git.sesse.net Git - movit/blob - effect_chain.cpp

   1 #define GL_GLEXT_PROTOTYPES 1
   2
   3 #include <GL/glew.h>
   4 #include <assert.h>
   5 #include <locale.h>
   6 #include <math.h>
   7 #include <stddef.h>
   8 #include <stdio.h>
   9 #include <stdlib.h>
  10 #include <string.h>
  11 #include <algorithm>
  12 #include <set>
  13 #include <stack>
  14 #include <vector>
  15
  16 #include "alpha_division_effect.h"
  17 #include "alpha_multiplication_effect.h"
  18 #include "colorspace_conversion_effect.h"
  19 #include "dither_effect.h"
  20 #include "effect.h"
  21 #include "effect_chain.h"
  22 #include "gamma_compression_effect.h"
  23 #include "gamma_expansion_effect.h"
  24 #include "init.h"
  25 #include "input.h"
  26 #include "resource_pool.h"
  27 #include "util.h"
  28
  29 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
  30         : aspect_nom(aspect_nom),
  31           aspect_denom(aspect_denom),
  32           dither_effect(NULL),
  33           num_dither_bits(0),
  34           finalized(false),
  35           resource_pool(resource_pool) {
  36         if (resource_pool == NULL) {
  37                 this->resource_pool = new ResourcePool();
  38                 owns_resource_pool = true;
  39         } else {
  40                 owns_resource_pool = false;
  41         }
  42 }
  43
  44 EffectChain::~EffectChain()
  45 {
  46         for (unsigned i = 0; i < nodes.size(); ++i) {
  47                 if (nodes[i]->output_texture != 0) {
  48                         glDeleteTextures(1, &nodes[i]->output_texture);
  49                 }
  50                 delete nodes[i]->effect;
  51                 delete nodes[i];
  52         }
  53         for (unsigned i = 0; i < phases.size(); ++i) {
  54                 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
  55                 delete phases[i];
  56         }
  57         if (owns_resource_pool) {
  58                 delete resource_pool;
  59         }
  60 }
  61
  62 Input *EffectChain::add_input(Input *input)
  63 {
  64         assert(!finalized);
  65         inputs.push_back(input);
  66         add_node(input);
  67         return input;
  68 }
  69
  70 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
  71 {
  72         assert(!finalized);
  73         output_format = format;
  74         output_alpha_format = alpha_format;
  75 }
  76
  77 Node *EffectChain::add_node(Effect *effect)
  78 {
  79         for (unsigned i = 0; i < nodes.size(); ++i) {
  80                 assert(nodes[i]->effect != effect);
  81         }
  82
  83         char effect_id[256];
  84         sprintf(effect_id, "eff%u", (unsigned)nodes.size());
  85
  86         Node *node = new Node;
  87         node->effect = effect;
  88         node->disabled = false;
  89         node->effect_id = effect_id;
  90         node->output_color_space = COLORSPACE_INVALID;
  91         node->output_gamma_curve = GAMMA_INVALID;
  92         node->output_alpha_type = ALPHA_INVALID;
  93         node->output_texture = 0;
  94
  95         nodes.push_back(node);
  96         node_map[effect] = node;
  97         return node;
  98 }
  99
 100 void EffectChain::connect_nodes(Node *sender, Node *receiver)
 101 {
 102         sender->outgoing_links.push_back(receiver);
 103         receiver->incoming_links.push_back(sender);
 104 }
 105
 106 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
 107 {
 108         new_receiver->incoming_links = old_receiver->incoming_links;
 109         old_receiver->incoming_links.clear();
 110
 111         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
 112                 Node *sender = new_receiver->incoming_links[i];
 113                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
 114                         if (sender->outgoing_links[j] == old_receiver) {
 115                                 sender->outgoing_links[j] = new_receiver;
 116                         }
 117                 }
 118         }
 119 }
 120
 121 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 122 {
 123         new_sender->outgoing_links = old_sender->outgoing_links;
 124         old_sender->outgoing_links.clear();
 125
 126         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 127                 Node *receiver = new_sender->outgoing_links[i];
 128                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 129                         if (receiver->incoming_links[j] == old_sender) {
 130                                 receiver->incoming_links[j] = new_sender;
 131                         }
 132                 }
 133         }
 134 }
 135
 136 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 137 {
 138         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 139                 if (sender->outgoing_links[i] == receiver) {
 140                         sender->outgoing_links[i] = middle;
 141                         middle->incoming_links.push_back(sender);
 142                 }
 143         }
 144         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 145                 if (receiver->incoming_links[i] == sender) {
 146                         receiver->incoming_links[i] = middle;
 147                         middle->outgoing_links.push_back(receiver);
 148                 }
 149         }
 150
 151         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 152 }
 153
 154 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
 155 {
 156         if (node->output_gamma_curve == GAMMA_LINEAR &&
 157             node->effect->effect_type_id() != "GammaCompressionEffect") {
 158                 return;
 159         }
 160         if (node->effect->num_inputs() == 0) {
 161                 nonlinear_inputs->push_back(node);
 162         } else {
 163                 assert(node->effect->num_inputs() == node->incoming_links.size());
 164                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 165                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 166                 }
 167         }
 168 }
 169
 170 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
 171 {
 172         assert(!finalized);
 173         assert(inputs.size() == effect->num_inputs());
 174         Node *node = add_node(effect);
 175         for (unsigned i = 0; i < inputs.size(); ++i) {
 176                 assert(node_map.count(inputs[i]) != 0);
 177                 connect_nodes(node_map[inputs[i]], node);
 178         }
 179         return effect;
 180 }
 181
 182 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 183 std::string replace_prefix(const std::string &text, const std::string &prefix)
 184 {
 185         std::string output;
 186         size_t start = 0;
 187
 188         while (start < text.size()) {
 189                 size_t pos = text.find("PREFIX(", start);
 190                 if (pos == std::string::npos) {
 191                         output.append(text.substr(start, std::string::npos));
 192                         break;
 193                 }
 194
 195                 output.append(text.substr(start, pos - start));
 196                 output.append(prefix);
 197                 output.append("_");
 198
 199                 pos += strlen("PREFIX(");
 200
 201                 // Output stuff until we find the matching ), which we then eat.
