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 <locale.h>
   7 #include <assert.h>
   8 #include <GL/glew.h>
   9
  10 #include <algorithm>
  11 #include <set>
  12 #include <stack>
  13 #include <vector>
  14
  15 #include "util.h"
  16 #include "effect_chain.h"
  17 #include "gamma_expansion_effect.h"
  18 #include "gamma_compression_effect.h"
  19 #include "colorspace_conversion_effect.h"
  20 #include "alpha_multiplication_effect.h"
  21 #include "alpha_division_effect.h"
  22 #include "dither_effect.h"
  23 #include "input.h"
  24 #include "init.h"
  25
  26 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
  27         : aspect_nom(aspect_nom),
  28           aspect_denom(aspect_denom),
  29           dither_effect(NULL),
  30           fbo(0),
  31           num_dither_bits(0),
  32           finalized(false) {}
  33
  34 EffectChain::~EffectChain()
  35 {
  36         for (unsigned i = 0; i < nodes.size(); ++i) {
  37                 if (nodes[i]->output_texture != 0) {
  38                         glDeleteTextures(1, &nodes[i]->output_texture);
  39                 }
  40                 delete nodes[i]->effect;
  41                 delete nodes[i];
  42         }
  43         for (unsigned i = 0; i < phases.size(); ++i) {
  44                 glDeleteProgram(phases[i]->glsl_program_num);
  45                 glDeleteShader(phases[i]->vertex_shader);
  46                 glDeleteShader(phases[i]->fragment_shader);
  47                 delete phases[i];
  48         }
  49         if (fbo != 0) {
  50                 glDeleteFramebuffers(1, &fbo);
  51         }
  52 }
  53
  54 Input *EffectChain::add_input(Input *input)
  55 {
  56         inputs.push_back(input);
  57         add_node(input);
  58         return input;
  59 }
  60
  61 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
  62 {
  63         output_format = format;
  64         output_alpha_format = alpha_format;
  65 }
  66
  67 Node *EffectChain::add_node(Effect *effect)
  68 {
  69         char effect_id[256];
  70         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
  71
  72         Node *node = new Node;
  73         node->effect = effect;
  74         node->disabled = false;
  75         node->effect_id = effect_id;
  76         node->output_color_space = COLORSPACE_INVALID;
  77         node->output_gamma_curve = GAMMA_INVALID;
  78         node->output_alpha_type = ALPHA_INVALID;
  79         node->output_texture = 0;
  80
  81         nodes.push_back(node);
  82         node_map[effect] = node;
  83         return node;
  84 }
  85
  86 void EffectChain::connect_nodes(Node *sender, Node *receiver)
  87 {
  88         sender->outgoing_links.push_back(receiver);
  89         receiver->incoming_links.push_back(sender);
  90 }
  91
  92 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
  93 {
  94         new_receiver->incoming_links = old_receiver->incoming_links;
  95         old_receiver->incoming_links.clear();
  96
  97         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
  98                 Node *sender = new_receiver->incoming_links[i];
  99                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
 100                         if (sender->outgoing_links[j] == old_receiver) {
 101                                 sender->outgoing_links[j] = new_receiver;
 102                         }
 103                 }
 104         }
 105 }
 106
 107 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 108 {
 109         new_sender->outgoing_links = old_sender->outgoing_links;
 110         old_sender->outgoing_links.clear();
 111
 112         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 113                 Node *receiver = new_sender->outgoing_links[i];
 114                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 115                         if (receiver->incoming_links[j] == old_sender) {
 116                                 receiver->incoming_links[j] = new_sender;
 117                         }
 118                 }
 119         }
 120 }
 121
 122 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 123 {
 124         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 125                 if (sender->outgoing_links[i] == receiver) {
 126                         sender->outgoing_links[i] = middle;
 127                         middle->incoming_links.push_back(sender);
 128                 }
 129         }
 130         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 131                 if (receiver->incoming_links[i] == sender) {
 132                         receiver->incoming_links[i] = middle;
 133                         middle->outgoing_links.push_back(receiver);
 134                 }
 135         }
 136
 137         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 138 }
 139
 140 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
 141 {
 142         if (node->output_gamma_curve == GAMMA_LINEAR &&
 143             node->effect->effect_type_id() != "GammaCompressionEffect") {
 144                 return;
 145         }
 146         if (node->effect->num_inputs() == 0) {
 147                 nonlinear_inputs->push_back(node);
 148         } else {
 149                 assert(node->effect->num_inputs() == node->incoming_links.size());
 150                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 151                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 152                 }
 153         }
 154 }
 155
 156 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
 157 {
 158         assert(inputs.size() == effect->num_inputs());
 159         Node *node = add_node(effect);
 160         for (unsigned i = 0; i < inputs.size(); ++i) {
 161                 assert(node_map.count(inputs[i]) != 0);
 162                 connect_nodes(node_map[inputs[i]], node);
 163         }
 164         return effect;
 165 }
 166
 167 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 168 std::string replace_prefix(const std::string &text, const std::string &prefix)
 169 {
 170         std::string output;
 171         size_t start = 0;
 172
 173         while (start < text.size()) {
 174                 size_t pos = text.find("PREFIX(", start);
 175                 if (pos == std::string::npos) {
 176                         output.append(text.substr(start, std::string::npos));
 177                         break;
 178                 }
 179
 180                 output.append(text.substr(start, pos - start));
 181                 output.append(prefix);
 182                 output.append("_");
 183
 184                 pos += strlen("PREFIX(");
 185
 186                 // Output stuff until we find the matching ), which we then eat.
 187                 int depth = 1;
 188                 size_t end_arg_pos = pos;
 189                 while (end_arg_pos < text.size()) {
 190                         if (text[end_arg_pos] == '(') {
 191                                 ++depth;
 192                         } else if (text[end_arg_pos] == ')') {
 193                                 --depth;
 194                                 if (depth == 0) {
 195                                         break;
 196                                 }
 197                         }
 198                         ++end_arg_pos;
 199                 }
 200                 output.append(text.substr(pos, end_arg_pos - pos));
 201                 ++end_arg_pos;
 202                 assert(depth == 0);
 203                 start = end_arg_pos;
 204         }
 205         return output;
 206 }
 207
 208 Phase *EffectChain::compile_glsl_program(
 209         const std::vector<Node *> &inputs,
 210         const std::vector<Node *> &effects)
 211 {
 212         assert(!effects.empty());
 213
 214         // Deduplicate the inputs.
