Output the chained shaders to temporary files instead of to stdout, since it's near...
[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 && deps[i]->effect->num_inputs() > 0) {
334                                         // More than one effect uses this as the input,
335                                         // and it is not a texture itself.
336                                         // The easiest thing to do (and probably also the safest
337                                         // performance-wise in most cases) is to bounce it to a texture
338                                         // and then let the next passes read from that.
339                                         start_new_phase = true;
340                                 }
341
342                                 if (deps[i]->effect->changes_output_size()) {
343                                         start_new_phase = true;
344                                 }
345
346                                 if (start_new_phase) {
347                                         effects_todo_other_phases.push(deps[i]);
348                                         this_phase_inputs.push_back(deps[i]);
349                                 } else {
350                                         effects_todo_this_phase.push(deps[i]);
351                                 }
352                         }
353                         continue;
354                 }
355
356                 // No more effects to do this phase. Take all the ones we have,
357                 // and create a GLSL program for it.
358                 if (!this_phase_effects.empty()) {
359                         reverse(this_phase_effects.begin(), this_phase_effects.end());
360                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
361                         this_phase_effects.back()->phase = phases.back();
362                         this_phase_inputs.clear();
363                         this_phase_effects.clear();
364                 }
365                 assert(this_phase_inputs.empty());
366                 assert(this_phase_effects.empty());
367
368                 // If we have no effects left, exit.
369                 if (effects_todo_other_phases.empty()) {
370                         break;
371                 }
372
373                 Node *node = effects_todo_other_phases.top();
374                 effects_todo_other_phases.pop();
375
376                 if (completed_effects.count(node) == 0) {
377                         // Start a new phase, calculating from this effect.
378                         effects_todo_this_phase.push(node);
379                 }
380         }
381
382         // Finally, since the phases are found from the output but must be executed
383         // from the input(s), reverse them, too.
384         std::reverse(phases.begin(), phases.end());
385 }
386
387 void EffectChain::output_dot(const char *filename)
388 {
389         FILE *fp = fopen(filename, "w");
390         if (fp == NULL) {
391                 perror(filename);
392                 exit(1);
393         }
394
395         fprintf(fp, "digraph G {\n");
396         for (unsigned i = 0; i < nodes.size(); ++i) {
397                 fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
398                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
399                         std::vector<std::string> labels;
400
401                         if (nodes[i]->outgoing_links[j]->effect->needs_texture_bounce()) {
402                                 labels.push_back("needs_bounce");
403                         }
404                         if (nodes[i]->effect->changes_output_size()) {
405                                 labels.push_back("resize");
406                         }
407
408                         switch (nodes[i]->output_color_space) {
409                         case COLORSPACE_INVALID:
410                                 labels.push_back("spc[invalid]");
411                                 break;
412                         case COLORSPACE_REC_601_525:
413                                 labels.push_back("spc[rec601-525]");
414                                 break;
415                         case COLORSPACE_REC_601_625:
416                                 labels.push_back("spc[rec601-625]");
417                                 break;
418                         default:
419                                 break;
420                         }
421
422                         switch (nodes[i]->output_gamma_curve) {
423                         case GAMMA_INVALID:
424                                 labels.push_back("gamma[invalid]");
425                                 break;
426                         case GAMMA_sRGB:
427                                 labels.push_back("gamma[sRGB]");
428                                 break;
429                         case GAMMA_REC_601:  // and GAMMA_REC_709
430                                 labels.push_back("gamma[rec601/709]");
431                                 break;
432                         default:
433                                 break;
434                         }
435
436                         if (labels.empty()) {
437                                 fprintf(fp, "  n%ld -> n%ld;\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j]);
438                         } else {
439                                 std::string label = labels[0];
440                                 for (unsigned k = 1; k < labels.size(); ++k) {
441                                         label += ", " + labels[k];
442                                 }
443                                 fprintf(fp, "  n%ld -> n%ld [label=\"%s\"];\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j], label.c_str());
444                         }
445                 }
446         }
447         fprintf(fp, "}\n");
448
449         fclose(fp);
450 }
451
452 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
453 {
454         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
455                 // Same aspect, or W/H > aspect (image is wider than the frame).
456                 // In either case, keep width.
457                 return width;
458         } else {
459                 // W/H < aspect (image is taller than the frame), so keep height,
460                 // and adjust width correspondingly.
