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

   1 #define GL_GLEXT_PROTOTYPES 1
   2
   3 #include <epoxy/gl.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 <utility>
  15 #include <vector>
  16 #if defined(__APPLE__)
  17 #include <xlocale.h>
  18 #endif
  19
  20 #include "alpha_division_effect.h"
  21 #include "alpha_multiplication_effect.h"
  22 #include "colorspace_conversion_effect.h"
  23 #include "dither_effect.h"
  24 #include "effect.h"
  25 #include "effect_chain.h"
  26 #include "gamma_compression_effect.h"
  27 #include "gamma_expansion_effect.h"
  28 #include "init.h"
  29 #include "input.h"
  30 #include "resource_pool.h"
  31 #include "util.h"
  32
  33 using namespace std;
  34
  35 namespace movit {
  36
  37 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
  38         : aspect_nom(aspect_nom),
  39           aspect_denom(aspect_denom),
  40           dither_effect(NULL),
  41           num_dither_bits(0),
  42           finalized(false),
  43           resource_pool(resource_pool) {
  44         if (resource_pool == NULL) {
  45                 this->resource_pool = new ResourcePool();
  46                 owns_resource_pool = true;
  47         } else {
  48                 owns_resource_pool = false;
  49         }
  50 }
  51
  52 EffectChain::~EffectChain()
  53 {
  54         for (unsigned i = 0; i < nodes.size(); ++i) {
  55                 delete nodes[i]->effect;
  56                 delete nodes[i];
  57         }
  58         for (unsigned i = 0; i < phases.size(); ++i) {
  59                 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
  60                 delete phases[i];
  61         }
  62         if (owns_resource_pool) {
  63                 delete resource_pool;
  64         }
  65 }
  66
  67 Input *EffectChain::add_input(Input *input)
  68 {
  69         assert(!finalized);
  70         inputs.push_back(input);
  71         add_node(input);
  72         return input;
  73 }
  74
  75 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
  76 {
  77         assert(!finalized);
  78         output_format = format;
  79         output_alpha_format = alpha_format;
  80 }
  81
  82 Node *EffectChain::add_node(Effect *effect)
  83 {
  84         for (unsigned i = 0; i < nodes.size(); ++i) {
  85                 assert(nodes[i]->effect != effect);
  86         }
  87
  88         Node *node = new Node;
  89         node->effect = effect;
  90         node->disabled = false;
  91         node->output_color_space = COLORSPACE_INVALID;
  92         node->output_gamma_curve = GAMMA_INVALID;
  93         node->output_alpha_type = ALPHA_INVALID;
  94
  95         nodes.push_back(node);
  96         node_map[effect] = node;
  97         effect->inform_added(this);
  98         return node;
  99 }
 100
 101 void EffectChain::connect_nodes(Node *sender, Node *receiver)
 102 {
 103         sender->outgoing_links.push_back(receiver);
 104         receiver->incoming_links.push_back(sender);
 105 }
 106
 107 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
 108 {
 109         new_receiver->incoming_links = old_receiver->incoming_links;
 110         old_receiver->incoming_links.clear();
 111
 112         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
 113                 Node *sender = new_receiver->incoming_links[i];
 114                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
 115                         if (sender->outgoing_links[j] == old_receiver) {
 116                                 sender->outgoing_links[j] = new_receiver;
 117                         }
 118                 }
 119         }
 120 }
 121
 122 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 123 {
 124         new_sender->outgoing_links = old_sender->outgoing_links;
 125         old_sender->outgoing_links.clear();
 126
 127         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 128                 Node *receiver = new_sender->outgoing_links[i];
 129                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 130                         if (receiver->incoming_links[j] == old_sender) {
 131                                 receiver->incoming_links[j] = new_sender;
 132                         }
 133                 }
 134         }
 135 }
 136
 137 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 138 {
 139         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 140                 if (sender->outgoing_links[i] == receiver) {
 141                         sender->outgoing_links[i] = middle;
 142                         middle->incoming_links.push_back(sender);
 143                 }
 144         }
 145         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 146                 if (receiver->incoming_links[i] == sender) {
 147                         receiver->incoming_links[i] = middle;
 148                         middle->outgoing_links.push_back(receiver);
 149                 }
 150         }
 151
 152         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 153 }
 154
 155 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
 156 {
 157         assert(node->effect->needs_texture_bounce());
 158         assert(input_num < node->incoming_links.size());
 159         assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
 160         assert(node->incoming_links[input_num]->bound_sampler_num < 8);
 161         return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
 162 }
 163
 164 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
 165 {
 166         if (node->output_gamma_curve == GAMMA_LINEAR &&
 167             node->effect->effect_type_id() != "GammaCompressionEffect") {
 168                 return;
 169         }
 170         if (node->effect->num_inputs() == 0) {
 171                 nonlinear_inputs->push_back(node);
 172         } else {
 173                 assert(node->effect->num_inputs() == node->incoming_links.size());
 174                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 175                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 176                 }
 177         }
 178 }
 179
 180 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
 181 {
 182         assert(!finalized);
 183         assert(inputs.size() == effect->num_inputs());
 184         Node *node = add_node(effect);
 185         for (unsigned i = 0; i < inputs.size(); ++i) {
 186                 assert(node_map.count(inputs[i]) != 0);
 187                 connect_nodes(node_map[inputs[i]], node);
 188         }
 189         return effect;
 190 }
 191
 192 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 193 string replace_prefix(const string &text, const string &prefix)
 194 {
 195         string output;
 196         size_t start = 0;
 197
 198         while (start < text.size()) {
 199                 size_t pos = text.find("PREFIX(", start);
 200                 if (pos == string::npos) {
 201                         output.append(text.substr(start, string::npos));
 202                         break;
 203                 }
 204
 205                 output.append(text.substr(start, pos - start));
 206                 output.append(prefix);
 207                 output.append("_");
 208
 209                 pos += strlen("PREFIX(");
 210
 211                 // Output stuff until we find the matching ), which we then eat.