 202                 int depth = 1;
 203                 size_t end_arg_pos = pos;
 204                 while (end_arg_pos < text.size()) {
 205                         if (text[end_arg_pos] == '(') {
 206                                 ++depth;
 207                         } else if (text[end_arg_pos] == ')') {
 208                                 --depth;
 209                                 if (depth == 0) {
 210                                         break;
 211                                 }
 212                         }
 213                         ++end_arg_pos;
 214                 }
 215                 output.append(text.substr(pos, end_arg_pos - pos));
 216                 ++end_arg_pos;
 217                 assert(depth == 0);
 218                 start = end_arg_pos;
 219         }
 220         return output;
 221 }
 222
 223 Phase *EffectChain::compile_glsl_program(
 224         const std::vector<Node *> &inputs,
 225         const std::vector<Node *> &effects)
 226 {
 227         assert(!effects.empty());
 228
 229         // Deduplicate the inputs.
 230         std::vector<Node *> true_inputs = inputs;
 231         std::sort(true_inputs.begin(), true_inputs.end());
 232         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
 233
 234         bool input_needs_mipmaps = false;
 235         std::string frag_shader = read_file("header.frag");
 236
 237         // Create functions for all the texture inputs that we need.
 238         for (unsigned i = 0; i < true_inputs.size(); ++i) {
 239                 Node *input = true_inputs[i];
 240
 241                 frag_shader += std::string("uniform sampler2D tex_") + input->effect_id + ";\n";
 242                 frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n";
 243                 frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n";
 244                 frag_shader += "}\n";
 245                 frag_shader += "\n";
 246         }
 247
 248         std::vector<Node *> sorted_effects = topological_sort(effects);
 249
 250         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 251                 Node *node = sorted_effects[i];
 252
 253                 if (node->incoming_links.size() == 1) {
 254                         frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
 255                 } else {
 256                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 257                                 char buf[256];
 258                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, node->incoming_links[j]->effect_id.c_str());
 259                                 frag_shader += buf;
 260                         }
 261                 }
 262
 263                 frag_shader += "\n";
 264                 frag_shader += std::string("#define FUNCNAME ") + node->effect_id + "\n";
 265                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id);
 266                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id);
 267                 frag_shader += "#undef PREFIX\n";
 268                 frag_shader += "#undef FUNCNAME\n";
 269                 if (node->incoming_links.size() == 1) {
 270                         frag_shader += "#undef INPUT\n";
 271                 } else {
 272                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 273                                 char buf[256];
 274                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 275                                 frag_shader += buf;
 276                         }
 277                 }
 278                 frag_shader += "\n";
 279
 280                 input_needs_mipmaps |= node->effect->needs_mipmaps();
 281         }
 282         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 283                 Node *node = sorted_effects[i];
 284                 if (node->effect->num_inputs() == 0) {
 285                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
 286                 }
 287         }
 288         frag_shader += std::string("#define INPUT ") + sorted_effects.back()->effect_id + "\n";
 289         frag_shader.append(read_file("footer.frag"));
 290
 291         Phase *phase = new Phase;
 292         phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader);
 293         phase->input_needs_mipmaps = input_needs_mipmaps;
 294         phase->inputs = true_inputs;
 295         phase->effects = sorted_effects;
 296
 297         return phase;
 298 }
 299
 300 // Construct GLSL programs, starting at the given effect and following
 301 // the chain from there. We end a program every time we come to an effect
 302 // marked as "needs texture bounce", one that is used by multiple other
 303 // effects, every time an effect wants to change the output size,
 304 // and of course at the end.
 305 //
 306 // We follow a quite simple depth-first search from the output, although
 307 // without any explicit recursion.
 308 void EffectChain::construct_glsl_programs(Node *output)
 309 {
 310         // Which effects have already been completed?
 311         // We need to keep track of it, as an effect with multiple outputs
 312         // could otherwise be calculated multiple times.
 313         std::set<Node *> completed_effects;
 314
 315         // Effects in the current phase, as well as inputs (outputs from other phases
 316         // that we depend on). Note that since we start iterating from the end,
 317         // the effect list will be in the reverse order.
 318         std::vector<Node *> this_phase_inputs;
 319         std::vector<Node *> this_phase_effects;
 320
 321         // Effects that we have yet to calculate, but that we know should
 322         // be in the current phase.
 323         std::stack<Node *> effects_todo_this_phase;
 324
 325         // Effects that we have yet to calculate, but that come from other phases.
 326         // We delay these until we have this phase done in its entirety,
 327         // at which point we pick any of them and start a new phase from that.
 328         std::stack<Node *> effects_todo_other_phases;
 329
 330         effects_todo_this_phase.push(output);
 331
 332         for ( ;; ) {  // Termination condition within loop.
 333                 if (!effects_todo_this_phase.empty()) {
 334                         // OK, we have more to do this phase.
 335                         Node *node = effects_todo_this_phase.top();
 336                         effects_todo_this_phase.pop();
 337
 338                         // This should currently only happen for effects that are inputs
 339                         // (either true inputs or phase outputs). We special-case inputs,
 340                         // and then deduplicate phase outputs in compile_glsl_program().
 341                         if (node->effect->num_inputs() == 0) {
 342                                 if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
 343                                         continue;
 344                                 }
 345                         } else {
 346                                 assert(completed_effects.count(node) == 0);
 347                         }
 348
 349                         this_phase_effects.push_back(node);
 350                         completed_effects.insert(node);
 351
 352                         // Find all the dependencies of this effect, and add them to the stack.
 353                         std::vector<Node *> deps = node->incoming_links;
 354                         assert(node->effect->num_inputs() == deps.size());
 355                         for (unsigned i = 0; i < deps.size(); ++i) {
 356                                 bool start_new_phase = false;
 357
 358                                 // FIXME: If we sample directly from a texture, we won't need this.
 359                                 if (node->effect->needs_texture_bounce()) {
 360                                         start_new_phase = true;
 361                                 }
 362
 363                                 if (deps[i]->outgoing_links.size() > 1) {
 364                                         if (deps[i]->effect->num_inputs() > 0) {
 365                                                 // More than one effect uses this as the input,
 366                                                 // and it is not a texture itself.
 367                                                 // The easiest thing to do (and probably also the safest
 368                                                 // performance-wise in most cases) is to bounce it to a texture
 369                                                 // and then let the next passes read from that.
 370                                                 start_new_phase = true;
 371                                         } else {
 372                                                 // For textures, we try to be slightly more clever;
 373                                                 // if none of our outputs need a bounce, we don't bounce
 374                                                 // but instead simply use the effect many times.
 375                                                 //
 376                                                 // Strictly speaking, we could bounce it for some outputs
 377                                                 // and use it directly for others, but the processing becomes
 378                                                 // somewhat simpler if the effect is only used in one such way.
