Be more consistent about how we handle textures and bouncing; either bounce them...
[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
8 #include <algorithm>
9 #include <set>
10 #include <stack>
11 #include <vector>
12
13 #include "util.h"
14 #include "effect_chain.h"
15 #include "gamma_expansion_effect.h"
16 #include "gamma_compression_effect.h"
17 #include "colorspace_conversion_effect.h"
18 #include "input.h"
19 #include "opengl.h"
20
21 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
22         : aspect_nom(aspect_nom),
23           aspect_denom(aspect_denom),
24           finalized(false) {}
25
26 Input *EffectChain::add_input(Input *input)
27 {
28         inputs.push_back(input);
29
30         Node *node = add_node(input);
31         node->output_color_space = input->get_color_space();
32         node->output_gamma_curve = input->get_gamma_curve();
33         return input;
34 }
35
36 void EffectChain::add_output(const ImageFormat &format)
37 {
38         output_format = format;
39 }
40
41 Node *EffectChain::add_node(Effect *effect)
42 {
43         char effect_id[256];
44         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
45
46         Node *node = new Node;
47         node->effect = effect;
48         node->disabled = false;
49         node->effect_id = effect_id;
50         node->output_color_space = COLORSPACE_INVALID;
51         node->output_gamma_curve = GAMMA_INVALID;
52
53         nodes.push_back(node);
54         node_map[effect] = node;
55         return node;
56 }
57
58 void EffectChain::connect_nodes(Node *sender, Node *receiver)
59 {
60         sender->outgoing_links.push_back(receiver);
61         receiver->incoming_links.push_back(sender);
62 }
63
64 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
65 {
66         new_receiver->incoming_links = old_receiver->incoming_links;
67         old_receiver->incoming_links.clear();
68         
69         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
70                 Node *sender = new_receiver->incoming_links[i];
71                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
72                         if (sender->outgoing_links[j] == old_receiver) {
73                                 sender->outgoing_links[j] = new_receiver;
74                         }
75                 }
76         }       
77 }
78
79 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
80 {
81         new_sender->outgoing_links = old_sender->outgoing_links;
82         old_sender->outgoing_links.clear();
83         
84         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
85                 Node *receiver = new_sender->outgoing_links[i];
86                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
87                         if (receiver->incoming_links[j] == old_sender) {
88                                 receiver->incoming_links[j] = new_sender;
89                         }
90                 }
91         }       
92 }
93
94 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
95 {
96         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
97                 if (sender->outgoing_links[i] == receiver) {
98                         sender->outgoing_links[i] = middle;
99                         middle->incoming_links.push_back(sender);
100                 }
101         }
102         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
103                 if (receiver->incoming_links[i] == sender) {
104                         receiver->incoming_links[i] = middle;
105                         middle->outgoing_links.push_back(receiver);
106                 }
107         }
108
109         assert(middle->incoming_links.size() == middle->effect->num_inputs());
110 }
111
112 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
113 {
114         if (node->output_gamma_curve == GAMMA_LINEAR &&
115             node->effect->effect_type_id() != "GammaCompressionEffect") {
116                 return;
117         }
118         if (node->effect->num_inputs() == 0) {
119                 nonlinear_inputs->push_back(node);
120         } else {
121                 assert(node->effect->num_inputs() == node->incoming_links.size());
122                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
123                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
124                 }
125         }
126 }
127
128 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
129 {
130         assert(inputs.size() == effect->num_inputs());
131         Node *node = add_node(effect);
132         for (unsigned i = 0; i < inputs.size(); ++i) {
133                 assert(node_map.count(inputs[i]) != 0);
134                 connect_nodes(node_map[inputs[i]], node);
135         }
136         return effect;
137 }
138
139 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
140 std::string replace_prefix(const std::string &text, const std::string &prefix)
141 {
142         std::string output;
143         size_t start = 0;
144
145         while (start < text.size()) {
146                 size_t pos = text.find("PREFIX(", start);
147                 if (pos == std::string::npos) {
148                         output.append(text.substr(start, std::string::npos));
149                         break;
150                 }
151
152                 output.append(text.substr(start, pos - start));
153                 output.append(prefix);
154                 output.append("_");
155
156                 pos += strlen("PREFIX(");
157         
158                 // Output stuff until we find the matching ), which we then eat.
