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

   1 #include <epoxy/gl.h>
   2 #include <assert.h>
   3 #include <math.h>
   4 #include <stddef.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <algorithm>
   9 #include <set>
  10 #include <stack>
  11 #include <utility>
  12 #include <vector>
  13 #include <Eigen/Core>
  14
  15 #include "alpha_division_effect.h"
  16 #include "alpha_multiplication_effect.h"
  17 #include "colorspace_conversion_effect.h"
  18 #include "dither_effect.h"
  19 #include "effect.h"
  20 #include "effect_chain.h"
  21 #include "effect_util.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 #include "ycbcr_conversion_effect.h"
  29
  30 using namespace Eigen;
  31 using namespace std;
  32
  33 namespace movit {
  34
  35 namespace {
  36
  37 // An effect whose only purpose is to sit in a phase on its own and take the
  38 // texture output from a compute shader and display it to the normal backbuffer
  39 // (or any FBO). That phase can be skipped when rendering using render_to_textures().
  40 class ComputeShaderOutputDisplayEffect : public Effect {
  41 public:
  42         ComputeShaderOutputDisplayEffect() {}
  43         string effect_type_id() const override { return "ComputeShaderOutputDisplayEffect"; }
  44         string output_fragment_shader() override { return read_file("identity.frag"); }
  45         bool needs_texture_bounce() const override { return true; }
  46 };
  47
  48 }  // namespace
  49
  50 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
  51         : aspect_nom(aspect_nom),
  52           aspect_denom(aspect_denom),
  53           output_color_rgba(false),
  54           num_output_color_ycbcr(0),
  55           dither_effect(nullptr),
  56           ycbcr_conversion_effect_node(nullptr),
  57           intermediate_format(GL_RGBA16F),
  58           intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION),
  59           num_dither_bits(0),
  60           output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
  61           finalized(false),
  62           resource_pool(resource_pool),
  63           do_phase_timing(false) {
  64         if (resource_pool == nullptr) {
  65                 this->resource_pool = new ResourcePool();
  66                 owns_resource_pool = true;
  67         } else {
  68                 owns_resource_pool = false;
  69         }
  70
  71         // Generate a VBO with some data in (shared position and texture coordinate data).
  72         float vertices[] = {
  73                 0.0f, 2.0f,
  74                 0.0f, 0.0f,
  75                 2.0f, 0.0f
  76         };
  77         vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
  78 }
  79
  80 EffectChain::~EffectChain()
  81 {
  82         for (unsigned i = 0; i < nodes.size(); ++i) {
  83                 delete nodes[i]->effect;
  84                 delete nodes[i];
  85         }
  86         for (unsigned i = 0; i < phases.size(); ++i) {
  87                 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
  88                 delete phases[i];
  89         }
  90         if (owns_resource_pool) {
  91                 delete resource_pool;
  92         }
  93         glDeleteBuffers(1, &vbo);
  94         check_error();
  95 }
  96
  97 Input *EffectChain::add_input(Input *input)
  98 {
  99         assert(!finalized);
 100         inputs.push_back(input);
 101         add_node(input);
 102         return input;
 103 }
 104
 105 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
 106 {
 107         assert(!finalized);
 108         assert(!output_color_rgba);
 109         output_format = format;
 110         output_alpha_format = alpha_format;
 111         output_color_rgba = true;
 112 }
 113
 114 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
 115                                    const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting,
 116                                    GLenum output_type)
 117 {
 118         assert(!finalized);
 119         assert(num_output_color_ycbcr < 2);
 120         output_format = format;
 121         output_alpha_format = alpha_format;
 122
 123         if (num_output_color_ycbcr == 1) {
 124                 // Check that the format is the same.
 125                 assert(output_ycbcr_format.luma_coefficients == ycbcr_format.luma_coefficients);
 126                 assert(output_ycbcr_format.full_range == ycbcr_format.full_range);
 127                 assert(output_ycbcr_format.num_levels == ycbcr_format.num_levels);
 128                 assert(output_ycbcr_format.chroma_subsampling_x == 1);
 129                 assert(output_ycbcr_format.chroma_subsampling_y == 1);
 130                 assert(output_ycbcr_type == output_type);
 131         } else {
 132                 output_ycbcr_format = ycbcr_format;
 133                 output_ycbcr_type = output_type;
 134         }
 135         output_ycbcr_splitting[num_output_color_ycbcr++] = output_splitting;
 136
 137         assert(ycbcr_format.chroma_subsampling_x == 1);
 138         assert(ycbcr_format.chroma_subsampling_y == 1);
 139 }
 140
 141 void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format)
 142 {
 143         assert(num_output_color_ycbcr > 0);
 144         assert(output_ycbcr_format.chroma_subsampling_x == 1);
 145         assert(output_ycbcr_format.chroma_subsampling_y == 1);
 146
 147         output_ycbcr_format = ycbcr_format;
 148         if (finalized) {
 149                 YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(ycbcr_conversion_effect_node->effect);
 150                 effect->change_output_format(ycbcr_format);
 151         }
 152 }
 153
 154 Node *EffectChain::add_node(Effect *effect)
 155 {
 156         for (unsigned i = 0; i < nodes.size(); ++i) {
 157                 assert(nodes[i]->effect != effect);
 158         }
 159
 160         Node *node = new Node;
 161         node->effect = effect;
 162         node->disabled = false;
 163         node->output_color_space = COLORSPACE_INVALID;
 164         node->output_gamma_curve = GAMMA_INVALID;
 165         node->output_alpha_type = ALPHA_INVALID;
 166         node->needs_mipmaps = Effect::DOES_NOT_NEED_MIPMAPS;
 167         node->one_to_one_sampling = false;
 168         node->strong_one_to_one_sampling = false;
 169
 170         nodes.push_back(node);
 171         node_map[effect] = node;
 172         effect->inform_added(this);
 173         return node;
 174 }
 175
 176 void EffectChain::connect_nodes(Node *sender, Node *receiver)
 177 {
 178         sender->outgoing_links.push_back(receiver);
 179         receiver->incoming_links.push_back(sender);
 180 }
 181
 182 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
 183 {
 184         new_receiver->incoming_links = old_receiver->incoming_links;
 185         old_receiver->incoming_links.clear();
 186
 187         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
 188                 Node *sender = new_receiver->incoming_links[i];
 189                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
 190                         if (sender->outgoing_links[j] == old_receiver) {
 191                                 sender->outgoing_links[j] = new_receiver;
 192                         }
 193                 }
 194         }
 195 }
 196
 197 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 198 {
 199         new_sender->outgoing_links = old_sender->outgoing_links;
 200         old_sender->outgoing_links.clear();
 201
 202         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
 203                 Node *receiver = new_sender->outgoing_links[i];
 204                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
 205                         if (receiver->incoming_links[j] == old_sender) {
 206                                 receiver->incoming_links[j] = new_sender;
 207                         }
 208                 }
 209         }
 210 }
 211
 212 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
 213 {
 214         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
 215                 if (sender->outgoing_links[i] == receiver) {
 216                         sender->outgoing_links[i] = middle;
 217                         middle->incoming_links.push_back(sender);
 218                 }
 219         }
 220         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
 221                 if (receiver->incoming_links[i] == sender) {
 222                         receiver->incoming_links[i] = middle;
 223                         middle->outgoing_links.push_back(receiver);
 224                 }
 225         }
 226
 227         assert(middle->incoming_links.size() == middle->effect->num_inputs());
 228 }
 229
 230 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
 231 {
 232         assert(node->effect->needs_texture_bounce());
 233         assert(input_num < node->incoming_links.size());
 234         assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
 235         assert(node->incoming_links[input_num]->bound_sampler_num < 8);
 236         return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
 237 }
 238
 239 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
 240 {
 241         assert(input_num < node->incoming_links.size());
 242         return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
 243                 node->incoming_links[input_num]->bound_sampler_num < 8;
 244 }
 245
 246 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
 247 {
 248         if (node->output_gamma_curve == GAMMA_LINEAR &&
 249             node->effect->effect_type_id() != "GammaCompressionEffect") {
 250                 return;
 251         }
 252         if (node->effect->num_inputs() == 0) {
 253                 nonlinear_inputs->push_back(node);
 254         } else {
 255                 assert(node->effect->num_inputs() == node->incoming_links.size());
 256                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
 257                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
 258                 }
 259         }
 260 }
 261
 262 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
 263 {
 264         assert(!finalized);
 265         assert(inputs.size() == effect->num_inputs());
 266         Node *node = add_node(effect);
 267         for (unsigned i = 0; i < inputs.size(); ++i) {
 268                 assert(node_map.count(inputs[i]) != 0);
 269                 connect_nodes(node_map[inputs[i]], node);
 270         }
 271         return effect;
 272 }
 273
 274 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 275 string replace_prefix(const string &text, const string &prefix)
 276 {
 277         string output;
 278         size_t start = 0;
 279
 280         while (start < text.size()) {
 281                 size_t pos = text.find("PREFIX(", start);
 282                 if (pos == string::npos) {
 283                         output.append(text.substr(start, string::npos));
 284                         break;
 285                 }
 286
 287                 output.append(text.substr(start, pos - start));
 288                 output.append(prefix);
 289                 output.append("_");
 290
 291                 pos += strlen("PREFIX(");
 292
 293                 // Output stuff until we find the matching ), which we then eat.