 379                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 380                                                         Node *rdep = deps[i]->outgoing_links[j];
 381                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
 382                                                 }
 383                                         }
 384                                 }
 385
 386                                 if (deps[i]->effect->changes_output_size()) {
 387                                         start_new_phase = true;
 388                                 }
 389
 390                                 if (start_new_phase) {
 391                                         effects_todo_other_phases.push(deps[i]);
 392                                         this_phase_inputs.push_back(deps[i]);
 393                                 } else {
 394                                         effects_todo_this_phase.push(deps[i]);
 395                                 }
 396                         }
 397                         continue;
 398                 }
 399
 400                 // No more effects to do this phase. Take all the ones we have,
 401                 // and create a GLSL program for it.
 402                 if (!this_phase_effects.empty()) {
 403                         reverse(this_phase_effects.begin(), this_phase_effects.end());
 404                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
 405                         this_phase_effects.back()->phase = phases.back();
 406                         this_phase_inputs.clear();
 407                         this_phase_effects.clear();
 408                 }
 409                 assert(this_phase_inputs.empty());
 410                 assert(this_phase_effects.empty());
 411
 412                 // If we have no effects left, exit.
 413                 if (effects_todo_other_phases.empty()) {
 414                         break;
 415                 }
 416
 417                 Node *node = effects_todo_other_phases.top();
 418                 effects_todo_other_phases.pop();
 419
 420                 if (completed_effects.count(node) == 0) {
 421                         // Start a new phase, calculating from this effect.
 422                         effects_todo_this_phase.push(node);
 423                 }
 424         }
 425
 426         // Finally, since the phases are found from the output but must be executed
 427         // from the input(s), reverse them, too.
 428         std::reverse(phases.begin(), phases.end());
 429 }
 430
 431 void EffectChain::output_dot(const char *filename)
 432 {
 433         if (movit_debug_level != MOVIT_DEBUG_ON) {
 434                 return;
 435         }
 436
 437         FILE *fp = fopen(filename, "w");
 438         if (fp == NULL) {
 439                 perror(filename);
 440                 exit(1);
 441         }
 442
 443         fprintf(fp, "digraph G {\n");
 444         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 445         for (unsigned i = 0; i < nodes.size(); ++i) {
 446                 // Find out which phase this event belongs to.
 447                 std::vector<int> in_phases;
 448                 for (unsigned j = 0; j < phases.size(); ++j) {
 449                         const Phase* p = phases[j];
 450                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 451                                 in_phases.push_back(j);
 452                         }
 453                 }
 454
 455                 if (in_phases.empty()) {
 456                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 457                 } else if (in_phases.size() == 1) {
 458                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 459                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 460                                 (in_phases[0] % 8) + 1);
 461                 } else {
 462                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 463                         // But alas.
 464                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 465                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 466                                 (in_phases[0] % 8) + 1);
 467                 }
 468
 469                 char from_node_id[256];
 470                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 471
 472                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 473                         char to_node_id[256];
 474                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 475
 476                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 477                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 478                 }
 479
 480                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 481                         // Output node.
 482                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
 483                         output_dot_edge(fp, from_node_id, "output", labels);
 484                 }
 485         }
 486         fprintf(fp, "}\n");
 487
 488         fclose(fp);
 489 }
 490
 491 std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 492 {
 493         std::vector<std::string> labels;
 494
 495         if (to != NULL && to->effect->needs_texture_bounce()) {
 496                 labels.push_back("needs_bounce");
 497         }
 498         if (from->effect->changes_output_size()) {
 499                 labels.push_back("resize");
 500         }
 501
 502         switch (from->output_color_space) {
 503         case COLORSPACE_INVALID:
 504                 labels.push_back("spc[invalid]");
 505                 break;
 506         case COLORSPACE_REC_601_525:
 507                 labels.push_back("spc[rec601-525]");
 508                 break;
 509         case COLORSPACE_REC_601_625:
 510                 labels.push_back("spc[rec601-625]");
 511                 break;
 512         default:
 513                 break;
 514         }
 515
 516         switch (from->output_gamma_curve) {
 517         case GAMMA_INVALID:
 518                 labels.push_back("gamma[invalid]");
 519                 break;
 520         case GAMMA_sRGB:
 521                 labels.push_back("gamma[sRGB]");
 522                 break;
 523         case GAMMA_REC_601:  // and GAMMA_REC_709
 524                 labels.push_back("gamma[rec601/709]");
 525                 break;
 526         default:
 527                 break;
 528         }
 529
 530         switch (from->output_alpha_type) {
 531         case ALPHA_INVALID:
 532                 labels.push_back("alpha[invalid]");
 533                 break;
 534         case ALPHA_BLANK:
 535                 labels.push_back("alpha[blank]");
 536                 break;
 537         case ALPHA_POSTMULTIPLIED:
 538                 labels.push_back("alpha[postmult]");
 539                 break;
 540         default:
 541                 break;
 542         }
 543
 544         return labels;
 545 }
 546
 547 void EffectChain::output_dot_edge(FILE *fp,
 548                                   const std::string &from_node_id,
 549                                   const std::string &to_node_id,
 550                                   const std::vector<std::string> &labels)
 551 {
 552         if (labels.empty()) {
 553                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 554         } else {
 555                 std::string label = labels[0];
 556                 for (unsigned k = 1; k < labels.size(); ++k) {
 557                         label += ", " + labels[k];
 558                 }
 559                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 560         }
 561 }
 562
 563 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
 564 {
 565         unsigned scaled_width, scaled_height;
 566
 567         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 568                 // Same aspect, or W/H > aspect (image is wider than the frame).
 569                 // In either case, keep width, and adjust height.
 570                 scaled_width = width;
 571                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
 572         } else {
 573                 // W/H < aspect (image is taller than the frame), so keep height,
 574                 // and adjust width.
 575                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
 576                 scaled_height = height;
 577         }
 578
 579         // We should be consistently larger or smaller then the existing choice,
 580         // since we have the same aspect.
 581         assert(!(scaled_width < *output_width && scaled_height > *output_height));
 582         assert(!(scaled_height < *output_height && scaled_width > *output_width));
 583
 584         if (scaled_width >= *output_width && scaled_height >= *output_height) {
 585                 *output_width = scaled_width;
 586                 *output_height = scaled_height;
 587         }
 588 }
 589
 590 // Propagate input texture sizes throughout, and inform effects downstream.
 591 // (Like a lot of other code, we depend on effects being in topological order.)
 592 void EffectChain::inform_input_sizes(Phase *phase)
 593 {
 594         // All effects that have a defined size (inputs and RTT inputs)
 595         // get that. Reset all others.
