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         Node *node = new Node;
  84         node->effect = effect;
  85         node->disabled = false;
  86         node->output_color_space = COLORSPACE_INVALID;
  87         node->output_gamma_curve = GAMMA_INVALID;
  88         node->output_alpha_type = ALPHA_INVALID;
  89         node->output_texture = 0;
  90
  91         nodes.push_back(node);
  92         node_map[effect] = node;
  93         effect->inform_added(this);
  94         return node;
  95 }
  96
  97 void EffectChain::connect_nodes(Node *sender, Node *receiver)
  98 {
  99         sender->outgoing_links.push_back(receiver);
 100         receiver->incoming_links.push_back(sender);
 101 }
 102
 103 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
 104 {
 105         new_receiver->incoming_links = old_receiver->incoming_links;
 106         old_receiver->incoming_links.clear();
 107
 108         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
 109                 Node *sender = new_receiver->incoming_links[i];
 110                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
 111                         if (sender->outgoing_links[j] == old_receiver) {
 112                                 sender->outgoing_links[j] = new_receiver;
 113                         }
 114                 }
 115         }
 116 }
 117
 118 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 119 {
 120         new_sender->outgoing_links = old_sender->outgoing_links;
 121         old_sender->outgoing_links.clear();
 122
 123         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 124                 Node *receiver = new_sender->outgoing_links[i];
 125                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 126                         if (receiver->incoming_links[j] == old_sender) {
 127                                 receiver->incoming_links[j] = new_sender;
 128                         }
 129                 }
 130         }
 131 }
 132
 133 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 134 {
 135         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 136                 if (sender->outgoing_links[i] == receiver) {
 137                         sender->outgoing_links[i] = middle;
 138                         middle->incoming_links.push_back(sender);
 139                 }
 140         }
 141         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 142                 if (receiver->incoming_links[i] == sender) {
 143                         receiver->incoming_links[i] = middle;
 144                         middle->outgoing_links.push_back(receiver);
 145                 }
 146         }
 147
 148         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 149 }
 150
 151 void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector<Node *> *nonlinear_inputs)
 152 {
 153         if (node->output_gamma_curve == GAMMA_LINEAR &&
 154             node->effect->effect_type_id() != "GammaCompressionEffect") {
 155                 return;
 156         }
 157         if (node->effect->num_inputs() == 0) {
 158                 nonlinear_inputs->push_back(node);
 159         } else {
 160                 assert(node->effect->num_inputs() == node->incoming_links.size());
 161                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 162                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 163                 }
 164         }
 165 }
 166
 167 Effect *EffectChain::add_effect(Effect *effect, const std::vector<Effect *> &inputs)
 168 {
 169         assert(!finalized);
 170         assert(inputs.size() == effect->num_inputs());
 171         Node *node = add_node(effect);
 172         for (unsigned i = 0; i < inputs.size(); ++i) {
 173                 assert(node_map.count(inputs[i]) != 0);
 174                 connect_nodes(node_map[inputs[i]], node);
 175         }
 176         return effect;
 177 }
 178
 179 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 180 std::string replace_prefix(const std::string &text, const std::string &prefix)
 181 {
 182         std::string output;
 183         size_t start = 0;
 184
 185         while (start < text.size()) {
 186                 size_t pos = text.find("PREFIX(", start);
 187                 if (pos == std::string::npos) {
 188                         output.append(text.substr(start, std::string::npos));
 189                         break;
 190                 }
 191
 192                 output.append(text.substr(start, pos - start));
 193                 output.append(prefix);
 194                 output.append("_");
 195
 196                 pos += strlen("PREFIX(");
 197
 198                 // Output stuff until we find the matching ), which we then eat.
 199                 int depth = 1;
 200                 size_t end_arg_pos = pos;
 201                 while (end_arg_pos < text.size()) {
 202                         if (text[end_arg_pos] == '(') {
 203                                 ++depth;
 204                         } else if (text[end_arg_pos] == ')') {
 205                                 --depth;
 206                                 if (depth == 0) {
 207                                         break;
 208                                 }
 209                         }
 210                         ++end_arg_pos;
 211                 }
 212                 output.append(text.substr(pos, end_arg_pos - pos));
 213                 ++end_arg_pos;
 214                 assert(depth == 0);
 215                 start = end_arg_pos;
 216         }
 217         return output;
 218 }
 219
 220 Phase *EffectChain::compile_glsl_program(
 221         const std::vector<Node *> &inputs,
 222         const std::vector<Node *> &effects)
 223 {
 224         Phase *phase = new Phase;
 225         assert(!effects.empty());
 226
 227         // Deduplicate the inputs.
 228         std::vector<Node *> true_inputs = inputs;
 229         std::sort(true_inputs.begin(), true_inputs.end());
 230         true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
 231
 232         bool input_needs_mipmaps = false;
 233         std::string frag_shader = read_file("header.frag");
 234
 235         // Create functions for all the texture inputs that we need.
 236         for (unsigned i = 0; i < true_inputs.size(); ++i) {
 237                 Node *input = true_inputs[i];
 238                 char effect_id[256];
 239                 sprintf(effect_id, "in%u", i);
 240                 phase->effect_ids.insert(std::make_pair(input, effect_id));
 241
 242                 frag_shader += std::string("uniform sampler2D tex_") + effect_id + ";\n";
 243                 frag_shader += std::string("vec4 ") + effect_id + "(vec2 tc) {\n";
 244                 frag_shader += "\treturn texture2D(tex_" + std::string(effect_id) + ", tc);\n";
 245                 frag_shader += "}\n";
 246                 frag_shader += "\n";
 247         }
 248
 249         std::vector<Node *> sorted_effects = topological_sort(effects);
 250
 251         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 252                 Node *node = sorted_effects[i];
 253                 char effect_id[256];
 254                 sprintf(effect_id, "eff%u", i);
 255                 phase->effect_ids.insert(std::make_pair(node, effect_id));
 256
 257                 if (node->incoming_links.size() == 1) {
 258                         frag_shader += std::string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
 259                 } else {
 260                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 261                                 char buf[256];
 262                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
 263                                 frag_shader += buf;
 264                         }
 265                 }
 266
 267                 frag_shader += "\n";
 268                 frag_shader += std::string("#define FUNCNAME ") + effect_id + "\n";
 269                 frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id);
 270                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
 271                 frag_shader += "#undef PREFIX\n";
 272                 frag_shader += "#undef FUNCNAME\n";
 273                 if (node->incoming_links.size() == 1) {
 274                         frag_shader += "#undef INPUT\n";
 275                 } else {
 276                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 277                                 char buf[256];
 278                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 279                                 frag_shader += buf;
 280                         }
 281                 }
 282                 frag_shader += "\n";
 283
 284                 input_needs_mipmaps |= node->effect->needs_mipmaps();
 285         }
 286         for (unsigned i = 0; i < sorted_effects.size(); ++i) {
 287                 Node *node = sorted_effects[i];
 288                 if (node->effect->num_inputs() == 0) {
 289                         CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
 290                 }
 291         }
 292         frag_shader += std::string("#define INPUT ") + phase->effect_ids[sorted_effects.back()] + "\n";
 293         frag_shader.append(read_file("footer.frag"));
 294
 295         phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader);
 296         phase->input_needs_mipmaps = input_needs_mipmaps;
 297         phase->inputs = true_inputs;
 298         phase->effects = sorted_effects;
 299
 300         return phase;
 301 }
 302
 303 // Construct GLSL programs, starting at the given effect and following
 304 // the chain from there. We end a program every time we come to an effect
 305 // marked as "needs texture bounce", one that is used by multiple other
 306 // effects, every time an effect wants to change the output size,
 307 // and of course at the end.
