In resizing effects, add the notion of a “virtual output size”.
[movit] / effect_chain.cpp
1 #define GL_GLEXT_PROTOTYPES 1
2
3 #include <stdio.h>
4 #include <math.h>
5 #include <string.h>
6 #include <locale.h>
7 #include <assert.h>
8 #include <GL/glew.h>
9
10 #include <algorithm>
11 #include <set>
12 #include <stack>
13 #include <vector>
14
15 #include "util.h"
16 #include "effect_chain.h"
17 #include "gamma_expansion_effect.h"
18 #include "gamma_compression_effect.h"
19 #include "colorspace_conversion_effect.h"
20 #include "alpha_multiplication_effect.h"
21 #include "alpha_division_effect.h"
22 #include "dither_effect.h"
23 #include "input.h"
24 #include "init.h"
25
26 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
27         : aspect_nom(aspect_nom),
28           aspect_denom(aspect_denom),
29           dither_effect(NULL),
30           fbo(0),
31           num_dither_bits(0),
32           finalized(false) {}
33
34 EffectChain::~EffectChain()
35 {
36         for (unsigned i = 0; i < nodes.size(); ++i) {
37                 if (nodes[i]->output_texture != 0) {
38                         glDeleteTextures(1, &nodes[i]->output_texture);
39                 }
40                 delete nodes[i]->effect;
41                 delete nodes[i];
42         }
43         for (unsigned i = 0; i < phases.size(); ++i) {
44                 glDeleteProgram(phases[i]->glsl_program_num);
45                 glDeleteShader(phases[i]->vertex_shader);
46                 glDeleteShader(phases[i]->fragment_shader);
47                 delete phases[i];
48         }
49         if (fbo != 0) {
50                 glDeleteFramebuffers(1, &fbo);
51         }
52 }
53
54 Input *EffectChain::add_input(Input *input)
55 {
56         inputs.push_back(input);
57         add_node(input);
58         return input;
59 }
60
61 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
62 {
63         output_format = format;
64         output_alpha_format = alpha_format;
65 }
66
67 Node *EffectChain::add_node(Effect *effect)
68 {
69         char effect_id[256];
70         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
71
72         Node *node = new Node;
73         node->effect = effect;
74         node->disabled = false;
75         node->effect_id = effect_id;
76         node->output_color_space = COLORSPACE_INVALID;
77         node->output_gamma_curve = GAMMA_INVALID;
78         node->output_alpha_type = ALPHA_INVALID;
79         node->output_texture = 0;
80
81         nodes.push_back(node);
82         node_map[effect] = node;
83         return node;
84 }
85
86 void EffectChain::connect_nodes(Node *sender, Node *receiver)
87 {
88         sender->outgoing_links.push_back(receiver);
89         receiver->incoming_links.push_back(sender);
90 }
91
92 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
93 {
94         new_receiver->incoming_links = old_receiver->incoming_links;
95         old_receiver->incoming_links.clear();
96         
97         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
98                 Node *sender = new_receiver->incoming_links[i];
99                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
100                         if (sender->outgoing_links[j] == old_receiver) {
101                                 sender->outgoing_links[j] = new_receiver;
102                         }
103                 }
104         }       
105 }
106
107 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
108 {
109         new_sender->outgoing_links = old_sender->outgoing_links;
110         old_sender->outgoing_links.clear();
111         
112         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
113                 Node *receiver = new_sender->outgoing_links[i];
114                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
115                         if (receiver->incoming_links[j] == old_sender) {
116                                 receiver->incoming_links[j] = new_sender;
117                         }
118                 }
119         }       
120 }
121
122 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
123 {
124         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
125                 if (sender->outgoing_links[i] == receiver) {
126                         sender->outgoing_links[i] = middle;
127                         middle->incoming_links.push_back(sender);
128                 }
129         }
130         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
131                 if (receiver->incoming_links[i] == sender) {
132                         receiver->incoming_links[i] = middle;
133                         middle->outgoing_links.push_back(receiver);
134                 }
135         }
136
137         assert(middle->incoming_links.size() == middle->effect->num_inputs());
138 }
139
140 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
141 {
142         if (node->output_gamma_curve == GAMMA_LINEAR &&
143             node->effect->effect_type_id() != "GammaCompressionEffect") {
144                 return;
145         }
146         if (node->effect->num_inputs() == 0) {
147                 nonlinear_inputs->push_back(node);
148         } else {
149                 assert(node->effect->num_inputs() == node->incoming_links.size());
150                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
151                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
152                 }
153         }
154 }
155
156 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
157 {
158         assert(inputs.size() == effect->num_inputs());
159         Node *node = add_node(effect);
160         for (unsigned i = 0; i < inputs.size(); ++i) {
161                 assert(node_map.count(inputs[i]) != 0);
162                 connect_nodes(node_map[inputs[i]], node);
163         }
164         return effect;
165 }
166
167 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
168 std::string replace_prefix(const std::string &text, const std::string &prefix)
169 {
170         std::string output;
171         size_t start = 0;
172
173         while (start < text.size()) {
174                 size_t pos = text.find("PREFIX(", start);
175                 if (pos == std::string::npos) {
176                         output.append(text.substr(start, std::string::npos));
177                         break;
178                 }
179
180                 output.append(text.substr(start, pos - start));
181                 output.append(prefix);
182                 output.append("_");
183
184                 pos += strlen("PREFIX(");
185         
186                 // Output stuff until we find the matching ), which we then eat.
187                 int depth = 1;
188                 size_t end_arg_pos = pos;
189                 while (end_arg_pos < text.size()) {
190                         if (text[end_arg_pos] == '(') {
191                                 ++depth;
192                         } else if (text[end_arg_pos] == ')') {
193                                 --depth;
194                                 if (depth == 0) {
195                                         break;
196                                 }
197                         }
198                         ++end_arg_pos;
199                 }
200                 output.append(text.substr(pos, end_arg_pos - pos));
201                 ++end_arg_pos;
202                 assert(depth == 0);
203                 start = end_arg_pos;
204         }
205         return output;
206 }
207
208 Phase *EffectChain::compile_glsl_program(
209         const std::vector<Node *> &inputs,
210         const std::vector<Node *> &effects)
211 {
212         assert(!effects.empty());
213
214         // Deduplicate the inputs.
