9aa092ba207db4f614abfe5f45617c0eff3f3849
[movit] / effect_chain.cpp
1 #define GL_GLEXT_PROTOTYPES 1
2
3 #include <GL/glew.h>
4 #include <assert.h>
5 #include <locale.h>
6 #include <math.h>
7 #include <stddef.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <algorithm>
12 #include <set>
13 #include <stack>
14 #include <utility>
15 #include <vector>
16
17 #include "alpha_division_effect.h"
18 #include "alpha_multiplication_effect.h"
19 #include "colorspace_conversion_effect.h"
20 #include "dither_effect.h"
21 #include "effect.h"
22 #include "effect_chain.h"
23 #include "gamma_compression_effect.h"
24 #include "gamma_expansion_effect.h"
25 #include "init.h"
26 #include "input.h"
27 #include "resource_pool.h"
28 #include "util.h"
29
30 using namespace std;
31
32 namespace movit {
33
34 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
35         : aspect_nom(aspect_nom),
36           aspect_denom(aspect_denom),
37           dither_effect(NULL),
38           num_dither_bits(0),
39           finalized(false),
40           resource_pool(resource_pool) {
41         if (resource_pool == NULL) {
42                 this->resource_pool = new ResourcePool();
43                 owns_resource_pool = true;
44         } else {
45                 owns_resource_pool = false;
46         }
47 }
48
49 EffectChain::~EffectChain()
50 {
51         for (unsigned i = 0; i < nodes.size(); ++i) {
52                 delete nodes[i]->effect;
53                 delete nodes[i];
54         }
55         for (unsigned i = 0; i < phases.size(); ++i) {
56                 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
57                 delete phases[i];
58         }
59         if (owns_resource_pool) {
60                 delete resource_pool;
61         }
62 }
63
64 Input *EffectChain::add_input(Input *input)
65 {
66         assert(!finalized);
67         inputs.push_back(input);
68         add_node(input);
69         return input;
70 }
71
72 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
73 {
74         assert(!finalized);
75         output_format = format;
76         output_alpha_format = alpha_format;
77 }
78
79 Node *EffectChain::add_node(Effect *effect)
80 {
81         for (unsigned i = 0; i < nodes.size(); ++i) {
82                 assert(nodes[i]->effect != effect);
83         }
84
85         Node *node = new Node;
86         node->effect = effect;
87         node->disabled = false;
88         node->output_color_space = COLORSPACE_INVALID;
89         node->output_gamma_curve = GAMMA_INVALID;
90         node->output_alpha_type = ALPHA_INVALID;
91
92         nodes.push_back(node);
93         node_map[effect] = node;
94         effect->inform_added(this);
95         return node;
96 }
97
98 void EffectChain::connect_nodes(Node *sender, Node *receiver)
99 {
100         sender->outgoing_links.push_back(receiver);
101         receiver->incoming_links.push_back(sender);
102 }
103
104 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
105 {
106         new_receiver->incoming_links = old_receiver->incoming_links;
107         old_receiver->incoming_links.clear();
108         
109         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
110                 Node *sender = new_receiver->incoming_links[i];
111                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
112                         if (sender->outgoing_links[j] == old_receiver) {
113                                 sender->outgoing_links[j] = new_receiver;
114                         }
115                 }
116         }       
117 }
118
119 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
120 {
121         new_sender->outgoing_links = old_sender->outgoing_links;
122         old_sender->outgoing_links.clear();
123         
124         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
125                 Node *receiver = new_sender->outgoing_links[i];
126                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
127                         if (receiver->incoming_links[j] == old_sender) {
128                                 receiver->incoming_links[j] = new_sender;
129                         }
130                 }
131         }       
132 }
133
134 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
135 {
136         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
137                 if (sender->outgoing_links[i] == receiver) {
138                         sender->outgoing_links[i] = middle;
139                         middle->incoming_links.push_back(sender);
140                 }
141         }
142         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
143                 if (receiver->incoming_links[i] == sender) {
144                         receiver->incoming_links[i] = middle;
145                         middle->outgoing_links.push_back(receiver);
146                 }
147         }
148
149         assert(middle->incoming_links.size() == middle->effect->num_inputs());
150 }
151
152 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
153 {
154         if (node->output_gamma_curve == GAMMA_LINEAR &&
155             node->effect->effect_type_id() != "GammaCompressionEffect") {
156                 return;
157         }
158         if (node->effect->num_inputs() == 0) {
159                 nonlinear_inputs->push_back(node);
160         } else {
161                 assert(node->effect->num_inputs() == node->incoming_links.size());
162                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
163                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
164                 }
165         }
166 }
167
168 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
169 {
170         assert(!finalized);
171         assert(inputs.size() == effect->num_inputs());
172         Node *node = add_node(effect);
173         for (unsigned i = 0; i < inputs.size(); ++i) {
174                 assert(node_map.count(inputs[i]) != 0);
175                 connect_nodes(node_map[inputs[i]], node);
176         }
177         return effect;
178 }
179
180 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
181 string replace_prefix(const string &text, const string &prefix)
182 {
183         string output;
184         size_t start = 0;
185
186         while (start < text.size()) {
187                 size_t pos = text.find("PREFIX(", start);
188                 if (pos == string::npos) {
189                         output.append(text.substr(start, string::npos));
190                         break;
191                 }
192
193                 output.append(text.substr(start, pos - start));
194                 output.append(prefix);
195                 output.append("_");
196
197                 pos += strlen("PREFIX(");
198         
199                 // Output stuff until we find the matching ), which we then eat.
200                 int depth = 1;
201                 size_t end_arg_pos = pos;
202                 while (end_arg_pos < text.size()) {
203                         if (text[end_arg_pos] == '(') {
204                                 ++depth;
205                         } else if (text[end_arg_pos] == ')') {
206                                 --depth;
207                                 if (depth == 0) {
208                                         break;
209                                 }
210                         }
211                         ++end_arg_pos;
212                 }
213                 output.append(text.substr(pos, end_arg_pos - pos));
214                 ++end_arg_pos;
215                 assert(depth == 0);
216                 start = end_arg_pos;
217         }
218         return output;
219 }
220
221 Phase *EffectChain::compile_glsl_program(
222         const vector<Node *> &inputs,
223         const vector<Node *> &effects)
224 {
225         Phase *phase = new Phase;
226         assert(!effects.empty());
227
228         // Deduplicate the inputs.