 308 //
 309 // We follow a quite simple depth-first search from the output, although
 310 // without any explicit recursion.
 311 void EffectChain::construct_glsl_programs(Node *output)
 312 {
 313         // Which effects have already been completed?
 314         // We need to keep track of it, as an effect with multiple outputs
 315         // could otherwise be calculated multiple times.
 316         std::set<Node *> completed_effects;
 317
 318         // Effects in the current phase, as well as inputs (outputs from other phases
 319         // that we depend on). Note that since we start iterating from the end,
 320         // the effect list will be in the reverse order.
 321         std::vector<Node *> this_phase_inputs;
 322         std::vector<Node *> this_phase_effects;
 323
 324         // Effects that we have yet to calculate, but that we know should
 325         // be in the current phase.
 326         std::stack<Node *> effects_todo_this_phase;
 327
 328         // Effects that we have yet to calculate, but that come from other phases.
 329         // We delay these until we have this phase done in its entirety,
 330         // at which point we pick any of them and start a new phase from that.
 331         std::stack<Node *> effects_todo_other_phases;
 332
 333         effects_todo_this_phase.push(output);
 334
 335         for ( ;; ) {  // Termination condition within loop.
 336                 if (!effects_todo_this_phase.empty()) {
 337                         // OK, we have more to do this phase.
 338                         Node *node = effects_todo_this_phase.top();
 339                         effects_todo_this_phase.pop();
 340
 341                         // This should currently only happen for effects that are inputs
 342                         // (either true inputs or phase outputs). We special-case inputs,
 343                         // and then deduplicate phase outputs in compile_glsl_program().
 344                         if (node->effect->num_inputs() == 0) {
 345                                 if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
 346                                         continue;
 347                                 }
 348                         } else {
 349                                 assert(completed_effects.count(node) == 0);
 350                         }
 351
 352                         this_phase_effects.push_back(node);
 353                         completed_effects.insert(node);
 354
 355                         // Find all the dependencies of this effect, and add them to the stack.
 356                         std::vector<Node *> deps = node->incoming_links;
 357                         assert(node->effect->num_inputs() == deps.size());
 358                         for (unsigned i = 0; i < deps.size(); ++i) {
 359                                 bool start_new_phase = false;
 360
 361                                 // FIXME: If we sample directly from a texture, we won't need this.
 362                                 if (node->effect->needs_texture_bounce()) {
 363                                         start_new_phase = true;
 364                                 }
 365
 366                                 if (deps[i]->outgoing_links.size() > 1) {
 367                                         if (deps[i]->effect->num_inputs() > 0) {
 368                                                 // More than one effect uses this as the input,
 369                                                 // and it is not a texture itself.
 370                                                 // The easiest thing to do (and probably also the safest
 371                                                 // performance-wise in most cases) is to bounce it to a texture
 372                                                 // and then let the next passes read from that.
 373                                                 start_new_phase = true;
 374                                         } else {
 375                                                 // For textures, we try to be slightly more clever;
 376                                                 // if none of our outputs need a bounce, we don't bounce
 377                                                 // but instead simply use the effect many times.
 378                                                 //
 379                                                 // Strictly speaking, we could bounce it for some outputs
 380                                                 // and use it directly for others, but the processing becomes
 381                                                 // somewhat simpler if the effect is only used in one such way.
 382                                                 for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 383                                                         Node *rdep = deps[i]->outgoing_links[j];
 384                                                         start_new_phase |= rdep->effect->needs_texture_bounce();
 385                                                 }
 386                                         }
 387                                 }
 388
 389                                 if (deps[i]->effect->changes_output_size()) {
 390                                         start_new_phase = true;
 391                                 }
 392
 393                                 if (start_new_phase) {
 394                                         effects_todo_other_phases.push(deps[i]);
 395                                         this_phase_inputs.push_back(deps[i]);
 396                                 } else {
 397                                         effects_todo_this_phase.push(deps[i]);
 398                                 }
 399                         }
 400                         continue;
 401                 }
 402
 403                 // No more effects to do this phase. Take all the ones we have,
 404                 // and create a GLSL program for it.
 405                 if (!this_phase_effects.empty()) {
 406                         reverse(this_phase_effects.begin(), this_phase_effects.end());
 407                         phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
 408                         this_phase_effects.back()->phase = phases.back();
 409                         this_phase_inputs.clear();
 410                         this_phase_effects.clear();
 411                 }
 412                 assert(this_phase_inputs.empty());
 413                 assert(this_phase_effects.empty());
 414
 415                 // If we have no effects left, exit.
 416                 if (effects_todo_other_phases.empty()) {
 417                         break;
 418                 }
 419
 420                 Node *node = effects_todo_other_phases.top();
 421                 effects_todo_other_phases.pop();
 422
 423                 if (completed_effects.count(node) == 0) {
 424                         // Start a new phase, calculating from this effect.
 425                         effects_todo_this_phase.push(node);
 426                 }
 427         }
 428
 429         // Finally, since the phases are found from the output but must be executed
 430         // from the input(s), reverse them, too.