461                 return lrintf(height * aspect_nom / aspect_denom);
462         }
463 }
464
465 // Propagate input texture sizes throughout, and inform effects downstream.
466 // (Like a lot of other code, we depend on effects being in topological order.)
467 void EffectChain::inform_input_sizes(Phase *phase)
468 {
469         // All effects that have a defined size (inputs and RTT inputs)
470         // get that. Reset all others.
471         for (unsigned i = 0; i < phase->effects.size(); ++i) {
472                 Node *node = phase->effects[i];
473                 if (node->effect->num_inputs() == 0) {
474                         Input *input = static_cast<Input *>(node->effect);
475                         node->output_width = input->get_width();
476                         node->output_height = input->get_height();
477                         assert(node->output_width != 0);
478                         assert(node->output_height != 0);
479                 } else {
480                         node->output_width = node->output_height = 0;
481                 }
482         }
483         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
484                 Node *input = phase->inputs[i];
485                 input->output_width = input->phase->output_width;
486                 input->output_height = input->phase->output_height;
487                 assert(input->output_width != 0);
488                 assert(input->output_height != 0);
489         }
490
491         // Now propagate from the inputs towards the end, and inform as we go.
492         // The rules are simple:
493         //
494         //   1. Don't touch effects that already have given sizes (ie., inputs).
495         //   2. If all of your inputs have the same size, that will be your output size.
496         //   3. Otherwise, your output size is 0x0.
497         for (unsigned i = 0; i < phase->effects.size(); ++i) {
498                 Node *node = phase->effects[i];
499                 if (node->effect->num_inputs() == 0) {
500                         continue;
501                 }
502                 unsigned this_output_width = 0;
503                 unsigned this_output_height = 0;
504                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
505                         Node *input = node->incoming_links[j];
506                         node->effect->inform_input_size(j, input->output_width, input->output_height);
507                         if (j == 0) {
508                                 this_output_width = input->output_width;
509                                 this_output_height = input->output_height;
510                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
511                                 // Inputs disagree.
512                                 this_output_width = 0;
513                                 this_output_height = 0;
514                         }
515                 }
516                 node->output_width = this_output_width;
517                 node->output_height = this_output_height;
518         }
519 }
520
521 // Note: You should call inform_input_sizes() before this, as the last effect's
522 // desired output size might change based on the inputs.
523 void EffectChain::find_output_size(Phase *phase)
524 {
525         Node *output_node = phase->effects.back();
526
527         // If the last effect explicitly sets an output size, use that.
528         if (output_node->effect->changes_output_size()) {
529                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
530                 return;
531         }
532
533         // If not, look at the input phases and textures.
534         // We select the largest one (by fit into the current aspect).
535         unsigned best_width = 0;
536         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
537                 Node *input = phase->inputs[i];
538                 assert(input->phase->output_width != 0);
539                 assert(input->phase->output_height != 0);
540                 unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height);
541                 if (width > best_width) {
542                         best_width = width;
543                 }
544         }
545         for (unsigned i = 0; i < phase->effects.size(); ++i) {
546                 Effect *effect = phase->effects[i]->effect;
547                 if (effect->num_inputs() != 0) {
548                         continue;
549                 }
550
551                 Input *input = static_cast<Input *>(effect);
552                 unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height());
553                 if (width > best_width) {
554                         best_width = width;
555                 }
556         }
557         assert(best_width != 0);
558         phase->output_width = best_width;
559         phase->output_height = best_width * aspect_denom / aspect_nom;
560 }
561
562 void EffectChain::sort_nodes_topologically()
563 {
564         std::set<Node *> visited_nodes;
565         std::vector<Node *> sorted_list;
566         for (unsigned i = 0; i < nodes.size(); ++i) {
567                 if (nodes[i]->incoming_links.size() == 0) {
568                         topological_sort_visit_node(nodes[i], &visited_nodes, &sorted_list);
569                 }
570         }
571         reverse(sorted_list.begin(), sorted_list.end());
572         nodes = sorted_list;
573 }
574
575 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *visited_nodes, std::vector<Node *> *sorted_list)
576 {
577         if (visited_nodes->count(node) != 0) {
578                 return;
579         }
580         visited_nodes->insert(node);
581         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
582                 topological_sort_visit_node(node->outgoing_links[i], visited_nodes, sorted_list);
583         }
584         sorted_list->push_back(node);
585 }
586
587 // Propagate gamma and color space information as far as we can in the graph.