 596         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 597                 Node *node = phase->effects[i];
 598                 if (node->effect->num_inputs() == 0) {
 599                         Input *input = static_cast<Input *>(node->effect);
 600                         node->output_width = input->get_width();
 601                         node->output_height = input->get_height();
 602                         assert(node->output_width != 0);
 603                         assert(node->output_height != 0);
 604                 } else {
 605                         node->output_width = node->output_height = 0;
 606                 }
 607         }
 608         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 609                 Node *input = phase->inputs[i];
 610                 input->output_width = input->phase->virtual_output_width;
 611                 input->output_height = input->phase->virtual_output_height;
 612                 assert(input->output_width != 0);
 613                 assert(input->output_height != 0);
 614         }
 615
 616         // Now propagate from the inputs towards the end, and inform as we go.
 617         // The rules are simple:
 618         //
 619         //   1. Don't touch effects that already have given sizes (ie., inputs).
 620         //   2. If all of your inputs have the same size, that will be your output size.
 621         //   3. Otherwise, your output size is 0x0.
 622         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 623                 Node *node = phase->effects[i];
 624                 if (node->effect->num_inputs() == 0) {
 625                         continue;
 626                 }
 627                 unsigned this_output_width = 0;
 628                 unsigned this_output_height = 0;
 629                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 630                         Node *input = node->incoming_links[j];
 631                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 632                         if (j == 0) {
 633                                 this_output_width = input->output_width;
 634                                 this_output_height = input->output_height;
 635                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 636                                 // Inputs disagree.
 637                                 this_output_width = 0;
 638                                 this_output_height = 0;
 639                         }
 640                 }
 641                 node->output_width = this_output_width;
 642                 node->output_height = this_output_height;
 643         }
 644 }
 645
 646 // Note: You should call inform_input_sizes() before this, as the last effect's
 647 // desired output size might change based on the inputs.
 648 void EffectChain::find_output_size(Phase *phase)
 649 {
 650         Node *output_node = phase->effects.back();
 651
 652         // If the last effect explicitly sets an output size, use that.
 653         if (output_node->effect->changes_output_size()) {
 654                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
 655                                                      &phase->virtual_output_width, &phase->virtual_output_height);
 656                 return;
 657         }
 658
 659         // If all effects have the same size, use that.
 660         unsigned output_width = 0, output_height = 0;
 661         bool all_inputs_same_size = true;
 662
 663         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 664                 Node *input = phase->inputs[i];
 665                 assert(input->phase->output_width != 0);
 666                 assert(input->phase->output_height != 0);
 667                 if (output_width == 0 && output_height == 0) {
 668                         output_width = input->phase->virtual_output_width;
 669                         output_height = input->phase->virtual_output_height;
 670                 } else if (output_width != input->phase->virtual_output_width ||
 671                            output_height != input->phase->virtual_output_height) {
 672                         all_inputs_same_size = false;
 673                 }
 674         }
 675         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 676                 Effect *effect = phase->effects[i]->effect;
 677                 if (effect->num_inputs() != 0) {
 678                         continue;
 679                 }
 680
 681                 Input *input = static_cast<Input *>(effect);
 682                 if (output_width == 0 && output_height == 0) {
 683                         output_width = input->get_width();
 684                         output_height = input->get_height();
 685                 } else if (output_width != input->get_width() ||
 686                            output_height != input->get_height()) {
 687                         all_inputs_same_size = false;
 688                 }
 689         }
 690
 691         if (all_inputs_same_size) {
 692                 assert(output_width != 0);
 693                 assert(output_height != 0);
 694                 phase->virtual_output_width = phase->output_width = output_width;
 695                 phase->virtual_output_height = phase->output_height = output_height;
 696                 return;
 697         }
 698
 699         // If not, fit all the inputs into the current aspect, and select the largest one.
 700         output_width = 0;
 701         output_height = 0;
 702         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 703                 Node *input = phase->inputs[i];
 704                 assert(input->phase->output_width != 0);
 705                 assert(input->phase->output_height != 0);
 706                 size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
 707         }
 708         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 709                 Effect *effect = phase->effects[i]->effect;
 710                 if (effect->num_inputs() != 0) {
 711                         continue;
 712                 }
 713
 714                 Input *input = static_cast<Input *>(effect);
 715                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
 716         }
 717         assert(output_width != 0);
 718         assert(output_height != 0);
 719         phase->virtual_output_width = phase->output_width = output_width;
 720         phase->virtual_output_height = phase->output_height = output_height;
 721 }
 722
 723 void EffectChain::sort_all_nodes_topologically()
 724 {
 725         nodes = topological_sort(nodes);
 726 }
 727
 728 std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
 729 {
 730         std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
 731         std::vector<Node *> sorted_list;
 732         for (unsigned i = 0; i < nodes.size(); ++i) {
 733                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
 734         }
 735         reverse(sorted_list.begin(), sorted_list.end());
 736         return sorted_list;
 737 }
 738
 739 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
 740 {
 741         if (nodes_left_to_visit->count(node) == 0) {
 742                 return;
 743         }
 744         nodes_left_to_visit->erase(node);
 745         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 746                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
 747         }
 748         sorted_list->push_back(node);
 749 }
 750
 751 void EffectChain::find_color_spaces_for_inputs()
 752 {
 753         for (unsigned i = 0; i < nodes.size(); ++i) {
 754                 Node *node = nodes[i];
 755                 if (node->disabled) {
 756                         continue;
 757                 }
 758                 if (node->incoming_links.size() == 0) {
 759                         Input *input = static_cast<Input *>(node->effect);
 760                         node->output_color_space = input->get_color_space();
 761                         node->output_gamma_curve = input->get_gamma_curve();
 762
 763                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
 764                         switch (alpha_handling) {
 765                         case Effect::OUTPUT_BLANK_ALPHA:
 766                                 node->output_alpha_type = ALPHA_BLANK;
 767                                 break;
 768                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
 769                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 770                                 break;
 771                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
 772                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 773                                 break;
 774                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
 775                         case Effect::DONT_CARE_ALPHA_TYPE:
 776                         default:
 777                                 assert(false);
 778                         }
 779
 780                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
 781                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
 782                         }
 783                 }
 784         }
 785 }
 786
 787 // Propagate gamma and color space information as far as we can in the graph.
 788 // The rules are simple: Anything where all the inputs agree, get that as
 789 // output as well. Anything else keeps having *_INVALID.
 790 void EffectChain::propagate_gamma_and_color_space()
 791 {
 792         // We depend on going through the nodes in order.