 431         std::reverse(phases.begin(), phases.end());
 432 }
 433
 434 void EffectChain::output_dot(const char *filename)
 435 {
 436         if (movit_debug_level != MOVIT_DEBUG_ON) {
 437                 return;
 438         }
 439
 440         FILE *fp = fopen(filename, "w");
 441         if (fp == NULL) {
 442                 perror(filename);
 443                 exit(1);
 444         }
 445
 446         fprintf(fp, "digraph G {\n");
 447         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 448         for (unsigned i = 0; i < nodes.size(); ++i) {
 449                 // Find out which phase this event belongs to.
 450                 std::vector<int> in_phases;
 451                 for (unsigned j = 0; j < phases.size(); ++j) {
 452                         const Phase* p = phases[j];
 453                         if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 454                                 in_phases.push_back(j);
 455                         }
 456                 }
 457
 458                 if (in_phases.empty()) {
 459                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 460                 } else if (in_phases.size() == 1) {
 461                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 462                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 463                                 (in_phases[0] % 8) + 1);
 464                 } else {
 465                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 466                         // But alas.
 467                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 468                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 469                                 (in_phases[0] % 8) + 1);
 470                 }
 471
 472                 char from_node_id[256];
 473                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 474
 475                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 476                         char to_node_id[256];
 477                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 478
 479                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 480                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 481                 }
 482
 483                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 484                         // Output node.
 485                         std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
 486                         output_dot_edge(fp, from_node_id, "output", labels);
 487                 }
 488         }
 489         fprintf(fp, "}\n");
 490
 491         fclose(fp);
 492 }
 493
 494 std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 495 {
 496         std::vector<std::string> labels;
 497
 498         if (to != NULL && to->effect->needs_texture_bounce()) {
 499                 labels.push_back("needs_bounce");
 500         }
 501         if (from->effect->changes_output_size()) {
 502                 labels.push_back("resize");
 503         }
 504
 505         switch (from->output_color_space) {
 506         case COLORSPACE_INVALID:
 507                 labels.push_back("spc[invalid]");
 508                 break;
 509         case COLORSPACE_REC_601_525:
 510                 labels.push_back("spc[rec601-525]");
 511                 break;
 512         case COLORSPACE_REC_601_625:
 513                 labels.push_back("spc[rec601-625]");
 514                 break;
 515         default:
 516                 break;
 517         }
 518
 519         switch (from->output_gamma_curve) {
 520         case GAMMA_INVALID:
 521                 labels.push_back("gamma[invalid]");
 522                 break;
 523         case GAMMA_sRGB:
 524                 labels.push_back("gamma[sRGB]");
 525                 break;
 526         case GAMMA_REC_601:  // and GAMMA_REC_709
 527                 labels.push_back("gamma[rec601/709]");
 528                 break;
 529         default:
 530                 break;
 531         }
 532
 533         switch (from->output_alpha_type) {
 534         case ALPHA_INVALID:
 535                 labels.push_back("alpha[invalid]");
 536                 break;
 537         case ALPHA_BLANK:
 538                 labels.push_back("alpha[blank]");
 539                 break;
 540         case ALPHA_POSTMULTIPLIED:
 541                 labels.push_back("alpha[postmult]");
 542                 break;
 543         default:
 544                 break;
 545         }
 546
 547         return labels;
 548 }
 549
 550 void EffectChain::output_dot_edge(FILE *fp,
 551                                   const std::string &from_node_id,
 552                                   const std::string &to_node_id,
 553                                   const std::vector<std::string> &labels)
 554 {
 555         if (labels.empty()) {
 556                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 557         } else {
 558                 std::string label = labels[0];
 559                 for (unsigned k = 1; k < labels.size(); ++k) {
 560                         label += ", " + labels[k];
 561                 }
 562                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 563         }
 564 }
 565
 566 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
 567 {
 568         unsigned scaled_width, scaled_height;
 569
 570         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 571                 // Same aspect, or W/H > aspect (image is wider than the frame).
 572                 // In either case, keep width, and adjust height.
 573                 scaled_width = width;
 574                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
 575         } else {
 576                 // W/H < aspect (image is taller than the frame), so keep height,
 577                 // and adjust width.
 578                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
 579                 scaled_height = height;
 580         }
 581
 582         // We should be consistently larger or smaller then the existing choice,
 583         // since we have the same aspect.
 584         assert(!(scaled_width < *output_width && scaled_height > *output_height));
 585         assert(!(scaled_height < *output_height && scaled_width > *output_width));
 586
 587         if (scaled_width >= *output_width && scaled_height >= *output_height) {
 588                 *output_width = scaled_width;
 589                 *output_height = scaled_height;
 590         }
 591 }
 592
 593 // Propagate input texture sizes throughout, and inform effects downstream.
 594 // (Like a lot of other code, we depend on effects being in topological order.)
 595 void EffectChain::inform_input_sizes(Phase *phase)
 596 {
 597         // All effects that have a defined size (inputs and RTT inputs)
 598         // get that. Reset all others.
 599         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 600                 Node *node = phase->effects[i];
 601                 if (node->effect->num_inputs() == 0) {
 602                         Input *input = static_cast<Input *>(node->effect);
 603                         node->output_width = input->get_width();
 604                         node->output_height = input->get_height();
 605                         assert(node->output_width != 0);
 606                         assert(node->output_height != 0);
 607                 } else {
 608                         node->output_width = node->output_height = 0;
 609                 }
 610         }
 611         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 612                 Node *input = phase->inputs[i];
 613                 input->output_width = input->phase->virtual_output_width;
 614                 input->output_height = input->phase->virtual_output_height;
 615                 assert(input->output_width != 0);
 616                 assert(input->output_height != 0);
 617         }
 618
 619         // Now propagate from the inputs towards the end, and inform as we go.
 620         // The rules are simple:
 621         //
 622         //   1. Don't touch effects that already have given sizes (ie., inputs).
 623         //   2. If all of your inputs have the same size, that will be your output size.
 624         //   3. Otherwise, your output size is 0x0.
 625         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 626                 Node *node = phase->effects[i];
 627                 if (node->effect->num_inputs() == 0) {
 628                         continue;
 629                 }
 630                 unsigned this_output_width = 0;
 631                 unsigned this_output_height = 0;
 632                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 633                         Node *input = node->incoming_links[j];
 634                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 635                         if (j == 0) {
 636                                 this_output_width = input->output_width;
 637                                 this_output_height = input->output_height;
 638                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 639                                 // Inputs disagree.