 212                 int depth = 1;
 213                 size_t end_arg_pos = pos;
 214                 while (end_arg_pos < text.size()) {
 215                         if (text[end_arg_pos] == '(') {
 216                                 ++depth;
 217                         } else if (text[end_arg_pos] == ')') {
 218                                 --depth;
 219                                 if (depth == 0) {
 220                                         break;
 221                                 }
 222                         }
 223                         ++end_arg_pos;
 224                 }
 225                 output.append(text.substr(pos, end_arg_pos - pos));
 226                 ++end_arg_pos;
 227                 assert(depth == 0);
 228                 start = end_arg_pos;
 229         }
 230         return output;
 231 }
 232
 233 void EffectChain::compile_glsl_program(Phase *phase)
 234 {
 235         string frag_shader = read_version_dependent_file("header", "frag");
 236
 237         // Create functions for all the texture inputs that we need.
 238         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 239                 Node *input = phase->inputs[i]->output_node;
 240                 char effect_id[256];
 241                 sprintf(effect_id, "in%u", i);
 242                 phase->effect_ids.insert(make_pair(input, effect_id));
 243
 244                 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
 245                 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
 246                 frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
 247                 frag_shader += "}\n";
 248                 frag_shader += "\n";
 249         }
 250
 251         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 252                 Node *node = phase->effects[i];
 253                 char effect_id[256];
 254                 sprintf(effect_id, "eff%u", i);
 255                 phase->effect_ids.insert(make_pair(node, effect_id));
 256
 257                 if (node->incoming_links.size() == 1) {
 258                         frag_shader += 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 += 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         frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
 285         frag_shader.append(read_version_dependent_file("footer", "frag"));
 286
 287         string vert_shader = read_version_dependent_file("vs", "vert");
 288         phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader);
 289 }
 290
 291 // Construct GLSL programs, starting at the given effect and following
 292 // the chain from there. We end a program every time we come to an effect
 293 // marked as "needs texture bounce", one that is used by multiple other
 294 // effects, every time an effect wants to change the output size,
 295 // and of course at the end.
 296 //
 297 // We follow a quite simple depth-first search from the output, although
 298 // without recursing explicitly within each phase.
 299 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
 300 {
 301         if (completed_effects->count(output)) {
 302                 return (*completed_effects)[output];
 303         }
 304
 305         Phase *phase = new Phase;
 306         phase->output_node = output;
 307
 308         // Effects that we have yet to calculate, but that we know should
 309         // be in the current phase.
 310         stack<Node *> effects_todo_this_phase;
 311         effects_todo_this_phase.push(output);
 312
 313         while (!effects_todo_this_phase.empty()) {
 314                 Node *node = effects_todo_this_phase.top();
 315                 effects_todo_this_phase.pop();
 316
 317                 // This should currently only happen for effects that are inputs
 318                 // (either true inputs or phase outputs). We special-case inputs,
 319                 // and then deduplicate phase outputs below.
 320                 if (node->effect->num_inputs() == 0) {
 321                         if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
 322                                 continue;
 323                         }
 324                 } else {
 325                         assert(completed_effects->count(node) == 0);
 326                 }
 327
 328                 phase->effects.push_back(node);
 329
 330                 // Find all the dependencies of this effect, and add them to the stack.
 331                 vector<Node *> deps = node->incoming_links;
 332                 assert(node->effect->num_inputs() == deps.size());
 333                 for (unsigned i = 0; i < deps.size(); ++i) {
 334                         bool start_new_phase = false;
 335
 336                         if (node->effect->needs_texture_bounce() &&
 337                             !deps[i]->effect->is_single_texture()) {
 338                                 start_new_phase = true;
 339                         }
 340
 341                         if (deps[i]->outgoing_links.size() > 1) {
 342                                 if (!deps[i]->effect->is_single_texture()) {
 343                                         // More than one effect uses this as the input,
 344                                         // and it is not a texture itself.
 345                                         // The easiest thing to do (and probably also the safest
 346                                         // performance-wise in most cases) is to bounce it to a texture
 347                                         // and then let the next passes read from that.
 348                                         start_new_phase = true;
 349                                 } else {
 350                                         assert(deps[i]->effect->num_inputs() == 0);
 351
 352                                         // For textures, we try to be slightly more clever;
 353                                         // if none of our outputs need a bounce, we don't bounce
 354                                         // but instead simply use the effect many times.
 355                                         //
 356                                         // Strictly speaking, we could bounce it for some outputs
 357                                         // and use it directly for others, but the processing becomes
 358                                         // somewhat simpler if the effect is only used in one such way.
 359                                         for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 360                                                 Node *rdep = deps[i]->outgoing_links[j];
 361                                                 start_new_phase |= rdep->effect->needs_texture_bounce();
 362                                         }
 363                                 }
 364                         }
 365
 366                         if (deps[i]->effect->changes_output_size()) {
 367                                 start_new_phase = true;
 368                         }
 369
 370                         if (start_new_phase) {
 371                                 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
 372                         } else {
 373                                 effects_todo_this_phase.push(deps[i]);
 374                         }
 375                 }
 376         }
 377
 378         // No more effects to do this phase. Take all the ones we have,
 379         // and create a GLSL program for it.
 380         assert(!phase->effects.empty());
 381
 382         // Deduplicate the inputs.
 383         sort(phase->inputs.begin(), phase->inputs.end());
 384         phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end());
 385
 386         // We added the effects from the output and back, but we need to output
 387         // them in topological sort order in the shader.
 388         phase->effects = topological_sort(phase->effects);
 389
 390         // Figure out if we need mipmaps or not, and if so, tell the inputs that.
 391         phase->input_needs_mipmaps = false;
 392         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 393                 Node *node = phase->effects[i];
 394                 phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
 395         }
 396         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 397                 Node *node = phase->effects[i];
 398                 if (node->effect->num_inputs() == 0) {
 399                         CHECK(node->effect->set_int("needs_mipmaps", phase->input_needs_mipmaps));
 400                 }
 401         }
 402
 403         // Actually make the shader for this phase.
 404         compile_glsl_program(phase);
 405
 406         assert(completed_effects->count(output) == 0);
 407         completed_effects->insert(make_pair(output, phase));
 408         phases.push_back(phase);
 409         return phase;
 410 }
 411
 412 void EffectChain::output_dot(const char *filename)
 413 {
 414         if (movit_debug_level != MOVIT_DEBUG_ON) {
 415                 return;
 416         }
 417
 418         FILE *fp = fopen(filename, "w");
 419         if (fp == NULL) {
 420                 perror(filename);
 421                 exit(1);
 422         }
 423
 424         fprintf(fp, "digraph G {\n");
 425         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 426         for (unsigned i = 0; i < nodes.size(); ++i) {
 427                 // Find out which phase this event belongs to.