 793         sort_all_nodes_topologically();
 794
 795         for (unsigned i = 0; i < nodes.size(); ++i) {
 796                 Node *node = nodes[i];
 797                 if (node->disabled) {
 798                         continue;
 799                 }
 800                 assert(node->incoming_links.size() == node->effect->num_inputs());
 801                 if (node->incoming_links.size() == 0) {
 802                         assert(node->output_color_space != COLORSPACE_INVALID);
 803                         assert(node->output_gamma_curve != GAMMA_INVALID);
 804                         continue;
 805                 }
 806
 807                 Colorspace color_space = node->incoming_links[0]->output_color_space;
 808                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 809                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 810                         if (node->incoming_links[j]->output_color_space != color_space) {
 811                                 color_space = COLORSPACE_INVALID;
 812                         }
 813                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 814                                 gamma_curve = GAMMA_INVALID;
 815                         }
 816                 }
 817
 818                 // The conversion effects already have their outputs set correctly,
 819                 // so leave them alone.
 820                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
 821                         node->output_color_space = color_space;
 822                 }
 823                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 824                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 825                         node->output_gamma_curve = gamma_curve;
 826                 }
 827         }
 828 }
 829
 830 // Propagate alpha information as far as we can in the graph.
 831 // Similar to propagate_gamma_and_color_space().
 832 void EffectChain::propagate_alpha()
 833 {
 834         // We depend on going through the nodes in order.
 835         sort_all_nodes_topologically();
 836
 837         for (unsigned i = 0; i < nodes.size(); ++i) {
 838                 Node *node = nodes[i];
 839                 if (node->disabled) {
 840                         continue;
 841                 }
 842                 assert(node->incoming_links.size() == node->effect->num_inputs());
 843                 if (node->incoming_links.size() == 0) {
 844                         assert(node->output_alpha_type != ALPHA_INVALID);
 845                         continue;
 846                 }
 847
 848                 // The alpha multiplication/division effects are special cases.
 849                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
 850                         assert(node->incoming_links.size() == 1);
 851                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
 852                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
 853                         continue;
 854                 }
 855                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
 856                         assert(node->incoming_links.size() == 1);
 857                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
 858                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 859                         continue;
 860                 }
 861
 862                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
 863                 // because they are the only one that _need_ postmultiplied alpha.
 864                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
 865                     node->effect->effect_type_id() == "GammaExpansionEffect") {
 866                         assert(node->incoming_links.size() == 1);
 867                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
 868                                 node->output_alpha_type = ALPHA_BLANK;
 869                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
 870                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 871                         } else {
 872                                 node->output_alpha_type = ALPHA_INVALID;
 873                         }
 874                         continue;
 875                 }
 876
 877                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
 878                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
 879                 // taken care of above. Rationale: Even if you could imagine
 880                 // e.g. an effect that took in an image and set alpha=1.0
 881                 // unconditionally, it wouldn't make any sense to have it as
 882                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
 883                 // got its input pre- or postmultiplied, so it wouldn't know
 884                 // whether to divide away the old alpha or not.
 885                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
 886                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 887                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
 888                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 889
 890                 // If the node has multiple inputs, check that they are all valid and
 891                 // the same.
 892                 bool any_invalid = false;
 893                 bool any_premultiplied = false;
 894                 bool any_postmultiplied = false;
 895
 896                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 897                         switch (node->incoming_links[j]->output_alpha_type) {
 898                         case ALPHA_INVALID:
 899                                 any_invalid = true;
 900                                 break;
 901                         case ALPHA_BLANK:
 902                                 // Blank is good as both pre- and postmultiplied alpha,
 903                                 // so just ignore it.
 904                                 break;
 905                         case ALPHA_PREMULTIPLIED:
 906                                 any_premultiplied = true;
 907                                 break;
 908                         case ALPHA_POSTMULTIPLIED:
 909                                 any_postmultiplied = true;
 910                                 break;
 911                         default:
 912                                 assert(false);
 913                         }
 914                 }
 915
 916                 if (any_invalid) {
 917                         node->output_alpha_type = ALPHA_INVALID;
 918                         continue;
 919                 }
 920
 921                 // Inputs must be of the same type.
 922                 if (any_premultiplied && any_postmultiplied) {
 923                         node->output_alpha_type = ALPHA_INVALID;
 924                         continue;
 925                 }
 926
 927                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 928                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 929                         // If the effect has asked for premultiplied alpha, check that it has got it.
 930                         if (any_postmultiplied) {
 931                                 node->output_alpha_type = ALPHA_INVALID;
 932                         } else if (!any_premultiplied &&
 933                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 934                                 // Blank input alpha, and the effect preserves blank alpha.
 935                                 node->output_alpha_type = ALPHA_BLANK;
 936                         } else {
 937                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 938                         }
 939                 } else {
 940                         // OK, all inputs are the same, and this effect is not going
 941                         // to change it.
 942                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 943                         if (any_premultiplied) {
 944                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 945                         } else if (any_postmultiplied) {
 946                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 947                         } else {
 948                                 node->output_alpha_type = ALPHA_BLANK;
 949                         }
 950                 }
 951         }
 952 }
 953
 954 bool EffectChain::node_needs_colorspace_fix(Node *node)
 955 {
 956         if (node->disabled) {
 957                 return false;
 958         }
 959         if (node->effect->num_inputs() == 0) {
 960                 return false;
 961         }
 962
 963         // propagate_gamma_and_color_space() has already set our output
 964         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 965         if (node->output_color_space == COLORSPACE_INVALID) {
 966                 return true;
 967         }
 968         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 969 }
 970
 971 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 972 // the graph. Our strategy is not always optimal, but quite simple:
 973 // Find an effect that's as early as possible where the inputs are of
 974 // unacceptable colorspaces (that is, either different, or, if the effect only
 975 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 976 // propagate the information anew, and repeat until there are no more such
 977 // effects.
 978 void EffectChain::fix_internal_color_spaces()
 979 {
 980         unsigned colorspace_propagation_pass = 0;
 981         bool found_any;
 982         do {
 983                 found_any = false;
 984                 for (unsigned i = 0; i < nodes.size(); ++i) {
 985                         Node *node = nodes[i];
 986                         if (!node_needs_colorspace_fix(node)) {
 987                                 continue;
 988                         }
 989
 990                         // Go through each input that is not sRGB, and insert
 991                         // a colorspace conversion after it.
 992                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 993                                 Node *input = node->incoming_links[j];
 994                                 assert(input->output_color_space != COLORSPACE_INVALID);
 995                                 if (input->output_color_space == COLORSPACE_sRGB) {
 996                                         continue;
 997                                 }
 998                                 Node *conversion = add_node(new ColorspaceConversionEffect());
 999                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1000                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1001                                 conversion->output_color_space = COLORSPACE_sRGB;
1002                                 replace_sender(input, conversion);
1003                                 connect_nodes(input, conversion);
1004                         }
1005
1006                         // Re-sort topologically, and propagate the new information.