159                 int depth = 1;
160                 size_t end_arg_pos = pos;
161                 while (end_arg_pos < text.size()) {
162                         if (text[end_arg_pos] == '(') {
163                                 ++depth;
164                         } else if (text[end_arg_pos] == ')') {
165                                 --depth;
166                                 if (depth == 0) {
167                                         break;
168                                 }
169                         }
170                         ++end_arg_pos;
171                 }
172                 output.append(text.substr(pos, end_arg_pos - pos));
173                 ++end_arg_pos;
174                 assert(depth == 0);
175                 start = end_arg_pos;
176         }
177         return output;
178 }
179
180 Phase *EffectChain::compile_glsl_program(
181         const std::vector<Node *> &inputs,
182         const std::vector<Node *> &effects)
183 {
184         assert(!effects.empty());
185
186         // Deduplicate the inputs.
187         std::vector<Node *> true_inputs = inputs;
188         std::sort(true_inputs.begin(), true_inputs.end());
189         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
190
191         bool input_needs_mipmaps = false;
192         std::string frag_shader = read_file("header.frag");
193
194         // Create functions for all the texture inputs that we need.
195         for (unsigned i = 0; i < true_inputs.size(); ++i) {
196                 Node *input = true_inputs[i];
197         
198                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
199                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
200                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
201                 frag_shader += "}\n";
202                 frag_shader += "\n";
203         }
204
205         for (unsigned i = 0; i < effects.size(); ++i) {
206                 Node *node = effects[i];
207
208                 if (node->incoming_links.size() == 1) {
209                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
210                 } else {
211                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
212                                 char buf[256];
213                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
214                                 frag_shader += buf;
215                         }
216                 }
217         
218                 frag_shader += "\n";
219                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
220                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
221                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
222                 frag_shader += "#undef PREFIX\n";
223                 frag_shader += "#undef FUNCNAME\n";
224                 if (node->incoming_links.size() == 1) {
225                         frag_shader += "#undef INPUT\n";
226                 } else {
227                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
228                                 char buf[256];
229                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
230                                 frag_shader += buf;
231                         }
232                 }
233                 frag_shader += "\n";
234
235                 input_needs_mipmaps |= node->effect->needs_mipmaps();
236         }
237         for (unsigned i = 0; i < effects.size(); ++i) {
238                 Node *node = effects[i];
239                 if (node->effect->num_inputs() == 0) {
240                         node->effect->set_int("needs_mipmaps", input_needs_mipmaps);
241                 }
242         }
243         frag_shader += std::string("#define INPUT ") + effects.back()->effect_id + "\n";
244         frag_shader.append(read_file("footer.frag"));
245
246         // Output shader to a temporary file, for easier debugging.
247         static int compiled_shader_num = 0;
248         char filename[256];
249         sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
250         FILE *fp = fopen(filename, "w");
251         if (fp == NULL) {
252                 perror(filename);
253                 exit(1);
254         }
255         fprintf(fp, "%s\n", frag_shader.c_str());
256         fclose(fp);
257         
258         GLuint glsl_program_num = glCreateProgram();
259         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
260         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
261         glAttachShader(glsl_program_num, vs_obj);
262         check_error();
263         glAttachShader(glsl_program_num, fs_obj);
264         check_error();
265         glLinkProgram(glsl_program_num);
266         check_error();
267
268         Phase *phase = new Phase;
269         phase->glsl_program_num = glsl_program_num;
270         phase->input_needs_mipmaps = input_needs_mipmaps;
271         phase->inputs = true_inputs;
272         phase->effects = effects;
273
274         return phase;
275 }
276
277 // Construct GLSL programs, starting at the given effect and following
278 // the chain from there. We end a program every time we come to an effect
279 // marked as "needs texture bounce", one that is used by multiple other
280 // effects, every time an effect wants to change the output size,
281 // and of course at the end.
282 //
283 // We follow a quite simple depth-first search from the output, although
284 // without any explicit recursion.
285 void EffectChain::construct_glsl_programs(Node *output)
286 {
287         // Which effects have already been completed in this phase?
288         // We need to keep track of it, as an effect with multiple outputs
289         // could otherwise be calculate multiple times.
290         std::set<Node *> completed_effects;
291
292         // Effects in the current phase, as well as inputs (outputs from other phases
293         // that we depend on). Note that since we start iterating from the end,
294         // the effect list will be in the reverse order.
295         std::vector<Node *> this_phase_inputs;
296         std::vector<Node *> this_phase_effects;
297
298         // Effects that we have yet to calculate, but that we know should
299         // be in the current phase.
300         std::stack<Node *> effects_todo_this_phase;
301
302         // Effects that we have yet to calculate, but that come from other phases.
303         // We delay these until we have this phase done in its entirety,
304         // at which point we pick any of them and start a new phase from that.
305         std::stack<Node *> effects_todo_other_phases;
306
307         effects_todo_this_phase.push(output);
308
309         for ( ;; ) {  // Termination condition within loop.
310                 if (!effects_todo_this_phase.empty()) {
311                         // OK, we have more to do this phase.