 215         std::vector<Node *> true_inputs = inputs;
 216         std::sort(true_inputs.begin(), true_inputs.end());
 217         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
 218
 219         bool input_needs_mipmaps = false;
 220         std::string frag_shader = read_file("header.frag");
 221
 222         // Create functions for all the texture inputs that we need.
 223         for (unsigned i = 0; i < true_inputs.size(); ++i) {
 224                 Node *input = true_inputs[i];
 225
 226                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
 227                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
 228                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
 229                 frag_shader += "}\n";
 230                 frag_shader += "\n";
 231         }
 232
 233         std::vector<Node *> sorted_effects = topological_sort(effects);
 234
 235         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 236                 Node *node = sorted_effects[i];
 237
 238                 if (node->incoming_links.size() == 1) {
 239                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
 240                 } else {
 241                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 242                                 char buf[256];
 243                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
 244                                 frag_shader += buf;
 245                         }
 246                 }
 247
 248                 frag_shader += "\n";
 249                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
 250                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
 251                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
 252                 frag_shader += "#undef PREFIX\n";
 253                 frag_shader += "#undef FUNCNAME\n";
 254                 if (node->incoming_links.size() == 1) {
 255                         frag_shader += "#undef INPUT\n";
 256                 } else {
 257                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 258                                 char buf[256];
 259                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 260                                 frag_shader += buf;
 261                         }
 262                 }
 263                 frag_shader += "\n";
 264
 265                 input_needs_mipmaps |= node->effect->needs_mipmaps();
 266         }
 267         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 268                 Node *node = sorted_effects[i];
 269                 if (node->effect->num_inputs() == 0) {
 270                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
 271                 }
 272         }
 273         frag_shader += std::string("#define INPUT ") + sorted_effects.back()->effect_id + "\n";
 274         frag_shader.append(read_file("footer.frag"));
 275
 276         if (movit_debug_level == MOVIT_DEBUG_ON) {
 277                 // Output shader to a temporary file, for easier debugging.
 278                 static int compiled_shader_num = 0;
 279                 char filename[256];
 280                 sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
 281                 FILE *fp = fopen(filename, "w");
 282                 if (fp == NULL) {
 283                         perror(filename);
 284                         exit(1);
 285                 }
 286                 fprintf(fp, "%s\n", frag_shader.c_str());
 287                 fclose(fp);
 288         }
 289
 290         GLuint glsl_program_num = glCreateProgram();
 291         GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 292         GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
 293         glAttachShader(glsl_program_num, vs_obj);
 294         check_error();
 295         glAttachShader(glsl_program_num, fs_obj);
 296         check_error();
 297         glLinkProgram(glsl_program_num);
 298         check_error();
 299
 300         Phase *phase = new Phase;
 301         phase->glsl_program_num = glsl_program_num;
 302         phase->vertex_shader = vs_obj;
 303         phase->fragment_shader = fs_obj;
 304         phase->input_needs_mipmaps = input_needs_mipmaps;
 305         phase->inputs = true_inputs;
 306         phase->effects = sorted_effects;
 307
 308         return phase;
 309 }
 310
 311 // Construct GLSL programs, starting at the given effect and following
 312 // the chain from there. We end a program every time we come to an effect
 313 // marked as "needs texture bounce", one that is used by multiple other
 314 // effects, every time an effect wants to change the output size,
 315 // and of course at the end.
 316 //
 317 // We follow a quite simple depth-first search from the output, although
 318 // without any explicit recursion.
 319 void EffectChain::construct_glsl_programs(Node *output)
 320 {
 321         // Which effects have already been completed?
 322         // We need to keep track of it, as an effect with multiple outputs
 323         // could otherwise be calculated multiple times.
 324         std::set<Node *> completed_effects;
 325
 326         // Effects in the current phase, as well as inputs (outputs from other phases
 327         // that we depend on). Note that since we start iterating from the end,
 328         // the effect list will be in the reverse order.
 329         std::vector<Node *> this_phase_inputs;
 330         std::vector<Node *> this_phase_effects;
 331
 332         // Effects that we have yet to calculate, but that we know should
 333         // be in the current phase.
 334         std::stack<Node *> effects_todo_this_phase;
 335
 336         // Effects that we have yet to calculate, but that come from other phases.
 337         // We delay these until we have this phase done in its entirety,
 338         // at which point we pick any of them and start a new phase from that.
 339         std::stack<Node *> effects_todo_other_phases;
 340
 341         effects_todo_this_phase.push(output);
 342
 343         for ( ;; ) {  // Termination condition within loop.
 344                 if (!effects_todo_this_phase.empty()) {
 345                         // OK, we have more to do this phase.
 346                         Node *node = effects_todo_this_phase.top();
 347                         effects_todo_this_phase.pop();
 348
 349                         // This should currently only happen for effects that are inputs
 350                         // (either true inputs or phase outputs). We special-case inputs,
 351                         // and then deduplicate phase outputs in compile_glsl_program().
 352                         if (node->effect->num_inputs() == 0) {
 353                                 if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
 354                                         continue;
 355                                 }
 356                         } else {
 357                                 assert(completed_effects.count(node) == 0);
 358                         }
 359
 360                         this_phase_effects.push_back(node);
 361                         completed_effects.insert(node);
 362
 363                         // Find all the dependencies of this effect, and add them to the stack.
 364                         std::vector<Node *> deps = node->incoming_links;
 365                         assert(node->effect->num_inputs() == deps.size());
 366                         for (unsigned i = 0; i < deps.size(); ++i) {
 367                                 bool start_new_phase = false;
 368
 369                                 // FIXME: If we sample directly from a texture, we won't need this.
 370                                 if (node->effect->needs_texture_bounce()) {
 371                                         start_new_phase = true;
 372                                 }
 373
 374                                 if (deps[i]->outgoing_links.size() > 1) {
 375                                         if (deps[i]->effect->num_inputs() > 0) {
 376                                                 // More than one effect uses this as the input,
 377                                                 // and it is not a texture itself.
 378                                                 // The easiest thing to do (and probably also the safest
 379                                                 // performance-wise in most cases) is to bounce it to a texture
 380                                                 // and then let the next passes read from that.
 381                                                 start_new_phase = true;
 382                                         } else {
 383                                                 // For textures, we try to be slightly more clever;
 384                                                 // if none of our outputs need a bounce, we don't bounce
 385                                                 // but instead simply use the effect many times.
