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 "dither_effect.h"
  19 #include "input.h"
  20 #include "opengl.h"
  21
  22 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
  23         : aspect_nom(aspect_nom),
  24           aspect_denom(aspect_denom),
  25           dither_effect(NULL),
  26           fbo(0),
  27           num_dither_bits(0),
  28           finalized(false) {}
  29
  30 EffectChain::~EffectChain()
  31 {
  32         for (unsigned i = 0; i < nodes.size(); ++i) {
  33                 if (nodes[i]->output_texture != 0) {
  34                         glDeleteTextures(1, &nodes[i]->output_texture);
  35                 }
  36                 delete nodes[i]->effect;
  37                 delete nodes[i];
  38         }
  39         for (unsigned i = 0; i < phases.size(); ++i) {
  40                 glDeleteProgram(phases[i]->glsl_program_num);
  41                 glDeleteShader(phases[i]->vertex_shader);
  42                 glDeleteShader(phases[i]->fragment_shader);
  43                 delete phases[i];
  44         }
  45         if (fbo != 0) {
  46                 glDeleteFramebuffers(1, &fbo);
  47         }
  48 }
  49
  50 Input *EffectChain::add_input(Input *input)
  51 {
  52         inputs.push_back(input);
  53
  54         Node *node = add_node(input);
  55         node->output_color_space = input->get_color_space();
  56         node->output_gamma_curve = input->get_gamma_curve();
  57         return input;
  58 }
  59
  60 void EffectChain::add_output(const ImageFormat &format)
  61 {
  62         output_format = format;
  63 }
  64
  65 Node *EffectChain::add_node(Effect *effect)
  66 {
  67         char effect_id[256];
  68         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
  69
  70         Node *node = new Node;
  71         node->effect = effect;
  72         node->disabled = false;
  73         node->effect_id = effect_id;
  74         node->output_color_space = COLORSPACE_INVALID;
  75         node->output_gamma_curve = GAMMA_INVALID;
  76         node->output_texture = 0;
  77
  78         nodes.push_back(node);
  79         node_map[effect] = node;
  80         return node;
  81 }
  82
  83 void EffectChain::connect_nodes(Node *sender, Node *receiver)
  84 {
  85         sender->outgoing_links.push_back(receiver);
  86         receiver->incoming_links.push_back(sender);
  87 }
  88
  89 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
  90 {
  91         new_receiver->incoming_links = old_receiver->incoming_links;
  92         old_receiver->incoming_links.clear();
  93
  94         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
  95                 Node *sender = new_receiver->incoming_links[i];
  96                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
  97                         if (sender->outgoing_links[j] == old_receiver) {
  98                                 sender->outgoing_links[j] = new_receiver;
  99                         }
 100                 }
 101         }
 102 }
 103
 104 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 105 {
 106         new_sender->outgoing_links = old_sender->outgoing_links;
 107         old_sender->outgoing_links.clear();
 108
 109         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 110                 Node *receiver = new_sender->outgoing_links[i];
 111                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 112                         if (receiver->incoming_links[j] == old_sender) {
 113                                 receiver->incoming_links[j] = new_sender;
 114                         }
 115                 }
 116         }
 117 }
 118
 119 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 120 {
 121         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 122                 if (sender->outgoing_links[i] == receiver) {
 123                         sender->outgoing_links[i] = middle;
 124                         middle->incoming_links.push_back(sender);
 125                 }
 126         }
 127         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 128                 if (receiver->incoming_links[i] == sender) {
 129                         receiver->incoming_links[i] = middle;
 130                         middle->outgoing_links.push_back(receiver);
 131                 }
 132         }
 133
 134         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 135 }
 136
 137 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
 138 {
 139         if (node->output_gamma_curve == GAMMA_LINEAR &&
 140             node->effect->effect_type_id() != "GammaCompressionEffect") {
 141                 return;
 142         }
 143         if (node->effect->num_inputs() == 0) {
 144                 nonlinear_inputs->push_back(node);
 145         } else {
 146                 assert(node->effect->num_inputs() == node->incoming_links.size());
 147                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 148                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 149                 }
 150         }
 151 }
 152
 153 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
 154 {
 155         assert(inputs.size() == effect->num_inputs());
 156         Node *node = add_node(effect);
 157         for (unsigned i = 0; i < inputs.size(); ++i) {
 158                 assert(node_map.count(inputs[i]) != 0);
 159                 connect_nodes(node_map[inputs[i]], node);
 160         }
 161         return effect;
 162 }
 163
 164 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 165 std::string replace_prefix(const std::string &text, const std::string &prefix)
 166 {
 167         std::string output;
 168         size_t start = 0;
 169
 170         while (start < text.size()) {
 171                 size_t pos = text.find("PREFIX(", start);
 172                 if (pos == std::string::npos) {
 173                         output.append(text.substr(start, std::string::npos));
 174                         break;
 175                 }
 176
 177                 output.append(text.substr(start, pos - start));
 178                 output.append(prefix);
 179                 output.append("_");
 180
 181                 pos += strlen("PREFIX(");
 182
 183                 // Output stuff until we find the matching ), which we then eat.
