a9b892725358039173725feea5306692f3432988
[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 <assert.h>
7 #include <GL/glew.h>
8
9 #include <algorithm>
10 #include <set>
11 #include <stack>
12 #include <vector>
13
14 #include "util.h"
15 #include "effect_chain.h"
16 #include "gamma_expansion_effect.h"
17 #include "gamma_compression_effect.h"
18 #include "colorspace_conversion_effect.h"
19 #include "alpha_multiplication_effect.h"
20 #include "alpha_division_effect.h"
21 #include "dither_effect.h"
22 #include "input.h"
23 #include "init.h"
24
25 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
26         : aspect_nom(aspect_nom),
27           aspect_denom(aspect_denom),
28           dither_effect(NULL),
29           fbo(0),
30           num_dither_bits(0),
31           finalized(false) {}
32
33 EffectChain::~EffectChain()
34 {
35         for (unsigned i = 0; i < nodes.size(); ++i) {
36                 if (nodes[i]->output_texture != 0) {
37                         glDeleteTextures(1, &nodes[i]->output_texture);
38                 }
39                 delete nodes[i]->effect;
40                 delete nodes[i];
41         }
42         for (unsigned i = 0; i < phases.size(); ++i) {
43                 glDeleteProgram(phases[i]->glsl_program_num);
44                 glDeleteShader(phases[i]->vertex_shader);
45                 glDeleteShader(phases[i]->fragment_shader);
46                 delete phases[i];
47         }
48         if (fbo != 0) {
49                 glDeleteFramebuffers(1, &fbo);
50         }
51 }
52
53 Input *EffectChain::add_input(Input *input)
54 {
55         inputs.push_back(input);
56         add_node(input);
57         return input;
58 }
59
60 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
61 {
62         output_format = format;
63         output_alpha_format = alpha_format;
64 }
65
66 Node *EffectChain::add_node(Effect *effect)
67 {
68         char effect_id[256];
69         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
70
71         Node *node = new Node;
72         node->effect = effect;
73         node->disabled = false;
74         node->effect_id = effect_id;
75         node->output_color_space = COLORSPACE_INVALID;
76         node->output_gamma_curve = GAMMA_INVALID;
77         node->output_alpha_type = ALPHA_INVALID;
78         node->output_texture = 0;
79
80         nodes.push_back(node);
81         node_map[effect] = node;
82         return node;
83 }
84
85 void EffectChain::connect_nodes(Node *sender, Node *receiver)
86 {
87         sender->outgoing_links.push_back(receiver);
88         receiver->incoming_links.push_back(sender);
89 }
90
91 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
92 {
93         new_receiver->incoming_links = old_receiver->incoming_links;
94         old_receiver->incoming_links.clear();
95         
96         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
97                 Node *sender = new_receiver->incoming_links[i];
98                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
99                         if (sender->outgoing_links[j] == old_receiver) {
100                                 sender->outgoing_links[j] = new_receiver;
101                         }
102                 }
103         }       
104 }
105
106 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
107 {
108         new_sender->outgoing_links = old_sender->outgoing_links;
109         old_sender->outgoing_links.clear();
110         
111         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
112                 Node *receiver = new_sender->outgoing_links[i];
113                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
114                         if (receiver->incoming_links[j] == old_sender) {
115                                 receiver->incoming_links[j] = new_sender;
116                         }
117                 }
118         }       
119 }
120
121 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
122 {
123         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
124                 if (sender->outgoing_links[i] == receiver) {
125                         sender->outgoing_links[i] = middle;
126                         middle->incoming_links.push_back(sender);
127                 }
128         }
129         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
130                 if (receiver->incoming_links[i] == sender) {
131                         receiver->incoming_links[i] = middle;
132                         middle->outgoing_links.push_back(receiver);
133                 }
134         }
135
136         assert(middle->incoming_links.size() == middle->effect->num_inputs());
137 }
138
139 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
140 {
141         if (node->output_gamma_curve == GAMMA_LINEAR &&
142             node->effect->effect_type_id() != "GammaCompressionEffect") {
143                 return;
144         }
145         if (node->effect->num_inputs() == 0) {
146                 nonlinear_inputs->push_back(node);
147         } else {
148                 assert(node->effect->num_inputs() == node->incoming_links.size());
149                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
150                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
151                 }
152         }
153 }
154
155 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
156 {
157         assert(inputs.size() == effect->num_inputs());
158         Node *node = add_node(effect);
159         for (unsigned i = 0; i < inputs.size(); ++i) {
160                 assert(node_map.count(inputs[i]) != 0);
161                 connect_nodes(node_map[inputs[i]], node);
162         }
163         return effect;
164 }
165
166 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
167 std::string replace_prefix(const std::string &text, const std::string &prefix)
168 {
169         std::string output;
170         size_t start = 0;
171
172         while (start < text.size()) {
173                 size_t pos = text.find("PREFIX(", start);
174                 if (pos == std::string::npos) {
175                         output.append(text.substr(start, std::string::npos));
176                         break;
177                 }
178
179                 output.append(text.substr(start, pos - start));
180                 output.append(prefix);
181                 output.append("_");
182
183                 pos += strlen("PREFIX(");
184         
185                 // Output stuff until we find the matching ), which we then eat.
186                 int depth = 1;
187                 size_t end_arg_pos = pos;
188                 while (end_arg_pos < text.size()) {
189                         if (text[end_arg_pos] == '(') {
190                                 ++depth;
191                         } else if (text[end_arg_pos] == ')') {
192                                 --depth;
193                                 if (depth == 0) {
194                                         break;
195                                 }
196                         }
197                         ++end_arg_pos;
198                 }
199                 output.append(text.substr(pos, end_arg_pos - pos));
200                 ++end_arg_pos;
201                 assert(depth == 0);
202                 start = end_arg_pos;
203         }
204         return output;
205 }
206
207 Phase *EffectChain::compile_glsl_program(
208         const std::vector<Node *> &inputs,
209         const std::vector<Node *> &effects)
210 {
211         assert(!effects.empty());
212
213         // Deduplicate the inputs.