 640                                 this_output_width = 0;
 641                                 this_output_height = 0;
 642                         }
 643                 }
 644                 node->output_width = this_output_width;
 645                 node->output_height = this_output_height;
 646         }
 647 }
 648
 649 // Note: You should call inform_input_sizes() before this, as the last effect's
 650 // desired output size might change based on the inputs.
 651 void EffectChain::find_output_size(Phase *phase)
 652 {
 653         Node *output_node = phase->effects.back();
 654
 655         // If the last effect explicitly sets an output size, use that.
 656         if (output_node->effect->changes_output_size()) {
 657                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
 658                                                      &phase->virtual_output_width, &phase->virtual_output_height);
 659                 return;
 660         }
 661
 662         // If all effects have the same size, use that.
 663         unsigned output_width = 0, output_height = 0;
 664         bool all_inputs_same_size = true;
 665
 666         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 667                 Node *input = phase->inputs[i];
 668                 assert(input->phase->output_width != 0);
 669                 assert(input->phase->output_height != 0);
 670                 if (output_width == 0 && output_height == 0) {
 671                         output_width = input->phase->virtual_output_width;
 672                         output_height = input->phase->virtual_output_height;
 673                 } else if (output_width != input->phase->virtual_output_width ||
 674                            output_height != input->phase->virtual_output_height) {
 675                         all_inputs_same_size = false;
 676                 }
 677         }
 678         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 679                 Effect *effect = phase->effects[i]->effect;
 680                 if (effect->num_inputs() != 0) {
 681                         continue;
 682                 }
 683
 684                 Input *input = static_cast<Input *>(effect);
 685                 if (output_width == 0 && output_height == 0) {
 686                         output_width = input->get_width();
 687                         output_height = input->get_height();
 688                 } else if (output_width != input->get_width() ||
 689                            output_height != input->get_height()) {
 690                         all_inputs_same_size = false;
 691                 }
 692         }
 693
 694         if (all_inputs_same_size) {
 695                 assert(output_width != 0);
 696                 assert(output_height != 0);
 697                 phase->virtual_output_width = phase->output_width = output_width;
 698                 phase->virtual_output_height = phase->output_height = output_height;
 699                 return;
 700         }
 701
 702         // If not, fit all the inputs into the current aspect, and select the largest one.
 703         output_width = 0;
 704         output_height = 0;
 705         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 706                 Node *input = phase->inputs[i];
 707                 assert(input->phase->output_width != 0);
 708                 assert(input->phase->output_height != 0);
 709                 size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
 710         }
 711         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 712                 Effect *effect = phase->effects[i]->effect;
 713                 if (effect->num_inputs() != 0) {
 714                         continue;
 715                 }
 716
 717                 Input *input = static_cast<Input *>(effect);
 718                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
 719         }
 720         assert(output_width != 0);
 721         assert(output_height != 0);
 722         phase->virtual_output_width = phase->output_width = output_width;
 723         phase->virtual_output_height = phase->output_height = output_height;
 724 }
 725
 726 void EffectChain::sort_all_nodes_topologically()
 727 {
 728         nodes = topological_sort(nodes);
 729 }
 730
 731 std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
 732 {
 733         std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
 734         std::vector<Node *> sorted_list;
 735         for (unsigned i = 0; i < nodes.size(); ++i) {
 736                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
 737         }
 738         reverse(sorted_list.begin(), sorted_list.end());
 739         return sorted_list;
 740 }
 741
 742 void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
 743 {
 744         if (nodes_left_to_visit->count(node) == 0) {
 745                 return;
 746         }
 747         nodes_left_to_visit->erase(node);
 748         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 749                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
 750         }
 751         sorted_list->push_back(node);
 752 }
 753
 754 void EffectChain::find_color_spaces_for_inputs()
 755 {
 756         for (unsigned i = 0; i < nodes.size(); ++i) {
 757                 Node *node = nodes[i];
 758                 if (node->disabled) {
 759                         continue;
 760                 }
 761                 if (node->incoming_links.size() == 0) {
 762                         Input *input = static_cast<Input *>(node->effect);
 763                         node->output_color_space = input->get_color_space();
 764                         node->output_gamma_curve = input->get_gamma_curve();
 765
 766                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
 767                         switch (alpha_handling) {
 768                         case Effect::OUTPUT_BLANK_ALPHA:
 769                                 node->output_alpha_type = ALPHA_BLANK;
 770                                 break;
 771                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
 772                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 773                                 break;
 774                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
 775                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 776                                 break;
 777                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
 778                         case Effect::DONT_CARE_ALPHA_TYPE:
 779                         default:
 780                                 assert(false);
 781                         }
 782
 783                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
 784                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
 785                         }
 786                 }
 787         }
 788 }
 789
 790 // Propagate gamma and color space information as far as we can in the graph.
 791 // The rules are simple: Anything where all the inputs agree, get that as
 792 // output as well. Anything else keeps having *_INVALID.
 793 void EffectChain::propagate_gamma_and_color_space()
 794 {
 795         // We depend on going through the nodes in order.
 796         sort_all_nodes_topologically();
 797
 798         for (unsigned i = 0; i < nodes.size(); ++i) {
 799                 Node *node = nodes[i];
 800                 if (node->disabled) {
 801                         continue;
 802                 }
 803                 assert(node->incoming_links.size() == node->effect->num_inputs());
 804                 if (node->incoming_links.size() == 0) {
 805                         assert(node->output_color_space != COLORSPACE_INVALID);
 806                         assert(node->output_gamma_curve != GAMMA_INVALID);
 807                         continue;
 808                 }
 809
 810                 Colorspace color_space = node->incoming_links[0]->output_color_space;
 811                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 812                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 813                         if (node->incoming_links[j]->output_color_space != color_space) {
 814                                 color_space = COLORSPACE_INVALID;
 815                         }
 816                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 817                                 gamma_curve = GAMMA_INVALID;
 818                         }
 819                 }
 820
 821                 // The conversion effects already have their outputs set correctly,
 822                 // so leave them alone.
 823                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
 824                         node->output_color_space = color_space;
 825                 }
 826                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 827                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 828                         node->output_gamma_curve = gamma_curve;
 829                 }
 830         }
 831 }
 832
 833 // Propagate alpha information as far as we can in the graph.
 834 // Similar to propagate_gamma_and_color_space().