588 // The rules are simple: Anything where all the inputs agree, get that as
589 // output as well. Anything else keeps having *_INVALID.
590 void EffectChain::propagate_gamma_and_color_space()
591 {
592         // We depend on going through the nodes in order.
593         sort_nodes_topologically();
594
595         for (unsigned i = 0; i < nodes.size(); ++i) {
596                 Node *node = nodes[i];
597                 if (node->disabled) {
598                         continue;
599                 }
600                 assert(node->incoming_links.size() == node->effect->num_inputs());
601                 if (node->incoming_links.size() == 0) {
602                         assert(node->output_color_space != COLORSPACE_INVALID);
603                         assert(node->output_gamma_curve != GAMMA_INVALID);
604                         continue;
605                 }
606
607                 ColorSpace color_space = node->incoming_links[0]->output_color_space;
608                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
609                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
610                         if (node->incoming_links[j]->output_color_space != color_space) {
611                                 color_space = COLORSPACE_INVALID;
612                         }
613                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
614                                 gamma_curve = GAMMA_INVALID;
615                         }
616                 }
617
618                 // The conversion effects already have their outputs set correctly,
619                 // so leave them alone.
620                 if (node->effect->effect_type_id() != "ColorSpaceConversionEffect") {
621                         node->output_color_space = color_space;
622                 }               
623                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
624                     node->effect->effect_type_id() != "GammaExpansionEffect") {
625                         node->output_gamma_curve = gamma_curve;
626                 }               
627         }
628 }
629
630 bool EffectChain::node_needs_colorspace_fix(Node *node)
631 {
632         if (node->disabled) {
633                 return false;
634         }
635         if (node->effect->num_inputs() == 0) {
636                 return false;
637         }
638
639         // propagate_gamma_and_color_space() has already set our output
640         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
641         if (node->output_color_space == COLORSPACE_INVALID) {
642                 return true;
643         }
644         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
645 }
646
647 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
648 // the graph. Our strategy is not always optimal, but quite simple:
649 // Find an effect that's as early as possible where the inputs are of
650 // unacceptable colorspaces (that is, either different, or, if the effect only
651 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
652 // propagate the information anew, and repeat until there are no more such
653 // effects.
654 void EffectChain::fix_internal_color_spaces()
655 {
656         unsigned colorspace_propagation_pass = 0;
657         bool found_any;
658         do {
659                 found_any = false;
660                 for (unsigned i = 0; i < nodes.size(); ++i) {
661                         Node *node = nodes[i];
662                         if (!node_needs_colorspace_fix(node)) {
663                                 continue;
664                         }
665
666                         // Go through each input that is not sRGB, and insert
667                         // a colorspace conversion before it.
668                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
669                                 Node *input = node->incoming_links[j];
670                                 assert(input->output_color_space != COLORSPACE_INVALID);
671                                 if (input->output_color_space == COLORSPACE_sRGB) {
672                                         continue;
673                                 }
674                                 Node *conversion = add_node(new ColorSpaceConversionEffect());
675                                 conversion->effect->set_int("source_space", input->output_color_space);
676                                 conversion->effect->set_int("destination_space", COLORSPACE_sRGB);
677                                 conversion->output_color_space = COLORSPACE_sRGB;
678                                 insert_node_between(input, conversion, node);
679                         }
680
681                         // Re-sort topologically, and propagate the new information.
682                         propagate_gamma_and_color_space();
683                         
684                         found_any = true;
685                         break;
686                 }
687         
688                 char filename[256];
689                 sprintf(filename, "step3-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
690                 output_dot(filename);
691                 assert(colorspace_propagation_pass < 100);
692         } while (found_any);
693
694         for (unsigned i = 0; i < nodes.size(); ++i) {
695                 Node *node = nodes[i];
696                 if (node->disabled) {
697                         continue;
698                 }
699                 assert(node->output_color_space != COLORSPACE_INVALID);
700         }
701 }
702
703 // Make so that the output is in the desired color space.