214         std::vector<Node *> true_inputs = inputs;
215         std::sort(true_inputs.begin(), true_inputs.end());
216         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
217
218         bool input_needs_mipmaps = false;
219         std::string frag_shader = read_file("header.frag");
220
221         // Create functions for all the texture inputs that we need.
222         for (unsigned i = 0; i < true_inputs.size(); ++i) {
223                 Node *input = true_inputs[i];
224         
225                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
226                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
227                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
228                 frag_shader += "}\n";
229                 frag_shader += "\n";
230         }
231
232         std::vector<Node *> sorted_effects = topological_sort(effects);
233
234         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
235                 Node *node = sorted_effects[i];
236
237                 if (node->incoming_links.size() == 1) {
238                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
239                 } else {
240                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
241                                 char buf[256];
242                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
243                                 frag_shader += buf;
244                         }
245                 }
246         
247                 frag_shader += "\n";
248                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
249                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
250                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
251                 frag_shader += "#undef PREFIX\n";
252                 frag_shader += "#undef FUNCNAME\n";
253                 if (node->incoming_links.size() == 1) {
254                         frag_shader += "#undef INPUT\n";
255                 } else {
256                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
257                                 char buf[256];
258                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
259                                 frag_shader += buf;
260                         }
261                 }
262                 frag_shader += "\n";
263
264                 input_needs_mipmaps |= node->effect->needs_mipmaps();
265         }
266         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
267                 Node *node = sorted_effects[i];
268                 if (node->effect->num_inputs() == 0) {
269                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
270                 }
271         }
272         frag_shader += std::string("#define INPUT ") + sorted_effects.back()->effect_id + "\n";
273         frag_shader.append(read_file("footer.frag"));
274
275         if (movit_debug_level == MOVIT_DEBUG_ON) {
276                 // Output shader to a temporary file, for easier debugging.
277                 static int compiled_shader_num = 0;
278                 char filename[256];
279                 sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
280                 FILE *fp = fopen(filename, "w");
281                 if (fp == NULL) {
282                         perror(filename);
283                         exit(1);
284                 }
285                 fprintf(fp, "%s\n", frag_shader.c_str());
286                 fclose(fp);
287         }
288         
289         GLuint glsl_program_num = glCreateProgram();
290         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
291         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
292         glAttachShader(glsl_program_num, vs_obj);
293         check_error();
294         glAttachShader(glsl_program_num, fs_obj);
295         check_error();
296         glLinkProgram(glsl_program_num);
297         check_error();
298
299         Phase *phase = new Phase;
300         phase->glsl_program_num = glsl_program_num;
301         phase->vertex_shader = vs_obj;
302         phase->fragment_shader = fs_obj;
303         phase->input_needs_mipmaps = input_needs_mipmaps;
304         phase->inputs = true_inputs;
305         phase->effects = sorted_effects;
306
307         return phase;
308 }
309
310 // Construct GLSL programs, starting at the given effect and following
311 // the chain from there. We end a program every time we come to an effect
312 // marked as "needs texture bounce", one that is used by multiple other
313 // effects, every time an effect wants to change the output size,
314 // and of course at the end.
315 //
316 // We follow a quite simple depth-first search from the output, although
317 // without any explicit recursion.
318 void EffectChain::construct_glsl_programs(Node *output)
319 {
320         // Which effects have already been completed in this phase?
321         // We need to keep track of it, as an effect with multiple outputs
322         // could otherwise be calculated multiple times.
323         std::set<Node *> completed_effects;
324
325         // Effects in the current phase, as well as inputs (outputs from other phases
326         // that we depend on). Note that since we start iterating from the end,
327         // the effect list will be in the reverse order.
328         std::vector<Node *> this_phase_inputs;
329         std::vector<Node *> this_phase_effects;
330
331         // Effects that we have yet to calculate, but that we know should
332         // be in the current phase.
333         std::stack<Node *> effects_todo_this_phase;
334
335         // Effects that we have yet to calculate, but that come from other phases.
336         // We delay these until we have this phase done in its entirety,
337         // at which point we pick any of them and start a new phase from that.
338         std::stack<Node *> effects_todo_other_phases;
339
340         effects_todo_this_phase.push(output);
341
342         for ( ;; ) {  // Termination condition within loop.
343                 if (!effects_todo_this_phase.empty()) {
344                         // OK, we have more to do this phase.
345                         Node *node = effects_todo_this_phase.top();
346                         effects_todo_this_phase.pop();
347
348                         // This should currently only happen for effects that are inputs
349                         // (either true inputs or phase outputs). We special-case inputs,
350                         // and then deduplicate phase outputs in compile_glsl_program().
351                         if (node->effect->num_inputs() == 0 && completed_effects.count(node)) {
352                                 continue;
353                         }
354                         assert(completed_effects.count(node) == 0);
355
356                         this_phase_effects.push_back(node);
357                         completed_effects.insert(node);
358
359                         // Find all the dependencies of this effect, and add them to the stack.
360                         std::vector<Node *> deps = node->incoming_links;
361                         assert(node->effect->num_inputs() == deps.size());
362                         for (unsigned i = 0; i < deps.size(); ++i) {
363                                 bool start_new_phase = false;
364
365                                 // FIXME: If we sample directly from a texture, we won't need this.
366                                 if (node->effect->needs_texture_bounce()) {
367                                         start_new_phase = true;
368                                 }
369
370                                 if (deps[i]->outgoing_links.size() > 1) {
371                                         if (deps[i]->effect->num_inputs() > 0) {
372                                                 // More than one effect uses this as the input,
373                                                 // and it is not a texture itself.
374                                                 // The easiest thing to do (and probably also the safest
375                                                 // performance-wise in most cases) is to bounce it to a texture
376                                                 // and then let the next passes read from that.
377                                                 start_new_phase = true;
378                                         } else {
379                                                 // For textures, we try to be slightly more clever;
380                                                 // if none of our outputs need a bounce, we don't bounce
381                                                 // but instead simply use the effect many times.
382                                                 //
383                                                 // Strictly speaking, we could bounce it for some outputs
384                                                 // and use it directly for others, but the processing becomes
385                                                 // somewhat simpler if the effect is only used in one such way.