 835 void EffectChain::propagate_alpha()
 836 {
 837         // We depend on going through the nodes in order.
 838         sort_all_nodes_topologically();
 839
 840         for (unsigned i = 0; i < nodes.size(); ++i) {
 841                 Node *node = nodes[i];
 842                 if (node->disabled) {
 843                         continue;
 844                 }
 845                 assert(node->incoming_links.size() == node->effect->num_inputs());
 846                 if (node->incoming_links.size() == 0) {
 847                         assert(node->output_alpha_type != ALPHA_INVALID);
 848                         continue;
 849                 }
 850
 851                 // The alpha multiplication/division effects are special cases.
 852                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
 853                         assert(node->incoming_links.size() == 1);
 854                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
 855                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
 856                         continue;
 857                 }
 858                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
 859                         assert(node->incoming_links.size() == 1);
 860                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
 861                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 862                         continue;
 863                 }
 864
 865                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
 866                 // because they are the only one that _need_ postmultiplied alpha.
 867                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
 868                     node->effect->effect_type_id() == "GammaExpansionEffect") {
 869                         assert(node->incoming_links.size() == 1);
 870                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
 871                                 node->output_alpha_type = ALPHA_BLANK;
 872                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
 873                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 874                         } else {
 875                                 node->output_alpha_type = ALPHA_INVALID;
 876                         }
 877                         continue;
 878                 }
 879
 880                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
 881                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
 882                 // taken care of above. Rationale: Even if you could imagine
 883                 // e.g. an effect that took in an image and set alpha=1.0
 884                 // unconditionally, it wouldn't make any sense to have it as
 885                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
 886                 // got its input pre- or postmultiplied, so it wouldn't know
 887                 // whether to divide away the old alpha or not.
 888                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
 889                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 890                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
 891                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 892
 893                 // If the node has multiple inputs, check that they are all valid and
 894                 // the same.
 895                 bool any_invalid = false;
 896                 bool any_premultiplied = false;
 897                 bool any_postmultiplied = false;
 898
 899                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 900                         switch (node->incoming_links[j]->output_alpha_type) {
 901                         case ALPHA_INVALID:
 902                                 any_invalid = true;
 903                                 break;
 904                         case ALPHA_BLANK:
 905                                 // Blank is good as both pre- and postmultiplied alpha,
 906                                 // so just ignore it.
 907                                 break;
 908                         case ALPHA_PREMULTIPLIED:
 909                                 any_premultiplied = true;
 910                                 break;
 911                         case ALPHA_POSTMULTIPLIED:
 912                                 any_postmultiplied = true;
 913                                 break;
 914                         default:
 915                                 assert(false);
 916                         }
 917                 }
 918
 919                 if (any_invalid) {
 920                         node->output_alpha_type = ALPHA_INVALID;
 921                         continue;
 922                 }
 923
 924                 // Inputs must be of the same type.
 925                 if (any_premultiplied && any_postmultiplied) {
 926                         node->output_alpha_type = ALPHA_INVALID;
 927                         continue;
 928                 }
 929
 930                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 931                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 932                         // If the effect has asked for premultiplied alpha, check that it has got it.
 933                         if (any_postmultiplied) {
 934                                 node->output_alpha_type = ALPHA_INVALID;
 935                         } else if (!any_premultiplied &&
 936                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 937                                 // Blank input alpha, and the effect preserves blank alpha.
 938                                 node->output_alpha_type = ALPHA_BLANK;
 939                         } else {
 940                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 941                         }
 942                 } else {
 943                         // OK, all inputs are the same, and this effect is not going
 944                         // to change it.
 945                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 946                         if (any_premultiplied) {
 947                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 948                         } else if (any_postmultiplied) {
 949                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 950                         } else {
 951                                 node->output_alpha_type = ALPHA_BLANK;
 952                         }
 953                 }
 954         }
 955 }
 956
 957 bool EffectChain::node_needs_colorspace_fix(Node *node)
 958 {
 959         if (node->disabled) {
 960                 return false;
 961         }
 962         if (node->effect->num_inputs() == 0) {
 963                 return false;
 964         }
 965
 966         // propagate_gamma_and_color_space() has already set our output
 967         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 968         if (node->output_color_space == COLORSPACE_INVALID) {
 969                 return true;
 970         }
 971         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 972 }
 973
 974 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 975 // the graph. Our strategy is not always optimal, but quite simple:
 976 // Find an effect that's as early as possible where the inputs are of
 977 // unacceptable colorspaces (that is, either different, or, if the effect only
 978 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 979 // propagate the information anew, and repeat until there are no more such
 980 // effects.
 981 void EffectChain::fix_internal_color_spaces()
 982 {
 983         unsigned colorspace_propagation_pass = 0;
 984         bool found_any;
 985         do {
 986                 found_any = false;
 987                 for (unsigned i = 0; i < nodes.size(); ++i) {
 988                         Node *node = nodes[i];
 989                         if (!node_needs_colorspace_fix(node)) {
 990                                 continue;
 991                         }
 992
 993                         // Go through each input that is not sRGB, and insert
 994                         // a colorspace conversion after it.
 995                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 996                                 Node *input = node->incoming_links[j];
 997                                 assert(input->output_color_space != COLORSPACE_INVALID);
 998                                 if (input->output_color_space == COLORSPACE_sRGB) {
 999                                         continue;
1000                                 }
1001                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1002                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1003                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1004                                 conversion->output_color_space = COLORSPACE_sRGB;
1005                                 replace_sender(input, conversion);
1006                                 connect_nodes(input, conversion);
1007                         }
1008
1009                         // Re-sort topologically, and propagate the new information.
1010                         propagate_gamma_and_color_space();
1011
1012                         found_any = true;
1013                         break;
1014                 }
1015
1016                 char filename[256];
1017                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1018                 output_dot(filename);
1019                 assert(colorspace_propagation_pass < 100);
1020         } while (found_any);
1021
1022         for (unsigned i = 0; i < nodes.size(); ++i) {
1023                 Node *node = nodes[i];
1024                 if (node->disabled) {
1025                         continue;
1026                 }
1027                 assert(node->output_color_space != COLORSPACE_INVALID);
1028         }
1029 }
1030
1031 bool EffectChain::node_needs_alpha_fix(Node *node)
1032 {
1033         if (node->disabled) {
1034                 return false;
1035         }
1036
1037         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1038         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1039         return (node->output_alpha_type == ALPHA_INVALID);
1040 }
1041
1042 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1043 // the graph. Similar to fix_internal_color_spaces().
