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                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 416                         if (node->incoming_links.size() == 1) {
 417                                 frag_shader += "#define INPUT";
 418                         } else {
 419                                 char buf[256];
 420                                 sprintf(buf, "#define INPUT%d", j + 1);
 421                                 frag_shader += buf;
 422                         }
 423
 424                         Node *input = node->incoming_links[j];
 425                         NodeLinkType link_type = node->incoming_link_type[j];
 426                         if (i != 0 &&
 427                             input->effect->is_compute_shader() &&
 428                             node->incoming_link_type[j] == IN_SAME_PHASE) {
 429                                 // First effect after the compute shader reads the value
 430                                 // that cs_output() wrote to a global variable,
 431                                 // ignoring the tc (since all such effects have to be
 432                                 // strong one-to-one).
 433                                 frag_shader += "(tc) CS_OUTPUT_VAL\n";
 434                         } else {
 435                                 assert(phase->effect_ids.count(make_pair(input, link_type)));
 436                                 frag_shader += string(" ") + phase->effect_ids[make_pair(input, link_type)] + "\n";
 437                         }
 438                 }
 439
 440                 frag_shader += "\n";
 441                 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
 442                 if (node->effect->is_compute_shader()) {
 443                         frag_shader += string("#define NORMALIZE_TEXTURE_COORDS(tc) ((tc) * ") + effect_id + "_inv_output_size + " + effect_id + "_output_texcoord_adjust)\n";
 444                 }
 445                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
 446                 frag_shader += "#undef FUNCNAME\n";
 447                 if (node->incoming_links.size() == 1) {
 448                         frag_shader += "#undef INPUT\n";
 449                 } else {
 450                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
 451                                 char buf[256];
 452                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
 453                                 frag_shader += buf;
 454                         }
 455                 }
 456                 frag_shader += "\n";
 457         }
 458         if (phase->is_compute_shader) {
 459                 assert(phase->effect_ids.count(make_pair(phase->compute_shader_node, IN_SAME_PHASE)));
 460                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->compute_shader_node, IN_SAME_PHASE)] + "\n";
 461                 if (phase->compute_shader_node == phase->effects.back()) {
 462                         // No postprocessing.
 463                         frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n";
 464                 } else {
 465                         frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
 466                 }
 467         } else {
 468                 assert(phase->effect_ids.count(make_pair(phase->effects.back(), IN_SAME_PHASE)));
 469                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
 470         }
 471
 472         // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
 473         vector<string> frag_shader_outputs;  // In order.
 474         if (phase->output_node->outgoing_links.empty() && num_output_color_ycbcr > 0) {
 475                 switch (output_ycbcr_splitting[0]) {
 476                 case YCBCR_OUTPUT_INTERLEAVED:
 477                         // No #defines set.
 478                         frag_shader_outputs.push_back("FragColor");
 479                         break;
 480                 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
 481                         frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
 482                         frag_shader_outputs.push_back("Y");
 483                         frag_shader_outputs.push_back("Chroma");
 484                         break;
 485                 case YCBCR_OUTPUT_PLANAR:
 486                         frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
 487                         frag_shader_outputs.push_back("Y");
 488                         frag_shader_outputs.push_back("Cb");
 489                         frag_shader_outputs.push_back("Cr");
 490                         break;
 491                 default:
 492                         assert(false);
 493                 }
 494
 495                 if (num_output_color_ycbcr > 1) {
 496                         switch (output_ycbcr_splitting[1]) {
 497                         case YCBCR_OUTPUT_INTERLEAVED:
 498                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_INTERLEAVED 1\n";
 499                                 frag_shader_outputs.push_back("YCbCr2");
 500                                 break;
 501                         case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
 502                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
 503                                 frag_shader_outputs.push_back("Y2");
 504                                 frag_shader_outputs.push_back("Chroma2");
 505                                 break;
 506                         case YCBCR_OUTPUT_PLANAR:
 507                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_PLANAR 1\n";
 508                                 frag_shader_outputs.push_back("Y2");
 509                                 frag_shader_outputs.push_back("Cb2");
 510                                 frag_shader_outputs.push_back("Cr2");
 511                                 break;
 512                         default:
 513                                 assert(false);
 514                         }
 515                 }
 516
 517                 if (output_color_rgba) {
 518                         // Note: Needs to come in the header, because not only the
 519                         // output needs to see it (YCbCrConversionEffect and DitherEffect
 520                         // do, too).