229         vector<Node *> true_inputs = inputs;
230         sort(true_inputs.begin(), true_inputs.end());
231         true_inputs.erase(unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
232
233         bool input_needs_mipmaps = false;
234         string frag_shader = read_file("header.frag");
235
236         // Create functions for all the texture inputs that we need.
237         for (unsigned i = 0; i < true_inputs.size(); ++i) {
238                 Node *input = true_inputs[i];
239                 char effect_id[256];
240                 sprintf(effect_id, "in%u", i);
241                 phase->effect_ids.insert(make_pair(input, effect_id));
242         
243                 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
244                 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
245                 frag_shader += "\treturn texture2D(tex_" + string(effect_id) + ", tc);\n";
246                 frag_shader += "}\n";
247                 frag_shader += "\n";
248         }
249
250         vector<Node *> sorted_effects = topological_sort(effects);
251
252         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
253                 Node *node = sorted_effects[i];
254                 char effect_id[256];
255                 sprintf(effect_id, "eff%u", i);
256                 phase->effect_ids.insert(make_pair(node, effect_id));
257
258                 if (node->incoming_links.size() == 1) {
259                         frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
260                 } else {
261                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
262                                 char buf[256];
263                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
264                                 frag_shader += buf;
265                         }
266                 }
267         
268                 frag_shader += "\n";
269                 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
270                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id);
271                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
272                 frag_shader += "#undef PREFIX\n";
273                 frag_shader += "#undef FUNCNAME\n";
274                 if (node->incoming_links.size() == 1) {
275                         frag_shader += "#undef INPUT\n";
276                 } else {
277                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
278                                 char buf[256];
279                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
280                                 frag_shader += buf;
281                         }
282                 }
283                 frag_shader += "\n";
284
285                 input_needs_mipmaps |= node->effect->needs_mipmaps();
286         }
287         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
288                 Node *node = sorted_effects[i];
289                 if (node->effect->num_inputs() == 0) {
290                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
291                 }
292         }
293         frag_shader += string("#define INPUT ") + phase->effect_ids[sorted_effects.back()] + "\n";
294         frag_shader.append(read_file("footer.frag"));
295
296         phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader);
297         phase->input_needs_mipmaps = input_needs_mipmaps;
298         phase->inputs = true_inputs;
299         phase->effects = sorted_effects;
300
301         return phase;
302 }
303
304 // Construct GLSL programs, starting at the given effect and following
305 // the chain from there. We end a program every time we come to an effect
306 // marked as "needs texture bounce", one that is used by multiple other
307 // effects, every time an effect wants to change the output size,
308 // and of course at the end.
309 //
310 // We follow a quite simple depth-first search from the output, although
311 // without any explicit recursion.
312 void EffectChain::construct_glsl_programs(Node *output)
313 {
314         // Which effects have already been completed?
315         // We need to keep track of it, as an effect with multiple outputs
316         // could otherwise be calculated multiple times.
317         set<Node *> completed_effects;
318
319         // Effects in the current phase, as well as inputs (outputs from other phases
320         // that we depend on). Note that since we start iterating from the end,
321         // the effect list will be in the reverse order.
322         vector<Node *> this_phase_inputs;
323         vector<Node *> this_phase_effects;
324
325         // Effects that we have yet to calculate, but that we know should
326         // be in the current phase.
327         stack<Node *> effects_todo_this_phase;
328
329         // Effects that we have yet to calculate, but that come from other phases.
330         // We delay these until we have this phase done in its entirety,
331         // at which point we pick any of them and start a new phase from that.
332         stack<Node *> effects_todo_other_phases;
333
334         effects_todo_this_phase.push(output);
335
336         for ( ;; ) {  // Termination condition within loop.
337                 if (!effects_todo_this_phase.empty()) {
338                         // OK, we have more to do this phase.
339                         Node *node = effects_todo_this_phase.top();
340                         effects_todo_this_phase.pop();
341
342                         // This should currently only happen for effects that are inputs
343                         // (either true inputs or phase outputs). We special-case inputs,
344                         // and then deduplicate phase outputs in compile_glsl_program().
345                         if (node->effect->num_inputs() == 0) {
346                                 if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
347                                         continue;
348                                 }
349                         } else {
350                                 assert(completed_effects.count(node) == 0);
351                         }
352
353                         this_phase_effects.push_back(node);
354                         completed_effects.insert(node);
355
356                         // Find all the dependencies of this effect, and add them to the stack.
357                         vector<Node *> deps = node->incoming_links;
358                         assert(node->effect->num_inputs() == deps.size());
359                         for (unsigned i = 0; i < deps.size(); ++i) {
360                                 bool start_new_phase = false;
361
362                                 // FIXME: If we sample directly from a texture, we won't need this.
363                                 if (node->effect->needs_texture_bounce()) {
364                                         start_new_phase = true;
365                                 }
366
367                                 if (deps[i]->outgoing_links.size() > 1) {
368                                         if (deps[i]->effect->num_inputs() > 0) {
369                                                 // More than one effect uses this as the input,
370                                                 // and it is not a texture itself.
371                                                 // The easiest thing to do (and probably also the safest
372                                                 // performance-wise in most cases) is to bounce it to a texture
373                                                 // and then let the next passes read from that.
374                                                 start_new_phase = true;
375                                         } else {
376                                                 // For textures, we try to be slightly more clever;
377                                                 // if none of our outputs need a bounce, we don't bounce
378                                                 // but instead simply use the effect many times.
379                                                 //
380                                                 // Strictly speaking, we could bounce it for some outputs
381                                                 // and use it directly for others, but the processing becomes
382                                                 // somewhat simpler if the effect is only used in one such way.
383                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
384                                                         Node *rdep = deps[i]->outgoing_links[j];
385                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
386                                                 }
387                                         }
388                                 }
389
390                                 if (deps[i]->effect->changes_output_size()) {
391                                         start_new_phase = true;
392                                 }
393
394                                 if (start_new_phase) {
395                                         effects_todo_other_phases.push(deps[i]);
396                                         this_phase_inputs.push_back(deps[i]);
397                                 } else {
398                                         effects_todo_this_phase.push(deps[i]);
399                                 }
400                         }
401                         continue;
402                 }
403
404                 // No more effects to do this phase. Take all the ones we have,
405                 // and create a GLSL program for it.