 294                 int depth = 1;
 295                 size_t end_arg_pos = pos;
 296                 while (end_arg_pos < text.size()) {
 297                         if (text[end_arg_pos] == '(') {
 298                                 ++depth;
 299                         } else if (text[end_arg_pos] == ')') {
 300                                 --depth;
 301                                 if (depth == 0) {
 302                                         break;
 303                                 }
 304                         }
 305                         ++end_arg_pos;
 306                 }
 307                 output.append(text.substr(pos, end_arg_pos - pos));
 308                 ++end_arg_pos;
 309                 assert(depth == 0);
 310                 start = end_arg_pos;
 311         }
 312         return output;
 313 }
 314
 315 namespace {
 316
 317 template<class T>
 318 void extract_uniform_declarations(const vector<Uniform<T>> &effect_uniforms,
 319                                   const string &type_specifier,
 320                                   const string &effect_id,
 321                                   vector<Uniform<T>> *phase_uniforms,
 322                                   string *glsl_string)
 323 {
 324         for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
 325                 phase_uniforms->push_back(effect_uniforms[i]);
 326                 phase_uniforms->back().prefix = effect_id;
 327
 328                 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
 329                         + "_" + effect_uniforms[i].name + ";\n";
 330         }
 331 }
 332
 333 template<class T>
 334 void extract_uniform_array_declarations(const vector<Uniform<T>> &effect_uniforms,
 335                                         const string &type_specifier,
 336                                         const string &effect_id,
 337                                         vector<Uniform<T>> *phase_uniforms,
 338                                         string *glsl_string)
 339 {
 340         for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
 341                 phase_uniforms->push_back(effect_uniforms[i]);
 342                 phase_uniforms->back().prefix = effect_id;
 343
 344                 char buf[256];
 345                 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
 346                         type_specifier.c_str(), effect_id.c_str(),
 347                         effect_uniforms[i].name.c_str(),
 348                         int(effect_uniforms[i].num_values));
 349                 *glsl_string += buf;
 350         }
 351 }
 352
 353 template<class T>
 354 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T>> *phase_uniforms)
 355 {
 356         for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
 357                 Uniform<T> &uniform = (*phase_uniforms)[i];
 358                 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
 359         }
 360 }
 361
 362 }  // namespace
 363
 364 void EffectChain::compile_glsl_program(Phase *phase)
 365 {
 366         string frag_shader_header;
 367         if (phase->is_compute_shader) {
 368                 frag_shader_header = read_file("header.comp");
 369         } else {
 370                 frag_shader_header = read_version_dependent_file("header", "frag");
 371         }
 372         string frag_shader = "";
 373
 374         // Create functions and uniforms for all the texture inputs that we need.
 375         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
 376                 Node *input = phase->inputs[i]->output_node;
 377                 char effect_id[256];
 378                 sprintf(effect_id, "in%u", i);
 379                 phase->effect_ids.insert(make_pair(make_pair(input, IN_ANOTHER_PHASE), effect_id));
 380
 381                 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
 382                 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
 383                 frag_shader += "\tvec4 tmp = tex2D(tex_" + string(effect_id) + ", tc);\n";
 384
 385                 if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
 386                     phase->inputs[i]->output_node->output_gamma_curve == GAMMA_LINEAR) {
 387                         frag_shader += "\ttmp.rgb *= tmp.rgb;\n";
 388                 }
 389
 390                 frag_shader += "\treturn tmp;\n";
 391                 frag_shader += "}\n";
 392                 frag_shader += "\n";
 393
 394                 Uniform<int> uniform;
 395                 uniform.name = effect_id;
 396                 uniform.value = &phase->input_samplers[i];
 397                 uniform.prefix = "tex";
 398                 uniform.num_values = 1;
 399                 uniform.location = -1;
 400                 phase->uniforms_sampler2d.push_back(uniform);
 401         }
 402
 403         // Give each effect in the phase its own ID.
 404         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 405                 Node *node = phase->effects[i];
 406                 char effect_id[256];
 407                 sprintf(effect_id, "eff%u", i);
 408                 bool inserted = phase->effect_ids.insert(make_pair(make_pair(node, IN_SAME_PHASE), effect_id)).second;
 409                 assert(inserted);
 410         }
 411
 412         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 413                 Node *node = phase->effects[i];
 414                 const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
 415                 if (node->incoming_links.size() == 1) {
 416                         Node *input = node->incoming_links[0];
 417                         NodeLinkType link_type = node->incoming_link_type[0];
 418                         if (i != 0 && input->effect->is_compute_shader()) {
 419                                 // First effect after the compute shader reads the value
 420                                 // that cs_output() wrote to a global variable.
 421                                 frag_shader += string("#define INPUT(tc) CS_OUTPUT_VAL\n");
 422                         } else {
 423                                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(input, link_type)] + "\n";
 424                         }
 425                 } else {
 426                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 427                                 assert(!node->incoming_links[j]->effect->is_compute_shader());
 428                                 char buf[256];
 429                                 string effect_id = phase->effect_ids[make_pair(node->incoming_links[j], node->incoming_link_type[j])];
 430                                 sprintf(buf, "#define INPUT%d %s\n", j + 1, effect_id.c_str());
 431                                 frag_shader += buf;
 432                         }
 433                 }
 434
 435                 frag_shader += "\n";
 436                 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
 437                 if (node->effect->is_compute_shader()) {
 438                         frag_shader += string("#define NORMALIZE_TEXTURE_COORDS(tc) ((tc) * ") + effect_id + "_inv_output_size + " + effect_id + "_output_texcoord_adjust)\n";
 439                 }
 440                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
 441                 frag_shader += "#undef FUNCNAME\n";
 442                 if (node->incoming_links.size() == 1) {
 443                         frag_shader += "#undef INPUT\n";
 444                 } else {
 445                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 446                                 char buf[256];
 447                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 448                                 frag_shader += buf;
 449                         }
 450                 }
 451                 frag_shader += "\n";
 452         }
 453         if (phase->is_compute_shader) {
 454                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->compute_shader_node, IN_SAME_PHASE)] + "\n";
 455                 if (phase->compute_shader_node == phase->effects.back()) {
 456                         // No postprocessing.
 457                         frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n";
 458                 } else {
 459                         frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
 460                 }
 461         } else {
 462                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
 463         }
 464
 465         // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
 466         vector<string> frag_shader_outputs;  // In order.
 467         if (phase->output_node->outgoing_links.empty() && num_output_color_ycbcr > 0) {
 468                 switch (output_ycbcr_splitting[0]) {
 469                 case YCBCR_OUTPUT_INTERLEAVED:
 470                         // No #defines set.
 471                         frag_shader_outputs.push_back("FragColor");
 472                         break;
 473                 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
 474                         frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
 475                         frag_shader_outputs.push_back("Y");
 476                         frag_shader_outputs.push_back("Chroma");
 477                         break;
 478                 case YCBCR_OUTPUT_PLANAR:
 479                         frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
 480                         frag_shader_outputs.push_back("Y");
 481                         frag_shader_outputs.push_back("Cb");
 482                         frag_shader_outputs.push_back("Cr");
 483                         break;
 484                 default:
 485                         assert(false);
 486                 }
 487
 488                 if (num_output_color_ycbcr > 1) {
 489                         switch (output_ycbcr_splitting[1]) {
 490                         case YCBCR_OUTPUT_INTERLEAVED:
 491                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_INTERLEAVED 1\n";
 492                                 frag_shader_outputs.push_back("YCbCr2");
 493                                 break;
 494                         case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
 495                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
 496                                 frag_shader_outputs.push_back("Y2");
 497                                 frag_shader_outputs.push_back("Chroma2");
 498                                 break;
 499                         case YCBCR_OUTPUT_PLANAR:
 500                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_PLANAR 1\n";
 501                                 frag_shader_outputs.push_back("Y2");
 502                                 frag_shader_outputs.push_back("Cb2");
 503                                 frag_shader_outputs.push_back("Cr2");
 504                                 break;
 505                         default:
 506                                 assert(false);
 507                         }
 508                 }
 509
 510                 if (output_color_rgba) {
 511                         // Note: Needs to come in the header, because not only the
 512                         // output needs to see it (YCbCrConversionEffect and DitherEffect
 513                         // do, too).
 514                         frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
 515                         frag_shader_outputs.push_back("RGBA");
 516                 }
 517         }
 518
 519         // If we're bouncing to a temporary texture, signal transformation if desired.
 520         if (!phase->output_node->outgoing_links.empty()) {
 521                 if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
 522                     phase->output_node->output_gamma_curve == GAMMA_LINEAR) {
 523                         frag_shader += "#define SQUARE_ROOT_TRANSFORMATION 1\n";
 524                 }
 525         }
 526
 527         if (phase->is_compute_shader) {
 528                 frag_shader.append(read_file("footer.comp"));
 529                 phase->compute_shader_node->effect->register_uniform_ivec2("output_size", phase->uniform_output_size);
 530                 phase->compute_shader_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size);
 531                 phase->compute_shader_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust);
 532         } else {
 533                 frag_shader.append(read_file("footer.frag"));
 534         }
 535
 536         // Collect uniforms from all effects and output them. Note that this needs
 537         // to happen after output_fragment_shader(), even though the uniforms come
 538         // before in the output source, since output_fragment_shader() is allowed
 539         // to register new uniforms (e.g. arrays that are of unknown length until
 540         // finalization time).
 541         // TODO: Make a uniform block for platforms that support it.
 542         string frag_shader_uniforms = "";
 543         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 544                 Node *node = phase->effects[i];
 545                 Effect *effect = node->effect;
 546                 const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
 547                 extract_uniform_declarations(effect->uniforms_image2d, "image2D", effect_id, &phase->uniforms_image2d, &frag_shader_uniforms);
 548                 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
 549                 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
 550                 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
 551                 extract_uniform_declarations(effect->uniforms_ivec2, "ivec2", effect_id, &phase->uniforms_ivec2, &frag_shader_uniforms);
 552                 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
 553                 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
 554                 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
 555                 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
 556                 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
 557                 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
 558                 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
 559                 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
 560                 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
 561         }
 562
 563         string vert_shader = read_version_dependent_file("vs", "vert");
 564
 565         // If we're the last phase and need to flip the picture to compensate for
 566         // the origin, tell the vertex or compute shader so.
 567         bool is_last_phase;
 568         if (has_dummy_effect) {
 569                 is_last_phase = (phase->output_node->outgoing_links.size() == 1 &&
 570                         phase->output_node->outgoing_links[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
 571         } else {
 572                 is_last_phase = phase->output_node->outgoing_links.empty();
 573         }
 574         if (is_last_phase && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
 575                 if (phase->is_compute_shader) {
 576                         frag_shader_header += "#define FLIP_ORIGIN 1\n";
 577                 } else {
 578                         const string needle = "#define FLIP_ORIGIN 0";
 579                         size_t pos = vert_shader.find(needle);
 580                         assert(pos != string::npos);
 581
 582                         vert_shader[pos + needle.size() - 1] = '1';
 583                 }
 584         }
 585
 586         frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
 587
 588         if (phase->is_compute_shader) {
 589                 phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader);
 590
 591                 Uniform<int> uniform;
 592                 uniform.name = "outbuf";
 593                 uniform.value = &phase->outbuf_image_unit;
 594                 uniform.prefix = "tex";
 595                 uniform.num_values = 1;
 596                 uniform.location = -1;
 597                 phase->uniforms_image2d.push_back(uniform);
 598         } else {
 599                 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
 600         }
 601         GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
 602         GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
 603         if (position_attribute_index != -1) {
 604                 phase->attribute_indexes.insert(position_attribute_index);
 605         }
 606         if (texcoord_attribute_index != -1) {
 607                 phase->attribute_indexes.insert(texcoord_attribute_index);
 608         }
 609
 610         // Collect the resulting location numbers for each uniform.