215         std::vector<Node *> true_inputs = inputs;
216         std::sort(true_inputs.begin(), true_inputs.end());
217         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
218
219         bool input_needs_mipmaps = false;
220         std::string frag_shader = read_file("header.frag");
221
222         // Create functions for all the texture inputs that we need.
223         for (unsigned i = 0; i < true_inputs.size(); ++i) {
224                 Node *input = true_inputs[i];
225         
226                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
227                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
228                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
229                 frag_shader += "}\n";
230                 frag_shader += "\n";
231         }
232
233         std::vector<Node *> sorted_effects = topological_sort(effects);
234
235         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
236                 Node *node = sorted_effects[i];
237
238                 if (node->incoming_links.size() == 1) {
239                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
240                 } else {
241                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
242                                 char buf[256];
243                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
244                                 frag_shader += buf;
245                         }
246                 }
247         
248                 frag_shader += "\n";
249                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
250                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
251                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
252                 frag_shader += "#undef PREFIX\n";
253                 frag_shader += "#undef FUNCNAME\n";
254                 if (node->incoming_links.size() == 1) {
255                         frag_shader += "#undef INPUT\n";
256                 } else {
257                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
258                                 char buf[256];
259                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
260                                 frag_shader += buf;
261                         }
262                 }
263                 frag_shader += "\n";
264
265                 input_needs_mipmaps |= node->effect->needs_mipmaps();
266         }
267         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
268                 Node *node = sorted_effects[i];
269                 if (node->effect->num_inputs() == 0) {
270                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
271                 }
272         }
273         frag_shader += std::string("#define INPUT ") + sorted_effects.back()->effect_id + "\n";
274         frag_shader.append(read_file("footer.frag"));
275
276         if (movit_debug_level == MOVIT_DEBUG_ON) {
277                 // Output shader to a temporary file, for easier debugging.
278                 static int compiled_shader_num = 0;
279                 char filename[256];
280                 sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
281                 FILE *fp = fopen(filename, "w");
282                 if (fp == NULL) {
283                         perror(filename);
284                         exit(1);
285                 }
286                 fprintf(fp, "%s\n", frag_shader.c_str());
287                 fclose(fp);
288         }
289         
290         GLuint glsl_program_num = glCreateProgram();
291         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
292         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
293         glAttachShader(glsl_program_num, vs_obj);
294         check_error();
295         glAttachShader(glsl_program_num, fs_obj);
296         check_error();
297         glLinkProgram(glsl_program_num);
298         check_error();
299
300         Phase *phase = new Phase;
301         phase->glsl_program_num = glsl_program_num;
302         phase->vertex_shader = vs_obj;
303         phase->fragment_shader = fs_obj;
304         phase->input_needs_mipmaps = input_needs_mipmaps;
305         phase->inputs = true_inputs;
306         phase->effects = sorted_effects;
307
308         return phase;
309 }
310
311 // Construct GLSL programs, starting at the given effect and following
312 // the chain from there. We end a program every time we come to an effect
313 // marked as "needs texture bounce", one that is used by multiple other
314 // effects, every time an effect wants to change the output size,
315 // and of course at the end.
316 //
317 // We follow a quite simple depth-first search from the output, although
318 // without any explicit recursion.
319 void EffectChain::construct_glsl_programs(Node *output)
320 {
321         // Which effects have already been completed?
322         // We need to keep track of it, as an effect with multiple outputs
323         // could otherwise be calculated multiple times.
324         std::set<Node *> completed_effects;
325
326         // Effects in the current phase, as well as inputs (outputs from other phases
327         // that we depend on). Note that since we start iterating from the end,
328         // the effect list will be in the reverse order.
329         std::vector<Node *> this_phase_inputs;
330         std::vector<Node *> this_phase_effects;
331
332         // Effects that we have yet to calculate, but that we know should
333         // be in the current phase.
334         std::stack<Node *> effects_todo_this_phase;
335
336         // Effects that we have yet to calculate, but that come from other phases.
337         // We delay these until we have this phase done in its entirety,
338         // at which point we pick any of them and start a new phase from that.
339         std::stack<Node *> effects_todo_other_phases;
340
341         effects_todo_this_phase.push(output);
342
343         for ( ;; ) {  // Termination condition within loop.
344                 if (!effects_todo_this_phase.empty()) {
345                         // OK, we have more to do this phase.
346                         Node *node = effects_todo_this_phase.top();
347                         effects_todo_this_phase.pop();
348
349                         // This should currently only happen for effects that are inputs
350                         // (either true inputs or phase outputs). We special-case inputs,
351                         // and then deduplicate phase outputs in compile_glsl_program().
352                         if (node->effect->num_inputs() == 0) {
353                                 if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
354                                         continue;
355                                 }
356                         } else {
357                                 assert(completed_effects.count(node) == 0);
358                         }
359
360                         this_phase_effects.push_back(node);
361                         completed_effects.insert(node);
362
363                         // Find all the dependencies of this effect, and add them to the stack.
364                         std::vector<Node *> deps = node->incoming_links;
365                         assert(node->effect->num_inputs() == deps.size());
366                         for (unsigned i = 0; i < deps.size(); ++i) {
367                                 bool start_new_phase = false;
368
369                                 // FIXME: If we sample directly from a texture, we won't need this.
370                                 if (node->effect->needs_texture_bounce()) {
371                                         start_new_phase = true;
372                                 }
373
374                                 if (deps[i]->outgoing_links.size() > 1) {
375                                         if (deps[i]->effect->num_inputs() > 0) {
376                                                 // More than one effect uses this as the input,
377                                                 // and it is not a texture itself.
378                                                 // The easiest thing to do (and probably also the safest
379                                                 // performance-wise in most cases) is to bounce it to a texture
380                                                 // and then let the next passes read from that.
381                                                 start_new_phase = true;
382                                         } else {
383                                                 // For textures, we try to be slightly more clever;
384                                                 // if none of our outputs need a bounce, we don't bounce
385                                                 // but instead simply use the effect many times.
386                                                 //
387                                                 // Strictly speaking, we could bounce it for some outputs
388                                                 // and use it directly for others, but the processing becomes
389                                                 // somewhat simpler if the effect is only used in one such way.
390                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
391                                                         Node *rdep = deps[i]->outgoing_links[j];
392                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
393                                                 }
394                                         }
395                                 }
396
397                                 if (deps[i]->effect->changes_output_size()) {
398                                         start_new_phase = true;
399                                 }
400
401                                 if (start_new_phase) {
402                                         effects_todo_other_phases.push(deps[i]);
403                                         this_phase_inputs.push_back(deps[i]);
404                                 } else {
405                                         effects_todo_this_phase.push(deps[i]);
406                                 }
407                         }
408                         continue;
409                 }
410
411                 // No more effects to do this phase. Take all the ones we have,
412                 // and create a GLSL program for it.