406                 if (!this_phase_effects.empty()) {
407                         reverse(this_phase_effects.begin(), this_phase_effects.end());
408                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
409                         this_phase_effects.back()->phase = phases.back();
410                         this_phase_inputs.clear();
411                         this_phase_effects.clear();
412                 }
413                 assert(this_phase_inputs.empty());
414                 assert(this_phase_effects.empty());
415
416                 // If we have no effects left, exit.
417                 if (effects_todo_other_phases.empty()) {
418                         break;
419                 }
420
421                 Node *node = effects_todo_other_phases.top();
422                 effects_todo_other_phases.pop();
423
424                 if (completed_effects.count(node) == 0) {
425                         // Start a new phase, calculating from this effect.
426                         effects_todo_this_phase.push(node);
427                 }
428         }
429
430         // Finally, since the phases are found from the output but must be executed
431         // from the input(s), reverse them, too.
432         reverse(phases.begin(), phases.end());
433 }
434
435 void EffectChain::output_dot(const char *filename)
436 {
437         if (movit_debug_level != MOVIT_DEBUG_ON) {
438                 return;
439         }
440
441         FILE *fp = fopen(filename, "w");
442         if (fp == NULL) {
443                 perror(filename);
444                 exit(1);
445         }
446
447         fprintf(fp, "digraph G {\n");
448         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
449         for (unsigned i = 0; i < nodes.size(); ++i) {
450                 // Find out which phase this event belongs to.
451                 vector<int> in_phases;
452                 for (unsigned j = 0; j < phases.size(); ++j) {
453                         const Phase* p = phases[j];
454                         if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
455                                 in_phases.push_back(j);
456                         }
457                 }
458
459                 if (in_phases.empty()) {
460                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
461                 } else if (in_phases.size() == 1) {
462                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
463                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
464                                 (in_phases[0] % 8) + 1);
465                 } else {
466                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
467                         // But alas.
468                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
469                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
470                                 (in_phases[0] % 8) + 1);
471                 }
472
473                 char from_node_id[256];
474                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
475
476                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
477                         char to_node_id[256];
478                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
479
480                         vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
481                         output_dot_edge(fp, from_node_id, to_node_id, labels);
482                 }
483
484                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
485                         // Output node.
486                         vector<string> labels = get_labels_for_edge(nodes[i], NULL);
487                         output_dot_edge(fp, from_node_id, "output", labels);
488                 }
489         }
490         fprintf(fp, "}\n");
491
492         fclose(fp);
493 }
494
495 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
496 {
497         vector<string> labels;
498
499         if (to != NULL && to->effect->needs_texture_bounce()) {
500                 labels.push_back("needs_bounce");
501         }
502         if (from->effect->changes_output_size()) {
503                 labels.push_back("resize");
504         }
505
506         switch (from->output_color_space) {
507         case COLORSPACE_INVALID:
508                 labels.push_back("spc[invalid]");
509                 break;
510         case COLORSPACE_REC_601_525:
511                 labels.push_back("spc[rec601-525]");
512                 break;
513         case COLORSPACE_REC_601_625:
514                 labels.push_back("spc[rec601-625]");
515                 break;
516         default:
517                 break;
518         }
519
520         switch (from->output_gamma_curve) {
521         case GAMMA_INVALID:
522                 labels.push_back("gamma[invalid]");
523                 break;
524         case GAMMA_sRGB:
525                 labels.push_back("gamma[sRGB]");
526                 break;
527         case GAMMA_REC_601:  // and GAMMA_REC_709
528                 labels.push_back("gamma[rec601/709]");
529                 break;
530         default:
531                 break;
532         }
533
534         switch (from->output_alpha_type) {
535         case ALPHA_INVALID:
536                 labels.push_back("alpha[invalid]");
537                 break;
538         case ALPHA_BLANK:
539                 labels.push_back("alpha[blank]");
540                 break;
541         case ALPHA_POSTMULTIPLIED:
542                 labels.push_back("alpha[postmult]");
543                 break;
544         default:
545                 break;
546         }
547
548         return labels;
549 }
550
551 void EffectChain::output_dot_edge(FILE *fp,
552                                   const string &from_node_id,
553                                   const string &to_node_id,
554                                   const vector<string> &labels)
555 {
556         if (labels.empty()) {
557                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
558         } else {
559                 string label = labels[0];
560                 for (unsigned k = 1; k < labels.size(); ++k) {
561                         label += ", " + labels[k];
562                 }
563                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
564         }
565 }
566
567 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
568 {
569         unsigned scaled_width, scaled_height;
570
571         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
572                 // Same aspect, or W/H > aspect (image is wider than the frame).
573                 // In either case, keep width, and adjust height.
574                 scaled_width = width;
575                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
576         } else {
577                 // W/H < aspect (image is taller than the frame), so keep height,
578                 // and adjust width.
579                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
580                 scaled_height = height;
581         }
582
583         // We should be consistently larger or smaller then the existing choice,
584         // since we have the same aspect.
585         assert(!(scaled_width < *output_width && scaled_height > *output_height));
586         assert(!(scaled_height < *output_height && scaled_width > *output_width));
587
588         if (scaled_width >= *output_width && scaled_height >= *output_height) {
589                 *output_width = scaled_width;
590                 *output_height = scaled_height;
591         }
592 }
593
594 // Propagate input texture sizes throughout, and inform effects downstream.
595 // (Like a lot of other code, we depend on effects being in topological order.)
596 void EffectChain::inform_input_sizes(Phase *phase)
597 {
598         // All effects that have a defined size (inputs and RTT inputs)
599         // get that. Reset all others.