 184                 int depth = 1;
 185                 size_t end_arg_pos = pos;
 186                 while (end_arg_pos < text.size()) {
 187                         if (text[end_arg_pos] == '(') {
 188                                 ++depth;
 189                         } else if (text[end_arg_pos] == ')') {
 190                                 --depth;
 191                                 if (depth == 0) {
 192                                         break;
 193                                 }
 194                         }
 195                         ++end_arg_pos;
 196                 }
 197                 output.append(text.substr(pos, end_arg_pos - pos));
 198                 ++end_arg_pos;
 199                 assert(depth == 0);
 200                 start = end_arg_pos;
 201         }
 202         return output;
 203 }
 204
 205 Phase *EffectChain::compile_glsl_program(
 206         const std::vector<Node *> &inputs,
 207         const std::vector<Node *> &effects)
 208 {
 209         assert(!effects.empty());
 210
 211         // Deduplicate the inputs.
 212         std::vector<Node *> true_inputs = inputs;
 213         std::sort(true_inputs.begin(), true_inputs.end());
 214         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
 215
 216         bool input_needs_mipmaps = false;
 217         std::string frag_shader = read_file("header.frag");
 218
 219         // Create functions for all the texture inputs that we need.
 220         for (unsigned i = 0; i < true_inputs.size(); ++i) {
 221                 Node *input = true_inputs[i];
 222
 223                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
 224                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
 225                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
 226                 frag_shader += "}\n";
 227                 frag_shader += "\n";
 228         }
 229
 230         for (unsigned i = 0; i < effects.size(); ++i) {
 231                 Node *node = effects[i];
 232
 233                 if (node->incoming_links.size() == 1) {
 234                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
 235                 } else {
 236                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 237                                 char buf[256];
 238                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
 239                                 frag_shader += buf;
 240                         }
 241                 }
 242
 243                 frag_shader += "\n";
 244                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
 245                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
 246                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
 247                 frag_shader += "#undef PREFIX\n";
 248                 frag_shader += "#undef FUNCNAME\n";
 249                 if (node->incoming_links.size() == 1) {
 250                         frag_shader += "#undef INPUT\n";
 251                 } else {
 252                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 253                                 char buf[256];
 254                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 255                                 frag_shader += buf;
 256                         }
 257                 }
 258                 frag_shader += "\n";
 259
 260                 input_needs_mipmaps |= node->effect->needs_mipmaps();
 261         }
 262         for (unsigned i = 0; i < effects.size(); ++i) {
 263                 Node *node = effects[i];
 264                 if (node->effect->num_inputs() == 0) {
 265                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
 266                 }
 267         }
 268         frag_shader += std::string("#define INPUT ") + effects.back()->effect_id + "\n";
 269         frag_shader.append(read_file("footer.frag"));
 270
 271         // Output shader to a temporary file, for easier debugging.
 272         static int compiled_shader_num = 0;
 273         char filename[256];
 274         sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
 275         FILE *fp = fopen(filename, "w");
 276         if (fp == NULL) {
 277                 perror(filename);
 278                 exit(1);
 279         }
 280         fprintf(fp, "%s\n", frag_shader.c_str());
 281         fclose(fp);
 282
 283         GLuint glsl_program_num = glCreateProgram();
 284         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 285         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
 286         glAttachShader(glsl_program_num, vs_obj);
 287         check_error();
 288         glAttachShader(glsl_program_num, fs_obj);
 289         check_error();
 290         glLinkProgram(glsl_program_num);
 291         check_error();
 292
 293         Phase *phase = new Phase;
 294         phase->glsl_program_num = glsl_program_num;
 295         phase->vertex_shader = vs_obj;
 296         phase->fragment_shader = fs_obj;
 297         phase->input_needs_mipmaps = input_needs_mipmaps;
 298         phase->inputs = true_inputs;
 299         phase->effects = effects;
 300
 301         return phase;
 302 }
 303
 304 // Construct GLSL programs, starting at the given effect and following
 305 // the chain from there. We end a program every time we come to an effect
 306 // marked as "needs texture bounce", one that is used by multiple other
 307 // effects, every time an effect wants to change the output size,
 308 // and of course at the end.
 309 //
 310 // We follow a quite simple depth-first search from the output, although
 311 // without any explicit recursion.
 312 void EffectChain::construct_glsl_programs(Node *output)
 313 {
 314         // Which effects have already been completed in this phase?
 315         // We need to keep track of it, as an effect with multiple outputs
 316         // could otherwise be calculate multiple times.
 317         std::set<Node *> completed_effects;
 318
 319         // Effects in the current phase, as well as inputs (outputs from other phases
 320         // that we depend on). Note that since we start iterating from the end,
 321         // the effect list will be in the reverse order.
 322         std::vector<Node *> this_phase_inputs;
 323         std::vector<Node *> this_phase_effects;
 324
 325         // Effects that we have yet to calculate, but that we know should
 326         // be in the current phase.
