git.sesse.net Git - movit/blob - 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                 return;
 116         }
 117         if (node->effect->num_inputs() == 0) {
 118                 nonlinear_inputs->push_back(node);
 119         } else {
 120                 assert(node->effect->num_inputs() == node->incoming_links.size());
 121                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 122                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 123                 }
 124         }
 125 }
 126
 127 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
 128 {
 129         assert(inputs.size() == effect->num_inputs());
 130         Node *node = add_node(effect);
 131         for (unsigned i = 0; i < inputs.size(); ++i) {
 132                 assert(node_map.count(inputs[i]) != 0);
 133                 connect_nodes(node_map[inputs[i]], node);
 134         }
 135         return effect;
 136 }
 137
 138 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 139 std::string replace_prefix(const std::string &text, const std::string &prefix)
 140 {
 141         std::string output;
 142         size_t start = 0;
 143
 144         while (start < text.size()) {
 145                 size_t pos = text.find("PREFIX(", start);
 146                 if (pos == std::string::npos) {
 147                         output.append(text.substr(start, std::string::npos));
 148                         break;
 149                 }
 150
 151                 output.append(text.substr(start, pos - start));
 152                 output.append(prefix);
 153                 output.append("_");
 154
 155                 pos += strlen("PREFIX(");
 156
 157                 // Output stuff until we find the matching ), which we then eat.
 158                 int depth = 1;
 159                 size_t end_arg_pos = pos;
 160                 while (end_arg_pos < text.size()) {
 161                         if (text[end_arg_pos] == '(') {
 162                                 ++depth;
 163                         } else if (text[end_arg_pos] == ')') {
 164                                 --depth;
 165                                 if (depth == 0) {
 166                                         break;
 167                                 }
 168                         }
 169                         ++end_arg_pos;
 170                 }
 171                 output.append(text.substr(pos, end_arg_pos - pos));
 172                 ++end_arg_pos;
 173                 assert(depth == 0);
 174                 start = end_arg_pos;
 175         }
 176         return output;
 177 }
 178
 179 Phase *EffectChain::compile_glsl_program(
 180         const std::vector<Node *> &inputs,
 181         const std::vector<Node *> &effects)
 182 {
 183         assert(!effects.empty());
 184
 185         // Deduplicate the inputs.
 186         std::vector<Node *> true_inputs = inputs;
 187         std::sort(true_inputs.begin(), true_inputs.end());
 188         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
 189
 190         bool input_needs_mipmaps = false;
 191         std::string frag_shader = read_file("header.frag");
 192
 193         // Create functions for all the texture inputs that we need.
 194         for (unsigned i = 0; i < true_inputs.size(); ++i) {
 195                 Node *input = true_inputs[i];
 196
 197                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
 198                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
 199                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
 200                 frag_shader += "}\n";
 201                 frag_shader += "\n";
 202         }
 203
 204         for (unsigned i = 0; i < effects.size(); ++i) {
 205                 Node *node = effects[i];
 206
 207                 if (node->incoming_links.size() == 1) {
 208                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
 209                 } else {
 210                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 211                                 char buf[256];
 212                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
 213                                 frag_shader += buf;
 214                         }
 215                 }
 216
 217                 frag_shader += "\n";
 218                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
 219                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
 220                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
 221                 frag_shader += "#undef PREFIX\n";
 222                 frag_shader += "#undef FUNCNAME\n";
 223                 if (node->incoming_links.size() == 1) {
 224                         frag_shader += "#undef INPUT\n";
 225                 } else {
 226                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 227                                 char buf[256];
 228                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 229                                 frag_shader += buf;
 230                         }
 231                 }
 232                 frag_shader += "\n";
 233
 234                 input_needs_mipmaps |= node->effect->needs_mipmaps();
 235         }
 236         for (unsigned i = 0; i < effects.size(); ++i) {
 237                 Node *node = effects[i];
 238                 if (node->effect->num_inputs() == 0) {
 239                         node->effect->set_int("needs_mipmaps", input_needs_mipmaps);
 240                 }
 241         }
 242         frag_shader += std::string("#define INPUT ") + effects.back()->effect_id + "\n";
 243         frag_shader.append(read_file("footer.frag"));
 244         printf("%s\n", frag_shader.c_str());
 245
 246         GLuint glsl_program_num = glCreateProgram();
 247         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 248         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
 249         glAttachShader(glsl_program_num, vs_obj);
 250         check_error();
 251         glAttachShader(glsl_program_num, fs_obj);
 252         check_error();
 253         glLinkProgram(glsl_program_num);
 254         check_error();
 255
 256         Phase *phase = new Phase;
 257         phase->glsl_program_num = glsl_program_num;
 258         phase->input_needs_mipmaps = input_needs_mipmaps;
 259         phase->inputs = true_inputs;
 260         phase->effects = effects;
 261
 262         return phase;
 263 }
 264
 265 // Construct GLSL programs, starting at the given effect and following
 266 // the chain from there. We end a program every time we come to an effect
 267 // marked as "needs texture bounce", one that is used by multiple other
 268 // effects, every time an effect wants to change the output size,
 269 // and of course at the end.