 386                                                 //
 387                                                 // Strictly speaking, we could bounce it for some outputs
 388                                                 // and use it directly for others, but the processing becomes
 389                                                 // somewhat simpler if the effect is only used in one such way.
 390                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 391                                                         Node *rdep = deps[i]->outgoing_links[j];
 392                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
 393                                                 }
 394                                         }
 395                                 }
 396
 397                                 if (deps[i]->effect->changes_output_size()) {
 398                                         start_new_phase = true;
 399                                 }
 400
 401                                 if (start_new_phase) {
 402                                         effects_todo_other_phases.push(deps[i]);
 403                                         this_phase_inputs.push_back(deps[i]);
 404                                 } else {
 405                                         effects_todo_this_phase.push(deps[i]);
 406                                 }
 407                         }
 408                         continue;
 409                 }
 410
 411                 // No more effects to do this phase. Take all the ones we have,
 412                 // and create a GLSL program for it.
 413                 if (!this_phase_effects.empty()) {
 414                         reverse(this_phase_effects.begin(), this_phase_effects.end());
 415                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
 416                         this_phase_effects.back()->phase = phases.back();
 417                         this_phase_inputs.clear();
 418                         this_phase_effects.clear();
 419                 }
 420                 assert(this_phase_inputs.empty());
 421                 assert(this_phase_effects.empty());
 422
 423                 // If we have no effects left, exit.
 424                 if (effects_todo_other_phases.empty()) {
 425                         break;
 426                 }
 427
 428                 Node *node = effects_todo_other_phases.top();
 429                 effects_todo_other_phases.pop();
 430
 431                 if (completed_effects.count(node) == 0) {
 432                         // Start a new phase, calculating from this effect.
 433                         effects_todo_this_phase.push(node);
 434                 }
 435         }
 436
 437         // Finally, since the phases are found from the output but must be executed
 438         // from the input(s), reverse them, too.
 439         std::reverse(phases.begin(), phases.end());
 440 }
 441
 442 void EffectChain::output_dot(const char *filename)
 443 {
 444         if (movit_debug_level != MOVIT_DEBUG_ON) {
 445                 return;
 446         }
 447
 448         FILE *fp = fopen(filename, "w");
 449         if (fp == NULL) {
 450                 perror(filename);
 451                 exit(1);
 452         }
 453
 454         fprintf(fp, "digraph G {\n");
 455         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 456         for (unsigned i = 0; i < nodes.size(); ++i) {
 457                 // Find out which phase this event belongs to.
 458                 std::vector<int> in_phases;
 459                 for (unsigned j = 0; j < phases.size(); ++j) {
 460                         const Phase* p = phases[j];
 461                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 462                                 in_phases.push_back(j);
 463                         }
 464                 }
 465
 466                 if (in_phases.empty()) {
 467                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 468                 } else if (in_phases.size() == 1) {
 469                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 470                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 471                                 (in_phases[0] % 8) + 1);
 472                 } else {
 473                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 474                         // But alas.
 475                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 476                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 477                                 (in_phases[0] % 8) + 1);
 478                 }
 479
 480                 char from_node_id[256];
 481                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 482
 483                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 484                         char to_node_id[256];
 485                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 486
 487                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 488                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 489                 }
 490
 491                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 492                         // Output node.
 493                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
 494                         output_dot_edge(fp, from_node_id, "output", labels);
 495                 }
 496         }
 497         fprintf(fp, "}\n");
 498
 499         fclose(fp);
 500 }
 501
 502 std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 503 {
 504         std::vector<std::string> labels;
 505
 506         if (to != NULL && to->effect->needs_texture_bounce()) {
 507                 labels.push_back("needs_bounce");
 508         }
 509         if (from->effect->changes_output_size()) {
 510                 labels.push_back("resize");
 511         }
 512
 513         switch (from->output_color_space) {
 514         case COLORSPACE_INVALID:
 515                 labels.push_back("spc[invalid]");
 516                 break;
 517         case COLORSPACE_REC_601_525:
 518                 labels.push_back("spc[rec601-525]");
 519                 break;
 520         case COLORSPACE_REC_601_625:
 521                 labels.push_back("spc[rec601-625]");
 522                 break;
 523         default:
 524                 break;
 525         }
 526
 527         switch (from->output_gamma_curve) {
 528         case GAMMA_INVALID:
 529                 labels.push_back("gamma[invalid]");
 530                 break;
 531         case GAMMA_sRGB:
 532                 labels.push_back("gamma[sRGB]");
 533                 break;
 534         case GAMMA_REC_601:  // and GAMMA_REC_709
 535                 labels.push_back("gamma[rec601/709]");
 536                 break;
 537         default:
 538                 break;
 539         }
 540
 541         switch (from->output_alpha_type) {
 542         case ALPHA_INVALID:
 543                 labels.push_back("alpha[invalid]");
 544                 break;
 545         case ALPHA_BLANK:
 546                 labels.push_back("alpha[blank]");
 547                 break;
 548         case ALPHA_POSTMULTIPLIED:
 549                 labels.push_back("alpha[postmult]");
 550                 break;
 551         default:
 552                 break;
 553         }
 554
 555         return labels;
 556 }
 557
 558 void EffectChain::output_dot_edge(FILE *fp,
 559                                   const std::string &from_node_id,
 560                                   const std::string &to_node_id,
 561                                   const std::vector<std::string> &labels)
 562 {
 563         if (labels.empty()) {
 564                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 565         } else {
 566                 std::string label = labels[0];
 567                 for (unsigned k = 1; k < labels.size(); ++k) {
 568                         label += ", " + labels[k];
 569                 }
 570                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 571         }
 572 }
 573
 574 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
 575 {
 576         unsigned scaled_width, scaled_height;
 577
 578         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 579                 // Same aspect, or W/H > aspect (image is wider than the frame).
 580                 // In either case, keep width, and adjust height.
 581                 scaled_width = width;
 582                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
 583         } else {
 584                 // W/H < aspect (image is taller than the frame), so keep height,
 585                 // and adjust width.
 586                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
 587                 scaled_height = height;
 588         }
 589
 590         // We should be consistently larger or smaller then the existing choice,
 591         // since we have the same aspect.
 592         assert(!(scaled_width < *output_width && scaled_height > *output_height));
 593         assert(!(scaled_height < *output_height && scaled_width > *output_width));
 594
 595         if (scaled_width >= *output_width && scaled_height >= *output_height) {
 596                 *output_width = scaled_width;
 597                 *output_height = scaled_height;
 598         }
 599 }
 600
 601 // Propagate input texture sizes throughout, and inform effects downstream.