386                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
387                                                         Node *rdep = deps[i]->outgoing_links[j];
388                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
389                                                 }
390                                         }
391                                 }
392
393                                 if (deps[i]->effect->changes_output_size()) {
394                                         start_new_phase = true;
395                                 }
396
397                                 if (start_new_phase) {
398                                         effects_todo_other_phases.push(deps[i]);
399                                         this_phase_inputs.push_back(deps[i]);
400                                 } else {
401                                         effects_todo_this_phase.push(deps[i]);
402                                 }
403                         }
404                         continue;
405                 }
406
407                 // No more effects to do this phase. Take all the ones we have,
408                 // and create a GLSL program for it.
409                 if (!this_phase_effects.empty()) {
410                         reverse(this_phase_effects.begin(), this_phase_effects.end());
411                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
412                         this_phase_effects.back()->phase = phases.back();
413                         this_phase_inputs.clear();
414                         this_phase_effects.clear();
415                 }
416                 assert(this_phase_inputs.empty());
417                 assert(this_phase_effects.empty());
418
419                 // If we have no effects left, exit.
420                 if (effects_todo_other_phases.empty()) {
421                         break;
422                 }
423
424                 Node *node = effects_todo_other_phases.top();
425                 effects_todo_other_phases.pop();
426
427                 if (completed_effects.count(node) == 0) {
428                         // Start a new phase, calculating from this effect.
429                         effects_todo_this_phase.push(node);
430                 }
431         }
432
433         // Finally, since the phases are found from the output but must be executed
434         // from the input(s), reverse them, too.
435         std::reverse(phases.begin(), phases.end());
436 }
437
438 void EffectChain::output_dot(const char *filename)
439 {
440         if (movit_debug_level != MOVIT_DEBUG_ON) {
441                 return;
442         }
443
444         FILE *fp = fopen(filename, "w");
445         if (fp == NULL) {
446                 perror(filename);
447                 exit(1);
448         }
449
450         fprintf(fp, "digraph G {\n");
451         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
452         for (unsigned i = 0; i < nodes.size(); ++i) {
453                 // Find out which phase this event belongs to.
454                 int in_phase = -1;
455                 for (unsigned j = 0; j < phases.size(); ++j) {
456                         const Phase* p = phases[j];
457                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
458                                 assert(in_phase == -1);
459                                 in_phase = j;
460                         }
461                 }
462
463                 if (in_phase == -1) {
464                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
465                 } else {
466                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
467                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
468                                 (in_phase % 8) + 1);
469                 }
470
471                 char from_node_id[256];
472                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
473
474                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
475                         char to_node_id[256];
476                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
477
478                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
479                         output_dot_edge(fp, from_node_id, to_node_id, labels);
480                 }
481
482                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
483                         // Output node.
484                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
485                         output_dot_edge(fp, from_node_id, "output", labels);
486                 }
487         }
488         fprintf(fp, "}\n");
489
490         fclose(fp);
491 }
492
493 std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
494 {
495         std::vector<std::string> labels;
496
497         if (to != NULL && to->effect->needs_texture_bounce()) {
498                 labels.push_back("needs_bounce");
499         }
500         if (from->effect->changes_output_size()) {
501                 labels.push_back("resize");
502         }
503
504         switch (from->output_color_space) {
505         case COLORSPACE_INVALID:
506                 labels.push_back("spc[invalid]");
507                 break;
508         case COLORSPACE_REC_601_525:
509                 labels.push_back("spc[rec601-525]");
510                 break;
511         case COLORSPACE_REC_601_625:
512                 labels.push_back("spc[rec601-625]");
513                 break;
514         default:
515                 break;
516         }
517
518         switch (from->output_gamma_curve) {
519         case GAMMA_INVALID:
520                 labels.push_back("gamma[invalid]");
521                 break;
522         case GAMMA_sRGB:
523                 labels.push_back("gamma[sRGB]");
524                 break;
525         case GAMMA_REC_601:  // and GAMMA_REC_709
526                 labels.push_back("gamma[rec601/709]");
527                 break;
528         default:
529                 break;
530         }
531
532         switch (from->output_alpha_type) {
533         case ALPHA_INVALID:
534                 labels.push_back("alpha[invalid]");
535                 break;
536         case ALPHA_BLANK:
537                 labels.push_back("alpha[blank]");
538                 break;
539         case ALPHA_POSTMULTIPLIED:
540                 labels.push_back("alpha[postmult]");
541                 break;
542         default:
543                 break;
544         }
545
546         return labels;
547 }
548
549 void EffectChain::output_dot_edge(FILE *fp,
550                                   const std::string &from_node_id,
551                                   const std::string &to_node_id,
552                                   const std::vector<std::string> &labels)
553 {
554         if (labels.empty()) {
555                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
556         } else {
557                 std::string label = labels[0];
558                 for (unsigned k = 1; k < labels.size(); ++k) {
559                         label += ", " + labels[k];
560                 }
561                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
562         }
563 }
564
565 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
566 {
567         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
568                 // Same aspect, or W/H > aspect (image is wider than the frame).
569                 // In either case, keep width.
570                 return width;
571         } else {
572                 // W/H < aspect (image is taller than the frame), so keep height,
573                 // and adjust width correspondingly.
574                 return lrintf(height * aspect_nom / aspect_denom);
575         }
576 }
577
578 // Propagate input texture sizes throughout, and inform effects downstream.
579 // (Like a lot of other code, we depend on effects being in topological order.)
580 void EffectChain::inform_input_sizes(Phase *phase)
581 {
582         // All effects that have a defined size (inputs and RTT inputs)
583         // get that. Reset all others.
584         for (unsigned i = 0; i < phase->effects.size(); ++i) {
585                 Node *node = phase->effects[i];
586                 if (node->effect->num_inputs() == 0) {
587                         Input *input = static_cast<Input *>(node->effect);
588                         node->output_width = input->get_width();
589                         node->output_height = input->get_height();
590                         assert(node->output_width != 0);
591                         assert(node->output_height != 0);
592                 } else {
593                         node->output_width = node->output_height = 0;
594                 }
595         }
596         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
597                 Node *input = phase->inputs[i];
598                 input->output_width = input->phase->output_width;
599                 input->output_height = input->phase->output_height;
600                 assert(input->output_width != 0);
601                 assert(input->output_height != 0);
602         }
603
604         // Now propagate from the inputs towards the end, and inform as we go.