704 void EffectChain::fix_output_color_space()
705 {
706         Node *output = find_output_node();
707         if (output->output_color_space != output_format.color_space) {
708                 Node *conversion = add_node(new ColorSpaceConversionEffect());
709                 conversion->effect->set_int("source_space", output->output_color_space);
710                 conversion->effect->set_int("destination_space", output_format.color_space);
711                 conversion->output_color_space = output_format.color_space;
712                 connect_nodes(output, conversion);
713         }
714 }
715
716 bool EffectChain::node_needs_gamma_fix(Node *node)
717 {
718         if (node->disabled) {
719                 return false;
720         }
721
722         // Small hack since the output is not an explicit node:
723         // If we are the last node and our output is in the wrong
724         // space compared to EffectChain's output, we need to fix it.
725         // This will only take us to linear, but fix_output_gamma()
726         // will come and take us to the desired output gamma
727         // if it is needed.
728         //
729         // This needs to be before everything else, since it could
730         // even apply to inputs (if they are the only effect).
731         if (node->outgoing_links.empty() &&
732             node->output_gamma_curve != output_format.gamma_curve) {
733                 return true;
734         }
735
736         if (node->effect->num_inputs() == 0) {
737                 return false;
738         }
739
740         // propagate_gamma_and_color_space() has already set our output
741         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
742         // except for GammaCompressionEffect.
743         if (node->output_gamma_curve == GAMMA_INVALID) {
744                 return true;
745         }
746         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
747                 assert(node->incoming_links.size() == 1);
748                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
749         }
750
751         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
752 }
753
754 // Very similar to fix_internal_color_spaces(), but for gamma.
755 // There is one difference, though; before we start adding conversion nodes,
756 // we see if we can get anything out of asking the sources to deliver
757 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
758 // does that part, while fix_internal_gamma_by_inserting_nodes()
759 // inserts nodes as needed afterwards.
760 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
761 {
762         unsigned gamma_propagation_pass = 0;
763         bool found_any;
764         do {
765                 found_any = false;
766                 for (unsigned i = 0; i < nodes.size(); ++i) {
767                         Node *node = nodes[i];
768                         if (!node_needs_gamma_fix(node)) {
769                                 continue;
770                         }
771
772                         // See if all inputs can give us linear gamma. If not, leave it.
773                         std::vector<Node *> nonlinear_inputs;
774                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
775                         assert(!nonlinear_inputs.empty());
776
777                         bool all_ok = true;
778                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
779                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
780                                 all_ok &= input->can_output_linear_gamma();
781                         }
782
783                         if (!all_ok) {
784                                 continue;
785                         }
786
787                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
788                                 nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1);
789                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
790                         }
791
792                         // Re-sort topologically, and propagate the new information.
793                         propagate_gamma_and_color_space();
794                         
795                         found_any = true;
796                         break;
797                 }
798         
799                 char filename[256];
800                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
801                 output_dot(filename);
802                 assert(gamma_propagation_pass < 100);
803         } while (found_any);
804 }
805
806 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
807 {
808         unsigned gamma_propagation_pass = 0;
809         bool found_any;
810         do {
811                 found_any = false;
812                 for (unsigned i = 0; i < nodes.size(); ++i) {
813                         Node *node = nodes[i];
814                         if (!node_needs_gamma_fix(node)) {
815                                 continue;
816                         }
817
818                         // Special case: We could be an input and still be asked to
819                         // fix our gamma; if so, we should be the only node
820                         // (as node_needs_gamma_fix() would only return true in
821                         // for an input in that case). That means we should insert
822                         // a conversion node _after_ ourselves.
823                         if (node->incoming_links.empty()) {
824                                 assert(node->outgoing_links.empty());
825                                 Node *conversion = add_node(new GammaExpansionEffect());
826                                 conversion->effect->set_int("source_curve", node->output_gamma_curve);
827                                 conversion->output_gamma_curve = GAMMA_LINEAR;
828                                 connect_nodes(node, conversion);
829                         }
830
831                         // If not, go through each input that is not linear gamma,
832                         // and insert a gamma conversion before it.
833                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
834                                 Node *input = node->incoming_links[j];
835                                 assert(input->output_gamma_curve != GAMMA_INVALID);
836                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
837                                         continue;
838                                 }
839                                 Node *conversion = add_node(new GammaExpansionEffect());
840                                 conversion->effect->set_int("source_curve", input->output_gamma_curve);
841                                 conversion->output_gamma_curve = GAMMA_LINEAR;
842                                 insert_node_between(input, conversion, node);
843                         }
844
845                         // Re-sort topologically, and propagate the new information.