 602 // (Like a lot of other code, we depend on effects being in topological order.)
 603 void EffectChain::inform_input_sizes(Phase *phase)
 604 {
 605         // All effects that have a defined size (inputs and RTT inputs)
 606         // get that. Reset all others.
 607         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 608                 Node *node = phase->effects[i];
 609                 if (node->effect->num_inputs() == 0) {
 610                         Input *input = static_cast<Input *>(node->effect);
 611                         node->output_width = input->get_width();
 612                         node->output_height = input->get_height();
 613                         assert(node->output_width != 0);
 614                         assert(node->output_height != 0);
 615                 } else {
 616                         node->output_width = node->output_height = 0;
 617                 }
 618         }
 619         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 620                 Node *input = phase->inputs[i];
 621                 input->output_width = input->phase->output_width;
 622                 input->output_height = input->phase->output_height;
 623                 assert(input->output_width != 0);
 624                 assert(input->output_height != 0);
 625         }
 626
 627         // Now propagate from the inputs towards the end, and inform as we go.
 628         // The rules are simple:
 629         //
 630         //   1. Don't touch effects that already have given sizes (ie., inputs).
 631         //   2. If all of your inputs have the same size, that will be your output size.
 632         //   3. Otherwise, your output size is 0x0.
 633         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 634                 Node *node = phase->effects[i];
 635                 if (node->effect->num_inputs() == 0) {
 636                         continue;
 637                 }
 638                 unsigned this_output_width = 0;
 639                 unsigned this_output_height = 0;
 640                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 641                         Node *input = node->incoming_links[j];
 642                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 643                         if (j == 0) {
 644                                 this_output_width = input->output_width;
 645                                 this_output_height = input->output_height;
 646                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 647                                 // Inputs disagree.
 648                                 this_output_width = 0;
 649                                 this_output_height = 0;
 650                         }
 651                 }
 652                 node->output_width = this_output_width;
 653                 node->output_height = this_output_height;
 654         }
 655 }
 656
 657 // Note: You should call inform_input_sizes() before this, as the last effect's
 658 // desired output size might change based on the inputs.
 659 void EffectChain::find_output_size(Phase *phase)
 660 {
 661         Node *output_node = phase->effects.back();
 662
 663         // If the last effect explicitly sets an output size, use that.
 664         if (output_node->effect->changes_output_size()) {
 665                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height);
 666                 return;
 667         }
 668
 669         unsigned output_width = 0, output_height = 0;
 670
 671         // If not, look at the input phases and textures.
 672         // We select the largest one (by fit into the current aspect).
 673         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 674                 Node *input = phase->inputs[i];
 675                 assert(input->phase->output_width != 0);
 676                 assert(input->phase->output_height != 0);
 677                 size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
 678         }
 679         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 680                 Effect *effect = phase->effects[i]->effect;
 681                 if (effect->num_inputs() != 0) {
 682                         continue;
 683                 }
 684
 685                 Input *input = static_cast<Input *>(effect);
 686                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
 687         }
 688         assert(output_width != 0);
 689         assert(output_height != 0);
 690         phase->output_width = output_width;
 691         phase->output_height = output_height;
 692 }
 693
 694 void EffectChain::sort_all_nodes_topologically()
 695 {
 696         nodes = topological_sort(nodes);
 697 }
 698
 699 std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
 700 {
 701         std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
 702         std::vector<Node *> sorted_list;
 703         for (unsigned i = 0; i < nodes.size(); ++i) {
 704                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
 705         }
 706         reverse(sorted_list.begin(), sorted_list.end());
 707         return sorted_list;
 708 }
 709
 710 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
 711 {
 712         if (nodes_left_to_visit->count(node) == 0) {
 713                 return;
 714         }
 715         nodes_left_to_visit->erase(node);
 716         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 717                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
 718         }
 719         sorted_list->push_back(node);
 720 }
 721
 722 void EffectChain::find_color_spaces_for_inputs()
 723 {
 724         for (unsigned i = 0; i < nodes.size(); ++i) {
 725                 Node *node = nodes[i];
 726                 if (node->disabled) {
 727                         continue;
 728                 }
 729                 if (node->incoming_links.size() == 0) {
 730                         Input *input = static_cast<Input *>(node->effect);
 731                         node->output_color_space = input->get_color_space();
 732                         node->output_gamma_curve = input->get_gamma_curve();
 733
 734                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
 735                         switch (alpha_handling) {
 736                         case Effect::OUTPUT_BLANK_ALPHA:
 737                                 node->output_alpha_type = ALPHA_BLANK;
 738                                 break;
 739                         case Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED:
 740                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 741                                 break;
 742                         case Effect::OUTPUT_ALPHA_POSTMULTIPLIED:
 743                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 744                                 break;
 745                         case Effect::DONT_CARE_ALPHA_TYPE:
 746                         default:
 747                                 assert(false);
 748                         }
 749
 750                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
 751                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
 752                         }
 753                 }
 754         }
 755 }
 756
 757 // Propagate gamma and color space information as far as we can in the graph.
 758 // The rules are simple: Anything where all the inputs agree, get that as
 759 // output as well. Anything else keeps having *_INVALID.
 760 void EffectChain::propagate_gamma_and_color_space()
 761 {
 762         // We depend on going through the nodes in order.
 763         sort_all_nodes_topologically();
 764
 765         for (unsigned i = 0; i < nodes.size(); ++i) {
 766                 Node *node = nodes[i];
 767                 if (node->disabled) {
 768                         continue;
 769                 }
 770                 assert(node->incoming_links.size() == node->effect->num_inputs());
 771                 if (node->incoming_links.size() == 0) {
 772                         assert(node->output_color_space != COLORSPACE_INVALID);
 773                         assert(node->output_gamma_curve != GAMMA_INVALID);
 774                         continue;
 775                 }
 776
 777                 Colorspace color_space = node->incoming_links[0]->output_color_space;
 778                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 779                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 780                         if (node->incoming_links[j]->output_color_space != color_space) {
 781                                 color_space = COLORSPACE_INVALID;
 782                         }
 783                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 784                                 gamma_curve = GAMMA_INVALID;
 785                         }
 786                 }
 787
 788                 // The conversion effects already have their outputs set correctly,
 789                 // so leave them alone.