 521                         frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
 522                         frag_shader_outputs.push_back("RGBA");
 523                 }
 524         }
 525
 526         // If we're bouncing to a temporary texture, signal transformation if desired.
 527         if (!phase->output_node->outgoing_links.empty()) {
 528                 if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
 529                     phase->output_node->output_gamma_curve == GAMMA_LINEAR) {
 530                         frag_shader += "#define SQUARE_ROOT_TRANSFORMATION 1\n";
 531                 }
 532         }
 533
 534         if (phase->is_compute_shader) {
 535                 frag_shader.append(read_file("footer.comp"));
 536                 phase->compute_shader_node->effect->register_uniform_ivec2("output_size", phase->uniform_output_size);
 537                 phase->compute_shader_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size);
 538                 phase->compute_shader_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust);
 539         } else {
 540                 frag_shader.append(read_file("footer.frag"));
 541         }
 542
 543         // Collect uniforms from all effects and output them. Note that this needs
 544         // to happen after output_fragment_shader(), even though the uniforms come
 545         // before in the output source, since output_fragment_shader() is allowed
 546         // to register new uniforms (e.g. arrays that are of unknown length until
 547         // finalization time).
 548         // TODO: Make a uniform block for platforms that support it.
 549         string frag_shader_uniforms = "";
 550         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 551                 Node *node = phase->effects[i];
 552                 Effect *effect = node->effect;
 553                 const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
 554                 extract_uniform_declarations(effect->uniforms_image2d, "image2D", effect_id, &phase->uniforms_image2d, &frag_shader_uniforms);
 555                 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
 556                 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
 557                 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
 558                 extract_uniform_declarations(effect->uniforms_ivec2, "ivec2", effect_id, &phase->uniforms_ivec2, &frag_shader_uniforms);
 559                 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
 560                 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
 561                 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
 562                 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
 563                 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
 564                 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
 565                 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
 566                 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
 567                 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
 568         }
 569
 570         string vert_shader = read_version_dependent_file("vs", "vert");
 571
 572         // If we're the last phase and need to flip the picture to compensate for
 573         // the origin, tell the vertex or compute shader so.
 574         bool is_last_phase;
 575         if (has_dummy_effect) {
 576                 is_last_phase = (phase->output_node->outgoing_links.size() == 1 &&
 577                         phase->output_node->outgoing_links[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
 578         } else {
 579                 is_last_phase = phase->output_node->outgoing_links.empty();
 580         }
 581         if (is_last_phase && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
 582                 if (phase->is_compute_shader) {
 583                         frag_shader_header += "#define FLIP_ORIGIN 1\n";
 584                 } else {
 585                         const string needle = "#define FLIP_ORIGIN 0";
 586                         size_t pos = vert_shader.find(needle);
 587                         assert(pos != string::npos);
 588
 589                         vert_shader[pos + needle.size() - 1] = '1';
 590                 }
 591         }
 592
 593         frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
 594
 595         if (phase->is_compute_shader) {
 596                 phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader);
 597
 598                 Uniform<int> uniform;
 599                 uniform.name = "outbuf";
 600                 uniform.value = &phase->outbuf_image_unit;
 601                 uniform.prefix = "tex";
 602                 uniform.num_values = 1;
 603                 uniform.location = -1;
 604                 phase->uniforms_image2d.push_back(uniform);
 605         } else {
 606                 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
 607         }
 608         GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
 609         GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
 610         if (position_attribute_index != -1) {
 611                 phase->attribute_indexes.insert(position_attribute_index);
 612         }
 613         if (texcoord_attribute_index != -1) {
 614                 phase->attribute_indexes.insert(texcoord_attribute_index);
 615         }
 616
 617         // Collect the resulting location numbers for each uniform.
 618         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_image2d);
 619         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
 620         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
 621         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
 622         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_ivec2);
 623         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
 624         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
 625         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
 626         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
 627         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
 628 }
 629
 630 // Construct GLSL programs, starting at the given effect and following
 631 // the chain from there. We end a program every time we come to an effect
 632 // marked as "needs texture bounce", one that is used by multiple other
 633 // effects, every time we need to bounce due to output size change
 634 // (not all size changes require ending), and of course at the end.