 327         std::stack<Node *> effects_todo_this_phase;
 328
 329         // Effects that we have yet to calculate, but that come from other phases.
 330         // We delay these until we have this phase done in its entirety,
 331         // at which point we pick any of them and start a new phase from that.
 332         std::stack<Node *> effects_todo_other_phases;
 333
 334         effects_todo_this_phase.push(output);
 335
 336         for ( ;; ) {  // Termination condition within loop.
 337                 if (!effects_todo_this_phase.empty()) {
 338                         // OK, we have more to do this phase.
 339                         Node *node = effects_todo_this_phase.top();
 340                         effects_todo_this_phase.pop();
 341
 342                         // This should currently only happen for effects that are phase outputs,
 343                         // and we throw those out separately below.
 344                         assert(completed_effects.count(node) == 0);
 345
 346                         this_phase_effects.push_back(node);
 347                         completed_effects.insert(node);
 348
 349                         // Find all the dependencies of this effect, and add them to the stack.
 350                         std::vector<Node *> deps = node->incoming_links;
 351                         assert(node->effect->num_inputs() == deps.size());
 352                         for (unsigned i = 0; i < deps.size(); ++i) {
 353                                 bool start_new_phase = false;
 354
 355                                 // FIXME: If we sample directly from a texture, we won't need this.
 356                                 if (node->effect->needs_texture_bounce()) {
 357                                         start_new_phase = true;
 358                                 }
 359
 360                                 if (deps[i]->outgoing_links.size() > 1) {
 361                                         if (deps[i]->effect->num_inputs() > 0) {
 362                                                 // More than one effect uses this as the input,
 363                                                 // and it is not a texture itself.
 364                                                 // The easiest thing to do (and probably also the safest
 365                                                 // performance-wise in most cases) is to bounce it to a texture
 366                                                 // and then let the next passes read from that.
 367                                                 start_new_phase = true;
 368                                         } else {
 369                                                 // For textures, we try to be slightly more clever;
 370                                                 // if none of our outputs need a bounce, we don't bounce
 371                                                 // but instead simply use the effect many times.
 372                                                 //
 373                                                 // Strictly speaking, we could bounce it for some outputs
 374                                                 // and use it directly for others, but the processing becomes
 375                                                 // somewhat simpler if the effect is only used in one such way.
 376                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 377                                                         Node *rdep = deps[i]->outgoing_links[j];
 378                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
 379                                                 }
 380                                         }
 381                                 }
 382
 383                                 if (deps[i]->effect->changes_output_size()) {
 384                                         start_new_phase = true;
 385                                 }
 386
 387                                 if (start_new_phase) {
 388                                         effects_todo_other_phases.push(deps[i]);
 389                                         this_phase_inputs.push_back(deps[i]);
 390                                 } else {
 391                                         effects_todo_this_phase.push(deps[i]);
 392                                 }
 393                         }
 394                         continue;
 395                 }
 396
 397                 // No more effects to do this phase. Take all the ones we have,
 398                 // and create a GLSL program for it.
 399                 if (!this_phase_effects.empty()) {
 400                         reverse(this_phase_effects.begin(), this_phase_effects.end());
 401                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
 402                         this_phase_effects.back()->phase = phases.back();
 403                         this_phase_inputs.clear();
 404                         this_phase_effects.clear();
 405                 }
 406                 assert(this_phase_inputs.empty());
 407                 assert(this_phase_effects.empty());
 408
 409                 // If we have no effects left, exit.
 410                 if (effects_todo_other_phases.empty()) {
 411                         break;
 412                 }
 413
 414                 Node *node = effects_todo_other_phases.top();
 415                 effects_todo_other_phases.pop();
 416
 417                 if (completed_effects.count(node) == 0) {
 418                         // Start a new phase, calculating from this effect.
 419                         effects_todo_this_phase.push(node);
 420                 }
 421         }
 422
 423         // Finally, since the phases are found from the output but must be executed
 424         // from the input(s), reverse them, too.
 425         std::reverse(phases.begin(), phases.end());
 426 }
 427
 428 void EffectChain::output_dot(const char *filename)
 429 {
 430         FILE *fp = fopen(filename, "w");
 431         if (fp == NULL) {
 432                 perror(filename);
 433                 exit(1);
 434         }
 435
 436         fprintf(fp, "digraph G {\n");
 437         for (unsigned i = 0; i < nodes.size(); ++i) {
 438                 // Find out which phase this event belongs to.