1044 void EffectChain::fix_internal_alpha(unsigned step)
1045 {
1046         unsigned alpha_propagation_pass = 0;
1047         bool found_any;
1048         do {
1049                 found_any = false;
1050                 for (unsigned i = 0; i < nodes.size(); ++i) {
1051                         Node *node = nodes[i];
1052                         if (!node_needs_alpha_fix(node)) {
1053                                 continue;
1054                         }
1055
1056                         // If we need to fix up GammaExpansionEffect, then clearly something
1057                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1058                         // is meaningless.
1059                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1060
1061                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1062
1063                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1064                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1065                                 assert(node->incoming_links.size() == 1);
1066                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1067                                 desired_type = ALPHA_POSTMULTIPLIED;
1068                         }
1069
1070                         // Go through each input that is not premultiplied alpha, and insert
1071                         // a conversion before it.
1072                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1073                                 Node *input = node->incoming_links[j];
1074                                 assert(input->output_alpha_type != ALPHA_INVALID);
1075                                 if (input->output_alpha_type == desired_type ||
1076                                     input->output_alpha_type == ALPHA_BLANK) {
1077                                         continue;
1078                                 }
1079                                 Node *conversion;
1080                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1081                                         conversion = add_node(new AlphaMultiplicationEffect());
1082                                 } else {
1083                                         conversion = add_node(new AlphaDivisionEffect());
1084                                 }
1085                                 conversion->output_alpha_type = desired_type;
1086                                 replace_sender(input, conversion);
1087                                 connect_nodes(input, conversion);
1088                         }
1089
1090                         // Re-sort topologically, and propagate the new information.
1091                         propagate_gamma_and_color_space();
1092                         propagate_alpha();
1093
1094                         found_any = true;
1095                         break;
1096                 }
1097
1098                 char filename[256];
1099                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1100                 output_dot(filename);
1101                 assert(alpha_propagation_pass < 100);
1102         } while (found_any);
1103
1104         for (unsigned i = 0; i < nodes.size(); ++i) {
1105                 Node *node = nodes[i];
1106                 if (node->disabled) {
1107                         continue;
1108                 }
1109                 assert(node->output_alpha_type != ALPHA_INVALID);
1110         }
1111 }
1112
1113 // Make so that the output is in the desired color space.
1114 void EffectChain::fix_output_color_space()
1115 {
1116         Node *output = find_output_node();
1117         if (output->output_color_space != output_format.color_space) {
1118                 Node *conversion = add_node(new ColorspaceConversionEffect());
1119                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1120                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1121                 conversion->output_color_space = output_format.color_space;
1122                 connect_nodes(output, conversion);
1123                 propagate_alpha();
1124                 propagate_gamma_and_color_space();
1125         }
1126 }
1127
1128 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1129 void EffectChain::fix_output_alpha()
1130 {
1131         Node *output = find_output_node();
1132         assert(output->output_alpha_type != ALPHA_INVALID);
1133         if (output->output_alpha_type == ALPHA_BLANK) {
1134                 // No alpha output, so we don't care.
1135                 return;
1136         }
1137         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1138             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1139                 Node *conversion = add_node(new AlphaDivisionEffect());
1140                 connect_nodes(output, conversion);
1141                 propagate_alpha();
1142                 propagate_gamma_and_color_space();
1143         }
1144         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1145             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1146                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1147                 connect_nodes(output, conversion);
1148                 propagate_alpha();
1149                 propagate_gamma_and_color_space();
1150         }
1151 }
1152
1153 bool EffectChain::node_needs_gamma_fix(Node *node)
1154 {
1155         if (node->disabled) {
1156                 return false;
1157         }
1158
1159         // Small hack since the output is not an explicit node:
1160         // If we are the last node and our output is in the wrong
1161         // space compared to EffectChain's output, we need to fix it.
1162         // This will only take us to linear, but fix_output_gamma()
1163         // will come and take us to the desired output gamma
1164         // if it is needed.
1165         //
1166         // This needs to be before everything else, since it could
1167         // even apply to inputs (if they are the only effect).
1168         if (node->outgoing_links.empty() &&
1169             node->output_gamma_curve != output_format.gamma_curve &&
1170             node->output_gamma_curve != GAMMA_LINEAR) {
1171                 return true;
1172         }
1173
1174         if (node->effect->num_inputs() == 0) {
1175                 return false;
1176         }
1177
1178         // propagate_gamma_and_color_space() has already set our output
1179         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1180         // except for GammaCompressionEffect.
1181         if (node->output_gamma_curve == GAMMA_INVALID) {
1182                 return true;
1183         }
1184         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1185                 assert(node->incoming_links.size() == 1);
1186                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1187         }
1188
1189         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1190 }
1191
1192 // Very similar to fix_internal_color_spaces(), but for gamma.
1193 // There is one difference, though; before we start adding conversion nodes,
1194 // we see if we can get anything out of asking the sources to deliver
1195 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1196 // does that part, while fix_internal_gamma_by_inserting_nodes()
1197 // inserts nodes as needed afterwards.
1198 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1199 {
1200         unsigned gamma_propagation_pass = 0;
1201         bool found_any;
1202         do {
1203                 found_any = false;
1204                 for (unsigned i = 0; i < nodes.size(); ++i) {
1205                         Node *node = nodes[i];
1206                         if (!node_needs_gamma_fix(node)) {
1207                                 continue;
1208                         }
1209
1210                         // See if all inputs can give us linear gamma. If not, leave it.
1211                         std::vector<Node *> nonlinear_inputs;
1212                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1213                         assert(!nonlinear_inputs.empty());
1214
1215                         bool all_ok = true;
1216                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1217                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1218                                 all_ok &= input->can_output_linear_gamma();
1219                         }
1220
1221                         if (!all_ok) {
1222                                 continue;
1223                         }
1224
1225                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1226                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1227                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1228                         }
1229
1230                         // Re-sort topologically, and propagate the new information.