 635 //
 636 // We follow a quite simple depth-first search from the output, although
 637 // without recursing explicitly within each phase.
 638 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
 639 {
 640         if (completed_effects->count(output)) {
 641                 return (*completed_effects)[output];
 642         }
 643
 644         Phase *phase = new Phase;
 645         phase->output_node = output;
 646         phase->is_compute_shader = false;
 647         phase->compute_shader_node = nullptr;
 648
 649         // If the output effect has one-to-one sampling, we try to trace this
 650         // status down through the dependency chain. This is important in case
 651         // we hit an effect that changes output size (and not sets a virtual
 652         // output size); if we have one-to-one sampling, we don't have to break
 653         // the phase.
 654         output->one_to_one_sampling = output->effect->one_to_one_sampling();
 655         output->strong_one_to_one_sampling = output->effect->strong_one_to_one_sampling();
 656
 657         // Effects that we have yet to calculate, but that we know should
 658         // be in the current phase.
 659         stack<Node *> effects_todo_this_phase;
 660         effects_todo_this_phase.push(output);
 661
 662         while (!effects_todo_this_phase.empty()) {
 663                 Node *node = effects_todo_this_phase.top();
 664                 effects_todo_this_phase.pop();
 665
 666                 assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling());
 667
 668                 if (node->effect->needs_mipmaps() != Effect::DOES_NOT_NEED_MIPMAPS) {
 669                         // Can't have incompatible requirements imposed on us from a dependent effect;
 670                         // if so, it should have started a new phase instead.
 671                         assert(node->needs_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS ||
 672                                node->needs_mipmaps == node->effect->needs_mipmaps());
 673                         node->needs_mipmaps = node->effect->needs_mipmaps();
 674                 }
 675
 676                 // This should currently only happen for effects that are inputs
 677                 // (either true inputs or phase outputs). We special-case inputs,
 678                 // and then deduplicate phase outputs below.
 679                 if (node->effect->num_inputs() == 0) {
 680                         if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
 681                                 continue;
 682                         }
 683                 } else {
 684                         assert(completed_effects->count(node) == 0);
 685                 }
 686
 687                 phase->effects.push_back(node);
 688                 if (node->effect->is_compute_shader()) {
 689                         assert(phase->compute_shader_node == nullptr ||
 690                                phase->compute_shader_node == node);
 691                         phase->is_compute_shader = true;
 692                         phase->compute_shader_node = node;
 693                 }
 694
 695                 // Find all the dependencies of this effect, and add them to the stack.
 696                 vector<Node *> deps = node->incoming_links;
 697                 assert(node->effect->num_inputs() == deps.size());
 698                 for (unsigned i = 0; i < deps.size(); ++i) {
 699                         bool start_new_phase = false;
 700
 701                         Effect::MipmapRequirements save_needs_mipmaps = deps[i]->needs_mipmaps;
 702
 703                         if (node->effect->needs_texture_bounce() &&
 704                             !deps[i]->effect->is_single_texture() &&
 705                             !deps[i]->effect->override_disable_bounce()) {
 706                                 start_new_phase = true;
 707                         }
 708
 709                         // Propagate information about needing mipmaps down the chain,
 710                         // breaking the phase if we notice an incompatibility.
 711                         //
 712                         // Note that we cannot do this propagation as a normal pass,
 713                         // because it needs information about where the phases end
 714                         // (we should not propagate the flag across phases).
 715                         if (node->needs_mipmaps != Effect::DOES_NOT_NEED_MIPMAPS) {
 716                                 // The node can have a value set (ie. not DOES_NOT_NEED_MIPMAPS)
 717                                 // if we have diamonds in the graph; if so, choose that.
 718                                 // If not, the effect on the node can also decide (this is the
 719                                 // more common case).