 439                 int in_phase = -1;
 440                 for (unsigned j = 0; j < phases.size(); ++j) {
 441                         const Phase* p = phases[j];
 442                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 443                                 assert(in_phase == -1);
 444                                 in_phase = j;
 445                         }
 446                 }
 447
 448                 if (in_phase == -1) {
 449                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 450                 } else {
 451                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 452                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 453                                 (in_phase % 8) + 1);
 454                 }
 455                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 456                         std::vector<std::string> labels;
 457
 458                         if (nodes[i]->outgoing_links[j]->effect->needs_texture_bounce()) {
 459                                 labels.push_back("needs_bounce");
 460                         }
 461                         if (nodes[i]->effect->changes_output_size()) {
 462                                 labels.push_back("resize");
 463                         }
 464
 465                         switch (nodes[i]->output_color_space) {
 466                         case COLORSPACE_INVALID:
 467                                 labels.push_back("spc[invalid]");
 468                                 break;
 469                         case COLORSPACE_REC_601_525:
 470                                 labels.push_back("spc[rec601-525]");
 471                                 break;
 472                         case COLORSPACE_REC_601_625:
 473                                 labels.push_back("spc[rec601-625]");
 474                                 break;
 475                         default:
 476                                 break;
 477                         }
 478
 479                         switch (nodes[i]->output_gamma_curve) {
 480                         case GAMMA_INVALID:
 481                                 labels.push_back("gamma[invalid]");
 482                                 break;
 483                         case GAMMA_sRGB:
 484                                 labels.push_back("gamma[sRGB]");
 485                                 break;
 486                         case GAMMA_REC_601:  // and GAMMA_REC_709
 487                                 labels.push_back("gamma[rec601/709]");
 488                                 break;
 489                         default:
 490                                 break;
 491                         }
 492
 493                         if (labels.empty()) {
 494                                 fprintf(fp, "  n%ld -> n%ld;\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j]);
 495                         } else {
 496                                 std::string label = labels[0];
 497                                 for (unsigned k = 1; k < labels.size(); ++k) {
 498                                         label += ", " + labels[k];
 499                                 }
 500                                 fprintf(fp, "  n%ld -> n%ld [label=\"%s\"];\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j], label.c_str());
 501                         }
 502                 }
 503         }
 504         fprintf(fp, "}\n");
 505
 506         fclose(fp);
 507 }
 508
 509 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
 510 {
 511         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 512                 // Same aspect, or W/H > aspect (image is wider than the frame).
 513                 // In either case, keep width.
 514                 return width;
 515         } else {
 516                 // W/H < aspect (image is taller than the frame), so keep height,
 517                 // and adjust width correspondingly.
 518                 return lrintf(height * aspect_nom / aspect_denom);
 519         }
 520 }
 521
 522 // Propagate input texture sizes throughout, and inform effects downstream.
 523 // (Like a lot of other code, we depend on effects being in topological order.)
 524 void EffectChain::inform_input_sizes(Phase *phase)
 525 {
 526         // All effects that have a defined size (inputs and RTT inputs)
 527         // get that. Reset all others.
 528         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 529                 Node *node = phase->effects[i];
 530                 if (node->effect->num_inputs() == 0) {
 531                         Input *input = static_cast<Input *>(node->effect);
 532                         node->output_width = input->get_width();
 533                         node->output_height = input->get_height();
 534                         assert(node->output_width != 0);
 535                         assert(node->output_height != 0);
 536                 } else {
 537                         node->output_width = node->output_height = 0;
 538                 }
 539         }
 540         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 541                 Node *input = phase->inputs[i];
 542                 input->output_width = input->phase->output_width;
 543                 input->output_height = input->phase->output_height;
 544                 assert(input->output_width != 0);
 545                 assert(input->output_height != 0);
 546         }
 547
 548         // Now propagate from the inputs towards the end, and inform as we go.
 549         // The rules are simple:
 550         //
 551         //   1. Don't touch effects that already have given sizes (ie., inputs).
 552         //   2. If all of your inputs have the same size, that will be your output size.
 553         //   3. Otherwise, your output size is 0x0.
 554         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 555                 Node *node = phase->effects[i];
 556                 if (node->effect->num_inputs() == 0) {
 557                         continue;
 558                 }
 559                 unsigned this_output_width = 0;
 560                 unsigned this_output_height = 0;
 561                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 562                         Node *input = node->incoming_links[j];
 563                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 564                         if (j == 0) {
 565                                 this_output_width = input->output_width;
 566                                 this_output_height = input->output_height;
 567                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 568                                 // Inputs disagree.
 569                                 this_output_width = 0;
 570                                 this_output_height = 0;
 571                         }
 572                 }
 573                 node->output_width = this_output_width;
 574                 node->output_height = this_output_height;
 575         }
 576 }
 577
 578 // Note: You should call inform_input_sizes() before this, as the last effect's
 579 // desired output size might change based on the inputs.
 580 void EffectChain::find_output_size(Phase *phase)
 581 {
 582         Node *output_node = phase->effects.back();
 583
 584         // If the last effect explicitly sets an output size, use that.
 585         if (output_node->effect->changes_output_size()) {
 586                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
 587                 return;
 588         }
 589
 590         // If not, look at the input phases and textures.
 591         // We select the largest one (by fit into the current aspect).