312                         Node *node = effects_todo_this_phase.top();
313                         effects_todo_this_phase.pop();
314
315                         // This should currently only happen for effects that are phase outputs,
316                         // and we throw those out separately below.
317                         assert(completed_effects.count(node) == 0);
318
319                         this_phase_effects.push_back(node);
320                         completed_effects.insert(node);
321
322                         // Find all the dependencies of this effect, and add them to the stack.
323                         std::vector<Node *> deps = node->incoming_links;
324                         assert(node->effect->num_inputs() == deps.size());
325                         for (unsigned i = 0; i < deps.size(); ++i) {
326                                 bool start_new_phase = false;
327
328                                 // FIXME: If we sample directly from a texture, we won't need this.
329                                 if (node->effect->needs_texture_bounce()) {
330                                         start_new_phase = true;
331                                 }
332
333                                 if (deps[i]->outgoing_links.size() > 1) {
334                                         if (deps[i]->effect->num_inputs() > 0) {
335                                                 // More than one effect uses this as the input,
336                                                 // and it is not a texture itself.
337                                                 // The easiest thing to do (and probably also the safest
338                                                 // performance-wise in most cases) is to bounce it to a texture
339                                                 // and then let the next passes read from that.
340                                                 start_new_phase = true;
341                                         } else {
342                                                 // For textures, we try to be slightly more clever;
343                                                 // if none of our outputs need a bounce, we don't bounce
344                                                 // but instead simply use the effect many times.
345                                                 //
346                                                 // Strictly speaking, we could bounce it for some outputs
347                                                 // and use it directly for others, but the processing becomes
348                                                 // somewhat simpler if the effect is only used in one such way.
349                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
350                                                         Node *rdep = deps[i]->outgoing_links[j];
351                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
352                                                 }
353                                         }
354                                 }
355
356                                 if (deps[i]->effect->changes_output_size()) {
357                                         start_new_phase = true;
358                                 }
359
360                                 if (start_new_phase) {
361                                         effects_todo_other_phases.push(deps[i]);
362                                         this_phase_inputs.push_back(deps[i]);
363                                 } else {
364                                         effects_todo_this_phase.push(deps[i]);
365                                 }
366                         }
367                         continue;
368                 }
369
370                 // No more effects to do this phase. Take all the ones we have,
371                 // and create a GLSL program for it.
372                 if (!this_phase_effects.empty()) {
373                         reverse(this_phase_effects.begin(), this_phase_effects.end());
374                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
375                         this_phase_effects.back()->phase = phases.back();
376                         this_phase_inputs.clear();
377                         this_phase_effects.clear();
378                 }
379                 assert(this_phase_inputs.empty());
380                 assert(this_phase_effects.empty());
381
382                 // If we have no effects left, exit.
383                 if (effects_todo_other_phases.empty()) {
384                         break;
385                 }
386
387                 Node *node = effects_todo_other_phases.top();
388                 effects_todo_other_phases.pop();
389
390                 if (completed_effects.count(node) == 0) {
391                         // Start a new phase, calculating from this effect.
392                         effects_todo_this_phase.push(node);
393                 }
394         }
395
396         // Finally, since the phases are found from the output but must be executed
397         // from the input(s), reverse them, too.
398         std::reverse(phases.begin(), phases.end());
399 }
400
401 void EffectChain::output_dot(const char *filename)
402 {
403         FILE *fp = fopen(filename, "w");
404         if (fp == NULL) {
405                 perror(filename);
406                 exit(1);
407         }
408
409         fprintf(fp, "digraph G {\n");
410         for (unsigned i = 0; i < nodes.size(); ++i) {
411                 // Find out which phase this event belongs to.