 428                 vector<int> in_phases;
 429                 for (unsigned j = 0; j < phases.size(); ++j) {
 430                         const Phase* p = phases[j];
 431                         if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 432                                 in_phases.push_back(j);
 433                         }
 434                 }
 435
 436                 if (in_phases.empty()) {
 437                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 438                 } else if (in_phases.size() == 1) {
 439                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 440                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 441                                 (in_phases[0] % 8) + 1);
 442                 } else {
 443                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 444                         // But alas.
 445                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 446                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 447                                 (in_phases[0] % 8) + 1);
 448                 }
 449
 450                 char from_node_id[256];
 451                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 452
 453                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 454                         char to_node_id[256];
 455                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 456
 457                         vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 458                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 459                 }
 460
 461                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 462                         // Output node.
 463                         vector<string> labels = get_labels_for_edge(nodes[i], NULL);
 464                         output_dot_edge(fp, from_node_id, "output", labels);
 465                 }
 466         }
 467         fprintf(fp, "}\n");
 468
 469         fclose(fp);
 470 }
 471
 472 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 473 {
 474         vector<string> labels;
 475
 476         if (to != NULL && to->effect->needs_texture_bounce()) {
 477                 labels.push_back("needs_bounce");
 478         }
 479         if (from->effect->changes_output_size()) {
 480                 labels.push_back("resize");
 481         }
 482
 483         switch (from->output_color_space) {
 484         case COLORSPACE_INVALID:
 485                 labels.push_back("spc[invalid]");
 486                 break;
 487         case COLORSPACE_REC_601_525:
 488                 labels.push_back("spc[rec601-525]");
 489                 break;
 490         case COLORSPACE_REC_601_625:
 491                 labels.push_back("spc[rec601-625]");
 492                 break;
 493         default:
 494                 break;
 495         }
 496
 497         switch (from->output_gamma_curve) {
 498         case GAMMA_INVALID:
 499                 labels.push_back("gamma[invalid]");
 500                 break;
 501         case GAMMA_sRGB:
 502                 labels.push_back("gamma[sRGB]");
 503                 break;
 504         case GAMMA_REC_601:  // and GAMMA_REC_709
 505                 labels.push_back("gamma[rec601/709]");
 506                 break;
 507         default:
 508                 break;
 509         }
 510
 511         switch (from->output_alpha_type) {
 512         case ALPHA_INVALID:
 513                 labels.push_back("alpha[invalid]");
 514                 break;
 515         case ALPHA_BLANK:
 516                 labels.push_back("alpha[blank]");
 517                 break;
 518         case ALPHA_POSTMULTIPLIED:
 519                 labels.push_back("alpha[postmult]");
 520                 break;
 521         default:
 522                 break;
 523         }
 524
 525         return labels;
 526 }
 527
 528 void EffectChain::output_dot_edge(FILE *fp,
 529                                   const string &from_node_id,
 530                                   const string &to_node_id,
 531                                   const vector<string> &labels)
 532 {
 533         if (labels.empty()) {
 534                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 535         } else {
 536                 string label = labels[0];
 537                 for (unsigned k = 1; k < labels.size(); ++k) {
 538                         label += ", " + labels[k];
 539                 }
 540                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 541         }
 542 }
 543
 544 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
 545 {
 546         unsigned scaled_width, scaled_height;
 547
 548         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 549                 // Same aspect, or W/H > aspect (image is wider than the frame).
 550                 // In either case, keep width, and adjust height.
 551                 scaled_width = width;
 552                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
 553         } else {
 554                 // W/H < aspect (image is taller than the frame), so keep height,
 555                 // and adjust width.
 556                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
 557                 scaled_height = height;
 558         }
 559
 560         // We should be consistently larger or smaller then the existing choice,
 561         // since we have the same aspect.
 562         assert(!(scaled_width < *output_width && scaled_height > *output_height));
 563         assert(!(scaled_height < *output_height && scaled_width > *output_width));
 564
 565         if (scaled_width >= *output_width && scaled_height >= *output_height) {
 566                 *output_width = scaled_width;
 567                 *output_height = scaled_height;
 568         }
 569 }
 570
 571 // Propagate input texture sizes throughout, and inform effects downstream.
 572 // (Like a lot of other code, we depend on effects being in topological order.)
 573 void EffectChain::inform_input_sizes(Phase *phase)
 574 {
 575         // All effects that have a defined size (inputs and RTT inputs)
 576         // get that. Reset all others.
 577         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 578                 Node *node = phase->effects[i];
 579                 if (node->effect->num_inputs() == 0) {
 580                         Input *input = static_cast<Input *>(node->effect);
 581                         node->output_width = input->get_width();
 582                         node->output_height = input->get_height();
 583                         assert(node->output_width != 0);
 584                         assert(node->output_height != 0);
 585                 } else {
 586                         node->output_width = node->output_height = 0;
 587                 }
 588         }
 589         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 590                 Phase *input = phase->inputs[i];
 591                 input->output_node->output_width = input->virtual_output_width;
 592                 input->output_node->output_height = input->virtual_output_height;
 593                 assert(input->output_node->output_width != 0);
 594                 assert(input->output_node->output_height != 0);
 595         }
 596
 597         // Now propagate from the inputs towards the end, and inform as we go.
 598         // The rules are simple:
 599         //
 600         //   1. Don't touch effects that already have given sizes (ie., inputs).
 601         //   2. If all of your inputs have the same size, that will be your output size.
 602         //   3. Otherwise, your output size is 0x0.
 603         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 604                 Node *node = phase->effects[i];
 605                 if (node->effect->num_inputs() == 0) {
 606                         continue;
 607                 }
 608                 unsigned this_output_width = 0;
 609                 unsigned this_output_height = 0;
 610                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 611                         Node *input = node->incoming_links[j];
 612                         node->effect->inform_input_size(j, input->output_width, input->output_height);
 613                         if (j == 0) {
 614                                 this_output_width = input->output_width;
 615                                 this_output_height = input->output_height;
 616                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
 617                                 // Inputs disagree.
 618                                 this_output_width = 0;
 619                                 this_output_height = 0;
 620                         }
 621                 }
 622                 node->output_width = this_output_width;
 623                 node->output_height = this_output_height;
 624         }
 625 }
 626
 627 // Note: You should call inform_input_sizes() before this, as the last effect's
 628 // desired output size might change based on the inputs.