 270 //
 271 // We follow a quite simple depth-first search from the output, although
 272 // without any explicit recursion.
 273 void EffectChain::construct_glsl_programs(Node *output)
 274 {
 275         // Which effects have already been completed in this phase?
 276         // We need to keep track of it, as an effect with multiple outputs
 277         // could otherwise be calculate multiple times.
 278         std::set<Node *> completed_effects;
 279
 280         // Effects in the current phase, as well as inputs (outputs from other phases
 281         // that we depend on). Note that since we start iterating from the end,
 282         // the effect list will be in the reverse order.
 283         std::vector<Node *> this_phase_inputs;
 284         std::vector<Node *> this_phase_effects;
 285
 286         // Effects that we have yet to calculate, but that we know should
 287         // be in the current phase.
 288         std::stack<Node *> effects_todo_this_phase;
 289
 290         // Effects that we have yet to calculate, but that come from other phases.
 291         // We delay these until we have this phase done in its entirety,
 292         // at which point we pick any of them and start a new phase from that.
 293         std::stack<Node *> effects_todo_other_phases;
 294
 295         effects_todo_this_phase.push(output);
 296
 297         for ( ;; ) {  // Termination condition within loop.
 298                 if (!effects_todo_this_phase.empty()) {
 299                         // OK, we have more to do this phase.
 300                         Node *node = effects_todo_this_phase.top();
 301                         effects_todo_this_phase.pop();
 302
 303                         // This should currently only happen for effects that are phase outputs,
 304                         // and we throw those out separately below.
 305                         assert(completed_effects.count(node) == 0);
 306
 307                         this_phase_effects.push_back(node);
 308                         completed_effects.insert(node);
 309
 310                         // Find all the dependencies of this effect, and add them to the stack.
 311                         std::vector<Node *> deps = node->incoming_links;
 312                         assert(node->effect->num_inputs() == deps.size());
 313                         for (unsigned i = 0; i < deps.size(); ++i) {
 314                                 bool start_new_phase = false;
 315
 316                                 // FIXME: If we sample directly from a texture, we won't need this.
 317                                 if (node->effect->needs_texture_bounce()) {
 318                                         start_new_phase = true;
 319                                 }
 320
 321                                 if (deps[i]->outgoing_links.size() > 1 && deps[i]->effect->num_inputs() > 0) {
 322                                         // More than one effect uses this as the input,
 323                                         // and it is not a texture itself.
 324                                         // The easiest thing to do (and probably also the safest
 325                                         // performance-wise in most cases) is to bounce it to a texture
 326                                         // and then let the next passes read from that.
 327                                         start_new_phase = true;
 328                                 }
 329
 330                                 if (deps[i]->effect->changes_output_size()) {
 331                                         start_new_phase = true;
 332                                 }
 333
 334                                 if (start_new_phase) {
 335                                         effects_todo_other_phases.push(deps[i]);
 336                                         this_phase_inputs.push_back(deps[i]);
 337                                 } else {
 338                                         effects_todo_this_phase.push(deps[i]);
 339                                 }
 340                         }
 341                         continue;
 342                 }
 343
 344                 // No more effects to do this phase. Take all the ones we have,
 345                 // and create a GLSL program for it.