605         // The rules are simple:
606         //
607         //   1. Don't touch effects that already have given sizes (ie., inputs).
608         //   2. If all of your inputs have the same size, that will be your output size.
609         //   3. Otherwise, your output size is 0x0.
610         for (unsigned i = 0; i < phase->effects.size(); ++i) {
611                 Node *node = phase->effects[i];
612                 if (node->effect->num_inputs() == 0) {
613                         continue;
614                 }
615                 unsigned this_output_width = 0;
616                 unsigned this_output_height = 0;
617                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
618                         Node *input = node->incoming_links[j];
619                         node->effect->inform_input_size(j, input->output_width, input->output_height);
620                         if (j == 0) {
621                                 this_output_width = input->output_width;
622                                 this_output_height = input->output_height;
623                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
624                                 // Inputs disagree.
625                                 this_output_width = 0;
626                                 this_output_height = 0;
627                         }
628                 }
629                 node->output_width = this_output_width;
630                 node->output_height = this_output_height;
631         }
632 }
633
634 // Note: You should call inform_input_sizes() before this, as the last effect's
635 // desired output size might change based on the inputs.
636 void EffectChain::find_output_size(Phase *phase)
637 {
638         Node *output_node = phase->effects.back();
639
640         // If the last effect explicitly sets an output size, use that.
641         if (output_node->effect->changes_output_size()) {
642                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
643                 return;
644         }
645
646         // If not, look at the input phases and textures.
647         // We select the largest one (by fit into the current aspect).
648         unsigned best_width = 0;
649         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
650                 Node *input = phase->inputs[i];
651                 assert(input->phase->output_width != 0);
652                 assert(input->phase->output_height != 0);
653                 unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height);
654                 if (width > best_width) {
655                         best_width = width;
656                 }
657         }
658         for (unsigned i = 0; i < phase->effects.size(); ++i) {
659                 Effect *effect = phase->effects[i]->effect;
660                 if (effect->num_inputs() != 0) {
661                         continue;
662                 }
663
664                 Input *input = static_cast<Input *>(effect);
665                 unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height());
666                 if (width > best_width) {
667                         best_width = width;
668                 }
669         }
670         assert(best_width != 0);
671         phase->output_width = best_width;
672         phase->output_height = best_width * aspect_denom / aspect_nom;
673 }
674
675 void EffectChain::sort_all_nodes_topologically()
676 {
677         nodes = topological_sort(nodes);
678 }
679
680 std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
681 {
682         std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
683         std::vector<Node *> sorted_list;
684         for (unsigned i = 0; i < nodes.size(); ++i) {
685                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
686         }
687         reverse(sorted_list.begin(), sorted_list.end());
688         return sorted_list;
689 }
690
691 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
692 {
693         if (nodes_left_to_visit->count(node) == 0) {
694                 return;
695         }
696         nodes_left_to_visit->erase(node);
697         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
698                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
699         }
700         sorted_list->push_back(node);
701 }
702
703 void EffectChain::find_color_spaces_for_inputs()
704 {
705         for (unsigned i = 0; i < nodes.size(); ++i) {
706                 Node *node = nodes[i];
707                 if (node->disabled) {
708                         continue;
709                 }
710                 if (node->incoming_links.size() == 0) {
711                         Input *input = static_cast<Input *>(node->effect);
712                         node->output_color_space = input->get_color_space();
713                         node->output_gamma_curve = input->get_gamma_curve();
714
715                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
716                         switch (alpha_handling) {
717                         case Effect::OUTPUT_BLANK_ALPHA:
718                                 node->output_alpha_type = ALPHA_BLANK;
719                                 break;
720                         case Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED:
721                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
722                                 break;
723                         case Effect::OUTPUT_ALPHA_POSTMULTIPLIED:
724                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
725                                 break;
726                         case Effect::DONT_CARE_ALPHA_TYPE:
727                         default:
728                                 assert(false);
729                         }
730
731                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
732                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
733                         }
734                 }
735         }
736 }
737
738 // Propagate gamma and color space information as far as we can in the graph.
739 // The rules are simple: Anything where all the inputs agree, get that as
740 // output as well. Anything else keeps having *_INVALID.
741 void EffectChain::propagate_gamma_and_color_space()
742 {
743         // We depend on going through the nodes in order.
744         sort_all_nodes_topologically();
745
746         for (unsigned i = 0; i < nodes.size(); ++i) {
747                 Node *node = nodes[i];
748                 if (node->disabled) {
749                         continue;
750                 }
751                 assert(node->incoming_links.size() == node->effect->num_inputs());
752                 if (node->incoming_links.size() == 0) {
753                         assert(node->output_color_space != COLORSPACE_INVALID);
754                         assert(node->output_gamma_curve != GAMMA_INVALID);
755                         continue;
756                 }
757
758                 Colorspace color_space = node->incoming_links[0]->output_color_space;
759                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
760                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
761                         if (node->incoming_links[j]->output_color_space != color_space) {
762                                 color_space = COLORSPACE_INVALID;
763                         }
764                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
765                                 gamma_curve = GAMMA_INVALID;
766                         }
767                 }
768
769                 // The conversion effects already have their outputs set correctly,
770                 // so leave them alone.
771                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
772                         node->output_color_space = color_space;
773                 }               
774                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
775                     node->effect->effect_type_id() != "GammaExpansionEffect") {
776                         node->output_gamma_curve = gamma_curve;
777                 }               
778         }
779 }
780
781 // Propagate alpha information as far as we can in the graph.
782 // Similar to propagate_gamma_and_color_space().
783 void EffectChain::propagate_alpha()
784 {
785         // We depend on going through the nodes in order.