 611         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_image2d);
 612         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
 613         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
 614         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
 615         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_ivec2);
 616         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
 617         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
 618         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
 619         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
 620         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
 621 }
 622
 623 // Construct GLSL programs, starting at the given effect and following
 624 // the chain from there. We end a program every time we come to an effect
 625 // marked as "needs texture bounce", one that is used by multiple other
 626 // effects, every time we need to bounce due to output size change
 627 // (not all size changes require ending), and of course at the end.
 628 //
 629 // We follow a quite simple depth-first search from the output, although
 630 // without recursing explicitly within each phase.
 631 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
 632 {
 633         if (completed_effects->count(output)) {
 634                 return (*completed_effects)[output];
 635         }
 636
 637         Phase *phase = new Phase;
 638         phase->output_node = output;
 639         phase->is_compute_shader = false;
 640         phase->compute_shader_node = nullptr;
 641
 642         // If the output effect has one-to-one sampling, we try to trace this
 643         // status down through the dependency chain. This is important in case
 644         // we hit an effect that changes output size (and not sets a virtual
 645         // output size); if we have one-to-one sampling, we don't have to break
 646         // the phase.
 647         output->one_to_one_sampling = output->effect->one_to_one_sampling();
 648         output->strong_one_to_one_sampling = output->effect->strong_one_to_one_sampling();
 649
 650         // Effects that we have yet to calculate, but that we know should
 651         // be in the current phase.
 652         stack<Node *> effects_todo_this_phase;
 653         effects_todo_this_phase.push(output);
 654
 655         while (!effects_todo_this_phase.empty()) {
 656                 Node *node = effects_todo_this_phase.top();
 657                 effects_todo_this_phase.pop();
 658
 659                 assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling());
 660
 661                 if (node->effect->needs_mipmaps() != Effect::DOES_NOT_NEED_MIPMAPS) {
 662                         // Can't have incompatible requirements imposed on us from a dependent effect;
 663                         // if so, it should have started a new phase instead.
 664                         assert(node->needs_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS ||
 665                                node->needs_mipmaps == node->effect->needs_mipmaps());
 666                         node->needs_mipmaps = node->effect->needs_mipmaps();
 667                 }
 668
 669                 // This should currently only happen for effects that are inputs
 670                 // (either true inputs or phase outputs). We special-case inputs,
 671                 // and then deduplicate phase outputs below.
 672                 if (node->effect->num_inputs() == 0) {
 673                         if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
 674                                 continue;
 675                         }
 676                 } else {
 677                         assert(completed_effects->count(node) == 0);
 678                 }
 679
 680                 phase->effects.push_back(node);
 681                 if (node->effect->is_compute_shader()) {
 682                         phase->is_compute_shader = true;
 683                         phase->compute_shader_node = node;
 684                 }
 685
 686                 // Find all the dependencies of this effect, and add them to the stack.
 687                 vector<Node *> deps = node->incoming_links;
 688                 assert(node->effect->num_inputs() == deps.size());
 689                 for (unsigned i = 0; i < deps.size(); ++i) {
 690                         bool start_new_phase = false;
 691
 692                         if (node->effect->needs_texture_bounce() &&
 693                             !deps[i]->effect->is_single_texture() &&
 694                             !deps[i]->effect->override_disable_bounce()) {
 695                                 start_new_phase = true;
 696                         }
 697
 698                         // Propagate information about needing mipmaps down the chain,
 699                         // breaking the phase if we notice an incompatibility.
 700                         //
 701                         // Note that we cannot do this propagation as a normal pass,
 702                         // because it needs information about where the phases end
 703                         // (we should not propagate the flag across phases).
 704                         if (node->needs_mipmaps != Effect::DOES_NOT_NEED_MIPMAPS) {
 705                                 // The node can have a value set (ie. not DOES_NOT_NEED_MIPMAPS)
 706                                 // if we have diamonds in the graph; if so, choose that.
 707                                 // If not, the effect on the node can also decide (this is the
 708                                 // more common case).
 709                                 Effect::MipmapRequirements dep_mipmaps = deps[i]->needs_mipmaps;
 710                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
 711                                         if (deps[i]->effect->num_inputs() == 0) {
 712                                                 Input *input = static_cast<Input *>(deps[i]->effect);
 713                                                 dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS;
 714                                         } else {
 715                                                 dep_mipmaps = deps[i]->effect->needs_mipmaps();
 716                                         }
 717                                 }
 718                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
 719                                         deps[i]->needs_mipmaps = node->needs_mipmaps;
 720                                 } else if (dep_mipmaps != node->needs_mipmaps) {
 721                                         // The dependency cannot supply our mipmap demands
 722                                         // (either because it's an input that can't do mipmaps,
 723                                         // or because there's a conflict between mipmap-needing
 724                                         // and mipmap-refusing effects somewhere in the graph),
 725                                         // so they cannot be in the same phase.
 726                                         start_new_phase = true;
 727                                 }
 728                         }
 729
 730                         if (deps[i]->outgoing_links.size() > 1) {
 731                                 if (!deps[i]->effect->is_single_texture()) {
 732                                         // More than one effect uses this as the input,
 733                                         // and it is not a texture itself.
 734                                         // The easiest thing to do (and probably also the safest
 735                                         // performance-wise in most cases) is to bounce it to a texture
 736                                         // and then let the next passes read from that.
 737                                         start_new_phase = true;
 738                                 } else {
 739                                         assert(deps[i]->effect->num_inputs() == 0);
 740
 741                                         // For textures, we try to be slightly more clever;
 742                                         // if none of our outputs need a bounce, we don't bounce
 743                                         // but instead simply use the effect many times.
 744                                         //
 745                                         // Strictly speaking, we could bounce it for some outputs
 746                                         // and use it directly for others, but the processing becomes
 747                                         // somewhat simpler if the effect is only used in one such way.
 748                                         for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 749                                                 Node *rdep = deps[i]->outgoing_links[j];
 750                                                 start_new_phase |= rdep->effect->needs_texture_bounce();
 751                                         }
 752                                 }
 753                         }
 754
 755                         if (deps[i]->effect->is_compute_shader()) {
 756                                 // Only one compute shader per phase; we should have been stopped
 757                                 // already due to the fact that compute shaders are not one-to-one.
 758                                 assert(!phase->is_compute_shader);
 759
 760                                 // If all nodes so far are strong one-to-one, we can put them after
 761                                 // the compute shader (ie., process them on the output).
 762                                 start_new_phase = !node->strong_one_to_one_sampling;
 763                         } else if (deps[i]->effect->sets_virtual_output_size()) {
 764                                 assert(deps[i]->effect->changes_output_size());
 765                                 // If the next effect sets a virtual size to rely on OpenGL's
 766                                 // bilinear sampling, we'll really need to break the phase here.
 767                                 start_new_phase = true;
 768                         } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
 769                                 // If the next effect changes size and we don't have one-to-one sampling,
 770                                 // we also need to break here.
 771                                 start_new_phase = true;
 772                         }
 773
 774                         if (start_new_phase) {
 775                                 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
 776                         } else {
 777                                 effects_todo_this_phase.push(deps[i]);
 778
 779                                 // Propagate the one-to-one status down through the dependency.
 780                                 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
 781                                         deps[i]->effect->one_to_one_sampling();
 782                                 deps[i]->strong_one_to_one_sampling = node->strong_one_to_one_sampling &&
 783                                         deps[i]->effect->strong_one_to_one_sampling();
 784                         }
 785
 786                         node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE);
 787                 }
 788         }
 789
 790         // No more effects to do this phase. Take all the ones we have,
 791         // and create a GLSL program for it.
 792         assert(!phase->effects.empty());
 793
 794         // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
 795         // that would be nondeterministic and thus reduce cacheability.
 796         // TODO: Make this even more deterministic.
 797         vector<Phase *> dedup_inputs;
 798         set<Phase *> seen_inputs;
 799         for (size_t i = 0; i < phase->inputs.size(); ++i) {
 800                 if (seen_inputs.insert(phase->inputs[i]).second) {
 801                         dedup_inputs.push_back(phase->inputs[i]);
 802                 }
 803         }
 804         swap(phase->inputs, dedup_inputs);
 805
 806         // Allocate samplers for each input.
 807         phase->input_samplers.resize(phase->inputs.size());
 808
 809         // We added the effects from the output and back, but we need to output
 810         // them in topological sort order in the shader.
 811         phase->effects = topological_sort(phase->effects);
 812
 813         // Figure out if we need mipmaps or not, and if so, tell the inputs that.
 814         // (RTT inputs have different logic, which is checked in execute_phase().)
 815         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 816                 Node *node = phase->effects[i];
 817                 if (node->effect->num_inputs() == 0) {
 818                         Input *input = static_cast<Input *>(node->effect);
 819                         assert(node->needs_mipmaps != Effect::NEEDS_MIPMAPS || input->can_supply_mipmaps());
 820                         CHECK(input->set_int("needs_mipmaps", node->needs_mipmaps == Effect::NEEDS_MIPMAPS));
 821                 }
 822         }
 823
 824         // Tell each node which phase it ended up in, so that the unit test
 825         // can check that the phases were split in the right place.
 826         // Note that this ignores that effects may be part of multiple phases;
 827         // if the unit tests need to test such cases, we'll reconsider.
 828         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 829                 phase->effects[i]->containing_phase = phase;
 830         }
 831
 832         // Actually make the shader for this phase.
 833         compile_glsl_program(phase);
 834
 835         // Initialize timers.
 836         if (movit_timer_queries_supported) {
 837                 phase->time_elapsed_ns = 0;
 838                 phase->num_measured_iterations = 0;
 839         }
 840
 841         assert(completed_effects->count(output) == 0);
 842         completed_effects->insert(make_pair(output, phase));
 843         phases.push_back(phase);
 844         return phase;
 845 }
 846
 847 void EffectChain::output_dot(const char *filename)
 848 {
 849         if (movit_debug_level != MOVIT_DEBUG_ON) {
 850                 return;
 851         }
 852
 853         FILE *fp = fopen(filename, "w");
 854         if (fp == nullptr) {
 855                 perror(filename);
 856                 exit(1);
 857         }
 858
 859         fprintf(fp, "digraph G {\n");
 860         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 861         for (unsigned i = 0; i < nodes.size(); ++i) {
 862                 // Find out which phase this event belongs to.