 346                 if (!this_phase_effects.empty()) {
 347                         reverse(this_phase_effects.begin(), this_phase_effects.end());
 348                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
 349                         this_phase_effects.back()->phase = phases.back();
 350                         this_phase_inputs.clear();
 351                         this_phase_effects.clear();
 352                 }
 353                 assert(this_phase_inputs.empty());
 354                 assert(this_phase_effects.empty());
 355
 356                 // If we have no effects left, exit.
 357                 if (effects_todo_other_phases.empty()) {
 358                         break;
 359                 }
 360
 361                 Node *node = effects_todo_other_phases.top();
 362                 effects_todo_other_phases.pop();
 363
 364                 if (completed_effects.count(node) == 0) {
 365                         // Start a new phase, calculating from this effect.
 366                         effects_todo_this_phase.push(node);
 367                 }
 368         }
 369
 370         // Finally, since the phases are found from the output but must be executed
 371         // from the input(s), reverse them, too.
 372         std::reverse(phases.begin(), phases.end());
 373 }
 374
 375 void EffectChain::output_dot(const char *filename)
 376 {
 377         FILE *fp = fopen(filename, "w");
 378         if (fp == NULL) {
 379                 perror(filename);
 380                 exit(1);
 381         }
 382
 383         fprintf(fp, "digraph G {\n");
 384         for (unsigned i = 0; i < nodes.size(); ++i) {
 385                 fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 386                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 387                         std::vector<std::string> labels;
 388
 389                         if (nodes[i]->outgoing_links[j]->effect->needs_texture_bounce()) {
 390                                 labels.push_back("needs_bounce");
 391                         }
 392                         if (nodes[i]->effect->changes_output_size()) {
 393                                 labels.push_back("resize");
 394                         }
 395
 396                         switch (nodes[i]->output_color_space) {
 397                         case COLORSPACE_INVALID:
 398                                 labels.push_back("spc[invalid]");
 399                                 break;
 400                         case COLORSPACE_REC_601_525:
 401                                 labels.push_back("spc[rec601-525]");
 402                                 break;
 403                         case COLORSPACE_REC_601_625:
 404                                 labels.push_back("spc[rec601-625]");
 405                                 break;
 406                         default:
 407                                 break;
 408                         }
 409
 410                         switch (nodes[i]->output_gamma_curve) {
 411                         case GAMMA_INVALID:
 412                                 labels.push_back("gamma[invalid]");
 413                                 break;
 414                         case GAMMA_sRGB:
 415                                 labels.push_back("gamma[sRGB]");
 416                                 break;
 417                         case GAMMA_REC_601:  // and GAMMA_REC_709
 418                                 labels.push_back("gamma[rec601/709]");
 419                                 break;
 420                         default:
 421                                 break;
 422                         }
 423
 424                         if (labels.empty()) {
 425                                 fprintf(fp, "  n%ld -> n%ld;\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j]);
 426                         } else {
 427                                 std::string label = labels[0];
 428                                 for (unsigned k = 1; k < labels.size(); ++k) {
 429                                         label += ", " + labels[k];
 430                                 }
 431                                 fprintf(fp, "  n%ld -> n%ld [label=\"%s\"];\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j], label.c_str());
 432                         }
 433                 }
 434         }
 435         fprintf(fp, "}\n");
 436
 437         fclose(fp);
 438 }
 439
 440 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
 441 {
 442         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 443                 // Same aspect, or W/H > aspect (image is wider than the frame).
 444                 // In either case, keep width.
 445                 return width;
 446         } else {
 447                 // W/H < aspect (image is taller than the frame), so keep height,
 448                 // and adjust width correspondingly.
 449                 return lrintf(height * aspect_nom / aspect_denom);
 450         }
 451 }
 452
 453 // Propagate input texture sizes throughout, and inform effects downstream.
 454 // (Like a lot of other code, we depend on effects being in topological order.)
 455 void EffectChain::inform_input_sizes(Phase *phase)
 456 {
 457         // All effects that have a defined size (inputs and RTT inputs)
 458         // get that. Reset all others.
