(sampler, phase->input_needs_mipmaps);
1863 phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform.
1866 // And now the output. (Already set up for us if it is the last phase.)
1868 fbo = resource_pool->create_fbo((*output_textures)[phase]);
1869 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1870 glViewport(0, 0, phase->output_width, phase->output_height);
1873 // Give the required parameters to all the effects.
1874 unsigned sampler_num = phase->inputs.size();
1875 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1876 Node *node = phase->effects[i];
1877 unsigned old_sampler_num = sampler_num;
1878 node->effect->set_gl_state(phase->glsl_program_num, phase->effect_ids[node], &sampler_num);
1881 if (node->effect->is_single_texture()) {
1882 assert(sampler_num - old_sampler_num == 1);
1883 node->bound_sampler_num = old_sampler_num;
1885 node->bound_sampler_num = -1;
1889 // Uniforms need to come after set_gl_state(), since they can be updated
1891 setup_uniforms(phase);
1893 // Clean up old attributes if they are no longer needed.
1894 for (set<GLint>::iterator attr_it = bound_attribute_indices->begin();
1895 attr_it != bound_attribute_indices->end(); ) {
1896 if (phase->attribute_indexes.count(*attr_it) == 0) {
1897 glDisableVertexAttribArray(*attr_it);
1899 bound_attribute_indices->erase(attr_it++);
1905 // Set up the new attributes, if needed.
1906 for (set<GLint>::iterator attr_it = phase->attribute_indexes.begin();
1907 attr_it != phase->attribute_indexes.end();
1909 if (bound_attribute_indices->count(*attr_it) == 0) {
1910 glEnableVertexAttribArray(*attr_it);
1912 glVertexAttribPointer(*attr_it, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1914 bound_attribute_indices->insert(*attr_it);
1918 glDrawArrays(GL_TRIANGLES, 0, 3);
1921 for (unsigned i = 0; i < phase->effects.size(); ++i) {
1922 Node *node = phase->effects[i];
1923 node->effect->clear_gl_state();
1927 resource_pool->release_fbo(fbo);
1931 void EffectChain::setup_uniforms(Phase *phase)
1933 // TODO: Use UBO blocks.
1934 for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
1935 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
1936 if (uniform.location != -1) {
1937 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1940 for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
1941 const Uniform<bool> &uniform = phase->uniforms_bool[i];
1942 assert(uniform.num_values == 1);
1943 if (uniform.location != -1) {
1944 glUniform1i(uniform.location, *uniform.value);
1947 for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
1948 const Uniform<int> &uniform = phase->uniforms_int[i];
1949 if (uniform.location != -1) {
1950 glUniform1iv(uniform.location, uniform.num_values, uniform.value);
1953 for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
1954 const Uniform<float> &uniform = phase->uniforms_float[i];
1955 if (uniform.location != -1) {
1956 glUniform1fv(uniform.location, uniform.num_values, uniform.value);
1959 for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
1960 const Uniform<float> &uniform = phase->uniforms_vec2[i];
1961 if (uniform.location != -1) {
1962 glUniform2fv(uniform.location, uniform.num_values, uniform.value);
1965 for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
1966 const Uniform<float> &uniform = phase->uniforms_vec3[i];
1967 if (uniform.location != -1) {
1968 glUniform3fv(uniform.location, uniform.num_values, uniform.value);
1971 for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
1972 const Uniform<float> &uniform = phase->uniforms_vec4[i];
1973 if (uniform.location != -1) {
1974 glUniform4fv(uniform.location, uniform.num_values, uniform.value);
1977 for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
1978 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
1979 assert(uniform.num_values == 1);
1980 if (uniform.location != -1) {
1981 // Convert to float (GLSL has no double matrices).
1983 for (unsigned y = 0; y < 3; ++y) {
1984 for (unsigned x = 0; x < 3; ++x) {
1985 matrixf[y + x * 3] = (*uniform.value)(y, x);
1988 glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
1993 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
1995 glActiveTexture(GL_TEXTURE0 + sampler_num);
1998 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2001 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2004 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2006 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2010 } // namespace movit