413                 if (!this_phase_effects.empty()) {
414                         reverse(this_phase_effects.begin(), this_phase_effects.end());
415                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
416                         this_phase_effects.back()->phase = phases.back();
417                         this_phase_inputs.clear();
418                         this_phase_effects.clear();
419                 }
420                 assert(this_phase_inputs.empty());
421                 assert(this_phase_effects.empty());
422
423                 // If we have no effects left, exit.
424                 if (effects_todo_other_phases.empty()) {
425                         break;
426                 }
427
428                 Node *node = effects_todo_other_phases.top();
429                 effects_todo_other_phases.pop();
430
431                 if (completed_effects.count(node) == 0) {
432                         // Start a new phase, calculating from this effect.
433                         effects_todo_this_phase.push(node);
434                 }
435         }
436
437         // Finally, since the phases are found from the output but must be executed
438         // from the input(s), reverse them, too.
439         std::reverse(phases.begin(), phases.end());
440 }
441
442 void EffectChain::output_dot(const char *filename)
443 {
444         if (movit_debug_level != MOVIT_DEBUG_ON) {
445                 return;
446         }
447
448         FILE *fp = fopen(filename, "w");
449         if (fp == NULL) {
450                 perror(filename);
451                 exit(1);
452         }
453
454         fprintf(fp, "digraph G {\n");
455         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
456         for (unsigned i = 0; i < nodes.size(); ++i) {
457                 // Find out which phase this event belongs to.
458                 std::vector<int> in_phases;
459                 for (unsigned j = 0; j < phases.size(); ++j) {
460                         const Phase* p = phases[j];
461                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
462                                 in_phases.push_back(j);
463                         }
464                 }
465
466                 if (in_phases.empty()) {
467                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
468                 } else if (in_phases.size() == 1) {
469                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
470                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
471                                 (in_phases[0] % 8) + 1);
472                 } else {
473                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
474                         // But alas.
475                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
476                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
477                                 (in_phases[0] % 8) + 1);
478                 }
479
480                 char from_node_id[256];
481                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
482
483                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
484                         char to_node_id[256];
485                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
486
487                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
488                         output_dot_edge(fp, from_node_id, to_node_id, labels);
489                 }
490
491                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
492                         // Output node.
493                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
494                         output_dot_edge(fp, from_node_id, "output", labels);
495                 }
496         }
497         fprintf(fp, "}\n");
498
499         fclose(fp);
500 }
501
502 std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
503 {
504         std::vector<std::string> labels;
505
506         if (to != NULL && to->effect->needs_texture_bounce()) {
507                 labels.push_back("needs_bounce");
508         }
509         if (from->effect->changes_output_size()) {
510                 labels.push_back("resize");
511         }
512
513         switch (from->output_color_space) {
514         case COLORSPACE_INVALID:
515                 labels.push_back("spc[invalid]");
516                 break;
517         case COLORSPACE_REC_601_525:
518                 labels.push_back("spc[rec601-525]");
519                 break;
520         case COLORSPACE_REC_601_625:
521                 labels.push_back("spc[rec601-625]");
522                 break;
523         default:
524                 break;
525         }
526
527         switch (from->output_gamma_curve) {
528         case GAMMA_INVALID:
529                 labels.push_back("gamma[invalid]");
530                 break;
531         case GAMMA_sRGB:
532                 labels.push_back("gamma[sRGB]");
533                 break;
534         case GAMMA_REC_601:  // and GAMMA_REC_709
535                 labels.push_back("gamma[rec601/709]");
536                 break;
537         default:
538                 break;
539         }
540
541         switch (from->output_alpha_type) {
542         case ALPHA_INVALID:
543                 labels.push_back("alpha[invalid]");
544                 break;
545         case ALPHA_BLANK:
546                 labels.push_back("alpha[blank]");
547                 break;
548         case ALPHA_POSTMULTIPLIED:
549                 labels.push_back("alpha[postmult]");
550                 break;
551         default:
552                 break;
553         }
554
555         return labels;
556 }
557
558 void EffectChain::output_dot_edge(FILE *fp,
559                                   const std::string &from_node_id,
560                                   const std::string &to_node_id,
561                                   const std::vector<std::string> &labels)
562 {
563         if (labels.empty()) {
564                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
565         } else {
566                 std::string label = labels[0];
567                 for (unsigned k = 1; k < labels.size(); ++k) {
568                         label += ", " + labels[k];
569                 }
570                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
571         }
572 }
573
574 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
575 {
576         unsigned scaled_width, scaled_height;
577
578         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
579                 // Same aspect, or W/H > aspect (image is wider than the frame).
580                 // In either case, keep width, and adjust height.
581                 scaled_width = width;
582                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
583         } else {
584                 // W/H < aspect (image is taller than the frame), so keep height,
585                 // and adjust width.
586                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
587                 scaled_height = height;
588         }
589
590         // We should be consistently larger or smaller then the existing choice,
591         // since we have the same aspect.
592         assert(!(scaled_width < *output_width && scaled_height > *output_height));
593         assert(!(scaled_height < *output_height && scaled_width > *output_width));
594
595         if (scaled_width >= *output_width && scaled_height >= *output_height) {
596                 *output_width = scaled_width;
597                 *output_height = scaled_height;
598         }
599 }
600
601 // Propagate input texture sizes throughout, and inform effects downstream.
602 // (Like a lot of other code, we depend on effects being in topological order.)
603 void EffectChain::inform_input_sizes(Phase *phase)
604 {
605         // All effects that have a defined size (inputs and RTT inputs)
606         // get that. Reset all others.
607         for (unsigned i = 0; i < phase->effects.size(); ++i) {
608                 Node *node = phase->effects[i];
609                 if (node->effect->num_inputs() == 0) {
610                         Input *input = static_cast<Input *>(node->effect);
611                         node->output_width = input->get_width();
612                         node->output_height = input->get_height();
613                         assert(node->output_width != 0);
614                         assert(node->output_height != 0);
615                 } else {
616                         node->output_width = node->output_height = 0;
617                 }
618         }
619         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
620                 Node *input = phase->inputs[i];
621                 input->output_width = input->phase->virtual_output_width;
622                 input->output_height = input->phase->virtual_output_height;
623                 assert(input->output_width != 0);
624                 assert(input->output_height != 0);
625         }
626
627         // Now propagate from the inputs towards the end, and inform as we go.