 720                                 Effect::MipmapRequirements dep_mipmaps = deps[i]->needs_mipmaps;
 721                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
 722                                         if (deps[i]->effect->num_inputs() == 0) {
 723                                                 Input *input = static_cast<Input *>(deps[i]->effect);
 724                                                 dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS;
 725                                         } else {
 726                                                 dep_mipmaps = deps[i]->effect->needs_mipmaps();
 727                                         }
 728                                 }
 729                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
 730                                         deps[i]->needs_mipmaps = node->needs_mipmaps;
 731                                 } else if (dep_mipmaps != node->needs_mipmaps) {
 732                                         // The dependency cannot supply our mipmap demands
 733                                         // (either because it's an input that can't do mipmaps,
 734                                         // or because there's a conflict between mipmap-needing
 735                                         // and mipmap-refusing effects somewhere in the graph),
 736                                         // so they cannot be in the same phase.
 737                                         start_new_phase = true;
 738                                 }
 739                         }
 740
 741                         if (deps[i]->outgoing_links.size() > 1) {
 742                                 if (!deps[i]->effect->is_single_texture()) {
 743                                         // More than one effect uses this as the input,
 744                                         // and it is not a texture itself.
 745                                         // The easiest thing to do (and probably also the safest
 746                                         // performance-wise in most cases) is to bounce it to a texture
 747                                         // and then let the next passes read from that.
 748                                         start_new_phase = true;
 749                                 } else {
 750                                         assert(deps[i]->effect->num_inputs() == 0);
 751
 752                                         // For textures, we try to be slightly more clever;
 753                                         // if none of our outputs need a bounce, we don't bounce
 754                                         // but instead simply use the effect many times.
 755                                         //
 756                                         // Strictly speaking, we could bounce it for some outputs
 757                                         // and use it directly for others, but the processing becomes
 758                                         // somewhat simpler if the effect is only used in one such way.
 759                                         for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
 760                                                 Node *rdep = deps[i]->outgoing_links[j];
 761                                                 start_new_phase |= rdep->effect->needs_texture_bounce();
 762                                         }
 763                                 }
 764                         }
 765
 766                         if (deps[i]->effect->is_compute_shader()) {
 767                                 if (phase->is_compute_shader) {
 768                                         // Only one compute shader per phase.
 769                                         start_new_phase = true;
 770                                 } else if (!node->strong_one_to_one_sampling) {
 771                                         // If all nodes so far are strong one-to-one, we can put them after
 772                                         // the compute shader (ie., process them on the output).
 773                                         start_new_phase = true;
 774                                 } else if (!start_new_phase) {
 775                                         phase->is_compute_shader = true;
 776                                         phase->compute_shader_node = deps[i];
 777                                 }
 778                         } else if (deps[i]->effect->sets_virtual_output_size()) {
 779                                 assert(deps[i]->effect->changes_output_size());
 780                                 // If the next effect sets a virtual size to rely on OpenGL's
 781                                 // bilinear sampling, we'll really need to break the phase here.
 782                                 start_new_phase = true;
 783                         } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
 784                                 // If the next effect changes size and we don't have one-to-one sampling,
 785                                 // we also need to break here.
 786                                 start_new_phase = true;
 787                         }
 788
 789                         if (start_new_phase) {
 790                                 // Since we're starting a new phase here, we don't need to impose any
 791                                 // new demands on this effect. Restore the status we had before we
 792                                 // started looking at it.
 793                                 deps[i]->needs_mipmaps = save_needs_mipmaps;
 794
 795                                 phase->inputs.push_back(construct_phase(deps[i], completed_effects));
 796                         } else {
 797                                 effects_todo_this_phase.push(deps[i]);
 798
 799                                 // Propagate the one-to-one status down through the dependency.
 800                                 deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
 801                                         deps[i]->effect->one_to_one_sampling();
 802                                 deps[i]->strong_one_to_one_sampling = node->strong_one_to_one_sampling &&
 803                                         deps[i]->effect->strong_one_to_one_sampling();
 804                         }
 805
 806                         node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE);
 807                 }
 808         }
 809
 810         // No more effects to do this phase. Take all the ones we have,
 811         // and create a GLSL program for it.
 812         assert(!phase->effects.empty());
 813
 814         // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
 815         // that would be nondeterministic and thus reduce cacheability.
 816         // TODO: Make this even more deterministic.