600         for (unsigned i = 0; i < phase->effects.size(); ++i) {
601                 Node *node = phase->effects[i];
602                 if (node->effect->num_inputs() == 0) {
603                         Input *input = static_cast<Input *>(node->effect);
604                         node->output_width = input->get_width();
605                         node->output_height = input->get_height();
606                         assert(node->output_width != 0);
607                         assert(node->output_height != 0);
608                 } else {
609                         node->output_width = node->output_height = 0;
610                 }
611         }
612         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
613                 Node *input = phase->inputs[i];
614                 input->output_width = input->phase->virtual_output_width;
615                 input->output_height = input->phase->virtual_output_height;
616                 assert(input->output_width != 0);
617                 assert(input->output_height != 0);
618         }
619
620         // Now propagate from the inputs towards the end, and inform as we go.
621         // The rules are simple:
622         //
623         //   1. Don't touch effects that already have given sizes (ie., inputs).
624         //   2. If all of your inputs have the same size, that will be your output size.
625         //   3. Otherwise, your output size is 0x0.
626         for (unsigned i = 0; i < phase->effects.size(); ++i) {
627                 Node *node = phase->effects[i];
628                 if (node->effect->num_inputs() == 0) {
629                         continue;
630                 }
631                 unsigned this_output_width = 0;
632                 unsigned this_output_height = 0;
633                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
634                         Node *input = node->incoming_links[j];
635                         node->effect->inform_input_size(j, input->output_width, input->output_height);
636                         if (j == 0) {
637                                 this_output_width = input->output_width;
638                                 this_output_height = input->output_height;
639                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
640                                 // Inputs disagree.
641                                 this_output_width = 0;
642                                 this_output_height = 0;
643                         }
644                 }
645                 node->output_width = this_output_width;
646                 node->output_height = this_output_height;
647         }
648 }
649
650 // Note: You should call inform_input_sizes() before this, as the last effect's
651 // desired output size might change based on the inputs.
652 void EffectChain::find_output_size(Phase *phase)
653 {
654         Node *output_node = phase->effects.back();
655
656         // If the last effect explicitly sets an output size, use that.
657         if (output_node->effect->changes_output_size()) {
658                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
659                                                      &phase->virtual_output_width, &phase->virtual_output_height);
660                 return;
661         }
662
663         // If all effects have the same size, use that.
664         unsigned output_width = 0, output_height = 0;
665         bool all_inputs_same_size = true;
666
667         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
668                 Node *input = phase->inputs[i];
669                 assert(input->phase->output_width != 0);
670                 assert(input->phase->output_height != 0);
671                 if (output_width == 0 && output_height == 0) {
672                         output_width = input->phase->virtual_output_width;
673                         output_height = input->phase->virtual_output_height;
674                 } else if (output_width != input->phase->virtual_output_width ||
675                            output_height != input->phase->virtual_output_height) {
676                         all_inputs_same_size = false;
677                 }
678         }
679         for (unsigned i = 0; i < phase->effects.size(); ++i) {
680                 Effect *effect = phase->effects[i]->effect;
681                 if (effect->num_inputs() != 0) {
682                         continue;
683                 }
684
685                 Input *input = static_cast<Input *>(effect);
686                 if (output_width == 0 && output_height == 0) {
687                         output_width = input->get_width();
688                         output_height = input->get_height();
689                 } else if (output_width != input->get_width() ||
690                            output_height != input->get_height()) {
691                         all_inputs_same_size = false;
692                 }
693         }
694
695         if (all_inputs_same_size) {
696                 assert(output_width != 0);
697                 assert(output_height != 0);
698                 phase->virtual_output_width = phase->output_width = output_width;
699                 phase->virtual_output_height = phase->output_height = output_height;
700                 return;
701         }
702
703         // If not, fit all the inputs into the current aspect, and select the largest one. 
704         output_width = 0;
705         output_height = 0;
706         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
707                 Node *input = phase->inputs[i];
708                 assert(input->phase->output_width != 0);
709                 assert(input->phase->output_height != 0);
710                 size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
711         }
712         for (unsigned i = 0; i < phase->effects.size(); ++i) {
713                 Effect *effect = phase->effects[i]->effect;
714                 if (effect->num_inputs() != 0) {
715                         continue;
716                 }
717
718                 Input *input = static_cast<Input *>(effect);
719                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
720         }
721         assert(output_width != 0);
722         assert(output_height != 0);
723         phase->virtual_output_width = phase->output_width = output_width;
724         phase->virtual_output_height = phase->output_height = output_height;
725 }
726
727 void EffectChain::sort_all_nodes_topologically()
728 {
729         nodes = topological_sort(nodes);
730 }
731
732 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
733 {
734         set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
735         vector<Node *> sorted_list;
736         for (unsigned i = 0; i < nodes.size(); ++i) {
737                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
738         }
739         reverse(sorted_list.begin(), sorted_list.end());
740         return sorted_list;
741 }
742
743 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
744 {
745         if (nodes_left_to_visit->count(node) == 0) {
746                 return;
747         }
748         nodes_left_to_visit->erase(node);
749         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
750                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
751         }
752         sorted_list->push_back(node);
753 }
754
755 void EffectChain::find_color_spaces_for_inputs()
756 {
757         for (unsigned i = 0; i < nodes.size(); ++i) {
758                 Node *node = nodes[i];
759                 if (node->disabled) {
760                         continue;
761                 }
762                 if (node->incoming_links.size() == 0) {
763                         Input *input = static_cast<Input *>(node->effect);
764                         node->output_color_space = input->get_color_space();
765                         node->output_gamma_curve = input->get_gamma_curve();
766
767                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
768                         switch (alpha_handling) {
769                         case Effect::OUTPUT_BLANK_ALPHA:
770                                 node->output_alpha_type = ALPHA_BLANK;
771                                 break;
772                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
773                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
774                                 break;
775                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
776                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
777                                 break;
778                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
779                         case Effect::DONT_CARE_ALPHA_TYPE:
780                         default:
781                                 assert(false);
782                         }
783
784                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
785                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
786                         }
787                 }
788         }
789 }
790
791 // Propagate gamma and color space information as far as we can in the graph.
792 // The rules are simple: Anything where all the inputs agree, get that as
793 // output as well. Anything else keeps having *_INVALID.
794 void EffectChain::propagate_gamma_and_color_space()
795 {
796         // We depend on going through the nodes in order.