628         // The rules are simple:
629         //
630         //   1. Don't touch effects that already have given sizes (ie., inputs).
631         //   2. If all of your inputs have the same size, that will be your output size.
632         //   3. Otherwise, your output size is 0x0.
633         for (unsigned i = 0; i < phase->effects.size(); ++i) {
634                 Node *node = phase->effects[i];
635                 if (node->effect->num_inputs() == 0) {
636                         continue;
637                 }
638                 unsigned this_output_width = 0;
639                 unsigned this_output_height = 0;
640                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
641                         Node *input = node->incoming_links[j];
642                         node->effect->inform_input_size(j, input->output_width, input->output_height);
643                         if (j == 0) {
644                                 this_output_width = input->output_width;
645                                 this_output_height = input->output_height;
646                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
647                                 // Inputs disagree.
648                                 this_output_width = 0;
649                                 this_output_height = 0;
650                         }
651                 }
652                 node->output_width = this_output_width;
653                 node->output_height = this_output_height;
654         }
655 }
656
657 // Note: You should call inform_input_sizes() before this, as the last effect's
658 // desired output size might change based on the inputs.
659 void EffectChain::find_output_size(Phase *phase)
660 {
661         Node *output_node = phase->effects.back();
662
663         // If the last effect explicitly sets an output size, use that.
664         if (output_node->effect->changes_output_size()) {
665                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
666                                                      &phase->virtual_output_width, &phase->virtual_output_height);
667                 return;
668         }
669
670         // If all effects have the same size, use that.
671         unsigned output_width = 0, output_height = 0;
672         bool all_inputs_same_size = true;
673
674         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
675                 Node *input = phase->inputs[i];
676                 assert(input->phase->output_width != 0);
677                 assert(input->phase->output_height != 0);
678                 if (output_width == 0 && output_height == 0) {
679                         output_width = input->phase->virtual_output_width;
680                         output_height = input->phase->virtual_output_height;
681                 } else if (output_width != input->phase->virtual_output_width ||
682                            output_height != input->phase->virtual_output_height) {
683                         all_inputs_same_size = false;
684                 }
685         }
686         for (unsigned i = 0; i < phase->effects.size(); ++i) {
687                 Effect *effect = phase->effects[i]->effect;
688                 if (effect->num_inputs() != 0) {
689                         continue;
690                 }
691
692                 Input *input = static_cast<Input *>(effect);
693                 if (output_width == 0 && output_height == 0) {
694                         output_width = input->get_width();
695                         output_height = input->get_height();
696                 } else if (output_width != input->get_width() ||
697                            output_height != input->get_height()) {
698                         all_inputs_same_size = false;
699                 }
700         }
701
702         if (all_inputs_same_size) {
703                 assert(output_width != 0);
704                 assert(output_height != 0);
705                 phase->virtual_output_width = phase->output_width = output_width;
706                 phase->virtual_output_height = phase->output_height = output_height;
707                 return;
708         }
709
710         // If not, fit all the inputs into the current aspect, and select the largest one. 
711         output_width = 0;
712         output_height = 0;
713         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
714                 Node *input = phase->inputs[i];
715                 assert(input->phase->output_width != 0);
716                 assert(input->phase->output_height != 0);
717                 size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
718         }
719         for (unsigned i = 0; i < phase->effects.size(); ++i) {
720                 Effect *effect = phase->effects[i]->effect;
721                 if (effect->num_inputs() != 0) {
722                         continue;
723                 }
724
725                 Input *input = static_cast<Input *>(effect);
726                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
727         }
728         assert(output_width != 0);
729         assert(output_height != 0);
730         phase->virtual_output_width = phase->output_width = output_width;
731         phase->virtual_output_height = phase->output_height = output_height;
732 }
733
734 void EffectChain::sort_all_nodes_topologically()
735 {
736         nodes = topological_sort(nodes);
737 }
738
739 std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
740 {
741         std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
742         std::vector<Node *> sorted_list;
743         for (unsigned i = 0; i < nodes.size(); ++i) {
744                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
745         }
746         reverse(sorted_list.begin(), sorted_list.end());
747         return sorted_list;
748 }
749
750 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
751 {
752         if (nodes_left_to_visit->count(node) == 0) {
753                 return;
754         }
755         nodes_left_to_visit->erase(node);
756         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
757                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
758         }
759         sorted_list->push_back(node);
760 }
761
762 void EffectChain::find_color_spaces_for_inputs()
763 {
764         for (unsigned i = 0; i < nodes.size(); ++i) {
765                 Node *node = nodes[i];
766                 if (node->disabled) {
767                         continue;
768                 }
769                 if (node->incoming_links.size() == 0) {
770                         Input *input = static_cast<Input *>(node->effect);
771                         node->output_color_space = input->get_color_space();
772                         node->output_gamma_curve = input->get_gamma_curve();
773
774                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
775                         switch (alpha_handling) {
776                         case Effect::OUTPUT_BLANK_ALPHA:
777                                 node->output_alpha_type = ALPHA_BLANK;
778                                 break;
779                         case Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED:
780                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
781                                 break;
782                         case Effect::OUTPUT_ALPHA_POSTMULTIPLIED:
783                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
784                                 break;
785                         case Effect::DONT_CARE_ALPHA_TYPE:
786                         default:
787                                 assert(false);
788                         }
789
790                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
791                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
792                         }
793                 }
794         }
795 }
796
797 // Propagate gamma and color space information as far as we can in the graph.
798 // The rules are simple: Anything where all the inputs agree, get that as
799 // output as well. Anything else keeps having *_INVALID.
800 void EffectChain::propagate_gamma_and_color_space()
801 {
802         // We depend on going through the nodes in order.
803         sort_all_nodes_topologically();
804
805         for (unsigned i = 0; i < nodes.size(); ++i) {
806                 Node *node = nodes[i];
807                 if (node->disabled) {
808                         continue;
809                 }
810                 assert(node->incoming_links.size() == node->effect->num_inputs());
811                 if (node->incoming_links.size() == 0) {
812                         assert(node->output_color_space != COLORSPACE_INVALID);
813                         assert(node->output_gamma_curve != GAMMA_INVALID);
814                         continue;
815                 }
816
817                 Colorspace color_space = node->incoming_links[0]->output_color_space;
818                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
819                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
820                         if (node->incoming_links[j]->output_color_space != color_space) {
821                                 color_space = COLORSPACE_INVALID;
822                         }
823                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
824                                 gamma_curve = GAMMA_INVALID;
825                         }
826                 }
827
828                 // The conversion effects already have their outputs set correctly,
829                 // so leave them alone.