 863                 vector<int> in_phases;
 864                 for (unsigned j = 0; j < phases.size(); ++j) {
 865                         const Phase* p = phases[j];
 866                         if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 867                                 in_phases.push_back(j);
 868                         }
 869                 }
 870
 871                 if (in_phases.empty()) {
 872                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 873                 } else if (in_phases.size() == 1) {
 874                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 875                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 876                                 (in_phases[0] % 8) + 1);
 877                 } else {
 878                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 879                         // But alas.
 880                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 881                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 882                                 (in_phases[0] % 8) + 1);
 883                 }
 884
 885                 char from_node_id[256];
 886                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 887
 888                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 889                         char to_node_id[256];
 890                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 891
 892                         vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 893                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 894                 }
 895
 896                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 897                         // Output node.
 898                         vector<string> labels = get_labels_for_edge(nodes[i], nullptr);
 899                         output_dot_edge(fp, from_node_id, "output", labels);
 900                 }
 901         }
 902         fprintf(fp, "}\n");
 903
 904         fclose(fp);
 905 }
 906
 907 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 908 {
 909         vector<string> labels;
 910
 911         if (to != nullptr && to->effect->needs_texture_bounce()) {
 912                 labels.push_back("needs_bounce");
 913         }
 914         if (from->effect->changes_output_size()) {
 915                 labels.push_back("resize");
 916         }
 917
 918         switch (from->output_color_space) {
 919         case COLORSPACE_INVALID:
 920                 labels.push_back("spc[invalid]");
 921                 break;
 922         case COLORSPACE_REC_601_525:
 923                 labels.push_back("spc[rec601-525]");
 924                 break;
 925         case COLORSPACE_REC_601_625:
 926                 labels.push_back("spc[rec601-625]");
 927                 break;
 928         default:
 929                 break;
 930         }
 931
 932         switch (from->output_gamma_curve) {
 933         case GAMMA_INVALID:
 934                 labels.push_back("gamma[invalid]");
 935                 break;
 936         case GAMMA_sRGB:
 937                 labels.push_back("gamma[sRGB]");
 938                 break;
 939         case GAMMA_REC_601:  // and GAMMA_REC_709
 940                 labels.push_back("gamma[rec601/709]");
 941                 break;
 942         default:
 943                 break;
 944         }
 945
 946         switch (from->output_alpha_type) {
 947         case ALPHA_INVALID:
 948                 labels.push_back("alpha[invalid]");
 949                 break;
 950         case ALPHA_BLANK:
 951                 labels.push_back("alpha[blank]");
 952                 break;
 953         case ALPHA_POSTMULTIPLIED:
 954                 labels.push_back("alpha[postmult]");
 955                 break;
 956         default:
 957                 break;
 958         }
 959
 960         return labels;
 961 }
 962
 963 void EffectChain::output_dot_edge(FILE *fp,
 964                                   const string &from_node_id,
 965                                   const string &to_node_id,
 966                                   const vector<string> &labels)
 967 {
 968         if (labels.empty()) {
 969                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 970         } else {
 971                 string label = labels[0];
 972                 for (unsigned k = 1; k < labels.size(); ++k) {
 973                         label += ", " + labels[k];
 974                 }
 975                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 976         }
 977 }
 978
 979 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
 980 {
 981         unsigned scaled_width, scaled_height;
 982
 983         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
 984                 // Same aspect, or W/H > aspect (image is wider than the frame).
 985                 // In either case, keep width, and adjust height.
 986                 scaled_width = width;
 987                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
 988         } else {
 989                 // W/H < aspect (image is taller than the frame), so keep height,
 990                 // and adjust width.
 991                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
 992                 scaled_height = height;
 993         }
 994
 995         // We should be consistently larger or smaller then the existing choice,
 996         // since we have the same aspect.
 997         assert(!(scaled_width < *output_width && scaled_height > *output_height));
 998         assert(!(scaled_height < *output_height && scaled_width > *output_width));
 999
1000         if (scaled_width >= *output_width && scaled_height >= *output_height) {
1001                 *output_width = scaled_width;
1002                 *output_height = scaled_height;
1003         }
1004 }
1005
1006 // Propagate input texture sizes throughout, and inform effects downstream.
1007 // (Like a lot of other code, we depend on effects being in topological order.)
1008 void EffectChain::inform_input_sizes(Phase *phase)
1009 {
1010         // All effects that have a defined size (inputs and RTT inputs)
1011         // get that. Reset all others.
1012         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1013                 Node *node = phase->effects[i];
1014                 if (node->effect->num_inputs() == 0) {
1015                         Input *input = static_cast<Input *>(node->effect);
1016                         node->output_width = input->get_width();
1017                         node->output_height = input->get_height();
1018                         assert(node->output_width != 0);
1019                         assert(node->output_height != 0);
1020                 } else {
1021                         node->output_width = node->output_height = 0;
1022                 }
1023         }
1024         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1025                 Phase *input = phase->inputs[i];
1026                 input->output_node->output_width = input->virtual_output_width;
1027                 input->output_node->output_height = input->virtual_output_height;
1028                 assert(input->output_node->output_width != 0);
1029                 assert(input->output_node->output_height != 0);
1030         }
1031
1032         // Now propagate from the inputs towards the end, and inform as we go.
1033         // The rules are simple:
1034         //
1035         //   1. Don't touch effects that already have given sizes (ie., inputs
1036         //      or effects that change the output size).
1037         //   2. If all of your inputs have the same size, that will be your output size.
1038         //   3. Otherwise, your output size is 0x0.
1039         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1040                 Node *node = phase->effects[i];
1041                 if (node->effect->num_inputs() == 0) {
1042                         continue;
1043                 }
1044                 unsigned this_output_width = 0;
1045                 unsigned this_output_height = 0;
1046                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1047                         Node *input = node->incoming_links[j];
1048                         node->effect->inform_input_size(j, input->output_width, input->output_height);
1049                         if (j == 0) {
1050                                 this_output_width = input->output_width;
1051                                 this_output_height = input->output_height;
1052                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
1053                                 // Inputs disagree.
1054                                 this_output_width = 0;
1055                                 this_output_height = 0;
1056                         }
1057                 }
1058                 if (node->effect->changes_output_size()) {
1059                         // We cannot call get_output_size() before we've done inform_input_size()
1060                         // on all inputs.
1061                         unsigned real_width, real_height;
1062                         node->effect->get_output_size(&real_width, &real_height,
1063                                                       &node->output_width, &node->output_height);
1064                         assert(node->effect->sets_virtual_output_size() ||
1065                                (real_width == node->output_width &&
1066                                 real_height == node->output_height));
1067                 } else {
1068                         node->output_width = this_output_width;
1069                         node->output_height = this_output_height;
1070                 }
1071         }
1072 }
1073
1074 // Note: You should call inform_input_sizes() before this, as the last effect's
1075 // desired output size might change based on the inputs.
1076 void EffectChain::find_output_size(Phase *phase)
1077 {
1078         Node *output_node = phase->is_compute_shader ? phase->compute_shader_node : phase->effects.back();
1079
1080         // If the last effect explicitly sets an output size, use that.
1081         if (output_node->effect->changes_output_size()) {
1082                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
1083                                                      &phase->virtual_output_width, &phase->virtual_output_height);
1084                 assert(output_node->effect->sets_virtual_output_size() ||
1085                        (phase->output_width == phase->virtual_output_width &&
1086                         phase->output_height == phase->virtual_output_height));
1087                 return;
1088         }
1089
1090         // If all effects have the same size, use that.
1091         unsigned output_width = 0, output_height = 0;
1092         bool all_inputs_same_size = true;
1093
1094         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1095                 Phase *input = phase->inputs[i];
1096                 assert(input->output_width != 0);
1097                 assert(input->output_height != 0);
1098                 if (output_width == 0 && output_height == 0) {
1099                         output_width = input->virtual_output_width;
1100                         output_height = input->virtual_output_height;
1101                 } else if (output_width != input->virtual_output_width ||
1102                            output_height != input->virtual_output_height) {
1103                         all_inputs_same_size = false;
1104                 }
1105         }
1106         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1107                 Effect *effect = phase->effects[i]->effect;
1108                 if (effect->num_inputs() != 0) {
1109                         continue;
1110                 }
1111
1112                 Input *input = static_cast<Input *>(effect);
1113                 if (output_width == 0 && output_height == 0) {
1114                         output_width = input->get_width();
1115                         output_height = input->get_height();
1116                 } else if (output_width != input->get_width() ||
1117                            output_height != input->get_height()) {
1118                         all_inputs_same_size = false;
1119                 }
1120         }
1121
1122         if (all_inputs_same_size) {
1123                 assert(output_width != 0);
1124                 assert(output_height != 0);
1125                 phase->virtual_output_width = phase->output_width = output_width;
1126                 phase->virtual_output_height = phase->output_height = output_height;
1127                 return;
1128         }
1129
1130         // If not, fit all the inputs into the current aspect, and select the largest one.