 790                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
 791                         node->output_color_space = color_space;
 792                 }
 793                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 794                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 795                         node->output_gamma_curve = gamma_curve;
 796                 }
 797         }
 798 }
 799
 800 // Propagate alpha information as far as we can in the graph.
 801 // Similar to propagate_gamma_and_color_space().
 802 void EffectChain::propagate_alpha()
 803 {
 804         // We depend on going through the nodes in order.
 805         sort_all_nodes_topologically();
 806
 807         for (unsigned i = 0; i < nodes.size(); ++i) {
 808                 Node *node = nodes[i];
 809                 if (node->disabled) {
 810                         continue;
 811                 }
 812                 assert(node->incoming_links.size() == node->effect->num_inputs());
 813                 if (node->incoming_links.size() == 0) {
 814                         assert(node->output_alpha_type != ALPHA_INVALID);
 815                         continue;
 816                 }
 817
 818                 // The alpha multiplication/division effects are special cases.
 819                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
 820                         assert(node->incoming_links.size() == 1);
 821                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
 822                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
 823                         continue;
 824                 }
 825                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
 826                         assert(node->incoming_links.size() == 1);
 827                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
 828                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 829                         continue;
 830                 }
 831
 832                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
 833                 // because they are the only one that _need_ postmultiplied alpha.
 834                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
 835                     node->effect->effect_type_id() == "GammaExpansionEffect") {
 836                         assert(node->incoming_links.size() == 1);
 837                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
 838                                 node->output_alpha_type = ALPHA_BLANK;
 839                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
 840                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 841                         } else {
 842                                 node->output_alpha_type = ALPHA_INVALID;
 843                         }
 844                         continue;
 845                 }
 846
 847                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
 848                 // or OUTPUT_ALPHA_POSTMULTIPLIED), and they have already been
 849                 // taken care of above. Rationale: Even if you could imagine
 850                 // e.g. an effect that took in an image and set alpha=1.0
 851                 // unconditionally, it wouldn't make any sense to have it as
 852                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
 853                 // got its input pre- or postmultiplied, so it wouldn't know
 854                 // whether to divide away the old alpha or not.
 855                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
 856                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED ||
 857                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 858
 859                 // If the node has multiple inputs, check that they are all valid and
 860                 // the same.
 861                 bool any_invalid = false;
 862                 bool any_premultiplied = false;
 863                 bool any_postmultiplied = false;
 864
 865                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 866                         switch (node->incoming_links[j]->output_alpha_type) {
 867                         case ALPHA_INVALID:
 868                                 any_invalid = true;
 869                                 break;
 870                         case ALPHA_BLANK:
 871                                 // Blank is good as both pre- and postmultiplied alpha,
 872                                 // so just ignore it.
 873                                 break;
 874                         case ALPHA_PREMULTIPLIED:
 875                                 any_premultiplied = true;
 876                                 break;
 877                         case ALPHA_POSTMULTIPLIED:
 878                                 any_postmultiplied = true;
 879                                 break;
 880                         default:
 881                                 assert(false);
 882                         }
 883                 }
 884
 885                 if (any_invalid) {
 886                         node->output_alpha_type = ALPHA_INVALID;
 887                         continue;
 888                 }
 889
 890                 // Inputs must be of the same type.
 891                 if (any_premultiplied && any_postmultiplied) {
 892                         node->output_alpha_type = ALPHA_INVALID;
 893                         continue;
 894                 }
 895
 896                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_ALPHA_PREMULTIPLIED) {
 897                         // If the effect has asked for premultiplied alpha, check that it has got it.
 898                         if (any_postmultiplied) {
 899                                 node->output_alpha_type = ALPHA_INVALID;
 900                         } else {
 901                                 // In some rare cases, it might be advantageous to say
 902                                 // that blank input alpha yields blank output alpha.
 903                                 // However, this would cause a more complex Effect interface
 904                                 // an effect would need to guarantee that it doesn't mess with
 905                                 // blank alpha), so this is the simplest.
 906                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 907                         }
 908                 } else {
 909                         // OK, all inputs are the same, and this effect is not going
 910                         // to change it.
 911                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 912                         if (any_premultiplied) {
 913                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 914                         } else if (any_postmultiplied) {
 915                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 916                         } else {
 917                                 node->output_alpha_type = ALPHA_BLANK;
 918                         }
 919                 }
 920         }
 921 }
 922
 923 bool EffectChain::node_needs_colorspace_fix(Node *node)
 924 {
 925         if (node->disabled) {
 926                 return false;
 927         }
 928         if (node->effect->num_inputs() == 0) {
 929                 return false;
 930         }
 931
 932         // propagate_gamma_and_color_space() has already set our output
 933         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 934         if (node->output_color_space == COLORSPACE_INVALID) {
 935                 return true;
 936         }
 937         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 938 }
 939
 940 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 941 // the graph. Our strategy is not always optimal, but quite simple:
 942 // Find an effect that's as early as possible where the inputs are of
 943 // unacceptable colorspaces (that is, either different, or, if the effect only
 944 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 945 // propagate the information anew, and repeat until there are no more such
 946 // effects.
 947 void EffectChain::fix_internal_color_spaces()
 948 {
 949         unsigned colorspace_propagation_pass = 0;
 950         bool found_any;
 951         do {
 952                 found_any = false;
 953                 for (unsigned i = 0; i < nodes.size(); ++i) {
 954                         Node *node = nodes[i];
 955                         if (!node_needs_colorspace_fix(node)) {
 956                                 continue;
 957                         }
 958
 959                         // Go through each input that is not sRGB, and insert
 960                         // a colorspace conversion after it.
 961                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 962                                 Node *input = node->incoming_links[j];
 963                                 assert(input->output_color_space != COLORSPACE_INVALID);
 964                                 if (input->output_color_space == COLORSPACE_sRGB) {
 965                                         continue;
 966                                 }
 967                                 Node *conversion = add_node(new ColorspaceConversionEffect());
 968                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
 969                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
 970                                 conversion->output_color_space = COLORSPACE_sRGB;
 971                                 replace_sender(input, conversion);
 972                                 connect_nodes(input, conversion);
 973                         }
 974
 975                         // Re-sort topologically, and propagate the new information.