 592         unsigned best_width = 0;
 593         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 594                 Node *input = phase->inputs[i];
 595                 assert(input->phase->output_width != 0);
 596                 assert(input->phase->output_height != 0);
 597                 unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height);
 598                 if (width > best_width) {
 599                         best_width = width;
 600                 }
 601         }
 602         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 603                 Effect *effect = phase->effects[i]->effect;
 604                 if (effect->num_inputs() != 0) {
 605                         continue;
 606                 }
 607
 608                 Input *input = static_cast<Input *>(effect);
 609                 unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height());
 610                 if (width > best_width) {
 611                         best_width = width;
 612                 }
 613         }
 614         assert(best_width != 0);
 615         phase->output_width = best_width;
 616         phase->output_height = best_width * aspect_denom / aspect_nom;
 617 }
 618
 619 void EffectChain::sort_nodes_topologically()
 620 {
 621         std::set<Node *> visited_nodes;
 622         std::vector<Node *> sorted_list;
 623         for (unsigned i = 0; i < nodes.size(); ++i) {
 624                 if (nodes[i]->incoming_links.size() == 0) {
 625                         topological_sort_visit_node(nodes[i], &visited_nodes, &sorted_list);
 626                 }
 627         }
 628         reverse(sorted_list.begin(), sorted_list.end());
 629         nodes = sorted_list;
 630 }
 631
 632 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *visited_nodes, std::vector<Node *> *sorted_list)
 633 {
 634         if (visited_nodes->count(node) != 0) {
 635                 return;
 636         }
 637         visited_nodes->insert(node);
 638         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 639                 topological_sort_visit_node(node->outgoing_links[i], visited_nodes, sorted_list);
 640         }
 641         sorted_list->push_back(node);
 642 }
 643
 644 // Propagate gamma and color space information as far as we can in the graph.
 645 // The rules are simple: Anything where all the inputs agree, get that as
 646 // output as well. Anything else keeps having *_INVALID.
 647 void EffectChain::propagate_gamma_and_color_space()
 648 {
 649         // We depend on going through the nodes in order.
 650         sort_nodes_topologically();
 651
 652         for (unsigned i = 0; i < nodes.size(); ++i) {
 653                 Node *node = nodes[i];
 654                 if (node->disabled) {
 655                         continue;
 656                 }
 657                 assert(node->incoming_links.size() == node->effect->num_inputs());
 658                 if (node->incoming_links.size() == 0) {
 659                         assert(node->output_color_space != COLORSPACE_INVALID);
 660                         assert(node->output_gamma_curve != GAMMA_INVALID);
 661                         continue;
 662                 }
 663
 664                 Colorspace color_space = node->incoming_links[0]->output_color_space;
 665                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 666                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 667                         if (node->incoming_links[j]->output_color_space != color_space) {
 668                                 color_space = COLORSPACE_INVALID;
 669                         }
 670                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 671                                 gamma_curve = GAMMA_INVALID;
 672                         }
 673                 }
 674
 675                 // The conversion effects already have their outputs set correctly,
 676                 // so leave them alone.
 677                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
 678                         node->output_color_space = color_space;
 679                 }
 680                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 681                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 682                         node->output_gamma_curve = gamma_curve;
 683                 }
 684         }
 685 }
 686
 687 bool EffectChain::node_needs_colorspace_fix(Node *node)
 688 {
 689         if (node->disabled) {
 690                 return false;
 691         }
 692         if (node->effect->num_inputs() == 0) {
 693                 return false;
 694         }
 695
 696         // propagate_gamma_and_color_space() has already set our output
 697         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 698         if (node->output_color_space == COLORSPACE_INVALID) {
 699                 return true;
 700         }
 701         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 702 }
 703
 704 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 705 // the graph. Our strategy is not always optimal, but quite simple:
 706 // Find an effect that's as early as possible where the inputs are of
 707 // unacceptable colorspaces (that is, either different, or, if the effect only
 708 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 709 // propagate the information anew, and repeat until there are no more such
 710 // effects.
 711 void EffectChain::fix_internal_color_spaces()
 712 {
 713         unsigned colorspace_propagation_pass = 0;
 714         bool found_any;
 715         do {
 716                 found_any = false;
 717                 for (unsigned i = 0; i < nodes.size(); ++i) {
 718                         Node *node = nodes[i];
 719                         if (!node_needs_colorspace_fix(node)) {
 720                                 continue;
 721                         }
 722
 723                         // Go through each input that is not sRGB, and insert
 724                         // a colorspace conversion before it.
 725                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 726                                 Node *input = node->incoming_links[j];
 727                                 assert(input->output_color_space != COLORSPACE_INVALID);
 728                                 if (input->output_color_space == COLORSPACE_sRGB) {
 729                                         continue;
 730                                 }
 731                                 Node *conversion = add_node(new ColorspaceConversionEffect());
 732                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
 733                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
 734                                 conversion->output_color_space = COLORSPACE_sRGB;
 735                                 insert_node_between(input, conversion, node);
 736                         }
 737
 738                         // Re-sort topologically, and propagate the new information.