 459         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 460                 Node *node = phase->effects[i];
 461                 if (node->effect->num_inputs() == 0) {
 462                         Input *input = static_cast<Input *>(node->effect);
 463                         node->output_width = input->get_width();
 464                         node->output_height = input->get_height();
 465                         assert(node->output_width != 0);
 466                         assert(node->output_height != 0);
 467                 } else {
 468                         node->output_width = node->output_height = 0;
 469                 }
 470         }
 471         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 472                 Node *input = phase->inputs[i];
 473                 input->output_width = input->phase->output_width;
 474                 input->output_height = input->phase->output_height;
 475                 assert(input->output_width != 0);
 476                 assert(input->output_height != 0);
 477         }
 478
 479         // Now propagate from the inputs towards the end, and inform as we go.
 480         // The rules are simple:
 481         //
 482         //   1. Don't touch effects that already have given sizes (ie., inputs).
 483         //   2. If all of your inputs have the same size, that will be your output size.
 484         //   3. Otherwise, your output size is 0x0.
 485         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 486                 Node *node = phase->effects[i];
 487                 if (node->effect->num_inputs() == 0) {
 488                         continue;
 489                 }
 490                 unsigned this_output_width = 0;
 491                 unsigned this_output_height = 0;
 492                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 493                         Node *input = node->incoming_links[j];
 494                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 495                         if (j == 0) {
 496                                 this_output_width = input->output_width;
 497                                 this_output_height = input->output_height;
 498                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 499                                 // Inputs disagree.
 500                                 this_output_width = 0;
 501                                 this_output_height = 0;
 502                         }
 503                 }
 504                 node->output_width = this_output_width;
 505                 node->output_height = this_output_height;
 506         }
 507 }
 508
 509 // Note: You should call inform_input_sizes() before this, as the last effect's
 510 // desired output size might change based on the inputs.
 511 void EffectChain::find_output_size(Phase *phase)
 512 {
 513         Node *output_node = phase->effects.back();
 514
 515         // If the last effect explicitly sets an output size, use that.
 516         if (output_node->effect->changes_output_size()) {
 517                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
 518                 return;
 519         }
 520
 521         // If not, look at the input phases and textures.
 522         // We select the largest one (by fit into the current aspect).
 523         unsigned best_width = 0;
 524         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 525                 Node *input = phase->inputs[i];
 526                 assert(input->phase->output_width != 0);
 527                 assert(input->phase->output_height != 0);
 528                 unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height);
 529                 if (width > best_width) {
 530                         best_width = width;
 531                 }
 532         }
 533         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 534                 Effect *effect = phase->effects[i]->effect;
 535                 if (effect->num_inputs() != 0) {
 536                         continue;
 537                 }
 538
 539                 Input *input = static_cast<Input *>(effect);
 540                 unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height());
 541                 if (width > best_width) {
 542                         best_width = width;
 543                 }
 544         }
 545         assert(best_width != 0);
 546         phase->output_width = best_width;
 547         phase->output_height = best_width * aspect_denom / aspect_nom;
 548 }
 549
 550 void EffectChain::sort_nodes_topologically()
 551 {
 552         std::set<Node *> visited_nodes;
 553         std::vector<Node *> sorted_list;
 554         for (unsigned i = 0; i < nodes.size(); ++i) {
 555                 if (nodes[i]->incoming_links.size() == 0) {
 556                         topological_sort_visit_node(nodes[i], &visited_nodes, &sorted_list);
 557                 }
 558         }
 559         reverse(sorted_list.begin(), sorted_list.end());
 560         nodes = sorted_list;
 561 }
 562
 563 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *visited_nodes, std::vector<Node *> *sorted_list)
 564 {
 565         if (visited_nodes->count(node) != 0) {
 566                 return;
 567         }
 568         visited_nodes->insert(node);
 569         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 570                 topological_sort_visit_node(node->outgoing_links[i], visited_nodes, sorted_list);
 571         }
 572         sorted_list->push_back(node);
 573 }
 574
 575 // Propagate gamma and color space information as far as we can in the graph.
 576 // The rules are simple: Anything where all the inputs agree, get that as
 577 // output as well. Anything else keeps having *_INVALID.
 578 void EffectChain::propagate_gamma_and_color_space()
 579 {
 580         // We depend on going through the nodes in order.