797         sort_all_nodes_topologically();
798
799         for (unsigned i = 0; i < nodes.size(); ++i) {
800                 Node *node = nodes[i];
801                 if (node->disabled) {
802                         continue;
803                 }
804                 assert(node->incoming_links.size() == node->effect->num_inputs());
805                 if (node->incoming_links.size() == 0) {
806                         assert(node->output_color_space != COLORSPACE_INVALID);
807                         assert(node->output_gamma_curve != GAMMA_INVALID);
808                         continue;
809                 }
810
811                 Colorspace color_space = node->incoming_links[0]->output_color_space;
812                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
813                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
814                         if (node->incoming_links[j]->output_color_space != color_space) {
815                                 color_space = COLORSPACE_INVALID;
816                         }
817                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
818                                 gamma_curve = GAMMA_INVALID;
819                         }
820                 }
821
822                 // The conversion effects already have their outputs set correctly,
823                 // so leave them alone.
824                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
825                         node->output_color_space = color_space;
826                 }               
827                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
828                     node->effect->effect_type_id() != "GammaExpansionEffect") {
829                         node->output_gamma_curve = gamma_curve;
830                 }               
831         }
832 }
833
834 // Propagate alpha information as far as we can in the graph.
835 // Similar to propagate_gamma_and_color_space().
836 void EffectChain::propagate_alpha()
837 {
838         // We depend on going through the nodes in order.
839         sort_all_nodes_topologically();
840
841         for (unsigned i = 0; i < nodes.size(); ++i) {
842                 Node *node = nodes[i];
843                 if (node->disabled) {
844                         continue;
845                 }
846                 assert(node->incoming_links.size() == node->effect->num_inputs());
847                 if (node->incoming_links.size() == 0) {
848                         assert(node->output_alpha_type != ALPHA_INVALID);
849                         continue;
850                 }
851
852                 // The alpha multiplication/division effects are special cases.
853                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
854                         assert(node->incoming_links.size() == 1);
855                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
856                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
857                         continue;
858                 }
859                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
860                         assert(node->incoming_links.size() == 1);
861                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
862                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
863                         continue;
864                 }
865
866                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
867                 // because they are the only one that _need_ postmultiplied alpha.
868                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
869                     node->effect->effect_type_id() == "GammaExpansionEffect") {
870                         assert(node->incoming_links.size() == 1);
871                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
872                                 node->output_alpha_type = ALPHA_BLANK;
873                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
874                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
875                         } else {
876                                 node->output_alpha_type = ALPHA_INVALID;
877                         }
878                         continue;
879                 }
880
881                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
882                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
883                 // taken care of above. Rationale: Even if you could imagine
884                 // e.g. an effect that took in an image and set alpha=1.0
885                 // unconditionally, it wouldn't make any sense to have it as
886                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
887                 // got its input pre- or postmultiplied, so it wouldn't know
888                 // whether to divide away the old alpha or not.
889                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
890                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
891                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
892                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
893
894                 // If the node has multiple inputs, check that they are all valid and
895                 // the same.
896                 bool any_invalid = false;
897                 bool any_premultiplied = false;
898                 bool any_postmultiplied = false;
899
900                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
901                         switch (node->incoming_links[j]->output_alpha_type) {
902                         case ALPHA_INVALID:
903                                 any_invalid = true;
904                                 break;
905                         case ALPHA_BLANK:
906                                 // Blank is good as both pre- and postmultiplied alpha,
907                                 // so just ignore it.
908                                 break;
909                         case ALPHA_PREMULTIPLIED:
910                                 any_premultiplied = true;
911                                 break;
912                         case ALPHA_POSTMULTIPLIED:
913                                 any_postmultiplied = true;
914                                 break;
915                         default:
916                                 assert(false);
917                         }
918                 }
919
920                 if (any_invalid) {
921                         node->output_alpha_type = ALPHA_INVALID;
922                         continue;
923                 }
924
925                 // Inputs must be of the same type.
926                 if (any_premultiplied && any_postmultiplied) {
927                         node->output_alpha_type = ALPHA_INVALID;
928                         continue;
929                 }
930
931                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
932                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
933                         // If the effect has asked for premultiplied alpha, check that it has got it.
934                         if (any_postmultiplied) {
935                                 node->output_alpha_type = ALPHA_INVALID;
936                         } else if (!any_premultiplied &&
937                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
938                                 // Blank input alpha, and the effect preserves blank alpha.
939                                 node->output_alpha_type = ALPHA_BLANK;
940                         } else {
941                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
942                         }
943                 } else {
944                         // OK, all inputs are the same, and this effect is not going
945                         // to change it.
946                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
947                         if (any_premultiplied) {
948                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
949                         } else if (any_postmultiplied) {
950                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
951                         } else {
952                                 node->output_alpha_type = ALPHA_BLANK;
953                         }
954                 }
955         }
956 }
957
958 bool EffectChain::node_needs_colorspace_fix(Node *node)
959 {
960         if (node->disabled) {
961                 return false;
962         }
963         if (node->effect->num_inputs() == 0) {
964                 return false;
965         }
966
967         // propagate_gamma_and_color_space() has already set our output
968         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
969         if (node->output_color_space == COLORSPACE_INVALID) {
970                 return true;
971         }
972         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
973 }
974
975 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
976 // the graph. Our strategy is not always optimal, but quite simple:
977 // Find an effect that's as early as possible where the inputs are of
978 // unacceptable colorspaces (that is, either different, or, if the effect only
979 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
980 // propagate the information anew, and repeat until there are no more such
981 // effects.
982 void EffectChain::fix_internal_color_spaces()
983 {
984         unsigned colorspace_propagation_pass = 0;
985         bool found_any;
986         do {
987                 found_any = false;
988                 for (unsigned i = 0; i < nodes.size(); ++i) {
989                         Node *node = nodes[i];
990                         if (!node_needs_colorspace_fix(node)) {
991                                 continue;
992                         }
993
994                         // Go through each input that is not sRGB, and insert
995                         // a colorspace conversion after it.
996                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
997                                 Node *input = node->incoming_links[j];
998                                 assert(input->output_color_space != COLORSPACE_INVALID);
999                                 if (input->output_color_space == COLORSPACE_sRGB) {
1000                                         continue;
1001                                 }
1002                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1003                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1004                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1005                                 conversion->output_color_space = COLORSPACE_sRGB;
1006                                 replace_sender(input, conversion);
1007                                 connect_nodes(input, conversion);
1008                         }
1009
1010                         // Re-sort topologically, and propagate the new information.