 629 void EffectChain::find_output_size(Phase *phase)
 630 {
 631         Node *output_node = phase->effects.back();
 632
 633         // If the last effect explicitly sets an output size, use that.
 634         if (output_node->effect->changes_output_size()) {
 635                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
 636                                                      &phase->virtual_output_width, &phase->virtual_output_height);
 637                 return;
 638         }
 639
 640         // If all effects have the same size, use that.
 641         unsigned output_width = 0, output_height = 0;
 642         bool all_inputs_same_size = true;
 643
 644         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 645                 Phase *input = phase->inputs[i];
 646                 assert(input->output_width != 0);
 647                 assert(input->output_height != 0);
 648                 if (output_width == 0 && output_height == 0) {
 649                         output_width = input->virtual_output_width;
 650                         output_height = input->virtual_output_height;
 651                 } else if (output_width != input->virtual_output_width ||
 652                            output_height != input->virtual_output_height) {
 653                         all_inputs_same_size = false;
 654                 }
 655         }
 656         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 657                 Effect *effect = phase->effects[i]->effect;
 658                 if (effect->num_inputs() != 0) {
 659                         continue;
 660                 }
 661
 662                 Input *input = static_cast<Input *>(effect);
 663                 if (output_width == 0 && output_height == 0) {
 664                         output_width = input->get_width();
 665                         output_height = input->get_height();
 666                 } else if (output_width != input->get_width() ||
 667                            output_height != input->get_height()) {
 668                         all_inputs_same_size = false;
 669                 }
 670         }
 671
 672         if (all_inputs_same_size) {
 673                 assert(output_width != 0);
 674                 assert(output_height != 0);
 675                 phase->virtual_output_width = phase->output_width = output_width;
 676                 phase->virtual_output_height = phase->output_height = output_height;
 677                 return;
 678         }
 679
 680         // If not, fit all the inputs into the current aspect, and select the largest one.
 681         output_width = 0;
 682         output_height = 0;
 683         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 684                 Phase *input = phase->inputs[i];
 685                 assert(input->output_width != 0);
 686                 assert(input->output_height != 0);
 687                 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
 688         }
 689         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 690                 Effect *effect = phase->effects[i]->effect;
 691                 if (effect->num_inputs() != 0) {
 692                         continue;
 693                 }
 694
 695                 Input *input = static_cast<Input *>(effect);
 696                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
 697         }
 698         assert(output_width != 0);
 699         assert(output_height != 0);
 700         phase->virtual_output_width = phase->output_width = output_width;
 701         phase->virtual_output_height = phase->output_height = output_height;
 702 }
 703
 704 void EffectChain::sort_all_nodes_topologically()
 705 {
 706         nodes = topological_sort(nodes);
 707 }
 708
 709 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
 710 {
 711         set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
 712         vector<Node *> sorted_list;
 713         for (unsigned i = 0; i < nodes.size(); ++i) {
 714                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
 715         }
 716         reverse(sorted_list.begin(), sorted_list.end());
 717         return sorted_list;
 718 }
 719
 720 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
 721 {
 722         if (nodes_left_to_visit->count(node) == 0) {
 723                 return;
 724         }
 725         nodes_left_to_visit->erase(node);
 726         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
 727                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
 728         }
 729         sorted_list->push_back(node);
 730 }
 731
 732 void EffectChain::find_color_spaces_for_inputs()
 733 {
 734         for (unsigned i = 0; i < nodes.size(); ++i) {
 735                 Node *node = nodes[i];
 736                 if (node->disabled) {
 737                         continue;
 738                 }
 739                 if (node->incoming_links.size() == 0) {
 740                         Input *input = static_cast<Input *>(node->effect);
 741                         node->output_color_space = input->get_color_space();
 742                         node->output_gamma_curve = input->get_gamma_curve();
 743
 744                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
 745                         switch (alpha_handling) {
 746                         case Effect::OUTPUT_BLANK_ALPHA:
 747                                 node->output_alpha_type = ALPHA_BLANK;
 748                                 break;
 749                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
 750                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 751                                 break;
 752                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
 753                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 754                                 break;
 755                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
 756                         case Effect::DONT_CARE_ALPHA_TYPE:
 757                         default:
 758                                 assert(false);
 759                         }
 760
 761                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
 762                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
 763                         }
 764                 }
 765         }
 766 }
 767
 768 // Propagate gamma and color space information as far as we can in the graph.
 769 // The rules are simple: Anything where all the inputs agree, get that as
 770 // output as well. Anything else keeps having *_INVALID.
 771 void EffectChain::propagate_gamma_and_color_space()
 772 {
 773         // We depend on going through the nodes in order.
 774         sort_all_nodes_topologically();
 775
 776         for (unsigned i = 0; i < nodes.size(); ++i) {
 777                 Node *node = nodes[i];
 778                 if (node->disabled) {
 779                         continue;
 780                 }
 781                 assert(node->incoming_links.size() == node->effect->num_inputs());
 782                 if (node->incoming_links.size() == 0) {
 783                         assert(node->output_color_space != COLORSPACE_INVALID);
 784                         assert(node->output_gamma_curve != GAMMA_INVALID);
 785                         continue;
 786                 }
 787
 788                 Colorspace color_space = node->incoming_links[0]->output_color_space;
 789                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
 790                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
 791                         if (node->incoming_links[j]->output_color_space != color_space) {
 792                                 color_space = COLORSPACE_INVALID;
 793                         }
 794                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
 795                                 gamma_curve = GAMMA_INVALID;
 796                         }
 797                 }
 798
 799                 // The conversion effects already have their outputs set correctly,
 800                 // so leave them alone.
 801                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
 802                         node->output_color_space = color_space;
 803                 }
 804                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
 805                     node->effect->effect_type_id() != "GammaExpansionEffect") {
 806                         node->output_gamma_curve = gamma_curve;
 807                 }
 808         }
 809 }
 810
 811 // Propagate alpha information as far as we can in the graph.
 812 // Similar to propagate_gamma_and_color_space().
 813 void EffectChain::propagate_alpha()
 814 {
 815         // We depend on going through the nodes in order.