830                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
831                         node->output_color_space = color_space;
832                 }               
833                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
834                     node->effect->effect_type_id() != "GammaExpansionEffect") {
835                         node->output_gamma_curve = gamma_curve;
836                 }               
837         }
838 }
839
840 // Propagate alpha information as far as we can in the graph.
841 // Similar to propagate_gamma_and_color_space().
842 void EffectChain::propagate_alpha()
843 {
844         // We depend on going through the nodes in order.
845         sort_all_nodes_topologically();
846
847         for (unsigned i = 0; i < nodes.size(); ++i) {
848                 Node *node = nodes[i];
849                 if (node->disabled) {
850                         continue;
851                 }
852                 assert(node->incoming_links.size() == node->effect->num_inputs());
853                 if (node->incoming_links.size() == 0) {
854                         assert(node->output_alpha_type != ALPHA_INVALID);
855                         continue;
856                 }
857
858                 // The alpha multiplication/division effects are special cases.
859                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
860                         assert(node->incoming_links.size() == 1);
861                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
862                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
863                         continue;
864                 }
865                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
866                         assert(node->incoming_links.size() == 1);
867                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
868                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
869                         continue;
870                 }
871
872                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
873                 // because they are the only one that _need_ postmultiplied alpha.
874                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
875                     node->effect->effect_type_id() == "GammaExpansionEffect") {
876                         assert(node->incoming_links.size() == 1);
877                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
878                                 node->output_alpha_type = ALPHA_BLANK;
879                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
880                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
881                         } else {
882                                 node->output_alpha_type = ALPHA_INVALID;
883                         }
884                         continue;
885                 }
886
887                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
888                 // or OUTPUT_ALPHA_POSTMULTIPLIED), and they have already been
889                 // taken care of above. Rationale: Even if you could imagine
890                 // e.g. an effect that took in an image and set alpha=1.0
891                 // unconditionally, it wouldn't make any sense to have it as
892                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
893                 // got its input pre- or postmultiplied, so it wouldn't know
894                 // whether to divide away the old alpha or not.
895                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
896                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED ||
897                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
898
899                 // If the node has multiple inputs, check that they are all valid and
900                 // the same.
901                 bool any_invalid = false;
902                 bool any_premultiplied = false;
903                 bool any_postmultiplied = false;
904
905                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
906                         switch (node->incoming_links[j]->output_alpha_type) {
907                         case ALPHA_INVALID:
908                                 any_invalid = true;
909                                 break;
910                         case ALPHA_BLANK:
911                                 // Blank is good as both pre- and postmultiplied alpha,
912                                 // so just ignore it.
913                                 break;
914                         case ALPHA_PREMULTIPLIED:
915                                 any_premultiplied = true;
916                                 break;
917                         case ALPHA_POSTMULTIPLIED:
918                                 any_postmultiplied = true;
919                                 break;
920                         default:
921                                 assert(false);
922                         }
923                 }
924
925                 if (any_invalid) {
926                         node->output_alpha_type = ALPHA_INVALID;
927                         continue;
928                 }
929
930                 // Inputs must be of the same type.
931                 if (any_premultiplied && any_postmultiplied) {
932                         node->output_alpha_type = ALPHA_INVALID;
933                         continue;
934                 }
935
936                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED) {
937                         // If the effect has asked for premultiplied alpha, check that it has got it.
938                         if (any_postmultiplied) {
939                                 node->output_alpha_type = ALPHA_INVALID;
940                         } else {
941                                 // In some rare cases, it might be advantageous to say
942                                 // that blank input alpha yields blank output alpha.
943                                 // However, this would cause a more complex Effect interface
944                                 // an effect would need to guarantee that it doesn't mess with
945                                 // blank alpha), so this is the simplest.
946                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
947                         }
948                 } else {
949                         // OK, all inputs are the same, and this effect is not going
950                         // to change it.
951                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
952                         if (any_premultiplied) {
953                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
954                         } else if (any_postmultiplied) {
955                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
956                         } else {
957                                 node->output_alpha_type = ALPHA_BLANK;
958                         }
959                 }
960         }
961 }
962
963 bool EffectChain::node_needs_colorspace_fix(Node *node)
964 {
965         if (node->disabled) {
966                 return false;
967         }
968         if (node->effect->num_inputs() == 0) {
969                 return false;
970         }
971
972         // propagate_gamma_and_color_space() has already set our output
973         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
974         if (node->output_color_space == COLORSPACE_INVALID) {
975                 return true;
976         }
977         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
978 }
979
980 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
981 // the graph. Our strategy is not always optimal, but quite simple:
982 // Find an effect that's as early as possible where the inputs are of
983 // unacceptable colorspaces (that is, either different, or, if the effect only
984 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
985 // propagate the information anew, and repeat until there are no more such
986 // effects.
987 void EffectChain::fix_internal_color_spaces()
988 {
989         unsigned colorspace_propagation_pass = 0;
990         bool found_any;
991         do {
992                 found_any = false;
993                 for (unsigned i = 0; i < nodes.size(); ++i) {
994                         Node *node = nodes[i];
995                         if (!node_needs_colorspace_fix(node)) {
996                                 continue;
997                         }
998
999                         // Go through each input that is not sRGB, and insert
1000                         // a colorspace conversion after it.
1001                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1002                                 Node *input = node->incoming_links[j];
1003                                 assert(input->output_color_space != COLORSPACE_INVALID);
1004                                 if (input->output_color_space == COLORSPACE_sRGB) {
1005                                         continue;
1006                                 }
1007                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1008                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1009                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1010                                 conversion->output_color_space = COLORSPACE_sRGB;
1011                                 replace_sender(input, conversion);
1012                                 connect_nodes(input, conversion);
1013                         }
1014
1015                         // Re-sort topologically, and propagate the new information.