 817         vector<Phase *> dedup_inputs;
 818         set<Phase *> seen_inputs;
 819         for (size_t i = 0; i < phase->inputs.size(); ++i) {
 820                 if (seen_inputs.insert(phase->inputs[i]).second) {
 821                         dedup_inputs.push_back(phase->inputs[i]);
 822                 }
 823         }
 824         swap(phase->inputs, dedup_inputs);
 825
 826         // Allocate samplers for each input.
 827         phase->input_samplers.resize(phase->inputs.size());
 828
 829         // We added the effects from the output and back, but we need to output
 830         // them in topological sort order in the shader.
 831         phase->effects = topological_sort(phase->effects);
 832
 833         // Figure out if we need mipmaps or not, and if so, tell the inputs that.
 834         // (RTT inputs have different logic, which is checked in execute_phase().)
 835         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 836                 Node *node = phase->effects[i];
 837                 if (node->effect->num_inputs() == 0) {
 838                         Input *input = static_cast<Input *>(node->effect);
 839                         assert(node->needs_mipmaps != Effect::NEEDS_MIPMAPS || input->can_supply_mipmaps());
 840                         CHECK(input->set_int("needs_mipmaps", node->needs_mipmaps == Effect::NEEDS_MIPMAPS));
 841                 }
 842         }
 843
 844         // Tell each node which phase it ended up in, so that the unit test
 845         // can check that the phases were split in the right place.
 846         // Note that this ignores that effects may be part of multiple phases;
 847         // if the unit tests need to test such cases, we'll reconsider.
 848         for (unsigned i = 0; i < phase->effects.size(); ++i) {
 849                 phase->effects[i]->containing_phase = phase;
 850         }
 851
 852         // Actually make the shader for this phase.
 853         compile_glsl_program(phase);
 854
 855         // Initialize timers.
 856         if (movit_timer_queries_supported) {
 857                 phase->time_elapsed_ns = 0;
 858                 phase->num_measured_iterations = 0;
 859         }
 860
 861         assert(completed_effects->count(output) == 0);
 862         completed_effects->insert(make_pair(output, phase));
 863         phases.push_back(phase);
 864         return phase;
 865 }
 866
 867 void EffectChain::output_dot(const char *filename)
 868 {
 869         if (movit_debug_level != MOVIT_DEBUG_ON) {
 870                 return;
 871         }
 872
 873         FILE *fp = fopen(filename, "w");
 874         if (fp == nullptr) {
 875                 perror(filename);
 876                 exit(1);
 877         }
 878
 879         fprintf(fp, "digraph G {\n");
 880         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
 881         for (unsigned i = 0; i < nodes.size(); ++i) {
 882                 // Find out which phase this event belongs to.
 883                 vector<int> in_phases;
 884                 for (unsigned j = 0; j < phases.size(); ++j) {
 885                         const Phase* p = phases[j];
 886                         if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
 887                                 in_phases.push_back(j);
 888                         }
 889                 }
 890
 891                 if (in_phases.empty()) {
 892                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
 893                 } else if (in_phases.size() == 1) {
 894                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 895                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 896                                 (in_phases[0] % 8) + 1);
 897                 } else {
 898                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
 899                         // But alas.
 900                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
 901                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
 902                                 (in_phases[0] % 8) + 1);
 903                 }
 904
 905                 char from_node_id[256];
 906                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 907
 908                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
 909                         char to_node_id[256];
 910                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
 911
 912                         vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
 913                         output_dot_edge(fp, from_node_id, to_node_id, labels);
 914                 }
 915
 916                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
 917                         // Output node.