 816         sort_all_nodes_topologically();
 817
 818         for (unsigned i = 0; i < nodes.size(); ++i) {
 819                 Node *node = nodes[i];
 820                 if (node->disabled) {
 821                         continue;
 822                 }
 823                 assert(node->incoming_links.size() == node->effect->num_inputs());
 824                 if (node->incoming_links.size() == 0) {
 825                         assert(node->output_alpha_type != ALPHA_INVALID);
 826                         continue;
 827                 }
 828
 829                 // The alpha multiplication/division effects are special cases.
 830                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
 831                         assert(node->incoming_links.size() == 1);
 832                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
 833                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
 834                         continue;
 835                 }
 836                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
 837                         assert(node->incoming_links.size() == 1);
 838                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
 839                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 840                         continue;
 841                 }
 842
 843                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
 844                 // because they are the only one that _need_ postmultiplied alpha.
 845                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
 846                     node->effect->effect_type_id() == "GammaExpansionEffect") {
 847                         assert(node->incoming_links.size() == 1);
 848                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
 849                                 node->output_alpha_type = ALPHA_BLANK;
 850                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
 851                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 852                         } else {
 853                                 node->output_alpha_type = ALPHA_INVALID;
 854                         }
 855                         continue;
 856                 }
 857
 858                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
 859                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
 860                 // taken care of above. Rationale: Even if you could imagine
 861                 // e.g. an effect that took in an image and set alpha=1.0
 862                 // unconditionally, it wouldn't make any sense to have it as
 863                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
 864                 // got its input pre- or postmultiplied, so it wouldn't know
 865                 // whether to divide away the old alpha or not.
 866                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
 867                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 868                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
 869                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 870
 871                 // If the node has multiple inputs, check that they are all valid and
 872                 // the same.
 873                 bool any_invalid = false;
 874                 bool any_premultiplied = false;
 875                 bool any_postmultiplied = false;
 876
 877                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 878                         switch (node->incoming_links[j]->output_alpha_type) {
 879                         case ALPHA_INVALID:
 880                                 any_invalid = true;
 881                                 break;
 882                         case ALPHA_BLANK:
 883                                 // Blank is good as both pre- and postmultiplied alpha,
 884                                 // so just ignore it.
 885                                 break;
 886                         case ALPHA_PREMULTIPLIED:
 887                                 any_premultiplied = true;
 888                                 break;
 889                         case ALPHA_POSTMULTIPLIED:
 890                                 any_postmultiplied = true;
 891                                 break;
 892                         default:
 893                                 assert(false);
 894                         }
 895                 }
 896
 897                 if (any_invalid) {
 898                         node->output_alpha_type = ALPHA_INVALID;
 899                         continue;
 900                 }
 901
 902                 // Inputs must be of the same type.
 903                 if (any_premultiplied && any_postmultiplied) {
 904                         node->output_alpha_type = ALPHA_INVALID;
 905                         continue;
 906                 }
 907
 908                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
 909                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 910                         // If the effect has asked for premultiplied alpha, check that it has got it.
 911                         if (any_postmultiplied) {
 912                                 node->output_alpha_type = ALPHA_INVALID;
 913                         } else if (!any_premultiplied &&
 914                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
 915                                 // Blank input alpha, and the effect preserves blank alpha.
 916                                 node->output_alpha_type = ALPHA_BLANK;
 917                         } else {
 918                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 919                         }
 920                 } else {
 921                         // OK, all inputs are the same, and this effect is not going
 922                         // to change it.
 923                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
 924                         if (any_premultiplied) {
 925                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
 926                         } else if (any_postmultiplied) {
 927                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
 928                         } else {
 929                                 node->output_alpha_type = ALPHA_BLANK;
 930                         }
 931                 }
 932         }
 933 }
 934
 935 bool EffectChain::node_needs_colorspace_fix(Node *node)
 936 {
 937         if (node->disabled) {
 938                 return false;
 939         }
 940         if (node->effect->num_inputs() == 0) {
 941                 return false;
 942         }
 943
 944         // propagate_gamma_and_color_space() has already set our output
 945         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
 946         if (node->output_color_space == COLORSPACE_INVALID) {
 947                 return true;
 948         }
 949         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
 950 }
 951
 952 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
 953 // the graph. Our strategy is not always optimal, but quite simple:
 954 // Find an effect that's as early as possible where the inputs are of
 955 // unacceptable colorspaces (that is, either different, or, if the effect only
 956 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
 957 // propagate the information anew, and repeat until there are no more such
 958 // effects.
 959 void EffectChain::fix_internal_color_spaces()
 960 {
 961         unsigned colorspace_propagation_pass = 0;
 962         bool found_any;
 963         do {
 964                 found_any = false;
 965                 for (unsigned i = 0; i < nodes.size(); ++i) {
 966                         Node *node = nodes[i];
 967                         if (!node_needs_colorspace_fix(node)) {
 968                                 continue;
 969                         }
 970
 971                         // Go through each input that is not sRGB, and insert
 972                         // a colorspace conversion after it.
 973                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 974                                 Node *input = node->incoming_links[j];
 975                                 assert(input->output_color_space != COLORSPACE_INVALID);
 976                                 if (input->output_color_space == COLORSPACE_sRGB) {
 977                                         continue;
 978                                 }
 979                                 Node *conversion = add_node(new ColorspaceConversionEffect());
 980                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
 981                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
 982                                 conversion->output_color_space = COLORSPACE_sRGB;
 983                                 replace_sender(input, conversion);
 984                                 connect_nodes(input, conversion);
 985                         }
 986
 987                         // Re-sort topologically, and propagate the new information.
 988                         propagate_gamma_and_color_space();
 989
 990                         found_any = true;
 991                         break;
 992                 }
 993
 994                 char filename[256];
 995                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
 996                 output_dot(filename);
 997                 assert(colorspace_propagation_pass < 100);
 998         } while (found_any);
 999
1000         for (unsigned i = 0; i < nodes.size(); ++i) {
1001                 Node *node = nodes[i];
1002                 if (node->disabled) {
1003                         continue;
1004                 }
1005                 assert(node->output_color_space != COLORSPACE_INVALID);
1006         }
1007 }
1008
1009 bool EffectChain::node_needs_alpha_fix(Node *node)
1010 {
1011         if (node->disabled) {
1012                 return false;
1013         }
1014
1015         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1016         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1017         return (node->output_alpha_type == ALPHA_INVALID);
1018 }
1019
1020 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1021 // the graph. Similar to fix_internal_color_spaces().