412                 int in_phase = -1;
413                 for (unsigned j = 0; j < phases.size(); ++j) {
414                         const Phase* p = phases[j];
415                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
416                                 assert(in_phase == -1);
417                                 in_phase = j;
418                         }
419                 }
420
421                 if (in_phase == -1) {
422                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
423                 } else {
424                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
425                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
426                                 (in_phase % 8) + 1);
427                 }
428                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
429                         std::vector<std::string> labels;
430
431                         if (nodes[i]->outgoing_links[j]->effect->needs_texture_bounce()) {
432                                 labels.push_back("needs_bounce");
433                         }
434                         if (nodes[i]->effect->changes_output_size()) {
435                                 labels.push_back("resize");
436                         }
437
438                         switch (nodes[i]->output_color_space) {
439                         case COLORSPACE_INVALID:
440                                 labels.push_back("spc[invalid]");
441                                 break;
442                         case COLORSPACE_REC_601_525:
443                                 labels.push_back("spc[rec601-525]");
444                                 break;
445                         case COLORSPACE_REC_601_625:
446                                 labels.push_back("spc[rec601-625]");
447                                 break;
448                         default:
449                                 break;
450                         }
451
452                         switch (nodes[i]->output_gamma_curve) {
453                         case GAMMA_INVALID:
454                                 labels.push_back("gamma[invalid]");
455                                 break;
456                         case GAMMA_sRGB:
457                                 labels.push_back("gamma[sRGB]");
458                                 break;
459                         case GAMMA_REC_601:  // and GAMMA_REC_709
460                                 labels.push_back("gamma[rec601/709]");
461                                 break;
462                         default:
463                                 break;
464                         }
465
466                         if (labels.empty()) {
467                                 fprintf(fp, "  n%ld -> n%ld;\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j]);
468                         } else {
469                                 std::string label = labels[0];
470                                 for (unsigned k = 1; k < labels.size(); ++k) {
471                                         label += ", " + labels[k];
472                                 }
473                                 fprintf(fp, "  n%ld -> n%ld [label=\"%s\"];\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j], label.c_str());
474                         }
475                 }
476         }
477         fprintf(fp, "}\n");
478
479         fclose(fp);
480 }
481
482 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
483 {
484         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
485                 // Same aspect, or W/H > aspect (image is wider than the frame).
486                 // In either case, keep width.
487                 return width;
488         } else {
489                 // W/H < aspect (image is taller than the frame), so keep height,
490                 // and adjust width correspondingly.
491                 return lrintf(height * aspect_nom / aspect_denom);
492         }
493 }
494
495 // Propagate input texture sizes throughout, and inform effects downstream.
496 // (Like a lot of other code, we depend on effects being in topological order.)
497 void EffectChain::inform_input_sizes(Phase *phase)
498 {
499         // All effects that have a defined size (inputs and RTT inputs)
500         // get that. Reset all others.
501         for (unsigned i = 0; i < phase->effects.size(); ++i) {
502                 Node *node = phase->effects[i];
503                 if (node->effect->num_inputs() == 0) {
504                         Input *input = static_cast<Input *>(node->effect);
505                         node->output_width = input->get_width();
506                         node->output_height = input->get_height();
507                         assert(node->output_width != 0);
508                         assert(node->output_height != 0);
509                 } else {
510                         node->output_width = node->output_height = 0;
511                 }
512         }
513         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
514                 Node *input = phase->inputs[i];
515                 input->output_width = input->phase->output_width;
516                 input->output_height = input->phase->output_height;
517                 assert(input->output_width != 0);
518                 assert(input->output_height != 0);
519         }
520
521         // Now propagate from the inputs towards the end, and inform as we go.
522         // The rules are simple:
523         //
524         //   1. Don't touch effects that already have given sizes (ie., inputs).
525         //   2. If all of your inputs have the same size, that will be your output size.
526         //   3. Otherwise, your output size is 0x0.
527         for (unsigned i = 0; i < phase->effects.size(); ++i) {
528                 Node *node = phase->effects[i];
529                 if (node->effect->num_inputs() == 0) {
530                         continue;
531                 }
532                 unsigned this_output_width = 0;
533                 unsigned this_output_height = 0;
534                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
535                         Node *input = node->incoming_links[j];
536                         node->effect->inform_input_size(j, input->output_width, input->output_height);
537                         if (j == 0) {
538                                 this_output_width = input->output_width;
539                                 this_output_height = input->output_height;
540                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
541                                 // Inputs disagree.
542                                 this_output_width = 0;
543                                 this_output_height = 0;
544                         }
545                 }
546                 node->output_width = this_output_width;
547                 node->output_height = this_output_height;
548         }
549 }
550
551 // Note: You should call inform_input_sizes() before this, as the last effect's
552 // desired output size might change based on the inputs.
553 void EffectChain::find_output_size(Phase *phase)
554 {
555         Node *output_node = phase->effects.back();
556
557         // If the last effect explicitly sets an output size, use that.
558         if (output_node->effect->changes_output_size()) {
559                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
560                 return;
561         }
562
563         // If not, look at the input phases and textures.
564         // We select the largest one (by fit into the current aspect).