1007                         propagate_gamma_and_color_space();
1008
1009                         found_any = true;
1010                         break;
1011                 }
1012
1013                 char filename[256];
1014                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1015                 output_dot(filename);
1016                 assert(colorspace_propagation_pass < 100);
1017         } while (found_any);
1018
1019         for (unsigned i = 0; i < nodes.size(); ++i) {
1020                 Node *node = nodes[i];
1021                 if (node->disabled) {
1022                         continue;
1023                 }
1024                 assert(node->output_color_space != COLORSPACE_INVALID);
1025         }
1026 }
1027
1028 bool EffectChain::node_needs_alpha_fix(Node *node)
1029 {
1030         if (node->disabled) {
1031                 return false;
1032         }
1033
1034         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1035         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1036         return (node->output_alpha_type == ALPHA_INVALID);
1037 }
1038
1039 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1040 // the graph. Similar to fix_internal_color_spaces().
1041 void EffectChain::fix_internal_alpha(unsigned step)
1042 {
1043         unsigned alpha_propagation_pass = 0;
1044         bool found_any;
1045         do {
1046                 found_any = false;
1047                 for (unsigned i = 0; i < nodes.size(); ++i) {
1048                         Node *node = nodes[i];
1049                         if (!node_needs_alpha_fix(node)) {
1050                                 continue;
1051                         }
1052
1053                         // If we need to fix up GammaExpansionEffect, then clearly something
1054                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1055                         // is meaningless.
1056                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1057
1058                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1059
1060                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1061                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1062                                 assert(node->incoming_links.size() == 1);
1063                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1064                                 desired_type = ALPHA_POSTMULTIPLIED;
1065                         }
1066
1067                         // Go through each input that is not premultiplied alpha, and insert
1068                         // a conversion before it.
1069                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1070                                 Node *input = node->incoming_links[j];
1071                                 assert(input->output_alpha_type != ALPHA_INVALID);
1072                                 if (input->output_alpha_type == desired_type ||
1073                                     input->output_alpha_type == ALPHA_BLANK) {
1074                                         continue;
1075                                 }
1076                                 Node *conversion;
1077                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1078                                         conversion = add_node(new AlphaMultiplicationEffect());
1079                                 } else {
1080                                         conversion = add_node(new AlphaDivisionEffect());
1081                                 }
1082                                 conversion->output_alpha_type = desired_type;
1083                                 replace_sender(input, conversion);
1084                                 connect_nodes(input, conversion);
1085                         }
1086
1087                         // Re-sort topologically, and propagate the new information.
1088                         propagate_gamma_and_color_space();
1089                         propagate_alpha();
1090
1091                         found_any = true;
1092                         break;
1093                 }
1094
1095                 char filename[256];
1096                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1097                 output_dot(filename);
1098                 assert(alpha_propagation_pass < 100);
1099         } while (found_any);
1100
1101         for (unsigned i = 0; i < nodes.size(); ++i) {
1102                 Node *node = nodes[i];
1103                 if (node->disabled) {
1104                         continue;
1105                 }
1106                 assert(node->output_alpha_type != ALPHA_INVALID);
1107         }
1108 }
1109
1110 // Make so that the output is in the desired color space.
1111 void EffectChain::fix_output_color_space()
1112 {
1113         Node *output = find_output_node();
1114         if (output->output_color_space != output_format.color_space) {
1115                 Node *conversion = add_node(new ColorspaceConversionEffect());
1116                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1117                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1118                 conversion->output_color_space = output_format.color_space;
1119                 connect_nodes(output, conversion);
1120                 propagate_alpha();
1121                 propagate_gamma_and_color_space();
1122         }
1123 }
1124
1125 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1126 void EffectChain::fix_output_alpha()
1127 {
1128         Node *output = find_output_node();
1129         assert(output->output_alpha_type != ALPHA_INVALID);
1130         if (output->output_alpha_type == ALPHA_BLANK) {
1131                 // No alpha output, so we don't care.
1132                 return;
1133         }
1134         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1135             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1136                 Node *conversion = add_node(new AlphaDivisionEffect());
1137                 connect_nodes(output, conversion);
1138                 propagate_alpha();
1139                 propagate_gamma_and_color_space();
1140         }
1141         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1142             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1143                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1144                 connect_nodes(output, conversion);
1145                 propagate_alpha();
1146                 propagate_gamma_and_color_space();
1147         }
1148 }
1149
1150 bool EffectChain::node_needs_gamma_fix(Node *node)
1151 {
1152         if (node->disabled) {
1153                 return false;
1154         }
1155
1156         // Small hack since the output is not an explicit node:
1157         // If we are the last node and our output is in the wrong
1158         // space compared to EffectChain's output, we need to fix it.
1159         // This will only take us to linear, but fix_output_gamma()
1160         // will come and take us to the desired output gamma
1161         // if it is needed.
1162         //
1163         // This needs to be before everything else, since it could
1164         // even apply to inputs (if they are the only effect).
1165         if (node->outgoing_links.empty() &&
1166             node->output_gamma_curve != output_format.gamma_curve &&
1167             node->output_gamma_curve != GAMMA_LINEAR) {
1168                 return true;
1169         }
1170
1171         if (node->effect->num_inputs() == 0) {
1172                 return false;
1173         }
1174
1175         // propagate_gamma_and_color_space() has already set our output
1176         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1177         // except for GammaCompressionEffect.
1178         if (node->output_gamma_curve == GAMMA_INVALID) {
1179                 return true;
1180         }
1181         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1182                 assert(node->incoming_links.size() == 1);
1183                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1184         }
1185
1186         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1187 }
1188
1189 // Very similar to fix_internal_color_spaces(), but for gamma.
1190 // There is one difference, though; before we start adding conversion nodes,
1191 // we see if we can get anything out of asking the sources to deliver
1192 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1193 // does that part, while fix_internal_gamma_by_inserting_nodes()
1194 // inserts nodes as needed afterwards.
1195 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1196 {
1197         unsigned gamma_propagation_pass = 0;
1198         bool found_any;
1199         do {
1200                 found_any = false;
1201                 for (unsigned i = 0; i < nodes.size(); ++i) {
1202                         Node *node = nodes[i];
1203                         if (!node_needs_gamma_fix(node)) {
1204                                 continue;
1205                         }
1206
1207                         // See if all inputs can give us linear gamma. If not, leave it.