1022 void EffectChain::fix_internal_alpha(unsigned step)
1023 {
1024         unsigned alpha_propagation_pass = 0;
1025         bool found_any;
1026         do {
1027                 found_any = false;
1028                 for (unsigned i = 0; i < nodes.size(); ++i) {
1029                         Node *node = nodes[i];
1030                         if (!node_needs_alpha_fix(node)) {
1031                                 continue;
1032                         }
1033
1034                         // If we need to fix up GammaExpansionEffect, then clearly something
1035                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1036                         // is meaningless.
1037                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1038
1039                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1040
1041                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1042                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1043                                 assert(node->incoming_links.size() == 1);
1044                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1045                                 desired_type = ALPHA_POSTMULTIPLIED;
1046                         }
1047
1048                         // Go through each input that is not premultiplied alpha, and insert
1049                         // a conversion before it.
1050                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1051                                 Node *input = node->incoming_links[j];
1052                                 assert(input->output_alpha_type != ALPHA_INVALID);
1053                                 if (input->output_alpha_type == desired_type ||
1054                                     input->output_alpha_type == ALPHA_BLANK) {
1055                                         continue;
1056                                 }
1057                                 Node *conversion;
1058                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1059                                         conversion = add_node(new AlphaMultiplicationEffect());
1060                                 } else {
1061                                         conversion = add_node(new AlphaDivisionEffect());
1062                                 }
1063                                 conversion->output_alpha_type = desired_type;
1064                                 replace_sender(input, conversion);
1065                                 connect_nodes(input, conversion);
1066                         }
1067
1068                         // Re-sort topologically, and propagate the new information.
1069                         propagate_gamma_and_color_space();
1070                         propagate_alpha();
1071
1072                         found_any = true;
1073                         break;
1074                 }
1075
1076                 char filename[256];
1077                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1078                 output_dot(filename);
1079                 assert(alpha_propagation_pass < 100);
1080         } while (found_any);
1081
1082         for (unsigned i = 0; i < nodes.size(); ++i) {
1083                 Node *node = nodes[i];
1084                 if (node->disabled) {
1085                         continue;
1086                 }
1087                 assert(node->output_alpha_type != ALPHA_INVALID);
1088         }
1089 }
1090
1091 // Make so that the output is in the desired color space.
1092 void EffectChain::fix_output_color_space()
1093 {
1094         Node *output = find_output_node();
1095         if (output->output_color_space != output_format.color_space) {
1096                 Node *conversion = add_node(new ColorspaceConversionEffect());
1097                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1098                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1099                 conversion->output_color_space = output_format.color_space;
1100                 connect_nodes(output, conversion);
1101                 propagate_alpha();
1102                 propagate_gamma_and_color_space();
1103         }
1104 }
1105
1106 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1107 void EffectChain::fix_output_alpha()
1108 {
1109         Node *output = find_output_node();
1110         assert(output->output_alpha_type != ALPHA_INVALID);
1111         if (output->output_alpha_type == ALPHA_BLANK) {
1112                 // No alpha output, so we don't care.
1113                 return;
1114         }
1115         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1116             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1117                 Node *conversion = add_node(new AlphaDivisionEffect());
1118                 connect_nodes(output, conversion);
1119                 propagate_alpha();
1120                 propagate_gamma_and_color_space();
1121         }
1122         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1123             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1124                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1125                 connect_nodes(output, conversion);
1126                 propagate_alpha();
1127                 propagate_gamma_and_color_space();
1128         }
1129 }
1130
1131 bool EffectChain::node_needs_gamma_fix(Node *node)
1132 {
1133         if (node->disabled) {
1134                 return false;
1135         }
1136
1137         // Small hack since the output is not an explicit node:
1138         // If we are the last node and our output is in the wrong
1139         // space compared to EffectChain's output, we need to fix it.
1140         // This will only take us to linear, but fix_output_gamma()
1141         // will come and take us to the desired output gamma
1142         // if it is needed.
1143         //
1144         // This needs to be before everything else, since it could
1145         // even apply to inputs (if they are the only effect).
1146         if (node->outgoing_links.empty() &&
1147             node->output_gamma_curve != output_format.gamma_curve &&
1148             node->output_gamma_curve != GAMMA_LINEAR) {
1149                 return true;
1150         }
1151
1152         if (node->effect->num_inputs() == 0) {
1153                 return false;
1154         }
1155
1156         // propagate_gamma_and_color_space() has already set our output
1157         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1158         // except for GammaCompressionEffect.
1159         if (node->output_gamma_curve == GAMMA_INVALID) {
1160                 return true;
1161         }
1162         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1163                 assert(node->incoming_links.size() == 1);
1164                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1165         }
1166
1167         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1168 }
1169
1170 // Very similar to fix_internal_color_spaces(), but for gamma.
1171 // There is one difference, though; before we start adding conversion nodes,
1172 // we see if we can get anything out of asking the sources to deliver
1173 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1174 // does that part, while fix_internal_gamma_by_inserting_nodes()
1175 // inserts nodes as needed afterwards.
1176 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1177 {
1178         unsigned gamma_propagation_pass = 0;
1179         bool found_any;
1180         do {
1181                 found_any = false;
1182                 for (unsigned i = 0; i < nodes.size(); ++i) {
1183                         Node *node = nodes[i];
1184                         if (!node_needs_gamma_fix(node)) {
1185                                 continue;
1186                         }
1187
1188                         // See if all inputs can give us linear gamma. If not, leave it.
1189                         vector<Node *> nonlinear_inputs;
1190                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1191                         assert(!nonlinear_inputs.empty());
1192
1193                         bool all_ok = true;
1194                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1195                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1196                                 all_ok &= input->can_output_linear_gamma();
1197                         }
1198
1199                         if (!all_ok) {
1200                                 continue;
1201                         }
1202
1203                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1204                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1205                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1206                         }
1207
1208                         // Re-sort topologically, and propagate the new information.
1209                         propagate_gamma_and_color_space();
1210
1211                         found_any = true;
1212                         break;
1213                 }
1214
1215                 char filename[256];
1216                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1217                 output_dot(filename);
1218                 assert(gamma_propagation_pass < 100);
1219         } while (found_any);
1220 }
1221
1222 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1223 {
1224         unsigned gamma_propagation_pass = 0;
1225         bool found_any;
1226         do {
1227                 found_any = false;
1228                 for (unsigned i = 0; i < nodes.size(); ++i) {
1229                         Node *node = nodes[i];
1230                         if (!node_needs_gamma_fix(node)) {
1231                                 continue;
1232                         }
1233
1234                         // Special case: We could be an input and still be asked to
1235                         // fix our gamma; if so, we should be the only node
1236                         // (as node_needs_gamma_fix() would only return true in
1237                         // for an input in that case). That means we should insert
1238                         // a conversion node _after_ ourselves.