565         unsigned best_width = 0;
566         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
567                 Node *input = phase->inputs[i];
568                 assert(input->phase->output_width != 0);
569                 assert(input->phase->output_height != 0);
570                 unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height);
571                 if (width > best_width) {
572                         best_width = width;
573                 }
574         }
575         for (unsigned i = 0; i < phase->effects.size(); ++i) {
576                 Effect *effect = phase->effects[i]->effect;
577                 if (effect->num_inputs() != 0) {
578                         continue;
579                 }
580
581                 Input *input = static_cast<Input *>(effect);
582                 unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height());
583                 if (width > best_width) {
584                         best_width = width;
585                 }
586         }
587         assert(best_width != 0);
588         phase->output_width = best_width;
589         phase->output_height = best_width * aspect_denom / aspect_nom;
590 }
591
592 void EffectChain::sort_nodes_topologically()
593 {
594         std::set<Node *> visited_nodes;
595         std::vector<Node *> sorted_list;
596         for (unsigned i = 0; i < nodes.size(); ++i) {
597                 if (nodes[i]->incoming_links.size() == 0) {
598                         topological_sort_visit_node(nodes[i], &visited_nodes, &sorted_list);
599                 }
600         }
601         reverse(sorted_list.begin(), sorted_list.end());
602         nodes = sorted_list;
603 }
604
605 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *visited_nodes, std::vector<Node *> *sorted_list)
606 {
607         if (visited_nodes->count(node) != 0) {
608                 return;
609         }
610         visited_nodes->insert(node);
611         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
612                 topological_sort_visit_node(node->outgoing_links[i], visited_nodes, sorted_list);
613         }
614         sorted_list->push_back(node);
615 }
616
617 // Propagate gamma and color space information as far as we can in the graph.
618 // The rules are simple: Anything where all the inputs agree, get that as
619 // output as well. Anything else keeps having *_INVALID.
620 void EffectChain::propagate_gamma_and_color_space()
621 {
622         // We depend on going through the nodes in order.
623         sort_nodes_topologically();
624
625         for (unsigned i = 0; i < nodes.size(); ++i) {
626                 Node *node = nodes[i];
627                 if (node->disabled) {
628                         continue;
629                 }
630                 assert(node->incoming_links.size() == node->effect->num_inputs());
631                 if (node->incoming_links.size() == 0) {
632                         assert(node->output_color_space != COLORSPACE_INVALID);
633                         assert(node->output_gamma_curve != GAMMA_INVALID);
634                         continue;
635                 }
636
637                 ColorSpace color_space = node->incoming_links[0]->output_color_space;
638                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
639                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
640                         if (node->incoming_links[j]->output_color_space != color_space) {
641                                 color_space = COLORSPACE_INVALID;
642                         }
643                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
644                                 gamma_curve = GAMMA_INVALID;
645                         }
646                 }
647
648                 // The conversion effects already have their outputs set correctly,
649                 // so leave them alone.
650                 if (node->effect->effect_type_id() != "ColorSpaceConversionEffect") {
651                         node->output_color_space = color_space;
652                 }               
653                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
654                     node->effect->effect_type_id() != "GammaExpansionEffect") {
655                         node->output_gamma_curve = gamma_curve;
656                 }               
657         }
658 }
659
660 bool EffectChain::node_needs_colorspace_fix(Node *node)
661 {
662         if (node->disabled) {
663                 return false;
664         }
665         if (node->effect->num_inputs() == 0) {
666                 return false;
667         }
668
669         // propagate_gamma_and_color_space() has already set our output
670         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
671         if (node->output_color_space == COLORSPACE_INVALID) {
672                 return true;
673         }
674         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
675 }
676
677 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
678 // the graph. Our strategy is not always optimal, but quite simple:
679 // Find an effect that's as early as possible where the inputs are of
680 // unacceptable colorspaces (that is, either different, or, if the effect only
681 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
682 // propagate the information anew, and repeat until there are no more such
683 // effects.
684 void EffectChain::fix_internal_color_spaces()
685 {
686         unsigned colorspace_propagation_pass = 0;
687         bool found_any;
688         do {
689                 found_any = false;
690                 for (unsigned i = 0; i < nodes.size(); ++i) {
691                         Node *node = nodes[i];
692                         if (!node_needs_colorspace_fix(node)) {
693                                 continue;
694                         }
695
696                         // Go through each input that is not sRGB, and insert
697                         // a colorspace conversion before it.
698                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
699                                 Node *input = node->incoming_links[j];
700                                 assert(input->output_color_space != COLORSPACE_INVALID);
701                                 if (input->output_color_space == COLORSPACE_sRGB) {
702                                         continue;
703                                 }
704                                 Node *conversion = add_node(new ColorSpaceConversionEffect());
705                                 conversion->effect->set_int("source_space", input->output_color_space);
706                                 conversion->effect->set_int("destination_space", COLORSPACE_sRGB);
707                                 conversion->output_color_space = COLORSPACE_sRGB;
708                                 insert_node_between(input, conversion, node);
709                         }
710
711                         // Re-sort topologically, and propagate the new information.