1208                         std::vector<Node *> nonlinear_inputs;
1209                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1210                         assert(!nonlinear_inputs.empty());
1211
1212                         bool all_ok = true;
1213                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1214                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1215                                 all_ok &= input->can_output_linear_gamma();
1216                         }
1217
1218                         if (!all_ok) {
1219                                 continue;
1220                         }
1221
1222                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1223                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1224                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1225                         }
1226
1227                         // Re-sort topologically, and propagate the new information.
1228                         propagate_gamma_and_color_space();
1229
1230                         found_any = true;
1231                         break;
1232                 }
1233
1234                 char filename[256];
1235                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1236                 output_dot(filename);
1237                 assert(gamma_propagation_pass < 100);
1238         } while (found_any);
1239 }
1240
1241 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1242 {
1243         unsigned gamma_propagation_pass = 0;
1244         bool found_any;
1245         do {
1246                 found_any = false;
1247                 for (unsigned i = 0; i < nodes.size(); ++i) {
1248                         Node *node = nodes[i];
1249                         if (!node_needs_gamma_fix(node)) {
1250                                 continue;
1251                         }
1252
1253                         // Special case: We could be an input and still be asked to
1254                         // fix our gamma; if so, we should be the only node
1255                         // (as node_needs_gamma_fix() would only return true in
1256                         // for an input in that case). That means we should insert
1257                         // a conversion node _after_ ourselves.
1258                         if (node->incoming_links.empty()) {
1259                                 assert(node->outgoing_links.empty());
1260                                 Node *conversion = add_node(new GammaExpansionEffect());
1261                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1262                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1263                                 connect_nodes(node, conversion);
1264                         }
1265
1266                         // If not, go through each input that is not linear gamma,
1267                         // and insert a gamma conversion after it.
1268                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1269                                 Node *input = node->incoming_links[j];
1270                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1271                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1272                                         continue;
1273                                 }
1274                                 Node *conversion = add_node(new GammaExpansionEffect());
1275                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1276                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1277                                 replace_sender(input, conversion);
1278                                 connect_nodes(input, conversion);
1279                         }
1280
1281                         // Re-sort topologically, and propagate the new information.
1282                         propagate_alpha();
1283                         propagate_gamma_and_color_space();
1284
1285                         found_any = true;
1286                         break;
1287                 }
1288
1289                 char filename[256];
1290                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1291                 output_dot(filename);
1292                 assert(gamma_propagation_pass < 100);
1293         } while (found_any);
1294
1295         for (unsigned i = 0; i < nodes.size(); ++i) {
1296                 Node *node = nodes[i];
1297                 if (node->disabled) {
1298                         continue;
1299                 }
1300                 assert(node->output_gamma_curve != GAMMA_INVALID);
1301         }
1302 }
1303
1304 // Make so that the output is in the desired gamma.
1305 // Note that this assumes linear input gamma, so it might create the need
1306 // for another pass of fix_internal_gamma().
1307 void EffectChain::fix_output_gamma()
1308 {
1309         Node *output = find_output_node();
1310         if (output->output_gamma_curve != output_format.gamma_curve) {
1311                 Node *conversion = add_node(new GammaCompressionEffect());
1312                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1313                 conversion->output_gamma_curve = output_format.gamma_curve;
1314                 connect_nodes(output, conversion);
1315         }
1316 }
1317
1318 // If the user has requested dither, add a DitherEffect right at the end
1319 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1320 // since dither is about the only effect that can _not_ be done in linear space.
1321 void EffectChain::add_dither_if_needed()
1322 {
1323         if (num_dither_bits == 0) {
1324                 return;
1325         }
1326         Node *output = find_output_node();
1327         Node *dither = add_node(new DitherEffect());
1328         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1329         connect_nodes(output, dither);
1330
1331         dither_effect = dither->effect;
1332 }
1333
1334 // Find the output node. This is, simply, one that has no outgoing links.
1335 // If there are multiple ones, the graph is malformed (we do not support
1336 // multiple outputs right now).
1337 Node *EffectChain::find_output_node()
1338 {
1339         std::vector<Node *> output_nodes;
1340         for (unsigned i = 0; i < nodes.size(); ++i) {
1341                 Node *node = nodes[i];
1342                 if (node->disabled) {
1343                         continue;
1344                 }
1345                 if (node->outgoing_links.empty()) {
1346                         output_nodes.push_back(node);
1347                 }
1348         }
1349         assert(output_nodes.size() == 1);
1350         return output_nodes[0];
1351 }
1352
1353 void EffectChain::finalize()
1354 {
1355         // Save the current locale, and set it to C, so that we can output decimal
1356         // numbers with printf and be sure to get them in the format mandated by GLSL.
1357         char *saved_locale = setlocale(LC_NUMERIC, "C");
1358
1359         // Output the graph as it is before we do any conversions on it.
1360         output_dot("step0-start.dot");
1361
1362         // Give each effect in turn a chance to rewrite its own part of the graph.
1363         // Note that if more effects are added as part of this, they will be
1364         // picked up as part of the same for loop, since they are added at the end.
1365         for (unsigned i = 0; i < nodes.size(); ++i) {
1366                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1367         }
1368         output_dot("step1-rewritten.dot");
1369
1370         find_color_spaces_for_inputs();
1371         output_dot("step2-input-colorspace.dot");
1372
1373         propagate_alpha();
1374         output_dot("step3-propagated-alpha.dot");
1375
1376         propagate_gamma_and_color_space();
1377         output_dot("step4-propagated-all.dot");
1378
1379         fix_internal_color_spaces();
1380         fix_internal_alpha(6);
1381         fix_output_color_space();
1382         output_dot("step7-output-colorspacefix.dot");
1383         fix_output_alpha();
1384         output_dot("step8-output-alphafix.dot");
1385
1386         // Note that we need to fix gamma after colorspace conversion,
1387         // because colorspace conversions might create needs for gamma conversions.
1388         // Also, we need to run an extra pass of fix_internal_gamma() after
1389         // fixing the output gamma, as we only have conversions to/from linear,
1390         // and fix_internal_alpha() since GammaCompressionEffect needs
1391         // postmultiplied input.
1392         fix_internal_gamma_by_asking_inputs(9);
1393         fix_internal_gamma_by_inserting_nodes(10);
1394         fix_output_gamma();
1395         output_dot("step11-output-gammafix.dot");
1396         propagate_alpha();
1397         output_dot("step12-output-alpha-propagated.dot");
1398         fix_internal_alpha(13);
1399         output_dot("step14-output-alpha-fixed.dot");
1400         fix_internal_gamma_by_asking_inputs(15);
1401         fix_internal_gamma_by_inserting_nodes(16);
1402
1403         output_dot("step17-before-dither.dot");
1404
1405         add_dither_if_needed();
1406
1407         output_dot("step18-final.dot");
1408
1409         // Construct all needed GLSL programs, starting at the output.