1011                         propagate_gamma_and_color_space();
1012                         
1013                         found_any = true;
1014                         break;
1015                 }
1016         
1017                 char filename[256];
1018                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1019                 output_dot(filename);
1020                 assert(colorspace_propagation_pass < 100);
1021         } while (found_any);
1022
1023         for (unsigned i = 0; i < nodes.size(); ++i) {
1024                 Node *node = nodes[i];
1025                 if (node->disabled) {
1026                         continue;
1027                 }
1028                 assert(node->output_color_space != COLORSPACE_INVALID);
1029         }
1030 }
1031
1032 bool EffectChain::node_needs_alpha_fix(Node *node)
1033 {
1034         if (node->disabled) {
1035                 return false;
1036         }
1037
1038         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1039         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1040         return (node->output_alpha_type == ALPHA_INVALID);
1041 }
1042
1043 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1044 // the graph. Similar to fix_internal_color_spaces().
1045 void EffectChain::fix_internal_alpha(unsigned step)
1046 {
1047         unsigned alpha_propagation_pass = 0;
1048         bool found_any;
1049         do {
1050                 found_any = false;
1051                 for (unsigned i = 0; i < nodes.size(); ++i) {
1052                         Node *node = nodes[i];
1053                         if (!node_needs_alpha_fix(node)) {
1054                                 continue;
1055                         }
1056
1057                         // If we need to fix up GammaExpansionEffect, then clearly something
1058                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1059                         // is meaningless.
1060                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1061
1062                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1063
1064                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1065                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1066                                 assert(node->incoming_links.size() == 1);
1067                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1068                                 desired_type = ALPHA_POSTMULTIPLIED;
1069                         }
1070
1071                         // Go through each input that is not premultiplied alpha, and insert
1072                         // a conversion before it.
1073                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1074                                 Node *input = node->incoming_links[j];
1075                                 assert(input->output_alpha_type != ALPHA_INVALID);
1076                                 if (input->output_alpha_type == desired_type ||
1077                                     input->output_alpha_type == ALPHA_BLANK) {
1078                                         continue;
1079                                 }
1080                                 Node *conversion;
1081                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1082                                         conversion = add_node(new AlphaMultiplicationEffect());
1083                                 } else {
1084                                         conversion = add_node(new AlphaDivisionEffect());
1085                                 }
1086                                 conversion->output_alpha_type = desired_type;
1087                                 replace_sender(input, conversion);
1088                                 connect_nodes(input, conversion);
1089                         }
1090
1091                         // Re-sort topologically, and propagate the new information.
1092                         propagate_gamma_and_color_space();
1093                         propagate_alpha();
1094                         
1095                         found_any = true;
1096                         break;
1097                 }
1098         
1099                 char filename[256];
1100                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1101                 output_dot(filename);
1102                 assert(alpha_propagation_pass < 100);
1103         } while (found_any);
1104
1105         for (unsigned i = 0; i < nodes.size(); ++i) {
1106                 Node *node = nodes[i];
1107                 if (node->disabled) {
1108                         continue;
1109                 }
1110                 assert(node->output_alpha_type != ALPHA_INVALID);
1111         }
1112 }
1113
1114 // Make so that the output is in the desired color space.
1115 void EffectChain::fix_output_color_space()
1116 {
1117         Node *output = find_output_node();
1118         if (output->output_color_space != output_format.color_space) {
1119                 Node *conversion = add_node(new ColorspaceConversionEffect());
1120                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1121                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1122                 conversion->output_color_space = output_format.color_space;
1123                 connect_nodes(output, conversion);
1124                 propagate_alpha();
1125                 propagate_gamma_and_color_space();
1126         }
1127 }
1128
1129 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1130 void EffectChain::fix_output_alpha()
1131 {
1132         Node *output = find_output_node();
1133         assert(output->output_alpha_type != ALPHA_INVALID);
1134         if (output->output_alpha_type == ALPHA_BLANK) {
1135                 // No alpha output, so we don't care.
1136                 return;
1137         }
1138         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1139             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1140                 Node *conversion = add_node(new AlphaDivisionEffect());
1141                 connect_nodes(output, conversion);
1142                 propagate_alpha();
1143                 propagate_gamma_and_color_space();
1144         }
1145         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1146             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1147                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1148                 connect_nodes(output, conversion);
1149                 propagate_alpha();
1150                 propagate_gamma_and_color_space();
1151         }
1152 }
1153
1154 bool EffectChain::node_needs_gamma_fix(Node *node)
1155 {
1156         if (node->disabled) {
1157                 return false;
1158         }
1159
1160         // Small hack since the output is not an explicit node:
1161         // If we are the last node and our output is in the wrong
1162         // space compared to EffectChain's output, we need to fix it.
1163         // This will only take us to linear, but fix_output_gamma()
1164         // will come and take us to the desired output gamma
1165         // if it is needed.
1166         //
1167         // This needs to be before everything else, since it could
1168         // even apply to inputs (if they are the only effect).
1169         if (node->outgoing_links.empty() &&
1170             node->output_gamma_curve != output_format.gamma_curve &&
1171             node->output_gamma_curve != GAMMA_LINEAR) {
1172                 return true;
1173         }
1174
1175         if (node->effect->num_inputs() == 0) {
1176                 return false;
1177         }
1178
1179         // propagate_gamma_and_color_space() has already set our output
1180         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1181         // except for GammaCompressionEffect.
1182         if (node->output_gamma_curve == GAMMA_INVALID) {
1183                 return true;
1184         }
1185         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1186                 assert(node->incoming_links.size() == 1);
1187                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1188         }
1189
1190         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1191 }
1192
1193 // Very similar to fix_internal_color_spaces(), but for gamma.
1194 // There is one difference, though; before we start adding conversion nodes,
1195 // we see if we can get anything out of asking the sources to deliver
1196 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1197 // does that part, while fix_internal_gamma_by_inserting_nodes()
1198 // inserts nodes as needed afterwards.