786         sort_all_nodes_topologically();
787
788         for (unsigned i = 0; i < nodes.size(); ++i) {
789                 Node *node = nodes[i];
790                 if (node->disabled) {
791                         continue;
792                 }
793                 assert(node->incoming_links.size() == node->effect->num_inputs());
794                 if (node->incoming_links.size() == 0) {
795                         assert(node->output_alpha_type != ALPHA_INVALID);
796                         continue;
797                 }
798
799                 // The alpha multiplication/division effects are special cases.
800                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
801                         assert(node->incoming_links.size() == 1);
802                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
803                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
804                         continue;
805                 }
806                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
807                         assert(node->incoming_links.size() == 1);
808                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
809                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
810                         continue;
811                 }
812
813                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
814                 // because they are the only one that _need_ postmultiplied alpha.
815                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
816                     node->effect->effect_type_id() == "GammaExpansionEffect") {
817                         assert(node->incoming_links.size() == 1);
818                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
819                                 node->output_alpha_type = ALPHA_BLANK;
820                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
821                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
822                         } else {
823                                 node->output_alpha_type = ALPHA_INVALID;
824                         }
825                         continue;
826                 }
827
828                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
829                 // or OUTPUT_ALPHA_POSTMULTIPLIED), and they have already been
830                 // taken care of above. Rationale: Even if you could imagine
831                 // e.g. an effect that took in an image and set alpha=1.0
832                 // unconditionally, it wouldn't make any sense to have it as
833                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
834                 // got its input pre- or postmultiplied, so it wouldn't know
835                 // whether to divide away the old alpha or not.
836                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
837                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED ||
838                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
839
840                 // If the node has multiple inputs, check that they are all valid and
841                 // the same.
842                 bool any_invalid = false;
843                 bool any_premultiplied = false;
844                 bool any_postmultiplied = false;
845
846                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
847                         switch (node->incoming_links[j]->output_alpha_type) {
848                         case ALPHA_INVALID:
849                                 any_invalid = true;
850                                 break;
851                         case ALPHA_BLANK:
852                                 // Blank is good as both pre- and postmultiplied alpha,
853                                 // so just ignore it.
854                                 break;
855                         case ALPHA_PREMULTIPLIED:
856                                 any_premultiplied = true;
857                                 break;
858                         case ALPHA_POSTMULTIPLIED:
859                                 any_postmultiplied = true;
860                                 break;
861                         default:
862                                 assert(false);
863                         }
864                 }
865
866                 if (any_invalid) {
867                         node->output_alpha_type = ALPHA_INVALID;
868                         continue;
869                 }
870
871                 // Inputs must be of the same type.
872                 if (any_premultiplied && any_postmultiplied) {
873                         node->output_alpha_type = ALPHA_INVALID;
874                         continue;
875                 }
876
877                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED) {
878                         // If the effect has asked for premultiplied alpha, check that it has got it.
879                         if (any_postmultiplied) {
880                                 node->output_alpha_type = ALPHA_INVALID;
881                         } else {
882                                 // In some rare cases, it might be advantageous to say
883                                 // that blank input alpha yields blank output alpha.
884                                 // However, this would cause a more complex Effect interface
885                                 // an effect would need to guarantee that it doesn't mess with
886                                 // blank alpha), so this is the simplest.
887                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
888                         }
889                 } else {
890                         // OK, all inputs are the same, and this effect is not going
891                         // to change it.
892                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
893                         if (any_premultiplied) {
894                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
895                         } else if (any_postmultiplied) {
896                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
897                         } else {
898                                 node->output_alpha_type = ALPHA_BLANK;
899                         }
900                 }
901         }
902 }
903
904 bool EffectChain::node_needs_colorspace_fix(Node *node)
905 {
906         if (node->disabled) {
907                 return false;
908         }
909         if (node->effect->num_inputs() == 0) {
910                 return false;
911         }
912
913         // propagate_gamma_and_color_space() has already set our output
914         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
915         if (node->output_color_space == COLORSPACE_INVALID) {
916                 return true;
917         }
918         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
919 }
920
921 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
922 // the graph. Our strategy is not always optimal, but quite simple:
923 // Find an effect that's as early as possible where the inputs are of
924 // unacceptable colorspaces (that is, either different, or, if the effect only
925 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
926 // propagate the information anew, and repeat until there are no more such
927 // effects.
928 void EffectChain::fix_internal_color_spaces()
929 {
930         unsigned colorspace_propagation_pass = 0;
931         bool found_any;
932         do {
933                 found_any = false;
934                 for (unsigned i = 0; i < nodes.size(); ++i) {
935                         Node *node = nodes[i];
936                         if (!node_needs_colorspace_fix(node)) {
937                                 continue;
938                         }
939
940                         // Go through each input that is not sRGB, and insert
941                         // a colorspace conversion before it.
942                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
943                                 Node *input = node->incoming_links[j];
944                                 assert(input->output_color_space != COLORSPACE_INVALID);
945                                 if (input->output_color_space == COLORSPACE_sRGB) {
946                                         continue;
947                                 }
948                                 Node *conversion = add_node(new ColorspaceConversionEffect());
949                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
950                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
951                                 conversion->output_color_space = COLORSPACE_sRGB;
952                                 insert_node_between(input, conversion, node);
953                         }
954
955                         // Re-sort topologically, and propagate the new information.
956                         propagate_gamma_and_color_space();
957                         
958                         found_any = true;
959                         break;
960                 }
961         
962                 char filename[256];
963                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
964                 output_dot(filename);
965                 assert(colorspace_propagation_pass < 100);
966         } while (found_any);
967
968         for (unsigned i = 0; i < nodes.size(); ++i) {
969                 Node *node = nodes[i];
970                 if (node->disabled) {
971                         continue;
972                 }
973                 assert(node->output_color_space != COLORSPACE_INVALID);
974         }
975 }
976
977 bool EffectChain::node_needs_alpha_fix(Node *node)
978 {
979         if (node->disabled) {
980                 return false;
981         }
982
983         // propagate_alpha() has already set our output to ALPHA_INVALID if the
984         // inputs differ or we are otherwise in mismatch, so we can rely on that.