 918                         vector<string> labels = get_labels_for_edge(nodes[i], nullptr);
 919                         output_dot_edge(fp, from_node_id, "output", labels);
 920                 }
 921         }
 922         fprintf(fp, "}\n");
 923
 924         fclose(fp);
 925 }
 926
 927 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 928 {
 929         vector<string> labels;
 930
 931         if (to != nullptr && to->effect->needs_texture_bounce()) {
 932                 labels.push_back("needs_bounce");
 933         }
 934         if (from->effect->changes_output_size()) {
 935                 labels.push_back("resize");
 936         }
 937
 938         switch (from->output_color_space) {
 939         case COLORSPACE_INVALID:
 940                 labels.push_back("spc[invalid]");
 941                 break;
 942         case COLORSPACE_REC_601_525:
 943                 labels.push_back("spc[rec601-525]");
 944                 break;
 945         case COLORSPACE_REC_601_625:
 946                 labels.push_back("spc[rec601-625]");
 947                 break;
 948         default:
 949                 break;
 950         }
 951
 952         switch (from->output_gamma_curve) {
 953         case GAMMA_INVALID:
 954                 labels.push_back("gamma[invalid]");
 955                 break;
 956         case GAMMA_sRGB:
 957                 labels.push_back("gamma[sRGB]");
 958                 break;
 959         case GAMMA_REC_601:  // and GAMMA_REC_709
 960                 labels.push_back("gamma[rec601/709]");
 961                 break;
 962         default:
 963                 break;
 964         }
 965
 966         switch (from->output_alpha_type) {
 967         case ALPHA_INVALID:
 968                 labels.push_back("alpha[invalid]");
 969                 break;
 970         case ALPHA_BLANK:
 971                 labels.push_back("alpha[blank]");
 972                 break;
 973         case ALPHA_POSTMULTIPLIED:
 974                 labels.push_back("alpha[postmult]");
 975                 break;
 976         default:
 977                 break;
 978         }
 979
 980         return labels;
 981 }
 982
 983 void EffectChain::output_dot_edge(FILE *fp,
 984                                   const string &from_node_id,
 985                                   const string &to_node_id,
 986                                   const vector<string> &labels)
 987 {
 988         if (labels.empty()) {
 989                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
 990         } else {
 991                 string label = labels[0];
 992                 for (unsigned k = 1; k < labels.size(); ++k) {
 993                         label += ", " + labels[k];
 994                 }
 995                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
 996         }
 997 }
 998
 999 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
1000 {
1001         unsigned scaled_width, scaled_height;
1002
1003         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
1004                 // Same aspect, or W/H > aspect (image is wider than the frame).
1005                 // In either case, keep width, and adjust height.
1006                 scaled_width = width;
1007                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
1008         } else {
1009                 // W/H < aspect (image is taller than the frame), so keep height,
1010                 // and adjust width.
1011                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
1012                 scaled_height = height;
1013         }
1014
1015         // We should be consistently larger or smaller then the existing choice,
1016         // since we have the same aspect.
1017         assert(!(scaled_width < *output_width && scaled_height > *output_height));
1018         assert(!(scaled_height < *output_height && scaled_width > *output_width));
1019
1020         if (scaled_width >= *output_width && scaled_height >= *output_height) {
1021                 *output_width = scaled_width;
1022                 *output_height = scaled_height;
1023         }
1024 }
1025
1026 // Propagate input texture sizes throughout, and inform effects downstream.
1027 // (Like a lot of other code, we depend on effects being in topological order.)
1028 void EffectChain::inform_input_sizes(Phase *phase)
1029 {
1030         // All effects that have a defined size (inputs and RTT inputs)
1031         // get that. Reset all others.
1032         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1033                 Node *node = phase->effects[i];
1034                 if (node->effect->num_inputs() == 0) {
1035                         Input *input = static_cast<Input *>(node->effect);
1036                         node->output_width = input->get_width();
1037                         node->output_height = input->get_height();
1038                         assert(node->output_width != 0);
1039                         assert(node->output_height != 0);
1040                 } else {
1041                         node->output_width = node->output_height = 0;
1042                 }
1043         }
1044         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1045                 Phase *input = phase->inputs[i];
1046                 input->output_node->output_width = input->virtual_output_width;
1047                 input->output_node->output_height = input->virtual_output_height;
1048                 assert(input->output_node->output_width != 0);
1049                 assert(input->output_node->output_height != 0);
1050         }
1051
1052         // Now propagate from the inputs towards the end, and inform as we go.
1053         // The rules are simple:
1054         //
1055         //   1. Don't touch effects that already have given sizes (ie., inputs
1056         //      or effects that change the output size).
1057         //   2. If all of your inputs have the same size, that will be your output size.
1058         //   3. Otherwise, your output size is 0x0.