846                         propagate_gamma_and_color_space();
847                         
848                         found_any = true;
849                         break;
850                 }
851         
852                 char filename[256];
853                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
854                 output_dot(filename);
855                 assert(gamma_propagation_pass < 100);
856         } while (found_any);
857
858         for (unsigned i = 0; i < nodes.size(); ++i) {
859                 Node *node = nodes[i];
860                 if (node->disabled) {
861                         continue;
862                 }
863                 assert(node->output_gamma_curve != GAMMA_INVALID);
864         }
865 }
866
867 // Make so that the output is in the desired gamma.
868 // Note that this assumes linear input gamma, so it might create the need
869 // for another pass of fix_internal_gamma().
870 void EffectChain::fix_output_gamma()
871 {
872         Node *output = find_output_node();
873         if (output->output_gamma_curve != output_format.gamma_curve) {
874                 Node *conversion = add_node(new GammaCompressionEffect());
875                 conversion->effect->set_int("destination_curve", output_format.gamma_curve);
876                 conversion->output_gamma_curve = output_format.gamma_curve;
877                 connect_nodes(output, conversion);
878         }
879 }
880
881 // Find the output node. This is, simply, one that has no outgoing links.
882 // If there are multiple ones, the graph is malformed (we do not support
883 // multiple outputs right now).
884 Node *EffectChain::find_output_node()
885 {
886         std::vector<Node *> output_nodes;
887         for (unsigned i = 0; i < nodes.size(); ++i) {
888                 Node *node = nodes[i];
889                 if (node->disabled) {
890                         continue;
891                 }
892                 if (node->outgoing_links.empty()) {
893                         output_nodes.push_back(node);
894                 }
895         }
896         assert(output_nodes.size() == 1);
897         return output_nodes[0];
898 }
899
900 void EffectChain::finalize()
901 {
902         // Output the graph as it is before we do any conversions on it.
903         output_dot("step0-start.dot");
904
905         // Give each effect in turn a chance to rewrite its own part of the graph.
906         // Note that if more effects are added as part of this, they will be
907         // picked up as part of the same for loop, since they are added at the end.
908         for (unsigned i = 0; i < nodes.size(); ++i) {
909                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
910         }
911         output_dot("step1-rewritten.dot");
912
913         propagate_gamma_and_color_space();
914         output_dot("step2-propagated.dot");
915
916         fix_internal_color_spaces();
917         fix_output_color_space();
918         output_dot("step4-output-colorspacefix.dot");
919
920         // Note that we need to fix gamma after colorspace conversion,
921         // because colorspace conversions might create needs for gamma conversions.
922         // Also, we need to run an extra pass of fix_internal_gamma() after 
923         // fixing the output gamma, as we only have conversions to/from linear.
924         fix_internal_gamma_by_asking_inputs(5);
925         fix_internal_gamma_by_inserting_nodes(6);
926         fix_output_gamma();
927         output_dot("step7-output-gammafix.dot");
928         fix_internal_gamma_by_asking_inputs(8);
929         fix_internal_gamma_by_inserting_nodes(9);
930
931         output_dot("step10-final.dot");
932         
933         // Construct all needed GLSL programs, starting at the output.
934         construct_glsl_programs(find_output_node());
935
936         // If we have more than one phase, we need intermediate render-to-texture.
937         // Construct an FBO, and then as many textures as we need.
938         // We choose the simplest option of having one texture per output,
939         // since otherwise this turns into an (albeit simple)
940         // register allocation problem.
941         if (phases.size() > 1) {
942                 glGenFramebuffers(1, &fbo);
943
944                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
945                         inform_input_sizes(phases[i]);
946                         find_output_size(phases[i]);
947
948                         Node *output_node = phases[i]->effects.back();
949                         glGenTextures(1, &output_node->output_texture);
950                         check_error();
951                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
952                         check_error();
953                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
954                         check_error();
955                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
956                         check_error();
957                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
958                         check_error();
959
960                         output_node->output_texture_width = phases[i]->output_width;
961                         output_node->output_texture_height = phases[i]->output_height;
962                 }
963                 inform_input_sizes(phases.back());
964         }
965                 
966         for (unsigned i = 0; i < inputs.size(); ++i) {
967                 inputs[i]->finalize();
968         }
969
970         assert(phases[0]->inputs.empty());
971         
972         finalized = true;
973 }
974
975 void EffectChain::render_to_fbo(GLuint fbo, unsigned width, unsigned height)
976 {
977         assert(finalized);
978
979         // Save original viewport.