1231                         propagate_gamma_and_color_space();
1232
1233                         found_any = true;
1234                         break;
1235                 }
1236
1237                 char filename[256];
1238                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1239                 output_dot(filename);
1240                 assert(gamma_propagation_pass < 100);
1241         } while (found_any);
1242 }
1243
1244 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1245 {
1246         unsigned gamma_propagation_pass = 0;
1247         bool found_any;
1248         do {
1249                 found_any = false;
1250                 for (unsigned i = 0; i < nodes.size(); ++i) {
1251                         Node *node = nodes[i];
1252                         if (!node_needs_gamma_fix(node)) {
1253                                 continue;
1254                         }
1255
1256                         // Special case: We could be an input and still be asked to
1257                         // fix our gamma; if so, we should be the only node
1258                         // (as node_needs_gamma_fix() would only return true in
1259                         // for an input in that case). That means we should insert
1260                         // a conversion node _after_ ourselves.
1261                         if (node->incoming_links.empty()) {
1262                                 assert(node->outgoing_links.empty());
1263                                 Node *conversion = add_node(new GammaExpansionEffect());
1264                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1265                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1266                                 connect_nodes(node, conversion);
1267                         }
1268
1269                         // If not, go through each input that is not linear gamma,
1270                         // and insert a gamma conversion after it.
1271                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1272                                 Node *input = node->incoming_links[j];
1273                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1274                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1275                                         continue;
1276                                 }
1277                                 Node *conversion = add_node(new GammaExpansionEffect());
1278                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1279                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1280                                 replace_sender(input, conversion);
1281                                 connect_nodes(input, conversion);
1282                         }
1283
1284                         // Re-sort topologically, and propagate the new information.
1285                         propagate_alpha();
1286                         propagate_gamma_and_color_space();
1287
1288                         found_any = true;
1289                         break;
1290                 }
1291
1292                 char filename[256];
1293                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1294                 output_dot(filename);
1295                 assert(gamma_propagation_pass < 100);
1296         } while (found_any);
1297
1298         for (unsigned i = 0; i < nodes.size(); ++i) {
1299                 Node *node = nodes[i];
1300                 if (node->disabled) {
1301                         continue;
1302                 }
1303                 assert(node->output_gamma_curve != GAMMA_INVALID);
1304         }
1305 }
1306
1307 // Make so that the output is in the desired gamma.
1308 // Note that this assumes linear input gamma, so it might create the need
1309 // for another pass of fix_internal_gamma().
1310 void EffectChain::fix_output_gamma()
1311 {
1312         Node *output = find_output_node();
1313         if (output->output_gamma_curve != output_format.gamma_curve) {
1314                 Node *conversion = add_node(new GammaCompressionEffect());
1315                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1316                 conversion->output_gamma_curve = output_format.gamma_curve;
1317                 connect_nodes(output, conversion);
1318         }
1319 }
1320
1321 // If the user has requested dither, add a DitherEffect right at the end
1322 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1323 // since dither is about the only effect that can _not_ be done in linear space.
1324 void EffectChain::add_dither_if_needed()
1325 {
1326         if (num_dither_bits == 0) {
1327                 return;
1328         }
1329         Node *output = find_output_node();
1330         Node *dither = add_node(new DitherEffect());
1331         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1332         connect_nodes(output, dither);
1333
1334         dither_effect = dither->effect;
1335 }
1336
1337 // Find the output node. This is, simply, one that has no outgoing links.
1338 // If there are multiple ones, the graph is malformed (we do not support
1339 // multiple outputs right now).
1340 Node *EffectChain::find_output_node()
1341 {
1342         std::vector<Node *> output_nodes;
1343         for (unsigned i = 0; i < nodes.size(); ++i) {
1344                 Node *node = nodes[i];
1345                 if (node->disabled) {
1346                         continue;
1347                 }
1348                 if (node->outgoing_links.empty()) {
1349                         output_nodes.push_back(node);
1350                 }
1351         }
1352         assert(output_nodes.size() == 1);
1353         return output_nodes[0];
1354 }
1355
1356 void EffectChain::finalize()
1357 {
1358         // Save the current locale, and set it to C, so that we can output decimal
1359         // numbers with printf and be sure to get them in the format mandated by GLSL.
1360         char *saved_locale = setlocale(LC_NUMERIC, "C");
1361
1362         // Output the graph as it is before we do any conversions on it.
1363         output_dot("step0-start.dot");
1364
1365         // Give each effect in turn a chance to rewrite its own part of the graph.
1366         // Note that if more effects are added as part of this, they will be
1367         // picked up as part of the same for loop, since they are added at the end.
1368         for (unsigned i = 0; i < nodes.size(); ++i) {
1369                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1370         }
1371         output_dot("step1-rewritten.dot");
1372
1373         find_color_spaces_for_inputs();
1374         output_dot("step2-input-colorspace.dot");
1375
1376         propagate_alpha();
1377         output_dot("step3-propagated-alpha.dot");
1378
1379         propagate_gamma_and_color_space();
1380         output_dot("step4-propagated-all.dot");
1381
1382         fix_internal_color_spaces();
1383         fix_internal_alpha(6);
1384         fix_output_color_space();
1385         output_dot("step7-output-colorspacefix.dot");
1386         fix_output_alpha();
1387         output_dot("step8-output-alphafix.dot");
1388
1389         // Note that we need to fix gamma after colorspace conversion,
1390         // because colorspace conversions might create needs for gamma conversions.
1391         // Also, we need to run an extra pass of fix_internal_gamma() after
1392         // fixing the output gamma, as we only have conversions to/from linear,
1393         // and fix_internal_alpha() since GammaCompressionEffect needs
1394         // postmultiplied input.
1395         fix_internal_gamma_by_asking_inputs(9);
1396         fix_internal_gamma_by_inserting_nodes(10);
1397         fix_output_gamma();
1398         output_dot("step11-output-gammafix.dot");
1399         propagate_alpha();
1400         output_dot("step12-output-alpha-propagated.dot");
1401         fix_internal_alpha(13);
1402         output_dot("step14-output-alpha-fixed.dot");
1403         fix_internal_gamma_by_asking_inputs(15);
1404         fix_internal_gamma_by_inserting_nodes(16);
1405
1406         output_dot("step17-before-dither.dot");
1407
1408         add_dither_if_needed();
1409
1410         output_dot("step18-final.dot");
1411
1412         // Construct all needed GLSL programs, starting at the output.
1413         construct_glsl_programs(find_output_node());
1414
1415         output_dot("step19-split-to-phases.dot");
1416
1417         // If we have more than one phase, we need intermediate render-to-texture.
1418         // Construct an FBO, and then as many textures as we need.
1419         // We choose the simplest option of having one texture per output,
1420         // since otherwise this turns into an (albeit simple)
1421         // register allocation problem.