985         return (node->output_alpha_type == ALPHA_INVALID);
986 }
987
988 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
989 // the graph. Similar to fix_internal_color_spaces().
990 void EffectChain::fix_internal_alpha(unsigned step)
991 {
992         unsigned alpha_propagation_pass = 0;
993         bool found_any;
994         do {
995                 found_any = false;
996                 for (unsigned i = 0; i < nodes.size(); ++i) {
997                         Node *node = nodes[i];
998                         if (!node_needs_alpha_fix(node)) {
999                                 continue;
1000                         }
1001
1002                         // If we need to fix up GammaExpansionEffect, then clearly something
1003                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1004                         // is meaningless.
1005                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1006
1007                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1008
1009                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1010                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1011                                 assert(node->incoming_links.size() == 1);
1012                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1013                                 desired_type = ALPHA_POSTMULTIPLIED;
1014                         }
1015
1016                         // Go through each input that is not premultiplied alpha, and insert
1017                         // a conversion before it.
1018                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1019                                 Node *input = node->incoming_links[j];
1020                                 assert(input->output_alpha_type != ALPHA_INVALID);
1021                                 if (input->output_alpha_type == desired_type ||
1022                                     input->output_alpha_type == ALPHA_BLANK) {
1023                                         continue;
1024                                 }
1025                                 Node *conversion;
1026                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1027                                         conversion = add_node(new AlphaMultiplicationEffect());
1028                                 } else {
1029                                         conversion = add_node(new AlphaDivisionEffect());
1030                                 }
1031                                 conversion->output_alpha_type = desired_type;
1032                                 insert_node_between(input, conversion, node);
1033                         }
1034
1035                         // Re-sort topologically, and propagate the new information.
1036                         propagate_gamma_and_color_space();
1037                         propagate_alpha();
1038                         
1039                         found_any = true;
1040                         break;
1041                 }
1042         
1043                 char filename[256];
1044                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1045                 output_dot(filename);
1046                 assert(alpha_propagation_pass < 100);
1047         } while (found_any);
1048
1049         for (unsigned i = 0; i < nodes.size(); ++i) {
1050                 Node *node = nodes[i];
1051                 if (node->disabled) {
1052                         continue;
1053                 }
1054                 assert(node->output_alpha_type != ALPHA_INVALID);
1055         }
1056 }
1057
1058 // Make so that the output is in the desired color space.
1059 void EffectChain::fix_output_color_space()
1060 {
1061         Node *output = find_output_node();
1062         if (output->output_color_space != output_format.color_space) {
1063                 Node *conversion = add_node(new ColorspaceConversionEffect());
1064                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1065                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1066                 conversion->output_color_space = output_format.color_space;
1067                 connect_nodes(output, conversion);
1068                 propagate_alpha();
1069                 propagate_gamma_and_color_space();
1070         }
1071 }
1072
1073 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1074 void EffectChain::fix_output_alpha()
1075 {
1076         Node *output = find_output_node();
1077         assert(output->output_alpha_type != ALPHA_INVALID);
1078         if (output->output_alpha_type == ALPHA_BLANK) {
1079                 // No alpha output, so we don't care.
1080                 return;
1081         }
1082         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1083             output_alpha_format == OUTPUT_ALPHA_POSTMULTIPLIED) {
1084                 Node *conversion = add_node(new AlphaDivisionEffect());
1085                 connect_nodes(output, conversion);
1086                 propagate_alpha();
1087                 propagate_gamma_and_color_space();
1088         }
1089         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1090             output_alpha_format == OUTPUT_ALPHA_PREMULTIPLIED) {
1091                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1092                 connect_nodes(output, conversion);
1093                 propagate_alpha();
1094                 propagate_gamma_and_color_space();
1095         }
1096 }
1097
1098 bool EffectChain::node_needs_gamma_fix(Node *node)
1099 {
1100         if (node->disabled) {
1101                 return false;
1102         }
1103
1104         // Small hack since the output is not an explicit node:
1105         // If we are the last node and our output is in the wrong
1106         // space compared to EffectChain's output, we need to fix it.
1107         // This will only take us to linear, but fix_output_gamma()
1108         // will come and take us to the desired output gamma
1109         // if it is needed.
1110         //
1111         // This needs to be before everything else, since it could
1112         // even apply to inputs (if they are the only effect).
1113         if (node->outgoing_links.empty() &&
1114             node->output_gamma_curve != output_format.gamma_curve &&
1115             node->output_gamma_curve != GAMMA_LINEAR) {
1116                 return true;
1117         }
1118
1119         if (node->effect->num_inputs() == 0) {
1120                 return false;
1121         }
1122
1123         // propagate_gamma_and_color_space() has already set our output
1124         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1125         // except for GammaCompressionEffect.
1126         if (node->output_gamma_curve == GAMMA_INVALID) {
1127                 return true;
1128         }
1129         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1130                 assert(node->incoming_links.size() == 1);
1131                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1132         }
1133
1134         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1135 }
1136
1137 // Very similar to fix_internal_color_spaces(), but for gamma.
1138 // There is one difference, though; before we start adding conversion nodes,
1139 // we see if we can get anything out of asking the sources to deliver
1140 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1141 // does that part, while fix_internal_gamma_by_inserting_nodes()
1142 // inserts nodes as needed afterwards.
1143 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1144 {
1145         unsigned gamma_propagation_pass = 0;
1146         bool found_any;
1147         do {
1148                 found_any = false;
1149                 for (unsigned i = 0; i < nodes.size(); ++i) {
1150                         Node *node = nodes[i];
1151                         if (!node_needs_gamma_fix(node)) {
1152                                 continue;
1153                         }
1154
1155                         // See if all inputs can give us linear gamma. If not, leave it.
1156                         std::vector<Node *> nonlinear_inputs;
1157                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1158                         assert(!nonlinear_inputs.empty());
1159
1160                         bool all_ok = true;
1161                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1162                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1163                                 all_ok &= input->can_output_linear_gamma();
1164                         }
1165
1166                         if (!all_ok) {
1167                                 continue;
1168                         }
1169
1170                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1171                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1172                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1173                         }
1174
1175                         // Re-sort topologically, and propagate the new information.