1016                         propagate_gamma_and_color_space();
1017                         
1018                         found_any = true;
1019                         break;
1020                 }
1021         
1022                 char filename[256];
1023                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1024                 output_dot(filename);
1025                 assert(colorspace_propagation_pass < 100);
1026         } while (found_any);
1027
1028         for (unsigned i = 0; i < nodes.size(); ++i) {
1029                 Node *node = nodes[i];
1030                 if (node->disabled) {
1031                         continue;
1032                 }
1033                 assert(node->output_color_space != COLORSPACE_INVALID);
1034         }
1035 }
1036
1037 bool EffectChain::node_needs_alpha_fix(Node *node)
1038 {
1039         if (node->disabled) {
1040                 return false;
1041         }
1042
1043         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1044         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1045         return (node->output_alpha_type == ALPHA_INVALID);
1046 }
1047
1048 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1049 // the graph. Similar to fix_internal_color_spaces().
1050 void EffectChain::fix_internal_alpha(unsigned step)
1051 {
1052         unsigned alpha_propagation_pass = 0;
1053         bool found_any;
1054         do {
1055                 found_any = false;
1056                 for (unsigned i = 0; i < nodes.size(); ++i) {
1057                         Node *node = nodes[i];
1058                         if (!node_needs_alpha_fix(node)) {
1059                                 continue;
1060                         }
1061
1062                         // If we need to fix up GammaExpansionEffect, then clearly something
1063                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1064                         // is meaningless.
1065                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1066
1067                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1068
1069                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1070                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1071                                 assert(node->incoming_links.size() == 1);
1072                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1073                                 desired_type = ALPHA_POSTMULTIPLIED;
1074                         }
1075
1076                         // Go through each input that is not premultiplied alpha, and insert
1077                         // a conversion before it.
1078                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1079                                 Node *input = node->incoming_links[j];
1080                                 assert(input->output_alpha_type != ALPHA_INVALID);
1081                                 if (input->output_alpha_type == desired_type ||
1082                                     input->output_alpha_type == ALPHA_BLANK) {
1083                                         continue;
1084                                 }
1085                                 Node *conversion;
1086                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1087                                         conversion = add_node(new AlphaMultiplicationEffect());
1088                                 } else {
1089                                         conversion = add_node(new AlphaDivisionEffect());
1090                                 }
1091                                 conversion->output_alpha_type = desired_type;
1092                                 replace_sender(input, conversion);
1093                                 connect_nodes(input, conversion);
1094                         }
1095
1096                         // Re-sort topologically, and propagate the new information.
1097                         propagate_gamma_and_color_space();
1098                         propagate_alpha();
1099                         
1100                         found_any = true;
1101                         break;
1102                 }
1103         
1104                 char filename[256];
1105                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1106                 output_dot(filename);
1107                 assert(alpha_propagation_pass < 100);
1108         } while (found_any);
1109
1110         for (unsigned i = 0; i < nodes.size(); ++i) {
1111                 Node *node = nodes[i];
1112                 if (node->disabled) {
1113                         continue;
1114                 }
1115                 assert(node->output_alpha_type != ALPHA_INVALID);
1116         }
1117 }
1118
1119 // Make so that the output is in the desired color space.
1120 void EffectChain::fix_output_color_space()
1121 {
1122         Node *output = find_output_node();
1123         if (output->output_color_space != output_format.color_space) {
1124                 Node *conversion = add_node(new ColorspaceConversionEffect());
1125                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1126                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1127                 conversion->output_color_space = output_format.color_space;
1128                 connect_nodes(output, conversion);
1129                 propagate_alpha();
1130                 propagate_gamma_and_color_space();
1131         }
1132 }
1133
1134 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1135 void EffectChain::fix_output_alpha()
1136 {
1137         Node *output = find_output_node();
1138         assert(output->output_alpha_type != ALPHA_INVALID);
1139         if (output->output_alpha_type == ALPHA_BLANK) {
1140                 // No alpha output, so we don't care.
1141                 return;
1142         }
1143         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1144             output_alpha_format == OUTPUT_ALPHA_POSTMULTIPLIED) {
1145                 Node *conversion = add_node(new AlphaDivisionEffect());
1146                 connect_nodes(output, conversion);
1147                 propagate_alpha();
1148                 propagate_gamma_and_color_space();
1149         }
1150         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1151             output_alpha_format == OUTPUT_ALPHA_PREMULTIPLIED) {
1152                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1153                 connect_nodes(output, conversion);
1154                 propagate_alpha();
1155                 propagate_gamma_and_color_space();
1156         }
1157 }
1158
1159 bool EffectChain::node_needs_gamma_fix(Node *node)
1160 {
1161         if (node->disabled) {
1162                 return false;
1163         }
1164
1165         // Small hack since the output is not an explicit node:
1166         // If we are the last node and our output is in the wrong
1167         // space compared to EffectChain's output, we need to fix it.
1168         // This will only take us to linear, but fix_output_gamma()
1169         // will come and take us to the desired output gamma
1170         // if it is needed.
1171         //
1172         // This needs to be before everything else, since it could
1173         // even apply to inputs (if they are the only effect).
1174         if (node->outgoing_links.empty() &&
1175             node->output_gamma_curve != output_format.gamma_curve &&
1176             node->output_gamma_curve != GAMMA_LINEAR) {
1177                 return true;
1178         }
1179
1180         if (node->effect->num_inputs() == 0) {
1181                 return false;
1182         }
1183
1184         // propagate_gamma_and_color_space() has already set our output
1185         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1186         // except for GammaCompressionEffect.
1187         if (node->output_gamma_curve == GAMMA_INVALID) {
1188                 return true;
1189         }
1190         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1191                 assert(node->incoming_links.size() == 1);
1192                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1193         }
1194
1195         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1196 }
1197
1198 // Very similar to fix_internal_color_spaces(), but for gamma.
1199 // There is one difference, though; before we start adding conversion nodes,
1200 // we see if we can get anything out of asking the sources to deliver
1201 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1202 // does that part, while fix_internal_gamma_by_inserting_nodes()
1203 // inserts nodes as needed afterwards.
1204 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1205 {
1206         unsigned gamma_propagation_pass = 0;
1207         bool found_any;
1208         do {
1209                 found_any = false;
1210                 for (unsigned i = 0; i < nodes.size(); ++i) {
1211                         Node *node = nodes[i];
1212                         if (!node_needs_gamma_fix(node)) {
1213                                 continue;
1214                         }
1215
1216                         // See if all inputs can give us linear gamma. If not, leave it.