 976                         propagate_gamma_and_color_space();
 977
 978                         found_any = true;
 979                         break;
 980                 }
 981
 982                 char filename[256];
 983                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
 984                 output_dot(filename);
 985                 assert(colorspace_propagation_pass < 100);
 986         } while (found_any);
 987
 988         for (unsigned i = 0; i < nodes.size(); ++i) {
 989                 Node *node = nodes[i];
 990                 if (node->disabled) {
 991                         continue;
 992                 }
 993                 assert(node->output_color_space != COLORSPACE_INVALID);
 994         }
 995 }
 996
 997 bool EffectChain::node_needs_alpha_fix(Node *node)
 998 {
 999         if (node->disabled) {
1000                 return false;
1001         }
1002
1003         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1004         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1005         return (node->output_alpha_type == ALPHA_INVALID);
1006 }
1007
1008 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1009 // the graph. Similar to fix_internal_color_spaces().
1010 void EffectChain::fix_internal_alpha(unsigned step)
1011 {
1012         unsigned alpha_propagation_pass = 0;
1013         bool found_any;
1014         do {
1015                 found_any = false;
1016                 for (unsigned i = 0; i < nodes.size(); ++i) {
1017                         Node *node = nodes[i];
1018                         if (!node_needs_alpha_fix(node)) {
1019                                 continue;
1020                         }
1021
1022                         // If we need to fix up GammaExpansionEffect, then clearly something
1023                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1024                         // is meaningless.
1025                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1026
1027                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1028
1029                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1030                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1031                                 assert(node->incoming_links.size() == 1);
1032                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1033                                 desired_type = ALPHA_POSTMULTIPLIED;
1034                         }
1035
1036                         // Go through each input that is not premultiplied alpha, and insert
1037                         // a conversion before it.
1038                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1039                                 Node *input = node->incoming_links[j];
1040                                 assert(input->output_alpha_type != ALPHA_INVALID);
1041                                 if (input->output_alpha_type == desired_type ||
1042                                     input->output_alpha_type == ALPHA_BLANK) {
1043                                         continue;
1044                                 }
1045                                 Node *conversion;
1046                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1047                                         conversion = add_node(new AlphaMultiplicationEffect());
1048                                 } else {
1049                                         conversion = add_node(new AlphaDivisionEffect());
1050                                 }
1051                                 conversion->output_alpha_type = desired_type;
1052                                 replace_sender(input, conversion);
1053                                 connect_nodes(input, conversion);
1054                         }
1055
1056                         // Re-sort topologically, and propagate the new information.
1057                         propagate_gamma_and_color_space();
1058                         propagate_alpha();
1059
1060                         found_any = true;
1061                         break;
1062                 }
1063
1064                 char filename[256];
1065                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1066                 output_dot(filename);
1067                 assert(alpha_propagation_pass < 100);
1068         } while (found_any);
1069
1070         for (unsigned i = 0; i < nodes.size(); ++i) {
1071                 Node *node = nodes[i];
1072                 if (node->disabled) {
1073                         continue;
1074                 }
1075                 assert(node->output_alpha_type != ALPHA_INVALID);
1076         }
1077 }
1078
1079 // Make so that the output is in the desired color space.
1080 void EffectChain::fix_output_color_space()
1081 {
1082         Node *output = find_output_node();
1083         if (output->output_color_space != output_format.color_space) {
1084                 Node *conversion = add_node(new ColorspaceConversionEffect());
1085                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1086                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1087                 conversion->output_color_space = output_format.color_space;
1088                 connect_nodes(output, conversion);
1089                 propagate_alpha();
1090                 propagate_gamma_and_color_space();
1091         }
1092 }
1093
1094 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1095 void EffectChain::fix_output_alpha()
1096 {
1097         Node *output = find_output_node();
1098         assert(output->output_alpha_type != ALPHA_INVALID);
1099         if (output->output_alpha_type == ALPHA_BLANK) {
1100                 // No alpha output, so we don't care.
1101                 return;
1102         }
1103         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1104             output_alpha_format == OUTPUT_ALPHA_POSTMULTIPLIED) {
1105                 Node *conversion = add_node(new AlphaDivisionEffect());
1106                 connect_nodes(output, conversion);
1107                 propagate_alpha();
1108                 propagate_gamma_and_color_space();
1109         }
1110         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1111             output_alpha_format == OUTPUT_ALPHA_PREMULTIPLIED) {
1112                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1113                 connect_nodes(output, conversion);
1114                 propagate_alpha();
1115                 propagate_gamma_and_color_space();
1116         }
1117 }
1118
1119 bool EffectChain::node_needs_gamma_fix(Node *node)
1120 {
1121         if (node->disabled) {
1122                 return false;
1123         }
1124
1125         // Small hack since the output is not an explicit node:
1126         // If we are the last node and our output is in the wrong
1127         // space compared to EffectChain's output, we need to fix it.
1128         // This will only take us to linear, but fix_output_gamma()
1129         // will come and take us to the desired output gamma
1130         // if it is needed.
1131         //
1132         // This needs to be before everything else, since it could
1133         // even apply to inputs (if they are the only effect).
1134         if (node->outgoing_links.empty() &&
1135             node->output_gamma_curve != output_format.gamma_curve &&
1136             node->output_gamma_curve != GAMMA_LINEAR) {
1137                 return true;
1138         }
1139
1140         if (node->effect->num_inputs() == 0) {
1141                 return false;
1142         }
1143
1144         // propagate_gamma_and_color_space() has already set our output
1145         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1146         // except for GammaCompressionEffect.
1147         if (node->output_gamma_curve == GAMMA_INVALID) {
1148                 return true;
1149         }
1150         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1151                 assert(node->incoming_links.size() == 1);
1152                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1153         }
1154
1155         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1156 }
1157
1158 // Very similar to fix_internal_color_spaces(), but for gamma.
1159 // There is one difference, though; before we start adding conversion nodes,
1160 // we see if we can get anything out of asking the sources to deliver
1161 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1162 // does that part, while fix_internal_gamma_by_inserting_nodes()
1163 // inserts nodes as needed afterwards.
1164 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1165 {
1166         unsigned gamma_propagation_pass = 0;
1167         bool found_any;
1168         do {
1169                 found_any = false;
1170                 for (unsigned i = 0; i < nodes.size(); ++i) {
1171                         Node *node = nodes[i];
1172                         if (!node_needs_gamma_fix(node)) {
1173                                 continue;
1174                         }
1175
1176                         // See if all inputs can give us linear gamma. If not, leave it.