1239                         if (node->incoming_links.empty()) {
1240                                 assert(node->outgoing_links.empty());
1241                                 Node *conversion = add_node(new GammaExpansionEffect());
1242                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1243                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1244                                 connect_nodes(node, conversion);
1245                         }
1246
1247                         // If not, go through each input that is not linear gamma,
1248                         // and insert a gamma conversion after it.
1249                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1250                                 Node *input = node->incoming_links[j];
1251                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1252                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1253                                         continue;
1254                                 }
1255                                 Node *conversion = add_node(new GammaExpansionEffect());
1256                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1257                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1258                                 replace_sender(input, conversion);
1259                                 connect_nodes(input, conversion);
1260                         }
1261
1262                         // Re-sort topologically, and propagate the new information.
1263                         propagate_alpha();
1264                         propagate_gamma_and_color_space();
1265
1266                         found_any = true;
1267                         break;
1268                 }
1269
1270                 char filename[256];
1271                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1272                 output_dot(filename);
1273                 assert(gamma_propagation_pass < 100);
1274         } while (found_any);
1275
1276         for (unsigned i = 0; i < nodes.size(); ++i) {
1277                 Node *node = nodes[i];
1278                 if (node->disabled) {
1279                         continue;
1280                 }
1281                 assert(node->output_gamma_curve != GAMMA_INVALID);
1282         }
1283 }
1284
1285 // Make so that the output is in the desired gamma.
1286 // Note that this assumes linear input gamma, so it might create the need
1287 // for another pass of fix_internal_gamma().
1288 void EffectChain::fix_output_gamma()
1289 {
1290         Node *output = find_output_node();
1291         if (output->output_gamma_curve != output_format.gamma_curve) {
1292                 Node *conversion = add_node(new GammaCompressionEffect());
1293                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1294                 conversion->output_gamma_curve = output_format.gamma_curve;
1295                 connect_nodes(output, conversion);
1296         }
1297 }
1298
1299 // If the user has requested dither, add a DitherEffect right at the end
1300 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1301 // since dither is about the only effect that can _not_ be done in linear space.
1302 void EffectChain::add_dither_if_needed()
1303 {
1304         if (num_dither_bits == 0) {
1305                 return;
1306         }
1307         Node *output = find_output_node();
1308         Node *dither = add_node(new DitherEffect());
1309         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1310         connect_nodes(output, dither);
1311
1312         dither_effect = dither->effect;
1313 }
1314
1315 // Find the output node. This is, simply, one that has no outgoing links.
1316 // If there are multiple ones, the graph is malformed (we do not support
1317 // multiple outputs right now).
1318 Node *EffectChain::find_output_node()
1319 {
1320         vector<Node *> output_nodes;
1321         for (unsigned i = 0; i < nodes.size(); ++i) {
1322                 Node *node = nodes[i];
1323                 if (node->disabled) {
1324                         continue;
1325                 }
1326                 if (node->outgoing_links.empty()) {
1327                         output_nodes.push_back(node);
1328                 }
1329         }
1330         assert(output_nodes.size() == 1);
1331         return output_nodes[0];
1332 }
1333
1334 void EffectChain::finalize()
1335 {
1336         // Save the current locale, and set it to C, so that we can output decimal
1337         // numbers with printf and be sure to get them in the format mandated by GLSL.
1338 #if defined(__MINGW32__)
1339         // Note that the OpenGL driver might call setlocale() behind-the-scenes,
1340         // and that might corrupt the returned pointer, so we need to take our own
1341         // copy of it here.
1342         char *saved_locale = strdup(setlocale(LC_NUMERIC, NULL));
1343         setlocale(LC_NUMERIC, "C");
1344 #else
1345         locale_t c_locale = newlocale(LC_NUMERIC_MASK, "C", (locale_t)0);
1346         locale_t saved_locale = uselocale(c_locale);
1347 #endif
1348
1349         // Output the graph as it is before we do any conversions on it.
1350         output_dot("step0-start.dot");
1351
1352         // Give each effect in turn a chance to rewrite its own part of the graph.
1353         // Note that if more effects are added as part of this, they will be
1354         // picked up as part of the same for loop, since they are added at the end.
1355         for (unsigned i = 0; i < nodes.size(); ++i) {
1356                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1357         }
1358         output_dot("step1-rewritten.dot");
1359
1360         find_color_spaces_for_inputs();
1361         output_dot("step2-input-colorspace.dot");
1362
1363         propagate_alpha();
1364         output_dot("step3-propagated-alpha.dot");
1365
1366         propagate_gamma_and_color_space();
1367         output_dot("step4-propagated-all.dot");
1368
1369         fix_internal_color_spaces();
1370         fix_internal_alpha(6);
1371         fix_output_color_space();
1372         output_dot("step7-output-colorspacefix.dot");
1373         fix_output_alpha();
1374         output_dot("step8-output-alphafix.dot");
1375
1376         // Note that we need to fix gamma after colorspace conversion,
1377         // because colorspace conversions might create needs for gamma conversions.
1378         // Also, we need to run an extra pass of fix_internal_gamma() after
1379         // fixing the output gamma, as we only have conversions to/from linear,
1380         // and fix_internal_alpha() since GammaCompressionEffect needs
1381         // postmultiplied input.
1382         fix_internal_gamma_by_asking_inputs(9);
1383         fix_internal_gamma_by_inserting_nodes(10);
1384         fix_output_gamma();
1385         output_dot("step11-output-gammafix.dot");
1386         propagate_alpha();
1387         output_dot("step12-output-alpha-propagated.dot");
1388         fix_internal_alpha(13);
1389         output_dot("step14-output-alpha-fixed.dot");
1390         fix_internal_gamma_by_asking_inputs(15);
1391         fix_internal_gamma_by_inserting_nodes(16);
1392
1393         output_dot("step17-before-dither.dot");
1394
1395         add_dither_if_needed();
1396
1397         output_dot("step18-final.dot");
1398
1399         // Construct all needed GLSL programs, starting at the output.