 581         sort_nodes_topologically();
 582
 583         for (unsigned i = 0; i < nodes.size(); ++i) {
 584                 Node *node = nodes[i];
 585                 if (node->disabled) {
 586                         continue;
 587                 }
 588                 assert(node->incoming_links.size() == node->effect->num_inputs());
 589                 if (node->incoming_links.size() == 0) {
 590                         assert(node->output_color_space != COLORSPACE_INVALID);
 591                         assert(node->output_gamma_curve != GAMMA_INVALID);
 592                         continue;
 593                 }
 594
 595                 ColorSpace color_space = node->incoming_links[0]->output_color_space;
 596                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 597                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 598                         if (node->incoming_links[j]->output_color_space != color_space) {
 599                                 color_space = COLORSPACE_INVALID;
 600                         }
 601                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 602                                 gamma_curve = GAMMA_INVALID;
 603                         }
 604                 }
 605
 606                 // The conversion effects already have their outputs set correctly,
 607                 // so leave them alone.
 608                 if (node->effect->effect_type_id() != "ColorSpaceConversionEffect") {
 609                         node->output_color_space = color_space;
 610                 }
 611                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 612                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 613                         node->output_gamma_curve = gamma_curve;
 614                 }
 615         }
 616 }
 617
 618 bool EffectChain::node_needs_colorspace_fix(Node *node)
 619 {
 620         if (node->disabled) {
 621                 return false;
 622         }
 623         if (node->effect->num_inputs() == 0) {
 624                 return false;
 625         }
 626
 627         // propagate_gamma_and_color_space() has already set our output
 628         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 629         if (node->output_color_space == COLORSPACE_INVALID) {
 630                 return true;
 631         }
 632         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 633 }
 634
 635 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 636 // the graph. Our strategy is not always optimal, but quite simple:
 637 // Find an effect that's as early as possible where the inputs are of
 638 // unacceptable colorspaces (that is, either different, or, if the effect only
 639 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 640 // propagate the information anew, and repeat until there are no more such
 641 // effects.
 642 void EffectChain::fix_internal_color_spaces()
 643 {
 644         unsigned colorspace_propagation_pass = 0;
 645         bool found_any;
 646         do {
 647                 found_any = false;
 648                 for (unsigned i = 0; i < nodes.size(); ++i) {
 649                         Node *node = nodes[i];
 650                         if (!node_needs_colorspace_fix(node)) {
 651                                 continue;
 652                         }
 653
 654                         // Go through each input that is not sRGB, and insert
 655                         // a colorspace conversion before it.
 656                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 657                                 Node *input = node->incoming_links[j];
 658                                 assert(input->output_color_space != COLORSPACE_INVALID);
 659                                 if (input->output_color_space == COLORSPACE_sRGB) {
 660                                         continue;
 661                                 }
 662                                 Node *conversion = add_node(new ColorSpaceConversionEffect());
 663                                 conversion->effect->set_int("source_space", input->output_color_space);
 664                                 conversion->effect->set_int("destination_space", COLORSPACE_sRGB);
 665                                 conversion->output_color_space = COLORSPACE_sRGB;
 666                                 insert_node_between(input, conversion, node);
 667                         }
 668
 669                         // Re-sort topologically, and propagate the new information.
 670                         propagate_gamma_and_color_space();
 671
 672                         found_any = true;
 673                         break;
 674                 }
 675
 676                 char filename[256];
 677                 sprintf(filename, "step3-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
 678                 output_dot(filename);
 679                 assert(colorspace_propagation_pass < 100);
 680         } while (found_any);
 681
 682         for (unsigned i = 0; i < nodes.size(); ++i) {
 683                 Node *node = nodes[i];
 684                 if (node->disabled) {
 685                         continue;
 686                 }
 687                 assert(node->output_color_space != COLORSPACE_INVALID);
 688         }
 689 }
 690
 691 // Make so that the output is in the desired color space.