1422         if (phases.size() > 1) {
1423                 for (unsigned i = 0; i < phases.size() - 1; ++i) {
1424                         inform_input_sizes(phases[i]);
1425                         find_output_size(phases[i]);
1426
1427                         Node *output_node = phases[i]->effects.back();
1428                         output_node->output_texture = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height);
1429                         check_error();
1430                         glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1431                         check_error();
1432                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1433                         check_error();
1434                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1435                         check_error();
1436
1437                         output_node->output_texture_width = phases[i]->output_width;
1438                         output_node->output_texture_height = phases[i]->output_height;
1439                 }
1440                 inform_input_sizes(phases.back());
1441         }
1442
1443         for (unsigned i = 0; i < inputs.size(); ++i) {
1444                 inputs[i]->finalize();
1445         }
1446
1447         assert(phases[0]->inputs.empty());
1448
1449         finalized = true;
1450         setlocale(LC_NUMERIC, saved_locale);
1451 }
1452
1453 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1454 {
1455         assert(finalized);
1456
1457         // Save original viewport.
1458         GLuint x = 0, y = 0;
1459         GLuint fbo = 0;
1460
1461         if (width == 0 && height == 0) {
1462                 GLint viewport[4];
1463                 glGetIntegerv(GL_VIEWPORT, viewport);
1464                 x = viewport[0];
1465                 y = viewport[1];
1466                 width = viewport[2];
1467                 height = viewport[3];
1468         }
1469
1470         // Basic state.
1471         glDisable(GL_BLEND);
1472         check_error();
1473         glDisable(GL_DEPTH_TEST);
1474         check_error();
1475         glDepthMask(GL_FALSE);
1476         check_error();
1477
1478         glMatrixMode(GL_PROJECTION);
1479         glLoadIdentity();
1480         glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0);
1481
1482         glMatrixMode(GL_MODELVIEW);
1483         glLoadIdentity();
1484
1485         if (phases.size() > 1) {
1486                 glGenFramebuffers(1, &fbo);
1487                 check_error();
1488                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1489                 check_error();
1490         }
1491
1492         std::set<Node *> generated_mipmaps;
1493
1494         for (unsigned phase = 0; phase < phases.size(); ++phase) {
1495                 // See if the requested output size has changed. If so, we need to recreate
1496                 // the texture (and before we start setting up inputs).
1497                 inform_input_sizes(phases[phase]);
1498                 if (phase != phases.size() - 1) {
1499                         find_output_size(phases[phase]);
1500
1501                         Node *output_node = phases[phase]->effects.back();
1502
1503                         if (output_node->output_texture_width != phases[phase]->output_width ||
1504                             output_node->output_texture_height != phases[phase]->output_height) {
1505                                 glActiveTexture(GL_TEXTURE0);
1506                                 check_error();
1507                                 glBindTexture(GL_TEXTURE_2D, output_node->output_texture);
1508                                 check_error();
1509                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1510                                 check_error();
1511                                 glBindTexture(GL_TEXTURE_2D, 0);
1512                                 check_error();
1513
1514                                 output_node->output_texture_width = phases[phase]->output_width;
1515                                 output_node->output_texture_height = phases[phase]->output_height;
1516                         }
1517                 }
1518
1519                 glUseProgram(phases[phase]->glsl_program_num);
1520                 check_error();
1521
1522                 // Set up RTT inputs for this phase.
1523                 for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
1524                         glActiveTexture(GL_TEXTURE0 + sampler);
1525                         Node *input = phases[phase]->inputs[sampler];
1526                         glBindTexture(GL_TEXTURE_2D, input->output_texture);
1527                         check_error();
1528                         if (phases[phase]->input_needs_mipmaps) {
1529                                 if (generated_mipmaps.count(input) == 0) {
1530                                         glGenerateMipmap(GL_TEXTURE_2D);
1531                                         check_error();
1532                                         generated_mipmaps.insert(input);
1533                                 }
1534                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1535                                 check_error();
1536                         } else {
1537                                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1538                                 check_error();
1539                         }
1540
1541                         std::string texture_name = std::string("tex_") + phases[phase]->effect_ids[input];
1542                         glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
1543                         check_error();
1544                 }
1545
1546                 // And now the output.
1547                 if (phase == phases.size() - 1) {
1548                         // Last phase goes to the output the user specified.
1549                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1550                         check_error();
1551                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1552                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1553                         glViewport(x, y, width, height);
1554                         if (dither_effect != NULL) {
1555                                 CHECK(dither_effect->set_int("output_width", width));
1556                                 CHECK(dither_effect->set_int("output_height", height));
1557                         }
1558                 } else {
1559                         Node *output_node = phases[phase]->effects.back();
1560                         glFramebufferTexture2D(
1561                                 GL_FRAMEBUFFER,
1562                                 GL_COLOR_ATTACHMENT0,
1563                                 GL_TEXTURE_2D,
1564                                 output_node->output_texture,
1565                                 0);
1566                         check_error();
1567                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1568                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1569                         glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
1570                 }
1571
1572                 // Give the required parameters to all the effects.
1573                 unsigned sampler_num = phases[phase]->inputs.size();
1574                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1575                         Node *node = phases[phase]->effects[i];
1576                         node->effect->set_gl_state(phases[phase]->glsl_program_num, phases[phase]->effect_ids[node], &sampler_num);
1577                         check_error();
1578                 }
1579
1580                 // Now draw!
1581                 glBegin(GL_QUADS);
1582
1583                 glTexCoord2f(0.0f, 0.0f);
1584                 glVertex2f(0.0f, 0.0f);
1585
1586                 glTexCoord2f(1.0f, 0.0f);
1587                 glVertex2f(1.0f, 0.0f);
1588
1589                 glTexCoord2f(1.0f, 1.0f);
1590                 glVertex2f(1.0f, 1.0f);
1591
1592                 glTexCoord2f(0.0f, 1.0f);
1593                 glVertex2f(0.0f, 1.0f);
1594
1595                 glEnd();
1596                 check_error();
1597
1598                 for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
1599                         Node *node = phases[phase]->effects[i];
1600                         node->effect->clear_gl_state();
1601                 }
1602         }
1603
1604         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1605         check_error();
1606
1607         if (fbo != 0) {
1608                 glDeleteFramebuffers(1, &fbo);
1609                 check_error();
1610         }
1611 }