1131         output_width = 0;
1132         output_height = 0;
1133         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1134                 Phase *input = phase->inputs[i];
1135                 assert(input->output_width != 0);
1136                 assert(input->output_height != 0);
1137                 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
1138         }
1139         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1140                 Effect *effect = phase->effects[i]->effect;
1141                 if (effect->num_inputs() != 0) {
1142                         continue;
1143                 }
1144
1145                 Input *input = static_cast<Input *>(effect);
1146                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
1147         }
1148         assert(output_width != 0);
1149         assert(output_height != 0);
1150         phase->virtual_output_width = phase->output_width = output_width;
1151         phase->virtual_output_height = phase->output_height = output_height;
1152 }
1153
1154 void EffectChain::sort_all_nodes_topologically()
1155 {
1156         nodes = topological_sort(nodes);
1157 }
1158
1159 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
1160 {
1161         set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
1162         vector<Node *> sorted_list;
1163         for (unsigned i = 0; i < nodes.size(); ++i) {
1164                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
1165         }
1166         reverse(sorted_list.begin(), sorted_list.end());
1167         return sorted_list;
1168 }
1169
1170 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
1171 {
1172         if (nodes_left_to_visit->count(node) == 0) {
1173                 return;
1174         }
1175         nodes_left_to_visit->erase(node);
1176         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
1177                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
1178         }
1179         sorted_list->push_back(node);
1180 }
1181
1182 void EffectChain::find_color_spaces_for_inputs()
1183 {
1184         for (unsigned i = 0; i < nodes.size(); ++i) {
1185                 Node *node = nodes[i];
1186                 if (node->disabled) {
1187                         continue;
1188                 }
1189                 if (node->incoming_links.size() == 0) {
1190                         Input *input = static_cast<Input *>(node->effect);
1191                         node->output_color_space = input->get_color_space();
1192                         node->output_gamma_curve = input->get_gamma_curve();
1193
1194                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
1195                         switch (alpha_handling) {
1196                         case Effect::OUTPUT_BLANK_ALPHA:
1197                                 node->output_alpha_type = ALPHA_BLANK;
1198                                 break;
1199                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1200                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1201                                 break;
1202                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1203                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1204                                 break;
1205                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1206                         case Effect::DONT_CARE_ALPHA_TYPE:
1207                         default:
1208                                 assert(false);
1209                         }
1210
1211                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1212                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
1213                         }
1214                 }
1215         }
1216 }
1217
1218 // Propagate gamma and color space information as far as we can in the graph.
1219 // The rules are simple: Anything where all the inputs agree, get that as
1220 // output as well. Anything else keeps having *_INVALID.
1221 void EffectChain::propagate_gamma_and_color_space()
1222 {
1223         // We depend on going through the nodes in order.
1224         sort_all_nodes_topologically();
1225
1226         for (unsigned i = 0; i < nodes.size(); ++i) {
1227                 Node *node = nodes[i];
1228                 if (node->disabled) {
1229                         continue;
1230                 }
1231                 assert(node->incoming_links.size() == node->effect->num_inputs());
1232                 if (node->incoming_links.size() == 0) {
1233                         assert(node->output_color_space != COLORSPACE_INVALID);
1234                         assert(node->output_gamma_curve != GAMMA_INVALID);
1235                         continue;
1236                 }
1237
1238                 Colorspace color_space = node->incoming_links[0]->output_color_space;
1239                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1240                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1241                         if (node->incoming_links[j]->output_color_space != color_space) {
1242                                 color_space = COLORSPACE_INVALID;
1243                         }
1244                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1245                                 gamma_curve = GAMMA_INVALID;
1246                         }
1247                 }
1248
1249                 // The conversion effects already have their outputs set correctly,
1250                 // so leave them alone.
1251                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1252                         node->output_color_space = color_space;
1253                 }
1254                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1255                     node->effect->effect_type_id() != "GammaExpansionEffect") {
1256                         node->output_gamma_curve = gamma_curve;
1257                 }
1258         }
1259 }
1260
1261 // Propagate alpha information as far as we can in the graph.
1262 // Similar to propagate_gamma_and_color_space().
1263 void EffectChain::propagate_alpha()
1264 {
1265         // We depend on going through the nodes in order.
1266         sort_all_nodes_topologically();
1267
1268         for (unsigned i = 0; i < nodes.size(); ++i) {
1269                 Node *node = nodes[i];
1270                 if (node->disabled) {
1271                         continue;
1272                 }
1273                 assert(node->incoming_links.size() == node->effect->num_inputs());
1274                 if (node->incoming_links.size() == 0) {
1275                         assert(node->output_alpha_type != ALPHA_INVALID);
1276                         continue;
1277                 }
1278
1279                 // The alpha multiplication/division effects are special cases.
1280                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1281                         assert(node->incoming_links.size() == 1);
1282                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1283                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
1284                         continue;
1285                 }
1286                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1287                         assert(node->incoming_links.size() == 1);
1288                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1289                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1290                         continue;
1291                 }
1292
1293                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1294                 // because they are the only one that _need_ postmultiplied alpha.
1295                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1296                     node->effect->effect_type_id() == "GammaExpansionEffect") {
1297                         assert(node->incoming_links.size() == 1);
1298                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1299                                 node->output_alpha_type = ALPHA_BLANK;
1300                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1301                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1302                         } else {
1303                                 node->output_alpha_type = ALPHA_INVALID;
1304                         }
1305                         continue;
1306                 }
1307
1308                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1309                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1310                 // taken care of above. Rationale: Even if you could imagine
1311                 // e.g. an effect that took in an image and set alpha=1.0
1312                 // unconditionally, it wouldn't make any sense to have it as
1313                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1314                 // got its input pre- or postmultiplied, so it wouldn't know
1315                 // whether to divide away the old alpha or not.
1316                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1317                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1318                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1319                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1320
1321                 // If the node has multiple inputs, check that they are all valid and
1322                 // the same.
1323                 bool any_invalid = false;
1324                 bool any_premultiplied = false;
1325                 bool any_postmultiplied = false;
1326
1327                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1328                         switch (node->incoming_links[j]->output_alpha_type) {
1329                         case ALPHA_INVALID:
1330                                 any_invalid = true;
1331                                 break;
1332                         case ALPHA_BLANK:
1333                                 // Blank is good as both pre- and postmultiplied alpha,
1334                                 // so just ignore it.
1335                                 break;
1336                         case ALPHA_PREMULTIPLIED:
1337                                 any_premultiplied = true;
1338                                 break;
1339                         case ALPHA_POSTMULTIPLIED:
1340                                 any_postmultiplied = true;
1341                                 break;
1342                         default:
1343                                 assert(false);
1344                         }
1345                 }
1346
1347                 if (any_invalid) {
1348                         node->output_alpha_type = ALPHA_INVALID;
1349                         continue;
1350                 }
1351
1352                 // Inputs must be of the same type.
1353                 if (any_premultiplied && any_postmultiplied) {
1354                         node->output_alpha_type = ALPHA_INVALID;
1355                         continue;
1356                 }
1357
1358                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1359                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1360                         // This combination (requiring premultiplied alpha, but _not_ requiring
1361                         // linear light) is illegal, since the combination of premultiplied alpha
1362                         // and nonlinear inputs is meaningless.
1363                         assert(node->effect->needs_linear_light());
1364
1365                         // If the effect has asked for premultiplied alpha, check that it has got it.
1366                         if (any_postmultiplied) {
1367                                 node->output_alpha_type = ALPHA_INVALID;
1368                         } else if (!any_premultiplied &&
1369                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1370                                 // Blank input alpha, and the effect preserves blank alpha.
1371                                 node->output_alpha_type = ALPHA_BLANK;
1372                         } else {
1373                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1374                         }
1375                 } else {
1376                         // OK, all inputs are the same, and this effect is not going
1377                         // to change it.
1378                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1379                         if (any_premultiplied) {
1380                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1381                         } else if (any_postmultiplied) {
1382                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1383                         } else {
1384                                 node->output_alpha_type = ALPHA_BLANK;
1385                         }
1386                 }
1387         }
1388 }
1389
1390 bool EffectChain::node_needs_colorspace_fix(Node *node)
1391 {
1392         if (node->disabled) {
1393                 return false;
1394         }
1395         if (node->effect->num_inputs() == 0) {
1396                 return false;
1397         }
1398
1399         // propagate_gamma_and_color_space() has already set our output
1400         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1401         if (node->output_color_space == COLORSPACE_INVALID) {
1402                 return true;
1403         }
1404         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1405 }
1406
1407 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1408 // the graph. Our strategy is not always optimal, but quite simple:
1409 // Find an effect that's as early as possible where the inputs are of
1410 // unacceptable colorspaces (that is, either different, or, if the effect only
1411 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1412 // propagate the information anew, and repeat until there are no more such
1413 // effects.
1414 void EffectChain::fix_internal_color_spaces()
1415 {
1416         unsigned colorspace_propagation_pass = 0;
1417         bool found_any;
1418         do {
1419                 found_any = false;
1420                 for (unsigned i = 0; i < nodes.size(); ++i) {
1421                         Node *node = nodes[i];
1422                         if (!node_needs_colorspace_fix(node)) {
1423                                 continue;
1424                         }
1425
1426                         // Go through each input that is not sRGB, and insert
1427                         // a colorspace conversion after it.
1428                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1429                                 Node *input = node->incoming_links[j];
1430                                 assert(input->output_color_space != COLORSPACE_INVALID);
1431                                 if (input->output_color_space == COLORSPACE_sRGB) {
1432                                         continue;
1433                                 }
1434                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1435                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1436                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1437                                 conversion->output_color_space = COLORSPACE_sRGB;
1438                                 replace_sender(input, conversion);
1439                                 connect_nodes(input, conversion);
1440                         }
1441
1442                         // Re-sort topologically, and propagate the new information.
1443                         propagate_gamma_and_color_space();
1444
1445                         found_any = true;
1446                         break;
1447                 }
1448
1449                 char filename[256];
1450                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1451                 output_dot(filename);
1452                 assert(colorspace_propagation_pass < 100);
1453         } while (found_any);
1454
1455         for (unsigned i = 0; i < nodes.size(); ++i) {
1456                 Node *node = nodes[i];
1457                 if (node->disabled) {
1458                         continue;
1459                 }
1460                 assert(node->output_color_space != COLORSPACE_INVALID);
1461         }
1462 }
1463
1464 bool EffectChain::node_needs_alpha_fix(Node *node)
1465 {
1466         if (node->disabled) {
1467                 return false;
1468         }
1469
1470         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1471         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1472         return (node->output_alpha_type == ALPHA_INVALID);
1473 }
1474
1475 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1476 // the graph. Similar to fix_internal_color_spaces().
1477 void EffectChain::fix_internal_alpha(unsigned step)
1478 {
1479         unsigned alpha_propagation_pass = 0;
1480         bool found_any;
1481         do {
1482                 found_any = false;
1483                 for (unsigned i = 0; i < nodes.size(); ++i) {
1484                         Node *node = nodes[i];
1485                         if (!node_needs_alpha_fix(node)) {
1486                                 continue;
1487                         }
1488
1489                         // If we need to fix up GammaExpansionEffect, then clearly something
1490                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1491                         // is meaningless.