712                         propagate_gamma_and_color_space();
713                         
714                         found_any = true;
715                         break;
716                 }
717         
718                 char filename[256];
719                 sprintf(filename, "step3-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
720                 output_dot(filename);
721                 assert(colorspace_propagation_pass < 100);
722         } while (found_any);
723
724         for (unsigned i = 0; i < nodes.size(); ++i) {
725                 Node *node = nodes[i];
726                 if (node->disabled) {
727                         continue;
728                 }
729                 assert(node->output_color_space != COLORSPACE_INVALID);
730         }
731 }
732
733 // Make so that the output is in the desired color space.
734 void EffectChain::fix_output_color_space()
735 {
736         Node *output = find_output_node();
737         if (output->output_color_space != output_format.color_space) {
738                 Node *conversion = add_node(new ColorSpaceConversionEffect());
739                 conversion->effect->set_int("source_space", output->output_color_space);
740                 conversion->effect->set_int("destination_space", output_format.color_space);
741                 conversion->output_color_space = output_format.color_space;
742                 connect_nodes(output, conversion);
743         }
744 }
745
746 bool EffectChain::node_needs_gamma_fix(Node *node)
747 {
748         if (node->disabled) {
749                 return false;
750         }
751
752         // Small hack since the output is not an explicit node:
753         // If we are the last node and our output is in the wrong
754         // space compared to EffectChain's output, we need to fix it.
755         // This will only take us to linear, but fix_output_gamma()
756         // will come and take us to the desired output gamma
757         // if it is needed.
758         //
759         // This needs to be before everything else, since it could
760         // even apply to inputs (if they are the only effect).
761         if (node->outgoing_links.empty() &&
762             node->output_gamma_curve != output_format.gamma_curve &&
763             node->output_gamma_curve != GAMMA_LINEAR) {
764                 return true;
765         }
766
767         if (node->effect->num_inputs() == 0) {
768                 return false;
769         }
770
771         // propagate_gamma_and_color_space() has already set our output
772         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
773         // except for GammaCompressionEffect.
774         if (node->output_gamma_curve == GAMMA_INVALID) {
775                 return true;
776         }
777         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
778                 assert(node->incoming_links.size() == 1);
779                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
780         }
781
782         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
783 }
784
785 // Very similar to fix_internal_color_spaces(), but for gamma.
786 // There is one difference, though; before we start adding conversion nodes,
787 // we see if we can get anything out of asking the sources to deliver
788 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
789 // does that part, while fix_internal_gamma_by_inserting_nodes()
790 // inserts nodes as needed afterwards.
791 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
792 {
793         unsigned gamma_propagation_pass = 0;
794         bool found_any;
795         do {
796                 found_any = false;
797                 for (unsigned i = 0; i < nodes.size(); ++i) {
798                         Node *node = nodes[i];
799                         if (!node_needs_gamma_fix(node)) {
800                                 continue;
801                         }
802
803                         // See if all inputs can give us linear gamma. If not, leave it.
804                         std::vector<Node *> nonlinear_inputs;
805                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
806                         assert(!nonlinear_inputs.empty());
807
808                         bool all_ok = true;
809                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
810                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
811                                 all_ok &= input->can_output_linear_gamma();
812                         }
813
814                         if (!all_ok) {
815                                 continue;
816                         }
817
818                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
819                                 nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1);
820                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
821                         }
822
823                         // Re-sort topologically, and propagate the new information.
824                         propagate_gamma_and_color_space();
825                         
826                         found_any = true;
827                         break;
828                 }
829         
830                 char filename[256];
831                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
832                 output_dot(filename);
833                 assert(gamma_propagation_pass < 100);
834         } while (found_any);
835 }
836
837 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
838 {
839         unsigned gamma_propagation_pass = 0;
840         bool found_any;
841         do {
842                 found_any = false;
843                 for (unsigned i = 0; i < nodes.size(); ++i) {
844                         Node *node = nodes[i];
845                         if (!node_needs_gamma_fix(node)) {
846                                 continue;
847                         }
848
849                         // Special case: We could be an input and still be asked to
850                         // fix our gamma; if so, we should be the only node
851                         // (as node_needs_gamma_fix() would only return true in
852                         // for an input in that case). That means we should insert
853                         // a conversion node _after_ ourselves.
854                         if (node->incoming_links.empty()) {
855                                 assert(node->outgoing_links.empty());
856                                 Node *conversion = add_node(new GammaExpansionEffect());
857                                 conversion->effect->set_int("source_curve", node->output_gamma_curve);
858                                 conversion->output_gamma_curve = GAMMA_LINEAR;
859                                 connect_nodes(node, conversion);
860                         }
861
862                         // If not, go through each input that is not linear gamma,
863                         // and insert a gamma conversion before it.