 739                         propagate_gamma_and_color_space();
 740
 741                         found_any = true;
 742                         break;
 743                 }
 744
 745                 char filename[256];
 746                 sprintf(filename, "step3-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
 747                 output_dot(filename);
 748                 assert(colorspace_propagation_pass < 100);
 749         } while (found_any);
 750
 751         for (unsigned i = 0; i < nodes.size(); ++i) {
 752                 Node *node = nodes[i];
 753                 if (node->disabled) {
 754                         continue;
 755                 }
 756                 assert(node->output_color_space != COLORSPACE_INVALID);
 757         }
 758 }
 759
 760 // Make so that the output is in the desired color space.
 761 void EffectChain::fix_output_color_space()
 762 {
 763         Node *output = find_output_node();
 764         if (output->output_color_space != output_format.color_space) {
 765                 Node *conversion = add_node(new ColorspaceConversionEffect());
 766                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
 767                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
 768                 conversion->output_color_space = output_format.color_space;
 769                 connect_nodes(output, conversion);
 770                 propagate_gamma_and_color_space();
 771         }
 772 }
 773
 774 bool EffectChain::node_needs_gamma_fix(Node *node)
 775 {
 776         if (node->disabled) {
 777                 return false;
 778         }
 779
 780         // Small hack since the output is not an explicit node:
 781         // If we are the last node and our output is in the wrong
 782         // space compared to EffectChain's output, we need to fix it.
 783         // This will only take us to linear, but fix_output_gamma()
 784         // will come and take us to the desired output gamma
 785         // if it is needed.
 786         //
 787         // This needs to be before everything else, since it could
 788         // even apply to inputs (if they are the only effect).
 789         if (node->outgoing_links.empty() &&
 790             node->output_gamma_curve != output_format.gamma_curve &&
 791             node->output_gamma_curve != GAMMA_LINEAR) {
 792                 return true;
 793         }
 794
 795         if (node->effect->num_inputs() == 0) {
 796                 return false;
 797         }
 798
 799         // propagate_gamma_and_color_space() has already set our output
 800         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
 801         // except for GammaCompressionEffect.
 802         if (node->output_gamma_curve == GAMMA_INVALID) {
 803                 return true;
 804         }
 805         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
 806                 assert(node->incoming_links.size() == 1);
 807                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
 808         }
 809
 810         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
 811 }
 812
 813 // Very similar to fix_internal_color_spaces(), but for gamma.
 814 // There is one difference, though; before we start adding conversion nodes,
 815 // we see if we can get anything out of asking the sources to deliver
 816 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
 817 // does that part, while fix_internal_gamma_by_inserting_nodes()
 818 // inserts nodes as needed afterwards.
 819 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
 820 {
 821         unsigned gamma_propagation_pass = 0;
 822         bool found_any;
 823         do {
 824                 found_any = false;
 825                 for (unsigned i = 0; i < nodes.size(); ++i) {
 826                         Node *node = nodes[i];
 827                         if (!node_needs_gamma_fix(node)) {
 828                                 continue;
 829                         }
 830
 831                         // See if all inputs can give us linear gamma. If not, leave it.
 832                         std::vector<Node *> nonlinear_inputs;
 833                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
 834                         assert(!nonlinear_inputs.empty());
 835
 836                         bool all_ok = true;
 837                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
 838                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
 839                                 all_ok &= input->can_output_linear_gamma();
 840                         }
 841
 842                         if (!all_ok) {
 843                                 continue;
 844                         }
 845
 846                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
 847                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
 848                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
 849                         }
 850
 851                         // Re-sort topologically, and propagate the new information.
 852                         propagate_gamma_and_color_space();
 853
 854                         found_any = true;
 855                         break;
 856                 }
 857
 858                 char filename[256];
 859                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
 860                 output_dot(filename);
 861                 assert(gamma_propagation_pass < 100);
 862         } while (found_any);
 863 }
 864
 865 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
 866 {
 867         unsigned gamma_propagation_pass = 0;
 868         bool found_any;
 869         do {
 870                 found_any = false;
 871                 for (unsigned i = 0; i < nodes.size(); ++i) {
 872                         Node *node = nodes[i];
 873                         if (!node_needs_gamma_fix(node)) {
 874                                 continue;
 875                         }
 876
 877                         // Special case: We could be an input and still be asked to
 878                         // fix our gamma; if so, we should be the only node
 879                         // (as node_needs_gamma_fix() would only return true in
 880                         // for an input in that case). That means we should insert
 881                         // a conversion node _after_ ourselves.
 882                         if (node->incoming_links.empty()) {
 883                                 assert(node->outgoing_links.empty());
 884                                 Node *conversion = add_node(new GammaExpansionEffect());
 885                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
 886                                 conversion->output_gamma_curve = GAMMA_LINEAR;
 887                                 connect_nodes(node, conversion);
 888                         }
 889
 890                         // If not, go through each input that is not linear gamma,
 891                         // and insert a gamma conversion before it.
 892                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 893                                 Node *input = node->incoming_links[j];
 894                                 assert(input->output_gamma_curve != GAMMA_INVALID);
 895                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
 896                                         continue;
 897                                 }
 898                                 Node *conversion = add_node(new GammaExpansionEffect());
 899                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
 900                                 conversion->output_gamma_curve = GAMMA_LINEAR;
 901                                 insert_node_between(input, conversion, node);
 902                         }
 903
 904                         // Re-sort topologically, and propagate the new information.