1400         // We need to keep track of which effects have already been computed,
1401         // as an effect with multiple users could otherwise be calculated
1402         // multiple times.
1403         map<Node *, Phase *> completed_effects;
1404         construct_phase(find_output_node(), &completed_effects);
1405
1406         output_dot("step19-split-to-phases.dot");
1407
1408         assert(phases[0]->inputs.empty());
1409
1410         finalized = true;
1411 #if defined(__MINGW32__)
1412         setlocale(LC_NUMERIC, saved_locale);
1413         free(saved_locale);
1414 #else
1415         uselocale(saved_locale);
1416         freelocale(c_locale);
1417 #endif
1418 }
1419
1420 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1421 {
1422         assert(finalized);
1423
1424         // Save original viewport.
1425         GLuint x = 0, y = 0;
1426
1427         if (width == 0 && height == 0) {
1428                 GLint viewport[4];
1429                 glGetIntegerv(GL_VIEWPORT, viewport);
1430                 x = viewport[0];
1431                 y = viewport[1];
1432                 width = viewport[2];
1433                 height = viewport[3];
1434         }
1435
1436         // Basic state.
1437         glDisable(GL_BLEND);
1438         check_error();
1439         glDisable(GL_DEPTH_TEST);
1440         check_error();
1441         glDepthMask(GL_FALSE);
1442         check_error();
1443
1444         set<Phase *> generated_mipmaps;
1445
1446         // We choose the simplest option of having one texture per output,
1447         // since otherwise this turns into an (albeit simple) register allocation problem.
1448         map<Phase *, GLuint> output_textures;
1449
1450         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
1451                 Phase *phase = phases[phase_num];
1452
1453                 if (phase_num == phases.size() - 1) {
1454                         // Last phase goes to the output the user specified.
1455                         glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
1456                         check_error();
1457                         GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
1458                         assert(status == GL_FRAMEBUFFER_COMPLETE);
1459                         glViewport(x, y, width, height);
1460                         if (dither_effect != NULL) {
1461                                 CHECK(dither_effect->set_int("output_width", width));
1462                                 CHECK(dither_effect->set_int("output_height", height));
1463                         }
1464                 }
1465                 execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
1466         }
1467
1468         for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
1469              texture_it != output_textures.end();
1470              ++texture_it) {
1471                 resource_pool->release_2d_texture(texture_it->second);
1472         }
1473
1474         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1475         check_error();
1476         glUseProgram(0);
1477         check_error();
1478 }
1479
1480 void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLuint> *output_textures, set<Phase *> *generated_mipmaps)
1481 {
1482         GLuint fbo = 0;
1483
1484         // Find a texture for this phase.
1485         inform_input_sizes(phase);
1486         if (!last_phase) {
1487                 find_output_size(phase);
1488
1489                 GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
1490                 output_textures->insert(make_pair(phase, tex_num));
1491         }
1492
1493         const GLuint glsl_program_num = phase->glsl_program_num;
1494         check_error();
1495         glUseProgram(glsl_program_num);
1496         check_error();
1497
1498         // Set up RTT inputs for this phase.
1499         for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
1500                 glActiveTexture(GL_TEXTURE0 + sampler);
1501                 Phase *input = phase->inputs[sampler];
1502                 input->output_node->bound_sampler_num = sampler;
1503                 glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
1504                 check_error();
1505                 if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
1506                         glGenerateMipmap(GL_TEXTURE_2D);
1507                         check_error();
1508                         generated_mipmaps->insert(input);
1509                 }
1510                 setup_rtt_sampler(glsl_program_num, sampler, phase->effect_ids[input->output_node], phase->input_needs_mipmaps);
1511         }
1512
1513         // And now the output. (Already set up for us if it is the last phase.)
1514         if (!last_phase) {
1515                 fbo = resource_pool->create_fbo((*output_textures)[phase]);
1516                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1517                 glViewport(0, 0, phase->output_width, phase->output_height);
1518         }
1519
1520         // Give the required parameters to all the effects.
1521         unsigned sampler_num = phase->inputs.size();
1522         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1523                 Node *node = phase->effects[i];
1524                 unsigned old_sampler_num = sampler_num;
1525                 node->effect->set_gl_state(glsl_program_num, phase->effect_ids[node], &sampler_num);
1526                 check_error();
1527
1528                 if (node->effect->is_single_texture()) {
1529                         assert(sampler_num - old_sampler_num == 1);
1530                         node->bound_sampler_num = old_sampler_num;
1531                 } else {
1532                         node->bound_sampler_num = -1;
1533                 }
1534         }
1535
1536         // Now draw!
1537         float vertices[] = {
1538                 0.0f, 1.0f,
1539                 0.0f, 0.0f,
1540                 1.0f, 1.0f,
1541                 1.0f, 0.0f
1542         };
1543
1544         GLuint vao;
1545         glGenVertexArrays(1, &vao);
1546         check_error();
1547         glBindVertexArray(vao);
1548         check_error();
1549
1550         GLuint position_vbo = fill_vertex_attribute(glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
1551         GLuint texcoord_vbo = fill_vertex_attribute(glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.
1552
1553         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1554         check_error();
1555
1556         cleanup_vertex_attribute(glsl_program_num, "position", position_vbo);
1557         cleanup_vertex_attribute(glsl_program_num, "texcoord", texcoord_vbo);
1558
1559         glUseProgram(0);
1560         check_error();
1561
1562         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1563                 Node *node = phase->effects[i];
1564                 node->effect->clear_gl_state();
1565         }
1566
1567         if (!last_phase) {
1568                 resource_pool->release_fbo(fbo);
1569         }
1570
1571         glDeleteVertexArrays(1, &vao);
1572         check_error();
1573 }
1574
1575 void EffectChain::setup_rtt_sampler(GLuint glsl_program_num, int sampler_num, const string &effect_id, bool use_mipmaps)
1576 {
1577         glActiveTexture(GL_TEXTURE0 + sampler_num);
1578         check_error();
1579         if (use_mipmaps) {
1580                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
1581                 check_error();
1582         } else {
1583                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1584                 check_error();
1585         }
1586         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1587         check_error();
1588         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1589         check_error();
1590
1591         string texture_name = string("tex_") + effect_id;
1592         glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler_num);
1593         check_error();
1594 }
1595
1596 }  // namespace movit