1176                         propagate_gamma_and_color_space();
1177                         
1178                         found_any = true;
1179                         break;
1180                 }
1181         
1182                 char filename[256];
1183                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1184                 output_dot(filename);
1185                 assert(gamma_propagation_pass < 100);
1186         } while (found_any);
1187 }
1188
1189 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1190 {
1191         unsigned gamma_propagation_pass = 0;
1192         bool found_any;
1193         do {
1194                 found_any = false;
1195                 for (unsigned i = 0; i < nodes.size(); ++i) {
1196                         Node *node = nodes[i];
1197                         if (!node_needs_gamma_fix(node)) {
1198                                 continue;
1199                         }
1200
1201                         // Special case: We could be an input and still be asked to
1202                         // fix our gamma; if so, we should be the only node
1203                         // (as node_needs_gamma_fix() would only return true in
1204                         // for an input in that case). That means we should insert
1205                         // a conversion node _after_ ourselves.
1206                         if (node->incoming_links.empty()) {
1207                                 assert(node->outgoing_links.empty());
1208                                 Node *conversion = add_node(new GammaExpansionEffect());
1209                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1210                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1211                                 connect_nodes(node, conversion);
1212                         }
1213
1214                         // If not, go through each input that is not linear gamma,
1215                         // and insert a gamma conversion before it.
1216                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1217                                 Node *input = node->incoming_links[j];
1218                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1219                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1220                                         continue;
1221                                 }
1222                                 Node *conversion = add_node(new GammaExpansionEffect());
1223                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1224                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1225                                 insert_node_between(input, conversion, node);
1226                         }
1227
1228                         // Re-sort topologically, and propagate the new information.
1229                         propagate_alpha();
1230                         propagate_gamma_and_color_space();
1231                         
1232                         found_any = true;
1233                         break;
1234                 }
1235         
1236                 char filename[256];
1237                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1238                 output_dot(filename);
1239                 assert(gamma_propagation_pass < 100);
1240         } while (found_any);
1241
1242         for (unsigned i = 0; i < nodes.size(); ++i) {
1243                 Node *node = nodes[i];
1244                 if (node->disabled) {
1245                         continue;
1246                 }
1247                 assert(node->output_gamma_curve != GAMMA_INVALID);
1248         }
1249 }
1250
1251 // Make so that the output is in the desired gamma.
1252 // Note that this assumes linear input gamma, so it might create the need
1253 // for another pass of fix_internal_gamma().
1254 void EffectChain::fix_output_gamma()
1255 {
1256         Node *output = find_output_node();
1257         if (output->output_gamma_curve != output_format.gamma_curve) {
1258                 Node *conversion = add_node(new GammaCompressionEffect());
1259                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1260                 conversion->output_gamma_curve = output_format.gamma_curve;
1261                 connect_nodes(output, conversion);
1262         }
1263 }
1264         
1265 // If the user has requested dither, add a DitherEffect right at the end
1266 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1267 // since dither is about the only effect that can _not_ be done in linear space.
1268 void EffectChain::add_dither_if_needed()
1269 {
1270         if (num_dither_bits == 0) {
1271                 return;
1272         }
1273         Node *output = find_output_node();
1274         Node *dither = add_node(new DitherEffect());
1275         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1276         connect_nodes(output, dither);
1277
1278         dither_effect = dither->effect;
1279 }
1280
1281 // Find the output node. This is, simply, one that has no outgoing links.
1282 // If there are multiple ones, the graph is malformed (we do not support
1283 // multiple outputs right now).
1284 Node *EffectChain::find_output_node()
1285 {
1286         std::vector<Node *> output_nodes;
1287         for (unsigned i = 0; i < nodes.size(); ++i) {
1288                 Node *node = nodes[i];
1289                 if (node->disabled) {
1290                         continue;
1291                 }
1292                 if (node->outgoing_links.empty()) {
1293                         output_nodes.push_back(node);
1294                 }
1295         }
1296         assert(output_nodes.size() == 1);
1297         return output_nodes[0];
1298 }
1299
1300 void EffectChain::finalize()
1301 {
1302         // Output the graph as it is before we do any conversions on it.
1303         output_dot("step0-start.dot");
1304
1305         // Give each effect in turn a chance to rewrite its own part of the graph.
1306         // Note that if more effects are added as part of this, they will be
1307         // picked up as part of the same for loop, since they are added at the end.
1308         for (unsigned i = 0; i < nodes.size(); ++i) {
1309                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1310         }
1311         output_dot("step1-rewritten.dot");
1312
1313         find_color_spaces_for_inputs();
1314         output_dot("step2-input-colorspace.dot");
1315
1316         propagate_alpha();
1317         output_dot("step3-propagated-alpha.dot");
1318
1319         propagate_gamma_and_color_space();
1320         output_dot("step4-propagated-all.dot");
1321
1322         fix_internal_color_spaces();
1323         fix_internal_alpha(6);
1324         fix_output_color_space();
1325         output_dot("step7-output-colorspacefix.dot");
1326         fix_output_alpha();
1327         output_dot("step8-output-alphafix.dot");
1328
1329         // Note that we need to fix gamma after colorspace conversion,
1330         // because colorspace conversions might create needs for gamma conversions.
1331         // Also, we need to run an extra pass of fix_internal_gamma() after 
1332         // fixing the output gamma, as we only have conversions to/from linear,
1333         // and fix_internal_alpha() since GammaCompressionEffect needs
1334         // postmultiplied input.
1335         fix_internal_gamma_by_asking_inputs(9);
1336         fix_internal_gamma_by_inserting_nodes(10);
1337         fix_output_gamma();
1338         output_dot("step11-output-gammafix.dot");
1339         propagate_alpha();
1340         output_dot("step12-output-alpha-propagated.dot");
1341         fix_internal_alpha(13);
1342         output_dot("step14-output-alpha-fixed.dot");
1343         fix_internal_gamma_by_asking_inputs(15);
1344         fix_internal_gamma_by_inserting_nodes(16);
1345
1346         output_dot("step17-before-dither.dot");
1347
1348         add_dither_if_needed();
1349
1350         output_dot("step18-final.dot");
1351         
1352         // Construct all needed GLSL programs, starting at the output.