1217                         std::vector<Node *> nonlinear_inputs;
1218                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1219                         assert(!nonlinear_inputs.empty());
1220
1221                         bool all_ok = true;
1222                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1223                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1224                                 all_ok &= input->can_output_linear_gamma();
1225                         }
1226
1227                         if (!all_ok) {
1228                                 continue;
1229                         }
1230
1231                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1232                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1233                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1234                         }
1235
1236                         // Re-sort topologically, and propagate the new information.
1237                         propagate_gamma_and_color_space();
1238                         
1239                         found_any = true;
1240                         break;
1241                 }
1242         
1243                 char filename[256];
1244                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1245                 output_dot(filename);
1246                 assert(gamma_propagation_pass < 100);
1247         } while (found_any);
1248 }
1249
1250 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1251 {
1252         unsigned gamma_propagation_pass = 0;
1253         bool found_any;
1254         do {
1255                 found_any = false;
1256                 for (unsigned i = 0; i < nodes.size(); ++i) {
1257                         Node *node = nodes[i];
1258                         if (!node_needs_gamma_fix(node)) {
1259                                 continue;
1260                         }
1261
1262                         // Special case: We could be an input and still be asked to
1263                         // fix our gamma; if so, we should be the only node
1264                         // (as node_needs_gamma_fix() would only return true in
1265                         // for an input in that case). That means we should insert
1266                         // a conversion node _after_ ourselves.
1267                         if (node->incoming_links.empty()) {
1268                                 assert(node->outgoing_links.empty());
1269                                 Node *conversion = add_node(new GammaExpansionEffect());
1270                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1271                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1272                                 connect_nodes(node, conversion);
1273                         }
1274
1275                         // If not, go through each input that is not linear gamma,
1276                         // and insert a gamma conversion after it.
1277                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1278                                 Node *input = node->incoming_links[j];
1279                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1280                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1281                                         continue;
1282                                 }
1283                                 Node *conversion = add_node(new GammaExpansionEffect());
1284                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1285                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1286                                 replace_sender(input, conversion);
1287                                 connect_nodes(input, conversion);
1288                         }
1289
1290                         // Re-sort topologically, and propagate the new information.
1291                         propagate_alpha();
1292                         propagate_gamma_and_color_space();
1293                         
1294                         found_any = true;
1295                         break;
1296                 }
1297         
1298                 char filename[256];
1299                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1300                 output_dot(filename);
1301                 assert(gamma_propagation_pass < 100);
1302         } while (found_any);
1303
1304         for (unsigned i = 0; i < nodes.size(); ++i) {
1305                 Node *node = nodes[i];
1306                 if (node->disabled) {
1307                         continue;
1308                 }
1309                 assert(node->output_gamma_curve != GAMMA_INVALID);
1310         }
1311 }
1312
1313 // Make so that the output is in the desired gamma.
1314 // Note that this assumes linear input gamma, so it might create the need
1315 // for another pass of fix_internal_gamma().
1316 void EffectChain::fix_output_gamma()
1317 {
1318         Node *output = find_output_node();
1319         if (output->output_gamma_curve != output_format.gamma_curve) {
1320                 Node *conversion = add_node(new GammaCompressionEffect());
1321                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1322                 conversion->output_gamma_curve = output_format.gamma_curve;
1323                 connect_nodes(output, conversion);
1324         }
1325 }
1326         
1327 // If the user has requested dither, add a DitherEffect right at the end
1328 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1329 // since dither is about the only effect that can _not_ be done in linear space.
1330 void EffectChain::add_dither_if_needed()
1331 {
1332         if (num_dither_bits == 0) {
1333                 return;
1334         }
1335         Node *output = find_output_node();
1336         Node *dither = add_node(new DitherEffect());
1337         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1338         connect_nodes(output, dither);
1339
1340         dither_effect = dither->effect;
1341 }
1342
1343 // Find the output node. This is, simply, one that has no outgoing links.
1344 // If there are multiple ones, the graph is malformed (we do not support
1345 // multiple outputs right now).
1346 Node *EffectChain::find_output_node()
1347 {
1348         std::vector<Node *> output_nodes;
1349         for (unsigned i = 0; i < nodes.size(); ++i) {
1350                 Node *node = nodes[i];
1351                 if (node->disabled) {
1352                         continue;
1353                 }
1354                 if (node->outgoing_links.empty()) {
1355                         output_nodes.push_back(node);
1356                 }
1357         }
1358         assert(output_nodes.size() == 1);
1359         return output_nodes[0];
1360 }
1361
1362 void EffectChain::finalize()
1363 {
1364         // Save the current locale, and set it to C, so that we can output decimal
1365         // numbers with printf and be sure to get them in the format mandated by GLSL.
1366         char *saved_locale = setlocale(LC_NUMERIC, "C");
1367
1368         // Output the graph as it is before we do any conversions on it.
1369         output_dot("step0-start.dot");
1370
1371         // Give each effect in turn a chance to rewrite its own part of the graph.
1372         // Note that if more effects are added as part of this, they will be
1373         // picked up as part of the same for loop, since they are added at the end.
1374         for (unsigned i = 0; i < nodes.size(); ++i) {
1375                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1376         }
1377         output_dot("step1-rewritten.dot");
1378
1379         find_color_spaces_for_inputs();
1380         output_dot("step2-input-colorspace.dot");
1381
1382         propagate_alpha();
1383         output_dot("step3-propagated-alpha.dot");
1384
1385         propagate_gamma_and_color_space();
1386         output_dot("step4-propagated-all.dot");
1387
1388         fix_internal_color_spaces();
1389         fix_internal_alpha(6);
1390         fix_output_color_space();
1391         output_dot("step7-output-colorspacefix.dot");
1392         fix_output_alpha();
1393         output_dot("step8-output-alphafix.dot");
1394
1395         // Note that we need to fix gamma after colorspace conversion,
1396         // because colorspace conversions might create needs for gamma conversions.
1397         // Also, we need to run an extra pass of fix_internal_gamma() after 
1398         // fixing the output gamma, as we only have conversions to/from linear,
1399         // and fix_internal_alpha() since GammaCompressionEffect needs
1400         // postmultiplied input.