980         GLuint x = 0, y = 0;
981
982         if (width == 0 && height == 0) {
983                 GLint viewport[4];
984                 glGetIntegerv(GL_VIEWPORT, viewport);
985                 x = viewport[0];
986                 y = viewport[1];
987                 width = viewport[2];
988                 height = viewport[3];
989         }
990
991         // Basic state.
992         glDisable(GL_BLEND);
993         check_error();
994         glDisable(GL_DEPTH_TEST);
995         check_error();
996         glDepthMask(GL_FALSE);
997         check_error();
998
999         glMatrixMode(GL_PROJECTION);
1000         glLoadIdentity();
1001         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1002
1003         glMatrixMode(GL_MODELVIEW);
1004         glLoadIdentity();
1005
1006         if (phases.size() > 1) {
1007                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1008                 check_error();
1009         }
1010
1011         std::set<Node *> generated_mipmaps;
1012
1013         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1014                 // See if the requested output size has changed. If so, we need to recreate
1015                 // the texture (and before we start setting up inputs).
1016                 inform_input_sizes(phases[phase]);
1017                 if (phase != phases.size() - 1) {
1018                         find_output_size(phases[phase]);
1019
1020                         Node *output_node = phases[phase]->effects.back();
1021
1022                         if (output_node->output_texture_width != phases[phase]->output_width ||
1023                             output_node->output_texture_height != phases[phase]->output_height) {
1024                                 glActiveTexture(GL_TEXTURE0);
1025                                 check_error();
1026                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1027                                 check_error();
1028                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1029                                 check_error();
1030                                 glBindTexture(GL_TEXTURE_2D, 0);
1031                                 check_error();
1032
1033                                 output_node->output_texture_width = phases[phase]->output_width;
1034                                 output_node->output_texture_height = phases[phase]->output_height;
1035                         }
1036                 }
1037
1038                 glUseProgram(phases[phase]->glsl_program_num);
1039                 check_error();
1040
1041                 // Set up RTT inputs for this phase.
1042                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1043                         glActiveTexture(GL_TEXTURE0 + sampler);
1044                         Node *input = phases[phase]->inputs[sampler];
1045                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1046                         check_error();
1047                         if (phases[phase]->input_needs_mipmaps) {
1048                                 if (generated_mipmaps.count(input) == 0) {
1049                                         glGenerateMipmap(GL_TEXTURE_2D);
1050                                         check_error();
1051                                         generated_mipmaps.insert(input);
1052                                 }
1053                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1054                                 check_error();
1055                         } else {
1056                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1057                                 check_error();
1058                         }
1059
1060                         std::string texture_name = std::string("tex_") + input->effect_id;
1061                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1062                         check_error();
1063                 }
1064
1065                 // And now the output.
1066                 if (phase == phases.size() - 1) {
1067                         // Last phase goes to the output the user specified.
1068                         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1069                         check_error();
1070                         glViewport(x, y, width, height);
1071                 } else {
1072                         Node *output_node = phases[phase]->effects.back();
1073                         glFramebufferTexture2D(
1074                                 GL_FRAMEBUFFER,
1075                                 GL_COLOR_ATTACHMENT0,
1076                                 GL_TEXTURE_2D,
1077                                 output_node->output_texture,
1078                                 0);
1079                         check_error();
1080                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1081                 }
1082
1083                 // Give the required parameters to all the effects.
1084                 unsigned sampler_num = phases[phase]->inputs.size();
1085                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1086                         Node *node = phases[phase]->effects[i];
1087                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1088                         check_error();
1089                 }
1090
1091                 // Now draw!
1092                 glBegin(GL_QUADS);
1093
1094                 glTexCoord2f(0.0f, 0.0f);
1095                 glVertex2f(0.0f, 0.0f);
1096
1097                 glTexCoord2f(1.0f, 0.0f);
1098                 glVertex2f(1.0f, 0.0f);
1099
1100                 glTexCoord2f(1.0f, 1.0f);
1101                 glVertex2f(1.0f, 1.0f);
1102
1103                 glTexCoord2f(0.0f, 1.0f);
1104                 glVertex2f(0.0f, 1.0f);
1105
1106                 glEnd();
1107                 check_error();
1108
1109                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1110                         Node *node = phases[phase]->effects[i];
1111                         node->effect->clear_gl_state();
1112                 }
1113         }
1114 }