1492                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1493
1494                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1495
1496                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1497                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1498                                 assert(node->incoming_links.size() == 1);
1499                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1500                                 desired_type = ALPHA_POSTMULTIPLIED;
1501                         }
1502
1503                         // Go through each input that is not premultiplied alpha, and insert
1504                         // a conversion before it.
1505                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1506                                 Node *input = node->incoming_links[j];
1507                                 assert(input->output_alpha_type != ALPHA_INVALID);
1508                                 if (input->output_alpha_type == desired_type ||
1509                                     input->output_alpha_type == ALPHA_BLANK) {
1510                                         continue;
1511                                 }
1512                                 Node *conversion;
1513                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1514                                         conversion = add_node(new AlphaMultiplicationEffect());
1515                                 } else {
1516                                         conversion = add_node(new AlphaDivisionEffect());
1517                                 }
1518                                 conversion->output_alpha_type = desired_type;
1519                                 replace_sender(input, conversion);
1520                                 connect_nodes(input, conversion);
1521                         }
1522
1523                         // Re-sort topologically, and propagate the new information.
1524                         propagate_gamma_and_color_space();
1525                         propagate_alpha();
1526
1527                         found_any = true;
1528                         break;
1529                 }
1530
1531                 char filename[256];
1532                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1533                 output_dot(filename);
1534                 assert(alpha_propagation_pass < 100);
1535         } while (found_any);
1536
1537         for (unsigned i = 0; i < nodes.size(); ++i) {
1538                 Node *node = nodes[i];
1539                 if (node->disabled) {
1540                         continue;
1541                 }
1542                 assert(node->output_alpha_type != ALPHA_INVALID);
1543         }
1544 }
1545
1546 // Make so that the output is in the desired color space.
1547 void EffectChain::fix_output_color_space()
1548 {
1549         Node *output = find_output_node();
1550         if (output->output_color_space != output_format.color_space) {
1551                 Node *conversion = add_node(new ColorspaceConversionEffect());
1552                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1553                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1554                 conversion->output_color_space = output_format.color_space;
1555                 connect_nodes(output, conversion);
1556                 propagate_alpha();
1557                 propagate_gamma_and_color_space();
1558         }
1559 }
1560
1561 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1562 void EffectChain::fix_output_alpha()
1563 {
1564         Node *output = find_output_node();
1565         assert(output->output_alpha_type != ALPHA_INVALID);
1566         if (output->output_alpha_type == ALPHA_BLANK) {
1567                 // No alpha output, so we don't care.
1568                 return;
1569         }
1570         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1571             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1572                 Node *conversion = add_node(new AlphaDivisionEffect());
1573                 connect_nodes(output, conversion);
1574                 propagate_alpha();
1575                 propagate_gamma_and_color_space();
1576         }
1577         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1578             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1579                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1580                 connect_nodes(output, conversion);
1581                 propagate_alpha();
1582                 propagate_gamma_and_color_space();
1583         }
1584 }
1585
1586 bool EffectChain::node_needs_gamma_fix(Node *node)
1587 {
1588         if (node->disabled) {
1589                 return false;
1590         }
1591
1592         // Small hack since the output is not an explicit node:
1593         // If we are the last node and our output is in the wrong
1594         // space compared to EffectChain's output, we need to fix it.
1595         // This will only take us to linear, but fix_output_gamma()
1596         // will come and take us to the desired output gamma
1597         // if it is needed.
1598         //
1599         // This needs to be before everything else, since it could
1600         // even apply to inputs (if they are the only effect).
1601         if (node->outgoing_links.empty() &&
1602             node->output_gamma_curve != output_format.gamma_curve &&
1603             node->output_gamma_curve != GAMMA_LINEAR) {
1604                 return true;
1605         }
1606
1607         if (node->effect->num_inputs() == 0) {
1608                 return false;
1609         }
1610
1611         // propagate_gamma_and_color_space() has already set our output
1612         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1613         // except for GammaCompressionEffect.
1614         if (node->output_gamma_curve == GAMMA_INVALID) {
1615                 return true;
1616         }
1617         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1618                 assert(node->incoming_links.size() == 1);
1619                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1620         }
1621
1622         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1623 }
1624
1625 // Very similar to fix_internal_color_spaces(), but for gamma.
1626 // There is one difference, though; before we start adding conversion nodes,
1627 // we see if we can get anything out of asking the sources to deliver
1628 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1629 // does that part, while fix_internal_gamma_by_inserting_nodes()
1630 // inserts nodes as needed afterwards.
1631 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1632 {
1633         unsigned gamma_propagation_pass = 0;
1634         bool found_any;
1635         do {
1636                 found_any = false;
1637                 for (unsigned i = 0; i < nodes.size(); ++i) {
1638                         Node *node = nodes[i];
1639                         if (!node_needs_gamma_fix(node)) {
1640                                 continue;
1641                         }
1642
1643                         // See if all inputs can give us linear gamma. If not, leave it.
1644                         vector<Node *> nonlinear_inputs;
1645                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1646                         assert(!nonlinear_inputs.empty());
1647
1648                         bool all_ok = true;
1649                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1650                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1651                                 all_ok &= input->can_output_linear_gamma();
1652                         }
1653
1654                         if (!all_ok) {
1655                                 continue;
1656                         }
1657
1658                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1659                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1660                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1661                         }
1662
1663                         // Re-sort topologically, and propagate the new information.
1664                         propagate_gamma_and_color_space();
1665
1666                         found_any = true;
1667                         break;
1668                 }
1669
1670                 char filename[256];
1671                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1672                 output_dot(filename);
1673                 assert(gamma_propagation_pass < 100);
1674         } while (found_any);
1675 }
1676
1677 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1678 {
1679         unsigned gamma_propagation_pass = 0;
1680         bool found_any;
1681         do {
1682                 found_any = false;
1683                 for (unsigned i = 0; i < nodes.size(); ++i) {
1684                         Node *node = nodes[i];
1685                         if (!node_needs_gamma_fix(node)) {
1686                                 continue;
1687                         }
1688
1689                         // Special case: We could be an input and still be asked to
1690                         // fix our gamma; if so, we should be the only node
1691                         // (as node_needs_gamma_fix() would only return true in
1692                         // for an input in that case). That means we should insert
1693                         // a conversion node _after_ ourselves.
1694                         if (node->incoming_links.empty()) {
1695                                 assert(node->outgoing_links.empty());
1696                                 Node *conversion = add_node(new GammaExpansionEffect());
1697                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1698                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1699                                 connect_nodes(node, conversion);
1700                         }
1701
1702                         // If not, go through each input that is not linear gamma,
1703                         // and insert a gamma conversion after it.
1704                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1705                                 Node *input = node->incoming_links[j];
1706                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1707                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1708                                         continue;
1709                                 }
1710                                 Node *conversion = add_node(new GammaExpansionEffect());
1711                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1712                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1713                                 replace_sender(input, conversion);
1714                                 connect_nodes(input, conversion);
1715                         }
1716
1717                         // Re-sort topologically, and propagate the new information.
1718                         propagate_alpha();
1719                         propagate_gamma_and_color_space();
1720
1721                         found_any = true;
1722                         break;
1723                 }
1724
1725                 char filename[256];
1726                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1727                 output_dot(filename);
1728                 assert(gamma_propagation_pass < 100);
1729         } while (found_any);
1730
1731         for (unsigned i = 0; i < nodes.size(); ++i) {
1732                 Node *node = nodes[i];
1733                 if (node->disabled) {
1734                         continue;
1735                 }
1736                 assert(node->output_gamma_curve != GAMMA_INVALID);
1737         }
1738 }
1739
1740 // Make so that the output is in the desired gamma.
1741 // Note that this assumes linear input gamma, so it might create the need
1742 // for another pass of fix_internal_gamma().
1743 void EffectChain::fix_output_gamma()
1744 {
1745         Node *output = find_output_node();
1746         if (output->output_gamma_curve != output_format.gamma_curve) {
1747                 Node *conversion = add_node(new GammaCompressionEffect());
1748                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1749                 conversion->output_gamma_curve = output_format.gamma_curve;
1750                 connect_nodes(output, conversion);
1751         }
1752 }
1753
1754 // If the user has requested Y'CbCr output, we need to do this conversion
1755 // _after_ GammaCompressionEffect etc., but before dither (see below).
1756 // This is because Y'CbCr, with the exception of a special optional mode
1757 // in Rec. 2020 (which we currently don't support), is defined to work on
1758 // gamma-encoded data.
1759 void EffectChain::add_ycbcr_conversion_if_needed()
1760 {
1761         assert(output_color_rgba || num_output_color_ycbcr > 0);
1762         if (num_output_color_ycbcr == 0) {
1763                 return;
1764         }
1765         Node *output = find_output_node();
1766         ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type));
1767         connect_nodes(output, ycbcr_conversion_effect_node);
1768 }
1769
1770 // If the user has requested dither, add a DitherEffect right at the end
1771 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1772 // since dither is about the only effect that can _not_ be done in linear space.
1773 void EffectChain::add_dither_if_needed()
1774 {
1775         if (num_dither_bits == 0) {
1776                 return;
1777         }
1778         Node *output = find_output_node();
1779         Node *dither = add_node(new DitherEffect());
1780         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1781         connect_nodes(output, dither);
1782
1783         dither_effect = dither->effect;
1784 }
1785
1786 // Compute shaders can't output to the framebuffer, so if the last
1787 // phase ends in a compute shader, add a dummy phase at the end that
1788 // only blits directly from the temporary texture.
1789 //
1790 // TODO: Add an API for rendering directly to textures, for the cases
1791 // where we're only rendering to an FBO anyway.
1792 void EffectChain::add_dummy_effect_if_needed()
1793 {
1794         Node *output = find_output_node();
1795
1796         // See if the last effect that's not strong one-to-one is a compute shader.
1797         Node *last_effect = output;
1798         while (last_effect->effect->num_inputs() == 1 &&
1799                last_effect->effect->strong_one_to_one_sampling()) {
1800                 last_effect = last_effect->incoming_links[0];
1801         }
1802         if (last_effect->effect->is_compute_shader()) {
1803                 Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
1804                 connect_nodes(output, dummy);
1805                 has_dummy_effect = true;
1806         }
1807 }
1808
1809 // Find the output node. This is, simply, one that has no outgoing links.
1810 // If there are multiple ones, the graph is malformed (we do not support
1811 // multiple outputs right now).