 692 void EffectChain::fix_output_color_space()
 693 {
 694         Node *output = find_output_node();
 695         if (output->output_color_space != output_format.color_space) {
 696                 Node *conversion = add_node(new ColorSpaceConversionEffect());
 697                 conversion->effect->set_int("source_space", output->output_color_space);
 698                 conversion->effect->set_int("destination_space", output_format.color_space);
 699                 conversion->output_color_space = output_format.color_space;
 700                 connect_nodes(output, conversion);
 701         }
 702 }
 703
 704 bool EffectChain::node_needs_gamma_fix(Node *node)
 705 {
 706         if (node->disabled) {
 707                 return false;
 708         }
 709         if (node->effect->num_inputs() == 0) {
 710                 return false;
 711         }
 712
 713         // propagate_gamma_and_color_space() has already set our output
 714         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
 715         // except for GammaCompressionEffect.
 716         if (node->output_gamma_curve == GAMMA_INVALID) {
 717                 return true;
 718         }
 719         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
 720                 assert(node->incoming_links.size() == 1);
 721                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
 722         }
 723         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
 724 }
 725
 726 // Very similar to fix_internal_color_spaces(), but for gamma.
 727 // There is one difference, though; before we start adding conversion nodes,
 728 // we see if we can get anything out of asking the sources to deliver
 729 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
 730 // does that part, while fix_internal_gamma_by_inserting_nodes()
 731 // inserts nodes as needed afterwards.
 732 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
 733 {
 734         unsigned gamma_propagation_pass = 0;
 735         bool found_any;
 736         do {
 737                 found_any = false;
 738                 for (unsigned i = 0; i < nodes.size(); ++i) {
 739                         Node *node = nodes[i];
 740                         if (!node_needs_gamma_fix(node)) {
 741                                 continue;
 742                         }
 743
 744                         // See if all inputs can give us linear gamma. If not, leave it.
 745                         std::vector<Node *> nonlinear_inputs;
 746                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
 747
 748                         bool all_ok = true;
 749                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
 750                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
 751                                 all_ok &= input->can_output_linear_gamma();
 752                         }
 753
 754                         if (!all_ok) {
 755                                 continue;
 756                         }
 757
 758                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
 759                                 nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1);
 760                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
 761                         }
 762
 763                         // Re-sort topologically, and propagate the new information.
 764                         propagate_gamma_and_color_space();
 765
 766                         found_any = true;
 767                         break;
 768                 }
 769
 770                 char filename[256];
 771                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
 772                 output_dot(filename);
 773                 assert(gamma_propagation_pass < 100);
 774         } while (found_any);
 775 }
 776
 777 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
 778 {
 779         unsigned gamma_propagation_pass = 0;
 780         bool found_any;
 781         do {
 782                 found_any = false;
 783                 for (unsigned i = 0; i < nodes.size(); ++i) {
 784                         Node *node = nodes[i];
 785                         if (!node_needs_gamma_fix(node)) {
 786                                 continue;
 787                         }
 788
 789                         // Go through each input that is not linear gamma, and insert
 790                         // a gamma conversion before it.
 791                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 792                                 Node *input = node->incoming_links[j];
 793                                 assert(input->output_gamma_curve != GAMMA_INVALID);
 794                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
 795                                         continue;
 796                                 }
 797                                 Node *conversion = add_node(new GammaExpansionEffect());
 798                                 conversion->effect->set_int("destination_curve", GAMMA_LINEAR);
 799                                 conversion->output_gamma_curve = GAMMA_LINEAR;
 800                                 insert_node_between(input, conversion, node);
 801                         }
 802
 803                         // Re-sort topologically, and propagate the new information.
 804                         propagate_gamma_and_color_space();
 805
 806                         found_any = true;
 807                         break;
 808                 }
 809
 810                 char filename[256];
 811                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
 812                 output_dot(filename);
 813                 assert(gamma_propagation_pass < 100);
 814         } while (found_any);
 815
 816         for (unsigned i = 0; i < nodes.size(); ++i) {
 817                 Node *node = nodes[i];
 818                 if (node->disabled) {
 819                         continue;
 820                 }
 821                 assert(node->output_gamma_curve != GAMMA_INVALID);
 822         }
 823 }
 824
 825 // Make so that the output is in the desired gamma.
 826 // Note that this assumes linear input gamma, so it might create the need
 827 // for another pass of fix_internal_gamma().