1410         construct_glsl_programs(find_output_node());
1411
1412         output_dot("step19-split-to-phases.dot");
1413
1414         // If we have more than one phase, we need intermediate render-to-texture.
1415         // Construct an FBO, and then as many textures as we need.
1416         // We choose the simplest option of having one texture per output,
1417         // since otherwise this turns into an (albeit simple)
1418         // register allocation problem.
1419         if (phases.size() > 1) {
1420                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1421                         inform_input_sizes(phases[i]);
1422                         find_output_size(phases[i]);
1423
1424                         Node *output_node = phases[i]->effects.back();
1425                         glGenTextures(1, &output_node->output_texture);
1426                         check_error();
1427                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1428                         check_error();
1429                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1430                         check_error();
1431                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1432                         check_error();
1433                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1434                         check_error();
1435
1436                         output_node->output_texture_width = phases[i]->output_width;
1437                         output_node->output_texture_height = phases[i]->output_height;
1438                 }
1439                 inform_input_sizes(phases.back());
1440         }
1441
1442         for (unsigned i = 0; i < inputs.size(); ++i) {
1443                 inputs[i]->finalize();
1444         }
1445
1446         assert(phases[0]->inputs.empty());
1447
1448         finalized = true;
1449         setlocale(LC_NUMERIC, saved_locale);
1450 }
1451
1452 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1453 {
1454         assert(finalized);
1455
1456         // Save original viewport.
1457         GLuint x = 0, y = 0;
1458         GLuint fbo = 0;
1459
1460         if (width == 0 && height == 0) {
1461                 GLint viewport[4];
1462                 glGetIntegerv(GL_VIEWPORT, viewport);
1463                 x = viewport[0];
1464                 y = viewport[1];
1465                 width = viewport[2];
1466                 height = viewport[3];
1467         }
1468
1469         // Basic state.
1470         glDisable(GL_BLEND);
1471         check_error();
1472         glDisable(GL_DEPTH_TEST);
1473         check_error();
1474         glDepthMask(GL_FALSE);
1475         check_error();
1476
1477         glMatrixMode(GL_PROJECTION);
1478         glLoadIdentity();
1479         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1480
1481         glMatrixMode(GL_MODELVIEW);
1482         glLoadIdentity();
1483
1484         if (phases.size() > 1) {
1485                 glGenFramebuffers(1, &fbo);
1486                 check_error();
1487                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1488                 check_error();
1489         }
1490
1491         std::set<Node *> generated_mipmaps;
1492
1493         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1494                 // See if the requested output size has changed. If so, we need to recreate
1495                 // the texture (and before we start setting up inputs).
1496                 inform_input_sizes(phases[phase]);
1497                 if (phase != phases.size() - 1) {
1498                         find_output_size(phases[phase]);
1499
1500                         Node *output_node = phases[phase]->effects.back();
1501
1502                         if (output_node->output_texture_width != phases[phase]->output_width ||
1503                             output_node->output_texture_height != phases[phase]->output_height) {
1504                                 glActiveTexture(GL_TEXTURE0);
1505                                 check_error();
1506                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1507                                 check_error();
1508                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1509                                 check_error();
1510                                 glBindTexture(GL_TEXTURE_2D, 0);
1511                                 check_error();
1512
1513                                 output_node->output_texture_width = phases[phase]->output_width;
1514                                 output_node->output_texture_height = phases[phase]->output_height;
1515                         }
1516                 }
1517
1518                 glUseProgram(phases[phase]->glsl_program_num);
1519                 check_error();
1520
1521                 // Set up RTT inputs for this phase.
1522                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1523                         glActiveTexture(GL_TEXTURE0 + sampler);
1524                         Node *input = phases[phase]->inputs[sampler];
1525                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1526                         check_error();
1527                         if (phases[phase]->input_needs_mipmaps) {
1528                                 if (generated_mipmaps.count(input) == 0) {
1529                                         glGenerateMipmap(GL_TEXTURE_2D);
1530                                         check_error();
1531                                         generated_mipmaps.insert(input);
1532                                 }
1533                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1534                                 check_error();
1535                         } else {
1536                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1537                                 check_error();
1538                         }
1539
1540                         std::string texture_name = std::string("tex_") + input->effect_id;
1541                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1542                         check_error();
1543                 }
1544
1545                 // And now the output.
1546                 if (phase == phases.size() - 1) {
1547                         // Last phase goes to the output the user specified.
1548                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1549                         check_error();
1550                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1551                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1552                         glViewport(x, y, width, height);
1553                         if (dither_effect != NULL) {
1554                                 CHECK(dither_effect->set_int("output_width", width));
1555                                 CHECK(dither_effect->set_int("output_height", height));
1556                         }
1557                 } else {
1558                         Node *output_node = phases[phase]->effects.back();
1559                         glFramebufferTexture2D(
1560                                 GL_FRAMEBUFFER,
1561                                 GL_COLOR_ATTACHMENT0,
1562                                 GL_TEXTURE_2D,
1563                                 output_node->output_texture,
1564                                 0);
1565                         check_error();
1566                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1567                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1568                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1569                 }
1570
1571                 // Give the required parameters to all the effects.
1572                 unsigned sampler_num = phases[phase]->inputs.size();
1573                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1574                         Node *node = phases[phase]->effects[i];
1575                         node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num);
1576                         check_error();
1577                 }
1578
1579                 // Now draw!
1580                 glBegin(GL_QUADS);
1581
1582                 glTexCoord2f(0.0f, 0.0f);
1583                 glVertex2f(0.0f, 0.0f);
1584
1585                 glTexCoord2f(1.0f, 0.0f);
1586                 glVertex2f(1.0f, 0.0f);
1587
1588                 glTexCoord2f(1.0f, 1.0f);
1589                 glVertex2f(1.0f, 1.0f);
1590
1591                 glTexCoord2f(0.0f, 1.0f);
1592                 glVertex2f(0.0f, 1.0f);
1593
1594                 glEnd();
1595                 check_error();
1596
1597                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1598                         Node *node = phases[phase]->effects[i];
1599                         node->effect->clear_gl_state();
1600                 }
1601         }
1602
1603         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1604         check_error();
1605
1606         if (fbo != 0) {
1607                 glDeleteFramebuffers(1, &fbo);
1608                 check_error();
1609         }
1610 }