1812 Node *EffectChain::find_output_node()
1813 {
1814         vector<Node *> output_nodes;
1815         for (unsigned i = 0; i < nodes.size(); ++i) {
1816                 Node *node = nodes[i];
1817                 if (node->disabled) {
1818                         continue;
1819                 }
1820                 if (node->outgoing_links.empty()) {
1821                         output_nodes.push_back(node);
1822                 }
1823         }
1824         assert(output_nodes.size() == 1);
1825         return output_nodes[0];
1826 }
1827
1828 void EffectChain::finalize()
1829 {
1830         // Output the graph as it is before we do any conversions on it.
1831         output_dot("step0-start.dot");
1832
1833         // Give each effect in turn a chance to rewrite its own part of the graph.
1834         // Note that if more effects are added as part of this, they will be
1835         // picked up as part of the same for loop, since they are added at the end.
1836         for (unsigned i = 0; i < nodes.size(); ++i) {
1837                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1838         }
1839         output_dot("step1-rewritten.dot");
1840
1841         find_color_spaces_for_inputs();
1842         output_dot("step2-input-colorspace.dot");
1843
1844         propagate_alpha();
1845         output_dot("step3-propagated-alpha.dot");
1846
1847         propagate_gamma_and_color_space();
1848         output_dot("step4-propagated-all.dot");
1849
1850         fix_internal_color_spaces();
1851         fix_internal_alpha(6);
1852         fix_output_color_space();
1853         output_dot("step7-output-colorspacefix.dot");
1854         fix_output_alpha();
1855         output_dot("step8-output-alphafix.dot");
1856
1857         // Note that we need to fix gamma after colorspace conversion,
1858         // because colorspace conversions might create needs for gamma conversions.
1859         // Also, we need to run an extra pass of fix_internal_gamma() after
1860         // fixing the output gamma, as we only have conversions to/from linear,
1861         // and fix_internal_alpha() since GammaCompressionEffect needs
1862         // postmultiplied input.
1863         fix_internal_gamma_by_asking_inputs(9);
1864         fix_internal_gamma_by_inserting_nodes(10);
1865         fix_output_gamma();
1866         output_dot("step11-output-gammafix.dot");
1867         propagate_alpha();
1868         output_dot("step12-output-alpha-propagated.dot");
1869         fix_internal_alpha(13);
1870         output_dot("step14-output-alpha-fixed.dot");
1871         fix_internal_gamma_by_asking_inputs(15);
1872         fix_internal_gamma_by_inserting_nodes(16);
1873
1874         output_dot("step17-before-ycbcr.dot");
1875         add_ycbcr_conversion_if_needed();
1876
1877         output_dot("step18-before-dither.dot");
1878         add_dither_if_needed();
1879
1880         output_dot("step19-before-dummy-effect.dot");
1881         add_dummy_effect_if_needed();
1882
1883         output_dot("step20-final.dot");
1884
1885         // Construct all needed GLSL programs, starting at the output.
1886         // We need to keep track of which effects have already been computed,
1887         // as an effect with multiple users could otherwise be calculated
1888         // multiple times.
1889         map<Node *, Phase *> completed_effects;
1890         construct_phase(find_output_node(), &completed_effects);
1891
1892         output_dot("step21-split-to-phases.dot");
1893
1894         assert(phases[0]->inputs.empty());
1895
1896         finalized = true;
1897 }
1898
1899 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1900 {
1901         // Save original viewport.
1902         GLuint x = 0, y = 0;
1903
1904         if (width == 0 && height == 0) {
1905                 GLint viewport[4];
1906                 glGetIntegerv(GL_VIEWPORT, viewport);
1907                 x = viewport[0];
1908                 y = viewport[1];
1909                 width = viewport[2];
1910                 height = viewport[3];
1911         }
1912
1913         render(dest_fbo, {}, x, y, width, height);
1914 }
1915
1916 void EffectChain::render_to_texture(const vector<DestinationTexture> &destinations, unsigned width, unsigned height)
1917 {
1918         assert(finalized);
1919         assert(!destinations.empty());
1920
1921         if (!has_dummy_effect) {
1922                 // We don't end in a compute shader, so there's nothing specific for us to do.
1923                 // Create an FBO for this set of textures, and just render to that.
1924                 GLuint texnums[4] = { 0, 0, 0, 0 };
1925                 for (unsigned i = 0; i < destinations.size() && i < 4; ++i) {
1926                         texnums[i] = destinations[i].texnum;
1927                 }
1928                 GLuint dest_fbo = resource_pool->create_fbo(texnums[0], texnums[1], texnums[2], texnums[3]);
1929                 render(dest_fbo, {}, 0, 0, width, height);
1930                 resource_pool->release_fbo(dest_fbo);
1931         } else {
1932                 render((GLuint)-1, destinations, 0, 0, width, height);
1933         }
1934 }
1935
1936 void EffectChain::render(GLuint dest_fbo, const vector<DestinationTexture> &destinations, unsigned x, unsigned y, unsigned width, unsigned height)
1937 {
1938         assert(finalized);
1939         assert(destinations.size() <= 1);
1940
1941         // This needs to be set anew, in case we are coming from a different context
1942         // from when we initialized.
1943         check_error();
1944         glDisable(GL_DITHER);
1945         check_error();
1946
1947         const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB);
1948         check_error();
1949         bool current_srgb = final_srgb;
1950
1951         // Basic state.
1952         check_error();
1953         glDisable(GL_BLEND);
1954         check_error();
1955         glDisable(GL_DEPTH_TEST);
1956         check_error();
1957         glDepthMask(GL_FALSE);
1958         check_error();
1959
1960         set<Phase *> generated_mipmaps;
1961
1962         // We keep one texture per output, but only for as long as we actually have any
1963         // phases that need it as an input. (We don't make any effort to reorder phases
1964         // to minimize the number of textures in play, as register allocation can be
1965         // complicated and we rarely have much to gain, since our graphs are typically
1966         // pretty linear.)
1967         map<Phase *, GLuint> output_textures;
1968         map<Phase *, int> ref_counts;
1969         for (Phase *phase : phases) {
1970                 for (Phase *input : phase->inputs) {
1971                         ++ref_counts[input];
1972                 }
1973         }
1974
1975         size_t num_phases = phases.size();
1976         if (destinations.empty()) {
1977                 assert(dest_fbo != (GLuint)-1);
1978         } else {
1979                 assert(has_dummy_effect);
1980                 assert(x == 0);
1981                 assert(y == 0);
1982                 assert(num_phases >= 2);
1983                 assert(!phases.back()->is_compute_shader);
1984                 assert(phases.back()->effects.size() == 1);
1985                 assert(phases.back()->effects[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
1986
1987                 // We are rendering to a set of textures, so we can run the compute shader
1988                 // directly and skip the dummy phase.
1989                 --num_phases;
1990         }
1991
1992         for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) {
1993                 Phase *phase = phases[phase_num];
1994
1995                 if (do_phase_timing) {
1996                         GLuint timer_query_object;
1997                         if (phase->timer_query_objects_free.empty()) {
1998                                 glGenQueries(1, &timer_query_object);
1999                         } else {
2000                                 timer_query_object = phase->timer_query_objects_free.front();
2001                                 phase->timer_query_objects_free.pop_front();
2002                         }
2003                         glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
2004                         phase->timer_query_objects_running.push_back(timer_query_object);
2005                 }
2006                 bool last_phase = (phase_num == num_phases - 1);
2007                 if (phase_num == num_phases - 1) {
2008                         // Last phase goes to the output the user specified.
2009                         if (!phase->is_compute_shader) {
2010                                 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
2011                                 check_error();
2012                                 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
2013                                 assert(status == GL_FRAMEBUFFER_COMPLETE);
2014                                 glViewport(x, y, width, height);
2015                         }
2016                         if (dither_effect != nullptr) {
2017                                 CHECK(dither_effect->set_int("output_width", width));
2018                                 CHECK(dither_effect->set_int("output_height", height));
2019                         }
2020                 }
2021
2022                 // Enable sRGB rendering for intermediates in case we are
2023                 // rendering to an sRGB format.
2024                 // TODO: Support this for compute shaders.
2025                 bool needs_srgb = last_phase ? final_srgb : true;
2026                 if (needs_srgb && !current_srgb) {
2027                         glEnable(GL_FRAMEBUFFER_SRGB);
2028                         check_error();
2029                         current_srgb = true;
2030                 } else if (!needs_srgb && current_srgb) {
2031                         glDisable(GL_FRAMEBUFFER_SRGB);
2032                         check_error();
2033                         current_srgb = true;
2034                 }
2035
2036                 // Find a texture for this phase.
2037                 inform_input_sizes(phase);
2038                 find_output_size(phase);
2039                 vector<DestinationTexture> phase_destinations;
2040                 if (!last_phase) {
2041                         GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
2042                         output_textures.insert(make_pair(phase, tex_num));
2043                         phase_destinations.push_back(DestinationTexture{ tex_num, intermediate_format });
2044
2045                         // The output texture needs to have valid state to be written to by a compute shader.
2046                         glActiveTexture(GL_TEXTURE0);
2047                         check_error();
2048                         glBindTexture(GL_TEXTURE_2D, tex_num);
2049                         check_error();
2050                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2051                         check_error();
2052                 } else if (phase->is_compute_shader) {
2053                         assert(!destinations.empty());
2054                         phase_destinations = destinations;
2055                 }
2056
2057                 execute_phase(phase, output_textures, phase_destinations, &generated_mipmaps);
2058                 if (do_phase_timing) {
2059                         glEndQuery(GL_TIME_ELAPSED);
2060                 }
2061
2062                 // Drop any input textures we don't need anymore.
2063                 for (Phase *input : phase->inputs) {
2064                         assert(ref_counts[input] > 0);
2065                         if (--ref_counts[input] == 0) {
2066                                 resource_pool->release_2d_texture(output_textures[input]);
2067                                 output_textures.erase(input);
2068                         }
2069                 }
2070         }
2071
2072         for (const auto &phase_and_texnum : output_textures) {
2073                 resource_pool->release_2d_texture(phase_and_texnum.second);
2074         }
2075
2076         glBindFramebuffer(GL_FRAMEBUFFER, 0);
2077         check_error();
2078         glUseProgram(0);
2079         check_error();
2080
2081         glBindBuffer(GL_ARRAY_BUFFER, 0);
2082         check_error();
2083         glBindVertexArray(0);
2084         check_error();
2085
2086         if (do_phase_timing) {
2087                 // Get back the timer queries.