1059         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1060                 Node *node = phase->effects[i];
1061                 if (node->effect->num_inputs() == 0) {
1062                         continue;
1063                 }
1064                 unsigned this_output_width = 0;
1065                 unsigned this_output_height = 0;
1066                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1067                         Node *input = node->incoming_links[j];
1068                         node->effect->inform_input_size(j, input->output_width, input->output_height);
1069                         if (j == 0) {
1070                                 this_output_width = input->output_width;
1071                                 this_output_height = input->output_height;
1072                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
1073                                 // Inputs disagree.
1074                                 this_output_width = 0;
1075                                 this_output_height = 0;
1076                         }
1077                 }
1078                 if (node->effect->changes_output_size()) {
1079                         // We cannot call get_output_size() before we've done inform_input_size()
1080                         // on all inputs.
1081                         unsigned real_width, real_height;
1082                         node->effect->get_output_size(&real_width, &real_height,
1083                                                       &node->output_width, &node->output_height);
1084                         assert(node->effect->sets_virtual_output_size() ||
1085                                (real_width == node->output_width &&
1086                                 real_height == node->output_height));
1087                 } else {
1088                         node->output_width = this_output_width;
1089                         node->output_height = this_output_height;
1090                 }
1091         }
1092 }
1093
1094 // Note: You should call inform_input_sizes() before this, as the last effect's
1095 // desired output size might change based on the inputs.
1096 void EffectChain::find_output_size(Phase *phase)
1097 {
1098         Node *output_node = phase->is_compute_shader ? phase->compute_shader_node : phase->effects.back();
1099
1100         // If the last effect explicitly sets an output size, use that.
1101         if (output_node->effect->changes_output_size()) {
1102                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
1103                                                      &phase->virtual_output_width, &phase->virtual_output_height);
1104                 assert(output_node->effect->sets_virtual_output_size() ||
1105                        (phase->output_width == phase->virtual_output_width &&
1106                         phase->output_height == phase->virtual_output_height));
1107                 return;
1108         }
1109
1110         // If all effects have the same size, use that.
1111         unsigned output_width = 0, output_height = 0;
1112         bool all_inputs_same_size = true;
1113
1114         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1115                 Phase *input = phase->inputs[i];
1116                 assert(input->output_width != 0);
1117                 assert(input->output_height != 0);
1118                 if (output_width == 0 && output_height == 0) {
1119                         output_width = input->virtual_output_width;
1120                         output_height = input->virtual_output_height;
1121                 } else if (output_width != input->virtual_output_width ||
1122                            output_height != input->virtual_output_height) {
1123                         all_inputs_same_size = false;
1124                 }
1125         }
1126         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1127                 Effect *effect = phase->effects[i]->effect;
1128                 if (effect->num_inputs() != 0) {
1129                         continue;
1130                 }
1131
1132                 Input *input = static_cast<Input *>(effect);
1133                 if (output_width == 0 && output_height == 0) {
1134                         output_width = input->get_width();
1135                         output_height = input->get_height();
1136                 } else if (output_width != input->get_width() ||
1137                            output_height != input->get_height()) {
1138                         all_inputs_same_size = false;
1139                 }
1140         }
1141
1142         if (all_inputs_same_size) {
1143                 assert(output_width != 0);
1144                 assert(output_height != 0);
1145                 phase->virtual_output_width = phase->output_width = output_width;
1146                 phase->virtual_output_height = phase->output_height = output_height;
1147                 return;
1148         }
1149
1150         // If not, fit all the inputs into the current aspect, and select the largest one.