864                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
865                                 Node *input = node->incoming_links[j];
866                                 assert(input->output_gamma_curve != GAMMA_INVALID);
867                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
868                                         continue;
869                                 }
870                                 Node *conversion = add_node(new GammaExpansionEffect());
871                                 conversion->effect->set_int("source_curve", input->output_gamma_curve);
872                                 conversion->output_gamma_curve = GAMMA_LINEAR;
873                                 insert_node_between(input, conversion, node);
874                         }
875
876                         // Re-sort topologically, and propagate the new information.
877                         propagate_gamma_and_color_space();
878                         
879                         found_any = true;
880                         break;
881                 }
882         
883                 char filename[256];
884                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
885                 output_dot(filename);
886                 assert(gamma_propagation_pass < 100);
887         } while (found_any);
888
889         for (unsigned i = 0; i < nodes.size(); ++i) {
890                 Node *node = nodes[i];
891                 if (node->disabled) {
892                         continue;
893                 }
894                 assert(node->output_gamma_curve != GAMMA_INVALID);
895         }
896 }
897
898 // Make so that the output is in the desired gamma.
899 // Note that this assumes linear input gamma, so it might create the need
900 // for another pass of fix_internal_gamma().
901 void EffectChain::fix_output_gamma()
902 {
903         Node *output = find_output_node();
904         if (output->output_gamma_curve != output_format.gamma_curve) {
905                 Node *conversion = add_node(new GammaCompressionEffect());
906                 conversion->effect->set_int("destination_curve", output_format.gamma_curve);
907                 conversion->output_gamma_curve = output_format.gamma_curve;
908                 connect_nodes(output, conversion);
909         }
910 }
911
912 // Find the output node. This is, simply, one that has no outgoing links.
913 // If there are multiple ones, the graph is malformed (we do not support
914 // multiple outputs right now).
915 Node *EffectChain::find_output_node()
916 {
917         std::vector<Node *> output_nodes;
918         for (unsigned i = 0; i < nodes.size(); ++i) {
919                 Node *node = nodes[i];
920                 if (node->disabled) {
921                         continue;
922                 }
923                 if (node->outgoing_links.empty()) {
924                         output_nodes.push_back(node);
925                 }
926         }
927         assert(output_nodes.size() == 1);
928         return output_nodes[0];
929 }
930
931 void EffectChain::finalize()
932 {
933         // Output the graph as it is before we do any conversions on it.
934         output_dot("step0-start.dot");
935
936         // Give each effect in turn a chance to rewrite its own part of the graph.
937         // Note that if more effects are added as part of this, they will be
938         // picked up as part of the same for loop, since they are added at the end.
939         for (unsigned i = 0; i < nodes.size(); ++i) {
940                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
941         }
942         output_dot("step1-rewritten.dot");
943
944         propagate_gamma_and_color_space();
945         output_dot("step2-propagated.dot");
946
947         fix_internal_color_spaces();
948         fix_output_color_space();
949         output_dot("step4-output-colorspacefix.dot");
950
951         // Note that we need to fix gamma after colorspace conversion,
952         // because colorspace conversions might create needs for gamma conversions.
953         // Also, we need to run an extra pass of fix_internal_gamma() after 
954         // fixing the output gamma, as we only have conversions to/from linear.
955         fix_internal_gamma_by_asking_inputs(5);
956         fix_internal_gamma_by_inserting_nodes(6);
957         fix_output_gamma();
958         output_dot("step7-output-gammafix.dot");
959         fix_internal_gamma_by_asking_inputs(8);
960         fix_internal_gamma_by_inserting_nodes(9);
961
962         output_dot("step10-final.dot");
963         
964         // Construct all needed GLSL programs, starting at the output.
965         construct_glsl_programs(find_output_node());
966
967         output_dot("step11-split-to-phases.dot");
968
969         // If we have more than one phase, we need intermediate render-to-texture.
970         // Construct an FBO, and then as many textures as we need.
971         // We choose the simplest option of having one texture per output,
972         // since otherwise this turns into an (albeit simple)
973         // register allocation problem.
974         if (phases.size() > 1) {
975                 glGenFramebuffers(1, &fbo);
976
977                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
978                         inform_input_sizes(phases[i]);
979                         find_output_size(phases[i]);
980
981                         Node *output_node = phases[i]->effects.back();
982                         glGenTextures(1, &output_node->output_texture);
983                         check_error();
984                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
985                         check_error();
986                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
987                         check_error();
988                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
989                         check_error();
990                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
991                         check_error();
992
993                         output_node->output_texture_width = phases[i]->output_width;
994                         output_node->output_texture_height = phases[i]->output_height;
995                 }
996                 inform_input_sizes(phases.back());
997         }
998                 
999         for (unsigned i = 0; i < inputs.size(); ++i) {
1000                 inputs[i]->finalize();
1001         }
1002
1003         assert(phases[0]->inputs.empty());
1004         
1005         finalized = true;
1006 }
1007
1008 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1009 {
1010         assert(finalized);
1011
1012         // Save original viewport.