1199 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1200 {
1201         unsigned gamma_propagation_pass = 0;
1202         bool found_any;
1203         do {
1204                 found_any = false;
1205                 for (unsigned i = 0; i < nodes.size(); ++i) {
1206                         Node *node = nodes[i];
1207                         if (!node_needs_gamma_fix(node)) {
1208                                 continue;
1209                         }
1210
1211                         // See if all inputs can give us linear gamma. If not, leave it.
1212                         vector<Node *> nonlinear_inputs;
1213                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1214                         assert(!nonlinear_inputs.empty());
1215
1216                         bool all_ok = true;
1217                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1218                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1219                                 all_ok &= input->can_output_linear_gamma();
1220                         }
1221
1222                         if (!all_ok) {
1223                                 continue;
1224                         }
1225
1226                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1227                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1228                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1229                         }
1230
1231                         // Re-sort topologically, and propagate the new information.
1232                         propagate_gamma_and_color_space();
1233                         
1234                         found_any = true;
1235                         break;
1236                 }
1237         
1238                 char filename[256];
1239                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1240                 output_dot(filename);
1241                 assert(gamma_propagation_pass < 100);
1242         } while (found_any);
1243 }
1244
1245 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1246 {
1247         unsigned gamma_propagation_pass = 0;
1248         bool found_any;
1249         do {
1250                 found_any = false;
1251                 for (unsigned i = 0; i < nodes.size(); ++i) {
1252                         Node *node = nodes[i];
1253                         if (!node_needs_gamma_fix(node)) {
1254                                 continue;
1255                         }
1256
1257                         // Special case: We could be an input and still be asked to
1258                         // fix our gamma; if so, we should be the only node
1259                         // (as node_needs_gamma_fix() would only return true in
1260                         // for an input in that case). That means we should insert
1261                         // a conversion node _after_ ourselves.
1262                         if (node->incoming_links.empty()) {
1263                                 assert(node->outgoing_links.empty());
1264                                 Node *conversion = add_node(new GammaExpansionEffect());
1265                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1266                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1267                                 connect_nodes(node, conversion);
1268                         }
1269
1270                         // If not, go through each input that is not linear gamma,
1271                         // and insert a gamma conversion after it.
1272                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1273                                 Node *input = node->incoming_links[j];
1274                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1275                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1276                                         continue;
1277                                 }
1278                                 Node *conversion = add_node(new GammaExpansionEffect());
1279                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1280                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1281                                 replace_sender(input, conversion);
1282                                 connect_nodes(input, conversion);
1283                         }
1284
1285                         // Re-sort topologically, and propagate the new information.
1286                         propagate_alpha();
1287                         propagate_gamma_and_color_space();
1288                         
1289                         found_any = true;
1290                         break;
1291                 }
1292         
1293                 char filename[256];
1294                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1295                 output_dot(filename);
1296                 assert(gamma_propagation_pass < 100);
1297         } while (found_any);
1298
1299         for (unsigned i = 0; i < nodes.size(); ++i) {
1300                 Node *node = nodes[i];
1301                 if (node->disabled) {
1302                         continue;
1303                 }
1304                 assert(node->output_gamma_curve != GAMMA_INVALID);
1305         }
1306 }
1307
1308 // Make so that the output is in the desired gamma.
1309 // Note that this assumes linear input gamma, so it might create the need
1310 // for another pass of fix_internal_gamma().
1311 void EffectChain::fix_output_gamma()
1312 {
1313         Node *output = find_output_node();
1314         if (output->output_gamma_curve != output_format.gamma_curve) {
1315                 Node *conversion = add_node(new GammaCompressionEffect());
1316                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1317                 conversion->output_gamma_curve = output_format.gamma_curve;
1318                 connect_nodes(output, conversion);
1319         }
1320 }
1321         
1322 // If the user has requested dither, add a DitherEffect right at the end
1323 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1324 // since dither is about the only effect that can _not_ be done in linear space.
1325 void EffectChain::add_dither_if_needed()
1326 {
1327         if (num_dither_bits == 0) {
1328                 return;
1329         }
1330         Node *output = find_output_node();
1331         Node *dither = add_node(new DitherEffect());
1332         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1333         connect_nodes(output, dither);
1334
1335         dither_effect = dither->effect;
1336 }
1337
1338 // Find the output node. This is, simply, one that has no outgoing links.
1339 // If there are multiple ones, the graph is malformed (we do not support
1340 // multiple outputs right now).
1341 Node *EffectChain::find_output_node()
1342 {
1343         vector<Node *> output_nodes;
1344         for (unsigned i = 0; i < nodes.size(); ++i) {
1345                 Node *node = nodes[i];
1346                 if (node->disabled) {
1347                         continue;
1348                 }
1349                 if (node->outgoing_links.empty()) {
1350                         output_nodes.push_back(node);
1351                 }
1352         }
1353         assert(output_nodes.size() == 1);
1354         return output_nodes[0];
1355 }
1356
1357 void EffectChain::finalize()
1358 {
1359         // Save the current locale, and set it to C, so that we can output decimal
1360         // numbers with printf and be sure to get them in the format mandated by GLSL.
1361         char *saved_locale = setlocale(LC_NUMERIC, "C");
1362
1363         // Output the graph as it is before we do any conversions on it.
1364         output_dot("step0-start.dot");
1365
1366         // Give each effect in turn a chance to rewrite its own part of the graph.
1367         // Note that if more effects are added as part of this, they will be
1368         // picked up as part of the same for loop, since they are added at the end.
1369         for (unsigned i = 0; i < nodes.size(); ++i) {
1370                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1371         }
1372         output_dot("step1-rewritten.dot");
1373
1374         find_color_spaces_for_inputs();
1375         output_dot("step2-input-colorspace.dot");
1376
1377         propagate_alpha();
1378         output_dot("step3-propagated-alpha.dot");
1379
1380         propagate_gamma_and_color_space();
1381         output_dot("step4-propagated-all.dot");
1382
1383         fix_internal_color_spaces();
1384         fix_internal_alpha(6);
1385         fix_output_color_space();
1386         output_dot("step7-output-colorspacefix.dot");
1387         fix_output_alpha();
1388         output_dot("step8-output-alphafix.dot");
1389
1390         // Note that we need to fix gamma after colorspace conversion,
1391         // because colorspace conversions might create needs for gamma conversions.