 828 void EffectChain::fix_output_gamma()
 829 {
 830         Node *output = find_output_node();
 831         if (output->output_gamma_curve != output_format.gamma_curve) {
 832                 Node *conversion = add_node(new GammaCompressionEffect());
 833                 conversion->effect->set_int("destination_curve", output_format.gamma_curve);
 834                 conversion->output_gamma_curve = output_format.gamma_curve;
 835                 connect_nodes(output, conversion);
 836         }
 837 }
 838
 839 // Find the output node. This is, simply, one that has no outgoing links.
 840 // If there are multiple ones, the graph is malformed (we do not support
 841 // multiple outputs right now).
 842 Node *EffectChain::find_output_node()
 843 {
 844         std::vector<Node *> output_nodes;
 845         for (unsigned i = 0; i < nodes.size(); ++i) {
 846                 Node *node = nodes[i];
 847                 if (node->disabled) {
 848                         continue;
 849                 }
 850                 if (node->outgoing_links.empty()) {
 851                         output_nodes.push_back(node);
 852                 }
 853         }
 854         assert(output_nodes.size() == 1);
 855         return output_nodes[0];
 856 }
 857
 858 void EffectChain::finalize()
 859 {
 860         // Output the graph as it is before we do any conversions on it.
 861         output_dot("step0-start.dot");
 862
 863         // Give each effect in turn a chance to rewrite its own part of the graph.
 864         // Note that if more effects are added as part of this, they will be
 865         // picked up as part of the same for loop, since they are added at the end.
 866         for (unsigned i = 0; i < nodes.size(); ++i) {
 867                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
 868         }
 869         output_dot("step1-rewritten.dot");
 870
 871         propagate_gamma_and_color_space();
 872         output_dot("step2-propagated.dot");
 873
 874         fix_internal_color_spaces();
 875         fix_output_color_space();
 876         output_dot("step4-output-colorspacefix.dot");
 877
 878         // Note that we need to fix gamma after colorspace conversion,
 879         // because colorspace conversions might create needs for gamma conversions.
 880         // Also, we need to run an extra pass of fix_internal_gamma() after
 881         // fixing the output gamma, as we only have conversions to/from linear.
 882         fix_internal_gamma_by_asking_inputs(5);
 883         fix_internal_gamma_by_inserting_nodes(6);
 884         fix_output_gamma();
 885         output_dot("step8-output-gammafix.dot");
 886         fix_internal_gamma_by_asking_inputs(9);
 887         fix_internal_gamma_by_inserting_nodes(10);
 888
 889         output_dot("step11-final.dot");
 890
 891         // Construct all needed GLSL programs, starting at the output.
 892         construct_glsl_programs(find_output_node());
 893
 894         // If we have more than one phase, we need intermediate render-to-texture.
 895         // Construct an FBO, and then as many textures as we need.
 896         // We choose the simplest option of having one texture per output,
 897         // since otherwise this turns into an (albeit simple)
 898         // register allocation problem.
 899         if (phases.size() > 1) {
 900                 glGenFramebuffers(1, &fbo);
 901
 902                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
 903                         inform_input_sizes(phases[i]);
 904                         find_output_size(phases[i]);
 905
 906                         Node *output_node = phases[i]->effects.back();
 907                         glGenTextures(1, &output_node->output_texture);
 908                         check_error();
 909                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
 910                         check_error();
 911                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 912                         check_error();
 913                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 914                         check_error();
 915                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
 916                         check_error();
 917
 918                         output_node->output_texture_width = phases[i]->output_width;
 919                         output_node->output_texture_height = phases[i]->output_height;
 920                 }
 921                 inform_input_sizes(phases.back());
 922         }
 923
 924         for (unsigned i = 0; i < inputs.size(); ++i) {
 925                 inputs[i]->finalize();
 926         }
 927
 928         assert(phases[0]->inputs.empty());
 929
 930         finalized = true;
 931 }
 932
 933 void EffectChain::render_to_screen()
 934 {
 935         assert(finalized);
 936
 937         // Save original viewport.
 938         GLint viewport[4];
 939         glGetIntegerv(GL_VIEWPORT, viewport);
 940
 941         // Basic state.