1177                         std::vector<Node *> nonlinear_inputs;
1178                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1179                         assert(!nonlinear_inputs.empty());
1180
1181                         bool all_ok = true;
1182                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1183                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1184                                 all_ok &= input->can_output_linear_gamma();
1185                         }
1186
1187                         if (!all_ok) {
1188                                 continue;
1189                         }
1190
1191                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1192                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1193                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1194                         }
1195
1196                         // Re-sort topologically, and propagate the new information.
1197                         propagate_gamma_and_color_space();
1198
1199                         found_any = true;
1200                         break;
1201                 }
1202
1203                 char filename[256];
1204                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1205                 output_dot(filename);
1206                 assert(gamma_propagation_pass < 100);
1207         } while (found_any);
1208 }
1209
1210 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1211 {
1212         unsigned gamma_propagation_pass = 0;
1213         bool found_any;
1214         do {
1215                 found_any = false;
1216                 for (unsigned i = 0; i < nodes.size(); ++i) {
1217                         Node *node = nodes[i];
1218                         if (!node_needs_gamma_fix(node)) {
1219                                 continue;
1220                         }
1221
1222                         // Special case: We could be an input and still be asked to
1223                         // fix our gamma; if so, we should be the only node
1224                         // (as node_needs_gamma_fix() would only return true in
1225                         // for an input in that case). That means we should insert
1226                         // a conversion node _after_ ourselves.
1227                         if (node->incoming_links.empty()) {
1228                                 assert(node->outgoing_links.empty());
1229                                 Node *conversion = add_node(new GammaExpansionEffect());
1230                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1231                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1232                                 connect_nodes(node, conversion);
1233                         }
1234
1235                         // If not, go through each input that is not linear gamma,
1236                         // and insert a gamma conversion after it.
1237                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1238                                 Node *input = node->incoming_links[j];
1239                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1240                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1241                                         continue;
1242                                 }
1243                                 Node *conversion = add_node(new GammaExpansionEffect());
1244                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1245                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1246                                 replace_sender(input, conversion);
1247                                 connect_nodes(input, conversion);
1248                         }
1249
1250                         // Re-sort topologically, and propagate the new information.
1251                         propagate_alpha();
1252                         propagate_gamma_and_color_space();
1253
1254                         found_any = true;
1255                         break;
1256                 }
1257
1258                 char filename[256];
1259                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1260                 output_dot(filename);
1261                 assert(gamma_propagation_pass < 100);
1262         } while (found_any);
1263
1264         for (unsigned i = 0; i < nodes.size(); ++i) {
1265                 Node *node = nodes[i];
1266                 if (node->disabled) {
1267                         continue;
1268                 }
1269                 assert(node->output_gamma_curve != GAMMA_INVALID);
1270         }
1271 }
1272
1273 // Make so that the output is in the desired gamma.
1274 // Note that this assumes linear input gamma, so it might create the need
1275 // for another pass of fix_internal_gamma().
1276 void EffectChain::fix_output_gamma()
1277 {
1278         Node *output = find_output_node();
1279         if (output->output_gamma_curve != output_format.gamma_curve) {
1280                 Node *conversion = add_node(new GammaCompressionEffect());
1281                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1282                 conversion->output_gamma_curve = output_format.gamma_curve;
1283                 connect_nodes(output, conversion);
1284         }
1285 }
1286
1287 // If the user has requested dither, add a DitherEffect right at the end
1288 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1289 // since dither is about the only effect that can _not_ be done in linear space.
1290 void EffectChain::add_dither_if_needed()
1291 {
1292         if (num_dither_bits == 0) {
1293                 return;
1294         }
1295         Node *output = find_output_node();
1296         Node *dither = add_node(new DitherEffect());
1297         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1298         connect_nodes(output, dither);
1299
1300         dither_effect = dither->effect;
1301 }
1302
1303 // Find the output node. This is, simply, one that has no outgoing links.
1304 // If there are multiple ones, the graph is malformed (we do not support
1305 // multiple outputs right now).
1306 Node *EffectChain::find_output_node()
1307 {
1308         std::vector<Node *> output_nodes;
1309         for (unsigned i = 0; i < nodes.size(); ++i) {
1310                 Node *node = nodes[i];
1311                 if (node->disabled) {
1312                         continue;
1313                 }
1314                 if (node->outgoing_links.empty()) {
1315                         output_nodes.push_back(node);
1316                 }
1317         }
1318         assert(output_nodes.size() == 1);
1319         return output_nodes[0];
1320 }
1321
1322 void EffectChain::finalize()
1323 {
1324         // Save the current locale, and set it to C, so that we can output decimal
1325         // numbers with printf and be sure to get them in the format mandated by GLSL.
1326         char *saved_locale = setlocale(LC_NUMERIC, "C");
1327
1328         // Output the graph as it is before we do any conversions on it.
1329         output_dot("step0-start.dot");
1330
1331         // Give each effect in turn a chance to rewrite its own part of the graph.
1332         // Note that if more effects are added as part of this, they will be
1333         // picked up as part of the same for loop, since they are added at the end.
1334         for (unsigned i = 0; i < nodes.size(); ++i) {
1335                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1336         }
1337         output_dot("step1-rewritten.dot");
1338
1339         find_color_spaces_for_inputs();
1340         output_dot("step2-input-colorspace.dot");
1341
1342         propagate_alpha();
1343         output_dot("step3-propagated-alpha.dot");
1344
1345         propagate_gamma_and_color_space();
1346         output_dot("step4-propagated-all.dot");
1347
1348         fix_internal_color_spaces();
1349         fix_internal_alpha(6);
1350         fix_output_color_space();
1351         output_dot("step7-output-colorspacefix.dot");
1352         fix_output_alpha();
1353         output_dot("step8-output-alphafix.dot");
1354
1355         // Note that we need to fix gamma after colorspace conversion,
1356         // because colorspace conversions might create needs for gamma conversions.
1357         // Also, we need to run an extra pass of fix_internal_gamma() after
1358         // fixing the output gamma, as we only have conversions to/from linear,
1359         // and fix_internal_alpha() since GammaCompressionEffect needs
1360         // postmultiplied input.