1013         GLuint x = 0, y = 0;
1014
1015         if (width == 0 && height == 0) {
1016                 GLint viewport[4];
1017                 glGetIntegerv(GL_VIEWPORT, viewport);
1018                 x = viewport[0];
1019                 y = viewport[1];
1020                 width = viewport[2];
1021                 height = viewport[3];
1022         }
1023
1024         // Basic state.
1025         glDisable(GL_BLEND);
1026         check_error();
1027         glDisable(GL_DEPTH_TEST);
1028         check_error();
1029         glDepthMask(GL_FALSE);
1030         check_error();
1031
1032         glMatrixMode(GL_PROJECTION);
1033         glLoadIdentity();
1034         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1035
1036         glMatrixMode(GL_MODELVIEW);
1037         glLoadIdentity();
1038
1039         if (phases.size() > 1) {
1040                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1041                 check_error();
1042         }
1043
1044         std::set<Node *> generated_mipmaps;
1045
1046         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1047                 // See if the requested output size has changed. If so, we need to recreate
1048                 // the texture (and before we start setting up inputs).
1049                 inform_input_sizes(phases[phase]);
1050                 if (phase != phases.size() - 1) {
1051                         find_output_size(phases[phase]);
1052
1053                         Node *output_node = phases[phase]->effects.back();
1054
1055                         if (output_node->output_texture_width != phases[phase]->output_width ||
1056                             output_node->output_texture_height != phases[phase]->output_height) {
1057                                 glActiveTexture(GL_TEXTURE0);
1058                                 check_error();
1059                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1060                                 check_error();
1061                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1062                                 check_error();
1063                                 glBindTexture(GL_TEXTURE_2D, 0);
1064                                 check_error();
1065
1066                                 output_node->output_texture_width = phases[phase]->output_width;
1067                                 output_node->output_texture_height = phases[phase]->output_height;
1068                         }
1069                 }
1070
1071                 glUseProgram(phases[phase]->glsl_program_num);
1072                 check_error();
1073
1074                 // Set up RTT inputs for this phase.
1075                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1076                         glActiveTexture(GL_TEXTURE0 + sampler);
1077                         Node *input = phases[phase]->inputs[sampler];
1078                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1079                         check_error();
1080                         if (phases[phase]->input_needs_mipmaps) {
1081                                 if (generated_mipmaps.count(input) == 0) {
1082                                         glGenerateMipmap(GL_TEXTURE_2D);
1083                                         check_error();
1084                                         generated_mipmaps.insert(input);
1085                                 }
1086                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1087                                 check_error();
1088                         } else {
1089                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1090                                 check_error();
1091                         }
1092
1093                         std::string texture_name = std::string("tex_") + input->effect_id;
1094                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1095                         check_error();
1096                 }
1097
1098                 // And now the output.
1099                 if (phase == phases.size() - 1) {
1100                         // Last phase goes to the output the user specified.
1101                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1102                         check_error();
1103                         glViewport(x, y, width, height);
1104                 } else {
1105                         Node *output_node = phases[phase]->effects.back();
1106                         glFramebufferTexture2D(
1107                                 GL_FRAMEBUFFER,
1108                                 GL_COLOR_ATTACHMENT0,
1109                                 GL_TEXTURE_2D,
1110                                 output_node->output_texture,
1111                                 0);
1112                         check_error();
1113                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1114                 }
1115
1116                 // Give the required parameters to all the effects.
1117                 unsigned sampler_num = phases[phase]->inputs.size();
1118                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1119                         Node *node = phases[phase]->effects[i];
1120                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1121                         check_error();
1122                 }
1123
1124                 // Now draw!
1125                 glBegin(GL_QUADS);
1126
1127                 glTexCoord2f(0.0f, 0.0f);
1128                 glVertex2f(0.0f, 0.0f);
1129
1130                 glTexCoord2f(1.0f, 0.0f);
1131                 glVertex2f(1.0f, 0.0f);
1132
1133                 glTexCoord2f(1.0f, 1.0f);
1134                 glVertex2f(1.0f, 1.0f);
1135
1136                 glTexCoord2f(0.0f, 1.0f);
1137                 glVertex2f(0.0f, 1.0f);
1138
1139                 glEnd();
1140                 check_error();
1141
1142                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1143                         Node *node = phases[phase]->effects[i];
1144                         node->effect->clear_gl_state();
1145                 }
1146         }
1147 }