 942         glDisable(GL_BLEND);
 943         check_error();
 944         glDisable(GL_DEPTH_TEST);
 945         check_error();
 946         glDepthMask(GL_FALSE);
 947         check_error();
 948
 949         glMatrixMode(GL_PROJECTION);
 950         glLoadIdentity();
 951         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
 952
 953         glMatrixMode(GL_MODELVIEW);
 954         glLoadIdentity();
 955
 956         if (phases.size() > 1) {
 957                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 958                 check_error();
 959         }
 960
 961         std::set<Node *> generated_mipmaps;
 962
 963         for (unsigned phase = 0; phase < phases.size(); ++phase) {
 964                 // See if the requested output size has changed. If so, we need to recreate
 965                 // the texture (and before we start setting up inputs).
 966                 inform_input_sizes(phases[phase]);
 967                 if (phase != phases.size() - 1) {
 968                         find_output_size(phases[phase]);
 969
 970                         Node *output_node = phases[phase]->effects.back();
 971
 972                         if (output_node->output_texture_width != phases[phase]->output_width ||
 973                             output_node->output_texture_height != phases[phase]->output_height) {
 974                                 glActiveTexture(GL_TEXTURE0);
 975                                 check_error();
 976                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
 977                                 check_error();
 978                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
 979                                 check_error();
 980                                 glBindTexture(GL_TEXTURE_2D, 0);
 981                                 check_error();
 982
 983                                 output_node->output_texture_width = phases[phase]->output_width;
 984                                 output_node->output_texture_height = phases[phase]->output_height;
 985                         }
 986                 }
 987
 988                 glUseProgram(phases[phase]->glsl_program_num);
 989                 check_error();
 990
 991                 // Set up RTT inputs for this phase.
 992                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
 993                         glActiveTexture(GL_TEXTURE0 + sampler);
 994                         Node *input = phases[phase]->inputs[sampler];
 995                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
 996                         check_error();
 997                         if (phases[phase]->input_needs_mipmaps) {
 998                                 if (generated_mipmaps.count(input) == 0) {
 999                                         glGenerateMipmap(GL_TEXTURE_2D);
1000                                         check_error();
1001                                         generated_mipmaps.insert(input);
1002                                 }
1003                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1004                                 check_error();
1005                         } else {
1006                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1007                                 check_error();
1008                         }
1009
1010                         std::string texture_name = std::string("tex_") + input->effect_id;
1011                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1012                         check_error();
1013                 }
1014
1015                 // And now the output.
1016                 if (phase == phases.size() - 1) {
1017                         // Last phase goes directly to the screen.
1018                         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1019                         check_error();
1020                         glViewport(viewport[0], viewport[1], viewport[2], viewport[3]);
1021                 } else {
1022                         Node *output_node = phases[phase]->effects.back();
1023                         glFramebufferTexture2D(
1024                                 GL_FRAMEBUFFER,
1025                                 GL_COLOR_ATTACHMENT0,
1026                                 GL_TEXTURE_2D,
1027                                 output_node->output_texture,
1028                                 0);
1029                         check_error();
1030                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1031                 }
1032
1033                 // Give the required parameters to all the effects.
1034                 unsigned sampler_num = phases[phase]->inputs.size();
1035                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1036                         Node *node = phases[phase]->effects[i];
1037                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1038                         check_error();
1039                 }
1040
1041                 // Now draw!
1042                 glBegin(GL_QUADS);
1043
1044                 glTexCoord2f(0.0f, 0.0f);
1045                 glVertex2f(0.0f, 0.0f);
1046
1047                 glTexCoord2f(1.0f, 0.0f);
1048                 glVertex2f(1.0f, 0.0f);
1049
1050                 glTexCoord2f(1.0f, 1.0f);
1051                 glVertex2f(1.0f, 1.0f);
1052
1053                 glTexCoord2f(0.0f, 1.0f);
1054                 glVertex2f(0.0f, 1.0f);
1055
1056                 glEnd();
1057                 check_error();
1058
1059                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1060                         Node *node = phases[phase]->effects[i];
1061                         node->effect->clear_gl_state();
1062                 }
1063         }
1064 }