1361         fix_internal_gamma_by_asking_inputs(9);
1362         fix_internal_gamma_by_inserting_nodes(10);
1363         fix_output_gamma();
1364         output_dot("step11-output-gammafix.dot");
1365         propagate_alpha();
1366         output_dot("step12-output-alpha-propagated.dot");
1367         fix_internal_alpha(13);
1368         output_dot("step14-output-alpha-fixed.dot");
1369         fix_internal_gamma_by_asking_inputs(15);
1370         fix_internal_gamma_by_inserting_nodes(16);
1371
1372         output_dot("step17-before-dither.dot");
1373
1374         add_dither_if_needed();
1375
1376         output_dot("step18-final.dot");
1377
1378         // Construct all needed GLSL programs, starting at the output.
1379         construct_glsl_programs(find_output_node());
1380
1381         output_dot("step19-split-to-phases.dot");
1382
1383         // If we have more than one phase, we need intermediate render-to-texture.
1384         // Construct an FBO, and then as many textures as we need.
1385         // We choose the simplest option of having one texture per output,
1386         // since otherwise this turns into an (albeit simple)
1387         // register allocation problem.
1388         if (phases.size() > 1) {
1389                 glGenFramebuffers(1, &fbo);
1390
1391                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1392                         inform_input_sizes(phases[i]);
1393                         find_output_size(phases[i]);
1394
1395                         Node *output_node = phases[i]->effects.back();
1396                         glGenTextures(1, &output_node->output_texture);
1397                         check_error();
1398                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1399                         check_error();
1400                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1401                         check_error();
1402                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1403                         check_error();
1404                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1405                         check_error();
1406
1407                         output_node->output_texture_width = phases[i]->output_width;
1408                         output_node->output_texture_height = phases[i]->output_height;
1409                 }
1410                 inform_input_sizes(phases.back());
1411         }
1412
1413         for (unsigned i = 0; i < inputs.size(); ++i) {
1414                 inputs[i]->finalize();
1415         }
1416
1417         assert(phases[0]->inputs.empty());
1418
1419         finalized = true;
1420         setlocale(LC_NUMERIC, saved_locale);
1421 }
1422
1423 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1424 {
1425         assert(finalized);
1426
1427         // Save original viewport.
1428         GLuint x = 0, y = 0;
1429
1430         if (width == 0 && height == 0) {
1431                 GLint viewport[4];
1432                 glGetIntegerv(GL_VIEWPORT, viewport);
1433                 x = viewport[0];
1434                 y = viewport[1];
1435                 width = viewport[2];
1436                 height = viewport[3];
1437         }
1438
1439         // Basic state.
1440         glDisable(GL_BLEND);
1441         check_error();
1442         glDisable(GL_DEPTH_TEST);
1443         check_error();
1444         glDepthMask(GL_FALSE);
1445         check_error();
1446
1447         glMatrixMode(GL_PROJECTION);
1448         glLoadIdentity();
1449         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1450
1451         glMatrixMode(GL_MODELVIEW);
1452         glLoadIdentity();
1453
1454         if (phases.size() > 1) {
1455                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1456                 check_error();
1457         }
1458
1459         std::set<Node *> generated_mipmaps;
1460
1461         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1462                 // See if the requested output size has changed. If so, we need to recreate
1463                 // the texture (and before we start setting up inputs).
1464                 inform_input_sizes(phases[phase]);
1465                 if (phase != phases.size() - 1) {
1466                         find_output_size(phases[phase]);
1467
1468                         Node *output_node = phases[phase]->effects.back();
1469
1470                         if (output_node->output_texture_width != phases[phase]->output_width ||
1471                             output_node->output_texture_height != phases[phase]->output_height) {
1472                                 glActiveTexture(GL_TEXTURE0);
1473                                 check_error();
1474                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1475                                 check_error();
1476                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1477                                 check_error();
1478                                 glBindTexture(GL_TEXTURE_2D, 0);
1479                                 check_error();
1480
1481                                 output_node->output_texture_width = phases[phase]->output_width;
1482                                 output_node->output_texture_height = phases[phase]->output_height;
1483                         }
1484                 }
1485
1486                 glUseProgram(phases[phase]->glsl_program_num);
1487                 check_error();
1488
1489                 // Set up RTT inputs for this phase.
1490                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1491                         glActiveTexture(GL_TEXTURE0 + sampler);
1492                         Node *input = phases[phase]->inputs[sampler];
1493                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1494                         check_error();
1495                         if (phases[phase]->input_needs_mipmaps) {
1496                                 if (generated_mipmaps.count(input) == 0) {
1497                                         glGenerateMipmap(GL_TEXTURE_2D);
1498                                         check_error();
1499                                         generated_mipmaps.insert(input);
1500                                 }
1501                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1502                                 check_error();
1503                         } else {
1504                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1505                                 check_error();
1506                         }
1507
1508                         std::string texture_name = std::string("tex_") + input->effect_id;
1509                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1510                         check_error();
1511                 }
1512
1513                 // And now the output.
1514                 if (phase == phases.size() - 1) {
1515                         // Last phase goes to the output the user specified.
1516                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1517                         check_error();
1518                         glViewport(x, y, width, height);
1519                         if (dither_effect != NULL) {
1520                                 CHECK(dither_effect->set_int("output_width", width));
1521                                 CHECK(dither_effect->set_int("output_height", height));
1522                         }
1523                 } else {
1524                         Node *output_node = phases[phase]->effects.back();
1525                         glFramebufferTexture2D(
1526                                 GL_FRAMEBUFFER,
1527                                 GL_COLOR_ATTACHMENT0,
1528                                 GL_TEXTURE_2D,
1529                                 output_node->output_texture,
1530                                 0);
1531                         check_error();
1532                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1533                 }
1534
1535                 // Give the required parameters to all the effects.
1536                 unsigned sampler_num = phases[phase]->inputs.size();
1537                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1538                         Node *node = phases[phase]->effects[i];
1539                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1540                         check_error();
1541                 }
1542
1543                 // Now draw!
1544                 glBegin(GL_QUADS);
1545
1546                 glTexCoord2f(0.0f, 0.0f);
1547                 glVertex2f(0.0f, 0.0f);
1548
1549                 glTexCoord2f(1.0f, 0.0f);
1550                 glVertex2f(1.0f, 0.0f);
1551
1552                 glTexCoord2f(1.0f, 1.0f);
1553                 glVertex2f(1.0f, 1.0f);
1554
1555                 glTexCoord2f(0.0f, 1.0f);
1556                 glVertex2f(0.0f, 1.0f);
1557
1558                 glEnd();
1559                 check_error();
1560
1561                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1562                         Node *node = phases[phase]->effects[i];
1563                         node->effect->clear_gl_state();
1564                 }
1565         }
1566 }