 905                         propagate_gamma_and_color_space();
 906
 907                         found_any = true;
 908                         break;
 909                 }
 910
 911                 char filename[256];
 912                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
 913                 output_dot(filename);
 914                 assert(gamma_propagation_pass < 100);
 915         } while (found_any);
 916
 917         for (unsigned i = 0; i < nodes.size(); ++i) {
 918                 Node *node = nodes[i];
 919                 if (node->disabled) {
 920                         continue;
 921                 }
 922                 assert(node->output_gamma_curve != GAMMA_INVALID);
 923         }
 924 }
 925
 926 // Make so that the output is in the desired gamma.
 927 // Note that this assumes linear input gamma, so it might create the need
 928 // for another pass of fix_internal_gamma().
 929 void EffectChain::fix_output_gamma()
 930 {
 931         Node *output = find_output_node();
 932         if (output->output_gamma_curve != output_format.gamma_curve) {
 933                 Node *conversion = add_node(new GammaCompressionEffect());
 934                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
 935                 conversion->output_gamma_curve = output_format.gamma_curve;
 936                 connect_nodes(output, conversion);
 937         }
 938 }
 939
 940 // If the user has requested dither, add a DitherEffect right at the end
 941 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
 942 // since dither is about the only effect that can _not_ be done in linear space.
 943 void EffectChain::add_dither_if_needed()
 944 {
 945         if (num_dither_bits == 0) {
 946                 return;
 947         }
 948         Node *output = find_output_node();
 949         Node *dither = add_node(new DitherEffect());
 950         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
 951         connect_nodes(output, dither);
 952
 953         dither_effect = dither->effect;
 954 }
 955
 956 // Find the output node. This is, simply, one that has no outgoing links.
 957 // If there are multiple ones, the graph is malformed (we do not support
 958 // multiple outputs right now).
 959 Node *EffectChain::find_output_node()
 960 {
 961         std::vector<Node *> output_nodes;
 962         for (unsigned i = 0; i < nodes.size(); ++i) {
 963                 Node *node = nodes[i];
 964                 if (node->disabled) {
 965                         continue;
 966                 }
 967                 if (node->outgoing_links.empty()) {
 968                         output_nodes.push_back(node);
 969                 }
 970         }
 971         assert(output_nodes.size() == 1);
 972         return output_nodes[0];
 973 }
 974
 975 void EffectChain::finalize()
 976 {
 977         // Output the graph as it is before we do any conversions on it.
 978         output_dot("step0-start.dot");
 979
 980         // Give each effect in turn a chance to rewrite its own part of the graph.
 981         // Note that if more effects are added as part of this, they will be
 982         // picked up as part of the same for loop, since they are added at the end.
 983         for (unsigned i = 0; i < nodes.size(); ++i) {
 984                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
 985         }
 986         output_dot("step1-rewritten.dot");
 987
 988         propagate_gamma_and_color_space();
 989         output_dot("step2-propagated.dot");
 990
 991         fix_internal_color_spaces();
 992         fix_output_color_space();
 993         output_dot("step4-output-colorspacefix.dot");
 994
 995         // Note that we need to fix gamma after colorspace conversion,
 996         // because colorspace conversions might create needs for gamma conversions.
 997         // Also, we need to run an extra pass of fix_internal_gamma() after
 998         // fixing the output gamma, as we only have conversions to/from linear.
 999         fix_internal_gamma_by_asking_inputs(5);
1000         fix_internal_gamma_by_inserting_nodes(6);
1001         fix_output_gamma();
1002         output_dot("step7-output-gammafix.dot");
1003         fix_internal_gamma_by_asking_inputs(8);
1004         fix_internal_gamma_by_inserting_nodes(9);
1005
1006         output_dot("step10-before-dither.dot");
1007
1008         add_dither_if_needed();
1009
1010         output_dot("step11-final.dot");
1011
1012         // Construct all needed GLSL programs, starting at the output.
1013         construct_glsl_programs(find_output_node());
1014
1015         output_dot("step12-split-to-phases.dot");
1016
1017         // If we have more than one phase, we need intermediate render-to-texture.
1018         // Construct an FBO, and then as many textures as we need.
1019         // We choose the simplest option of having one texture per output,
1020         // since otherwise this turns into an (albeit simple)
1021         // register allocation problem.
1022         if (phases.size() > 1) {
1023                 glGenFramebuffers(1, &fbo);
1024
1025                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1026                         inform_input_sizes(phases[i]);
1027                         find_output_size(phases[i]);
1028
1029                         Node *output_node = phases[i]->effects.back();
1030                         glGenTextures(1, &output_node->output_texture);
1031                         check_error();
1032                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1033                         check_error();
1034                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1035                         check_error();
1036                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1037                         check_error();
1038                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1039                         check_error();
1040
1041                         output_node->output_texture_width = phases[i]->output_width;
1042                         output_node->output_texture_height = phases[i]->output_height;
1043                 }
1044                 inform_input_sizes(phases.back());
1045         }
1046
1047         for (unsigned i = 0; i < inputs.size(); ++i) {
1048                 inputs[i]->finalize();
1049         }
1050
1051         assert(phases[0]->inputs.empty());
1052
1053         finalized = true;
1054 }
1055
1056 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1057 {
1058         assert(finalized);
1059
1060         // Save original viewport.