1353         construct_glsl_programs(find_output_node());
1354
1355         output_dot("step19-split-to-phases.dot");
1356
1357         // If we have more than one phase, we need intermediate render-to-texture.
1358         // Construct an FBO, and then as many textures as we need.
1359         // We choose the simplest option of having one texture per output,
1360         // since otherwise this turns into an (albeit simple)
1361         // register allocation problem.
1362         if (phases.size() > 1) {
1363                 glGenFramebuffers(1, &fbo);
1364
1365                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1366                         inform_input_sizes(phases[i]);
1367                         find_output_size(phases[i]);
1368
1369                         Node *output_node = phases[i]->effects.back();
1370                         glGenTextures(1, &output_node->output_texture);
1371                         check_error();
1372                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1373                         check_error();
1374                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1375                         check_error();
1376                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1377                         check_error();
1378                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1379                         check_error();
1380
1381                         output_node->output_texture_width = phases[i]->output_width;
1382                         output_node->output_texture_height = phases[i]->output_height;
1383                 }
1384                 inform_input_sizes(phases.back());
1385         }
1386                 
1387         for (unsigned i = 0; i < inputs.size(); ++i) {
1388                 inputs[i]->finalize();
1389         }
1390
1391         assert(phases[0]->inputs.empty());
1392         
1393         finalized = true;
1394 }
1395
1396 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1397 {
1398         assert(finalized);
1399
1400         // Save original viewport.
1401         GLuint x = 0, y = 0;
1402
1403         if (width == 0 && height == 0) {
1404                 GLint viewport[4];
1405                 glGetIntegerv(GL_VIEWPORT, viewport);
1406                 x = viewport[0];
1407                 y = viewport[1];
1408                 width = viewport[2];
1409                 height = viewport[3];
1410         }
1411
1412         // Basic state.
1413         glDisable(GL_BLEND);
1414         check_error();
1415         glDisable(GL_DEPTH_TEST);
1416         check_error();
1417         glDepthMask(GL_FALSE);
1418         check_error();
1419
1420         glMatrixMode(GL_PROJECTION);
1421         glLoadIdentity();
1422         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1423
1424         glMatrixMode(GL_MODELVIEW);
1425         glLoadIdentity();
1426
1427         if (phases.size() > 1) {
1428                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1429                 check_error();
1430         }
1431
1432         std::set<Node *> generated_mipmaps;
1433
1434         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1435                 // See if the requested output size has changed. If so, we need to recreate
1436                 // the texture (and before we start setting up inputs).
1437                 inform_input_sizes(phases[phase]);
1438                 if (phase != phases.size() - 1) {
1439                         find_output_size(phases[phase]);
1440
1441                         Node *output_node = phases[phase]->effects.back();
1442
1443                         if (output_node->output_texture_width != phases[phase]->output_width ||
1444                             output_node->output_texture_height != phases[phase]->output_height) {
1445                                 glActiveTexture(GL_TEXTURE0);
1446                                 check_error();
1447                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1448                                 check_error();
1449                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1450                                 check_error();
1451                                 glBindTexture(GL_TEXTURE_2D, 0);
1452                                 check_error();
1453
1454                                 output_node->output_texture_width = phases[phase]->output_width;
1455                                 output_node->output_texture_height = phases[phase]->output_height;
1456                         }
1457                 }
1458
1459                 glUseProgram(phases[phase]->glsl_program_num);
1460                 check_error();
1461
1462                 // Set up RTT inputs for this phase.
1463                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1464                         glActiveTexture(GL_TEXTURE0 + sampler);
1465                         Node *input = phases[phase]->inputs[sampler];
1466                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1467                         check_error();
1468                         if (phases[phase]->input_needs_mipmaps) {
1469                                 if (generated_mipmaps.count(input) == 0) {
1470                                         glGenerateMipmap(GL_TEXTURE_2D);
1471                                         check_error();
1472                                         generated_mipmaps.insert(input);
1473                                 }
1474                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1475                                 check_error();
1476                         } else {
1477                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1478                                 check_error();
1479                         }
1480
1481                         std::string texture_name = std::string("tex_") + input->effect_id;
1482                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1483                         check_error();
1484                 }
1485
1486                 // And now the output.
1487                 if (phase == phases.size() - 1) {
1488                         // Last phase goes to the output the user specified.
1489                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1490                         check_error();
1491                         glViewport(x, y, width, height);
1492                         if (dither_effect != NULL) {
1493                                 CHECK(dither_effect->set_int("output_width", width));
1494                                 CHECK(dither_effect->set_int("output_height", height));
1495                         }
1496                 } else {
1497                         Node *output_node = phases[phase]->effects.back();
1498                         glFramebufferTexture2D(
1499                                 GL_FRAMEBUFFER,
1500                                 GL_COLOR_ATTACHMENT0,
1501                                 GL_TEXTURE_2D,
1502                                 output_node->output_texture,
1503                                 0);
1504                         check_error();
1505                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1506                 }
1507
1508                 // Give the required parameters to all the effects.
1509                 unsigned sampler_num = phases[phase]->inputs.size();
1510                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1511                         Node *node = phases[phase]->effects[i];
1512                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1513                         check_error();
1514                 }
1515
1516                 // Now draw!
1517                 glBegin(GL_QUADS);
1518
1519                 glTexCoord2f(0.0f, 0.0f);
1520                 glVertex2f(0.0f, 0.0f);
1521
1522                 glTexCoord2f(1.0f, 0.0f);
1523                 glVertex2f(1.0f, 0.0f);
1524
1525                 glTexCoord2f(1.0f, 1.0f);
1526                 glVertex2f(1.0f, 1.0f);
1527
1528                 glTexCoord2f(0.0f, 1.0f);
1529                 glVertex2f(0.0f, 1.0f);
1530
1531                 glEnd();
1532                 check_error();
1533
1534                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1535                         Node *node = phases[phase]->effects[i];
1536                         node->effect->clear_gl_state();
1537                 }
1538         }
1539 }