1392         // Also, we need to run an extra pass of fix_internal_gamma() after 
1393         // fixing the output gamma, as we only have conversions to/from linear,
1394         // and fix_internal_alpha() since GammaCompressionEffect needs
1395         // postmultiplied input.
1396         fix_internal_gamma_by_asking_inputs(9);
1397         fix_internal_gamma_by_inserting_nodes(10);
1398         fix_output_gamma();
1399         output_dot("step11-output-gammafix.dot");
1400         propagate_alpha();
1401         output_dot("step12-output-alpha-propagated.dot");
1402         fix_internal_alpha(13);
1403         output_dot("step14-output-alpha-fixed.dot");
1404         fix_internal_gamma_by_asking_inputs(15);
1405         fix_internal_gamma_by_inserting_nodes(16);
1406
1407         output_dot("step17-before-dither.dot");
1408
1409         add_dither_if_needed();
1410
1411         output_dot("step18-final.dot");
1412         
1413         // Construct all needed GLSL programs, starting at the output.
1414         construct_glsl_programs(find_output_node());
1415
1416         output_dot("step19-split-to-phases.dot");
1417
1418         assert(phases[0]->inputs.empty());
1419         
1420         finalized = true;
1421         setlocale(LC_NUMERIC, saved_locale);
1422 }
1423
1424 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1425 {
1426         assert(finalized);
1427
1428         // Save original viewport.
1429         GLuint x = 0, y = 0;
1430         GLuint fbo = 0;
1431
1432         if (width == 0 && height == 0) {
1433                 GLint viewport[4];
1434                 glGetIntegerv(GL_VIEWPORT, viewport);
1435                 x = viewport[0];
1436                 y = viewport[1];
1437                 width = viewport[2];
1438                 height = viewport[3];
1439         }
1440
1441         // Basic state.
1442         glDisable(GL_BLEND);
1443         check_error();
1444         glDisable(GL_DEPTH_TEST);
1445         check_error();
1446         glDepthMask(GL_FALSE);
1447         check_error();
1448
1449         glMatrixMode(GL_PROJECTION);
1450         glLoadIdentity();
1451         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1452
1453         glMatrixMode(GL_MODELVIEW);
1454         glLoadIdentity();
1455
1456         if (phases.size() > 1) {
1457                 glGenFramebuffers(1, &fbo);
1458                 check_error();
1459                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1460                 check_error();
1461         }
1462
1463         set<Node *> generated_mipmaps;
1464
1465         // We choose the simplest option of having one texture per output,
1466         // since otherwise this turns into an (albeit simple) register allocation problem.
1467         map<Phase *, GLuint> output_textures;
1468
1469         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1470                 // Find a texture for this phase.
1471                 inform_input_sizes(phases[phase]);
1472                 if (phase != phases.size() - 1) {
1473                         find_output_size(phases[phase]);
1474
1475                         GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height);
1476                         output_textures.insert(make_pair(phases[phase], tex_num));
1477                 }
1478
1479                 glUseProgram(phases[phase]->glsl_program_num);
1480                 check_error();
1481
1482                 // Set up RTT inputs for this phase.
1483                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1484                         glActiveTexture(GL_TEXTURE0 + sampler);
1485                         Node *input = phases[phase]->inputs[sampler];
1486                         glBindTexture(GL_TEXTURE_2D, output_textures[input->phase]);
1487                         check_error();
1488                         if (phases[phase]->input_needs_mipmaps) {
1489                                 if (generated_mipmaps.count(input) == 0) {
1490                                         glGenerateMipmap(GL_TEXTURE_2D);
1491                                         check_error();
1492                                         generated_mipmaps.insert(input);
1493                                 }
1494                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1495                                 check_error();
1496                         } else {
1497                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1498                                 check_error();
1499                         }
1500                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1501                         check_error();
1502                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1503                         check_error();
1504
1505                         string texture_name = string("tex_") + phases[phase]->effect_ids[input];
1506                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1507                         check_error();
1508                 }
1509
1510                 // And now the output.
1511                 if (phase == phases.size() - 1) {
1512                         // Last phase goes to the output the user specified.
1513                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1514                         check_error();
1515                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1516                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1517                         glViewport(x, y, width, height);
1518                         if (dither_effect != NULL) {
1519                                 CHECK(dither_effect->set_int("output_width", width));
1520                                 CHECK(dither_effect->set_int("output_height", height));
1521                         }
1522                 } else {
1523                         glFramebufferTexture2D(
1524                                 GL_FRAMEBUFFER,
1525                                 GL_COLOR_ATTACHMENT0,
1526                                 GL_TEXTURE_2D,
1527                                 output_textures[phases[phase]],
1528                                 0);
1529                         check_error();
1530                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1531                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1532                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1533                 }
1534
1535                 // Give the required parameters to all the effects.
1536                 unsigned sampler_num = phases[phase]->inputs.size();
1537                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1538                         Node *node = phases[phase]->effects[i];
1539                         node->effect->set_gl_state(phases[phase]->glsl_program_num, phases[phase]->effect_ids[node], &sampler_num);
1540                         check_error();
1541                 }
1542
1543                 // Now draw!
1544                 glBegin(GL_QUADS);
1545
1546                 glTexCoord2f(0.0f, 0.0f);
1547                 glVertex2f(0.0f, 0.0f);
1548
1549                 glTexCoord2f(1.0f, 0.0f);
1550                 glVertex2f(1.0f, 0.0f);
1551
1552                 glTexCoord2f(1.0f, 1.0f);
1553                 glVertex2f(1.0f, 1.0f);
1554
1555                 glTexCoord2f(0.0f, 1.0f);
1556                 glVertex2f(0.0f, 1.0f);
1557
1558                 glEnd();
1559                 check_error();
1560
1561                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1562                         Node *node = phases[phase]->effects[i];
1563                         node->effect->clear_gl_state();
1564                 }
1565         }
1566
1567         for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1568              texture_it != output_textures.end();
1569              ++texture_it) {
1570                 resource_pool->release_2d_texture(texture_it->second);
1571         }
1572
1573         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1574         check_error();
1575
1576         if (fbo != 0) {
1577                 glDeleteFramebuffers(1, &fbo);
1578                 check_error();
1579         }
1580 }
1581
1582 }  // namespace movit