1061         GLuint x = 0, y = 0;
1062
1063         if (width == 0 && height == 0) {
1064                 GLint viewport[4];
1065                 glGetIntegerv(GL_VIEWPORT, viewport);
1066                 x = viewport[0];
1067                 y = viewport[1];
1068                 width = viewport[2];
1069                 height = viewport[3];
1070         }
1071
1072         // Basic state.
1073         glDisable(GL_BLEND);
1074         check_error();
1075         glDisable(GL_DEPTH_TEST);
1076         check_error();
1077         glDepthMask(GL_FALSE);
1078         check_error();
1079
1080         glMatrixMode(GL_PROJECTION);
1081         glLoadIdentity();
1082         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1083
1084         glMatrixMode(GL_MODELVIEW);
1085         glLoadIdentity();
1086
1087         if (phases.size() > 1) {
1088                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1089                 check_error();
1090         }
1091
1092         std::set<Node *> generated_mipmaps;
1093
1094         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1095                 // See if the requested output size has changed. If so, we need to recreate
1096                 // the texture (and before we start setting up inputs).
1097                 inform_input_sizes(phases[phase]);
1098                 if (phase != phases.size() - 1) {
1099                         find_output_size(phases[phase]);
1100
1101                         Node *output_node = phases[phase]->effects.back();
1102
1103                         if (output_node->output_texture_width != phases[phase]->output_width ||
1104                             output_node->output_texture_height != phases[phase]->output_height) {
1105                                 glActiveTexture(GL_TEXTURE0);
1106                                 check_error();
1107                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1108                                 check_error();
1109                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1110                                 check_error();
1111                                 glBindTexture(GL_TEXTURE_2D, 0);
1112                                 check_error();
1113
1114                                 output_node->output_texture_width = phases[phase]->output_width;
1115                                 output_node->output_texture_height = phases[phase]->output_height;
1116                         }
1117                 }
1118
1119                 glUseProgram(phases[phase]->glsl_program_num);
1120                 check_error();
1121
1122                 // Set up RTT inputs for this phase.
1123                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1124                         glActiveTexture(GL_TEXTURE0 + sampler);
1125                         Node *input = phases[phase]->inputs[sampler];
1126                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1127                         check_error();
1128                         if (phases[phase]->input_needs_mipmaps) {
1129                                 if (generated_mipmaps.count(input) == 0) {
1130                                         glGenerateMipmap(GL_TEXTURE_2D);
1131                                         check_error();
1132                                         generated_mipmaps.insert(input);
1133                                 }
1134                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1135                                 check_error();
1136                         } else {
1137                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1138                                 check_error();
1139                         }
1140
1141                         std::string texture_name = std::string("tex_") + input->effect_id;
1142                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1143                         check_error();
1144                 }
1145
1146                 // And now the output.
1147                 if (phase == phases.size() - 1) {
1148                         // Last phase goes to the output the user specified.
1149                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1150                         check_error();
1151                         glViewport(x, y, width, height);
1152                         if (dither_effect != NULL) {
1153                                 CHECK(dither_effect->set_int("output_width", width));
1154                                 CHECK(dither_effect->set_int("output_height", height));
1155                         }
1156                 } else {
1157                         Node *output_node = phases[phase]->effects.back();
1158                         glFramebufferTexture2D(
1159                                 GL_FRAMEBUFFER,
1160                                 GL_COLOR_ATTACHMENT0,
1161                                 GL_TEXTURE_2D,
1162                                 output_node->output_texture,
1163                                 0);
1164                         check_error();
1165                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1166                 }
1167
1168                 // Give the required parameters to all the effects.
1169                 unsigned sampler_num = phases[phase]->inputs.size();
1170                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1171                         Node *node = phases[phase]->effects[i];
1172                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1173                         check_error();
1174                 }
1175
1176                 // Now draw!
1177                 glBegin(GL_QUADS);
1178
1179                 glTexCoord2f(0.0f, 0.0f);
1180                 glVertex2f(0.0f, 0.0f);
1181
1182                 glTexCoord2f(1.0f, 0.0f);
1183                 glVertex2f(1.0f, 0.0f);
1184
1185                 glTexCoord2f(1.0f, 1.0f);
1186                 glVertex2f(1.0f, 1.0f);
1187
1188                 glTexCoord2f(0.0f, 1.0f);
1189                 glVertex2f(0.0f, 1.0f);
1190
1191                 glEnd();
1192                 check_error();
1193
1194                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1195                         Node *node = phases[phase]->effects[i];
1196                         node->effect->clear_gl_state();
1197                 }
1198         }
1199 }