1401         fix_internal_gamma_by_asking_inputs(9);
1402         fix_internal_gamma_by_inserting_nodes(10);
1403         fix_output_gamma();
1404         output_dot("step11-output-gammafix.dot");
1405         propagate_alpha();
1406         output_dot("step12-output-alpha-propagated.dot");
1407         fix_internal_alpha(13);
1408         output_dot("step14-output-alpha-fixed.dot");
1409         fix_internal_gamma_by_asking_inputs(15);
1410         fix_internal_gamma_by_inserting_nodes(16);
1411
1412         output_dot("step17-before-dither.dot");
1413
1414         add_dither_if_needed();
1415
1416         output_dot("step18-final.dot");
1417         
1418         // Construct all needed GLSL programs, starting at the output.
1419         construct_glsl_programs(find_output_node());
1420
1421         output_dot("step19-split-to-phases.dot");
1422
1423         // If we have more than one phase, we need intermediate render-to-texture.
1424         // Construct an FBO, and then as many textures as we need.
1425         // We choose the simplest option of having one texture per output,
1426         // since otherwise this turns into an (albeit simple)
1427         // register allocation problem.
1428         if (phases.size() > 1) {
1429                 glGenFramebuffers(1, &fbo);
1430
1431                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1432                         inform_input_sizes(phases[i]);
1433                         find_output_size(phases[i]);
1434
1435                         Node *output_node = phases[i]->effects.back();
1436                         glGenTextures(1, &output_node->output_texture);
1437                         check_error();
1438                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1439                         check_error();
1440                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1441                         check_error();
1442                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1443                         check_error();
1444                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1445                         check_error();
1446
1447                         output_node->output_texture_width = phases[i]->output_width;
1448                         output_node->output_texture_height = phases[i]->output_height;
1449                 }
1450                 inform_input_sizes(phases.back());
1451         }
1452                 
1453         for (unsigned i = 0; i < inputs.size(); ++i) {
1454                 inputs[i]->finalize();
1455         }
1456
1457         assert(phases[0]->inputs.empty());
1458         
1459         finalized = true;
1460         setlocale(LC_NUMERIC, saved_locale);
1461 }
1462
1463 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1464 {
1465         assert(finalized);
1466
1467         // Save original viewport.
1468         GLuint x = 0, y = 0;
1469
1470         if (width == 0 && height == 0) {
1471                 GLint viewport[4];
1472                 glGetIntegerv(GL_VIEWPORT, viewport);
1473                 x = viewport[0];
1474                 y = viewport[1];
1475                 width = viewport[2];
1476                 height = viewport[3];
1477         }
1478
1479         // Basic state.
1480         glDisable(GL_BLEND);
1481         check_error();
1482         glDisable(GL_DEPTH_TEST);
1483         check_error();
1484         glDepthMask(GL_FALSE);
1485         check_error();
1486
1487         glMatrixMode(GL_PROJECTION);
1488         glLoadIdentity();
1489         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1490
1491         glMatrixMode(GL_MODELVIEW);
1492         glLoadIdentity();
1493
1494         if (phases.size() > 1) {
1495                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1496                 check_error();
1497         }
1498
1499         std::set<Node *> generated_mipmaps;
1500
1501         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1502                 // See if the requested output size has changed. If so, we need to recreate
1503                 // the texture (and before we start setting up inputs).
1504                 inform_input_sizes(phases[phase]);
1505                 if (phase != phases.size() - 1) {
1506                         find_output_size(phases[phase]);
1507
1508                         Node *output_node = phases[phase]->effects.back();
1509
1510                         if (output_node->output_texture_width != phases[phase]->output_width ||
1511                             output_node->output_texture_height != phases[phase]->output_height) {
1512                                 glActiveTexture(GL_TEXTURE0);
1513                                 check_error();
1514                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1515                                 check_error();
1516                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1517                                 check_error();
1518                                 glBindTexture(GL_TEXTURE_2D, 0);
1519                                 check_error();
1520
1521                                 output_node->output_texture_width = phases[phase]->output_width;
1522                                 output_node->output_texture_height = phases[phase]->output_height;
1523                         }
1524                 }
1525
1526                 glUseProgram(phases[phase]->glsl_program_num);
1527                 check_error();
1528
1529                 // Set up RTT inputs for this phase.
1530                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1531                         glActiveTexture(GL_TEXTURE0 + sampler);
1532                         Node *input = phases[phase]->inputs[sampler];
1533                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1534                         check_error();
1535                         if (phases[phase]->input_needs_mipmaps) {
1536                                 if (generated_mipmaps.count(input) == 0) {
1537                                         glGenerateMipmap(GL_TEXTURE_2D);
1538                                         check_error();
1539                                         generated_mipmaps.insert(input);
1540                                 }
1541                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1542                                 check_error();
1543                         } else {
1544                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1545                                 check_error();
1546                         }
1547
1548                         std::string texture_name = std::string("tex_") + input->effect_id;
1549                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1550                         check_error();
1551                 }
1552
1553                 // And now the output.
1554                 if (phase == phases.size() - 1) {
1555                         // Last phase goes to the output the user specified.
1556                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1557                         check_error();
1558                         glViewport(x, y, width, height);
1559                         if (dither_effect != NULL) {
1560                                 CHECK(dither_effect->set_int("output_width", width));
1561                                 CHECK(dither_effect->set_int("output_height", height));
1562                         }
1563                 } else {
1564                         Node *output_node = phases[phase]->effects.back();
1565                         glFramebufferTexture2D(
1566                                 GL_FRAMEBUFFER,
1567                                 GL_COLOR_ATTACHMENT0,
1568                                 GL_TEXTURE_2D,
1569                                 output_node->output_texture,
1570                                 0);
1571                         check_error();
1572                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1573                 }
1574
1575                 // Give the required parameters to all the effects.
1576                 unsigned sampler_num = phases[phase]->inputs.size();
1577                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1578                         Node *node = phases[phase]->effects[i];
1579                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1580                         check_error();
1581                 }
1582
1583                 // Now draw!
1584                 glBegin(GL_QUADS);
1585
1586                 glTexCoord2f(0.0f, 0.0f);
1587                 glVertex2f(0.0f, 0.0f);
1588
1589                 glTexCoord2f(1.0f, 0.0f);
1590                 glVertex2f(1.0f, 0.0f);
1591
1592                 glTexCoord2f(1.0f, 1.0f);
1593                 glVertex2f(1.0f, 1.0f);
1594
1595                 glTexCoord2f(0.0f, 1.0f);
1596                 glVertex2f(0.0f, 1.0f);
1597
1598                 glEnd();
1599                 check_error();
1600
1601                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1602                         Node *node = phases[phase]->effects[i];
1603                         node->effect->clear_gl_state();
1604                 }
1605         }
1606 }