2088                 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2089                         Phase *phase = phases[phase_num];
2090                         for (auto timer_it = phase->timer_query_objects_running.cbegin();
2091                              timer_it != phase->timer_query_objects_running.cend(); ) {
2092                                 GLint timer_query_object = *timer_it;
2093                                 GLint available;
2094                                 glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
2095                                 if (available) {
2096                                         GLuint64 time_elapsed;
2097                                         glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
2098                                         phase->time_elapsed_ns += time_elapsed;
2099                                         ++phase->num_measured_iterations;
2100                                         phase->timer_query_objects_free.push_back(timer_query_object);
2101                                         phase->timer_query_objects_running.erase(timer_it++);
2102                                 } else {
2103                                         ++timer_it;
2104                                 }
2105                         }
2106                 }
2107         }
2108 }
2109
2110 void EffectChain::enable_phase_timing(bool enable)
2111 {
2112         if (enable) {
2113                 assert(movit_timer_queries_supported);
2114         }
2115         this->do_phase_timing = enable;
2116 }
2117
2118 void EffectChain::reset_phase_timing()
2119 {
2120         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2121                 Phase *phase = phases[phase_num];
2122                 phase->time_elapsed_ns = 0;
2123                 phase->num_measured_iterations = 0;
2124         }
2125 }
2126
2127 void EffectChain::print_phase_timing()
2128 {
2129         double total_time_ms = 0.0;
2130         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2131                 Phase *phase = phases[phase_num];
2132                 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
2133                 printf("Phase %d: %5.1f ms  [", phase_num, avg_time_ms);
2134                 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
2135                         if (effect_num != 0) {
2136                                 printf(", ");
2137                         }
2138                         printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
2139                 }
2140                 printf("]\n");
2141                 total_time_ms += avg_time_ms;
2142         }
2143         printf("Total:   %5.1f ms\n", total_time_ms);
2144 }
2145
2146 void EffectChain::execute_phase(Phase *phase,
2147                                 const map<Phase *, GLuint> &output_textures,
2148                                 const vector<DestinationTexture> &destinations,
2149                                 set<Phase *> *generated_mipmaps)
2150 {
2151         // Set up RTT inputs for this phase.
2152         for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
2153                 glActiveTexture(GL_TEXTURE0 + sampler);
2154                 Phase *input = phase->inputs[sampler];
2155                 input->output_node->bound_sampler_num = sampler;
2156                 const auto it = output_textures.find(input);
2157                 assert(it != output_textures.end());
2158                 glBindTexture(GL_TEXTURE_2D, it->second);
2159                 check_error();
2160
2161                 // See if anything using this RTT input (in this phase) needs mipmaps.
2162                 // TODO: It could be that we get conflicting logic here, if we have
2163                 // multiple effects with incompatible mipmaps using the same
2164                 // RTT input. However, that is obscure enough that we can deal
2165                 // with it at some future point (preferably when we have
2166                 // universal support for separate sampler objects!). For now,
2167                 // an assert is good enough. See also the TODO at bound_sampler_num.
2168                 bool any_needs_mipmaps = false, any_refuses_mipmaps = false;
2169                 for (Node *node : phase->effects) {
2170                         assert(node->incoming_links.size() == node->incoming_link_type.size());
2171                         for (size_t i = 0; i < node->incoming_links.size(); ++i) {
2172                                 if (node->incoming_links[i] == input->output_node &&
2173                                     node->incoming_link_type[i] == IN_ANOTHER_PHASE) {
2174                                         if (node->needs_mipmaps == Effect::NEEDS_MIPMAPS) {
2175                                                 any_needs_mipmaps = true;
2176                                         } else if (node->needs_mipmaps == Effect::CANNOT_ACCEPT_MIPMAPS) {
2177                                                 any_refuses_mipmaps = true;
2178                                         }
2179                                 }
2180                         }
2181                 }
2182                 assert(!(any_needs_mipmaps && any_refuses_mipmaps));
2183
2184                 if (any_needs_mipmaps && generated_mipmaps->count(input) == 0) {
2185                         glGenerateMipmap(GL_TEXTURE_2D);
2186                         check_error();
2187                         generated_mipmaps->insert(input);
2188                 }
2189                 setup_rtt_sampler(sampler, any_needs_mipmaps);
2190                 phase->input_samplers[sampler] = sampler;  // Bind the sampler to the right uniform.
2191         }
2192
2193         GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
2194         check_error();
2195
2196         // And now the output.
2197         GLuint fbo = 0;
2198         if (phase->is_compute_shader) {
2199                 assert(!destinations.empty());
2200
2201                 // This is currently the only place where we use image units,
2202                 // so we can always start at 0. TODO: Support multiple destinations.
2203                 phase->outbuf_image_unit = 0;
2204                 glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format);
2205                 check_error();
2206                 phase->uniform_output_size[0] = phase->output_width;
2207                 phase->uniform_output_size[1] = phase->output_height;
2208                 phase->inv_output_size.x = 1.0f / phase->output_width;
2209                 phase->inv_output_size.y = 1.0f / phase->output_height;
2210                 phase->output_texcoord_adjust.x = 0.5f / phase->output_width;
2211                 phase->output_texcoord_adjust.y = 0.5f / phase->output_height;
2212         } else if (!destinations.empty()) {
2213                 assert(destinations.size() == 1);
2214                 fbo = resource_pool->create_fbo(destinations[0].texnum);
2215                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
2216                 glViewport(0, 0, phase->output_width, phase->output_height);
2217         }
2218
2219         // Give the required parameters to all the effects.
2220         unsigned sampler_num = phase->inputs.size();
2221         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2222                 Node *node = phase->effects[i];
2223                 unsigned old_sampler_num = sampler_num;
2224                 node->effect->set_gl_state(instance_program_num, phase->effect_ids[make_pair(node, IN_SAME_PHASE)], &sampler_num);
2225                 check_error();
2226
2227                 if (node->effect->is_single_texture()) {
2228                         assert(sampler_num - old_sampler_num == 1);
2229                         node->bound_sampler_num = old_sampler_num;
2230                 } else {
2231                         node->bound_sampler_num = -1;
2232                 }
2233         }
2234
2235         if (phase->is_compute_shader) {
2236                 unsigned x, y, z;
2237                 phase->compute_shader_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z);
2238
2239                 // Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(),
2240                 // since they can be updated from there.
2241                 setup_uniforms(phase);
2242                 glDispatchCompute(x, y, z);
2243                 check_error();
2244                 glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT | GL_TEXTURE_UPDATE_BARRIER_BIT);
2245                 check_error();
2246         } else {
2247                 // Uniforms need to come after set_gl_state(), since they can be updated
2248                 // from there.
2249                 setup_uniforms(phase);
2250
2251                 // Bind the vertex data.
2252                 GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
2253                 glBindVertexArray(vao);
2254
2255                 glDrawArrays(GL_TRIANGLES, 0, 3);
2256                 check_error();
2257
2258                 resource_pool->release_vec2_vao(vao);
2259         }
2260
2261         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2262                 Node *node = phase->effects[i];
2263                 node->effect->clear_gl_state();
2264         }
2265
2266         resource_pool->unuse_glsl_program(instance_program_num);
2267
2268         if (fbo != 0) {
2269                 resource_pool->release_fbo(fbo);
2270         }
2271 }
2272
2273 void EffectChain::setup_uniforms(Phase *phase)
2274 {
2275         // TODO: Use UBO blocks.
2276         for (size_t i = 0; i < phase->uniforms_image2d.size(); ++i) {
2277                 const Uniform<int> &uniform = phase->uniforms_image2d[i];
2278                 if (uniform.location != -1) {
2279                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2280                 }
2281         }
2282         for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
2283                 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
2284                 if (uniform.location != -1) {
2285                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2286                 }
2287         }
2288         for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
2289                 const Uniform<bool> &uniform = phase->uniforms_bool[i];
2290                 assert(uniform.num_values == 1);
2291                 if (uniform.location != -1) {
2292                         glUniform1i(uniform.location, *uniform.value);
2293                 }
2294         }
2295         for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
2296                 const Uniform<int> &uniform = phase->uniforms_int[i];
2297                 if (uniform.location != -1) {
2298                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2299                 }
2300         }
2301         for (size_t i = 0; i < phase->uniforms_ivec2.size(); ++i) {
2302                 const Uniform<int> &uniform = phase->uniforms_ivec2[i];
2303                 if (uniform.location != -1) {
2304                         glUniform2iv(uniform.location, uniform.num_values, uniform.value);
2305                 }
2306         }
2307         for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
2308                 const Uniform<float> &uniform = phase->uniforms_float[i];
2309                 if (uniform.location != -1) {
2310                         glUniform1fv(uniform.location, uniform.num_values, uniform.value);
2311                 }
2312         }
2313         for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
2314                 const Uniform<float> &uniform = phase->uniforms_vec2[i];
2315                 if (uniform.location != -1) {
2316                         glUniform2fv(uniform.location, uniform.num_values, uniform.value);
2317                 }
2318         }
2319         for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
2320                 const Uniform<float> &uniform = phase->uniforms_vec3[i];
2321                 if (uniform.location != -1) {
2322                         glUniform3fv(uniform.location, uniform.num_values, uniform.value);
2323                 }
2324         }
2325         for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
2326                 const Uniform<float> &uniform = phase->uniforms_vec4[i];
2327                 if (uniform.location != -1) {
2328                         glUniform4fv(uniform.location, uniform.num_values, uniform.value);
2329                 }
2330         }
2331         for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
2332                 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
2333                 assert(uniform.num_values == 1);
2334                 if (uniform.location != -1) {
2335                         // Convert to float (GLSL has no double matrices).
2336                         float matrixf[9];
2337                         for (unsigned y = 0; y < 3; ++y) {
2338                                 for (unsigned x = 0; x < 3; ++x) {
2339                                         matrixf[y + x * 3] = (*uniform.value)(y, x);
2340                                 }
2341                         }
2342                         glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
2343                 }
2344         }
2345 }
2346
2347 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
2348 {
2349         glActiveTexture(GL_TEXTURE0 + sampler_num);
2350         check_error();
2351         if (use_mipmaps) {
2352                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2353                 check_error();
2354         } else {
2355                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2356                 check_error();
2357         }
2358         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2359         check_error();
2360         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2361         check_error();
2362 }
2363
2364 }  // namespace movit