1151         output_width = 0;
1152         output_height = 0;
1153         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1154                 Phase *input = phase->inputs[i];
1155                 assert(input->output_width != 0);
1156                 assert(input->output_height != 0);
1157                 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
1158         }
1159         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1160                 Effect *effect = phase->effects[i]->effect;
1161                 if (effect->num_inputs() != 0) {
1162                         continue;
1163                 }
1164
1165                 Input *input = static_cast<Input *>(effect);
1166                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
1167         }
1168         assert(output_width != 0);
1169         assert(output_height != 0);
1170         phase->virtual_output_width = phase->output_width = output_width;
1171         phase->virtual_output_height = phase->output_height = output_height;
1172 }
1173
1174 void EffectChain::sort_all_nodes_topologically()
1175 {
1176         nodes = topological_sort(nodes);
1177 }
1178
1179 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
1180 {
1181         set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
1182         vector<Node *> sorted_list;
1183         for (unsigned i = 0; i < nodes.size(); ++i) {
1184                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
1185         }
1186         reverse(sorted_list.begin(), sorted_list.end());
1187         return sorted_list;
1188 }
1189
1190 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
1191 {
1192         if (nodes_left_to_visit->count(node) == 0) {
1193                 return;
1194         }
1195         nodes_left_to_visit->erase(node);
1196         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
1197                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
1198         }
1199         sorted_list->push_back(node);
1200 }
1201
1202 void EffectChain::find_color_spaces_for_inputs()
1203 {
1204         for (unsigned i = 0; i < nodes.size(); ++i) {
1205                 Node *node = nodes[i];
1206                 if (node->disabled) {
1207                         continue;
1208                 }
1209                 if (node->incoming_links.size() == 0) {
1210                         Input *input = static_cast<Input *>(node->effect);
1211                         node->output_color_space = input->get_color_space();
1212                         node->output_gamma_curve = input->get_gamma_curve();
1213
1214                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
1215                         switch (alpha_handling) {
1216                         case Effect::OUTPUT_BLANK_ALPHA:
1217                                 node->output_alpha_type = ALPHA_BLANK;
1218                                 break;
1219                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1220                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1221                                 break;
1222                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1223                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1224                                 break;
1225                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1226                         case Effect::DONT_CARE_ALPHA_TYPE:
1227                         default:
1228                                 assert(false);
1229                         }
1230
1231                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1232                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
1233                         }
1234                 }
1235         }
1236 }
1237
1238 // Propagate gamma and color space information as far as we can in the graph.
1239 // The rules are simple: Anything where all the inputs agree, get that as
1240 // output as well. Anything else keeps having *_INVALID.
1241 void EffectChain::propagate_gamma_and_color_space()
1242 {
1243         // We depend on going through the nodes in order.
1244         sort_all_nodes_topologically();
1245
1246         for (unsigned i = 0; i < nodes.size(); ++i) {
1247                 Node *node = nodes[i];
1248                 if (node->disabled) {
1249                         continue;
1250                 }
1251                 assert(node->incoming_links.size() == node->effect->num_inputs());
1252                 if (node->incoming_links.size() == 0) {
1253                         assert(node->output_color_space != COLORSPACE_INVALID);
1254                         assert(node->output_gamma_curve != GAMMA_INVALID);
1255                         continue;
1256                 }
1257
1258                 Colorspace color_space = node->incoming_links[0]->output_color_space;
1259                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1260                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1261                         if (node->incoming_links[j]->output_color_space != color_space) {
1262                                 color_space = COLORSPACE_INVALID;
1263                         }
1264                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1265                                 gamma_curve = GAMMA_INVALID;
1266                         }
1267                 }
1268
1269                 // The conversion effects already have their outputs set correctly,
1270                 // so leave them alone.
1271                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1272                         node->output_color_space = color_space;
1273                 }
1274                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1275                     node->effect->effect_type_id() != "GammaExpansionEffect") {
1276                         node->output_gamma_curve = gamma_curve;
1277                 }
1278         }
1279 }
1280
1281 // Propagate alpha information as far as we can in the graph.
1282 // Similar to propagate_gamma_and_color_space().
1283 void EffectChain::propagate_alpha()
1284 {
1285         // We depend on going through the nodes in order.
1286         sort_all_nodes_topologically();
1287
1288         for (unsigned i = 0; i < nodes.size(); ++i) {
1289                 Node *node = nodes[i];
1290                 if (node->disabled) {
1291                         continue;
1292                 }
1293                 assert(node->incoming_links.size() == node->effect->num_inputs());
1294                 if (node->incoming_links.size() == 0) {
1295                         assert(node->output_alpha_type != ALPHA_INVALID);
1296                         continue;
1297                 }
1298
1299                 // The alpha multiplication/division effects are special cases.
1300                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1301                         assert(node->incoming_links.size() == 1);
1302                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1303                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
1304                         continue;
1305                 }
1306                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1307                         assert(node->incoming_links.size() == 1);
1308                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1309                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1310                         continue;
1311                 }
1312
1313                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1314                 // because they are the only one that _need_ postmultiplied alpha.
1315                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1316                     node->effect->effect_type_id() == "GammaExpansionEffect") {
1317                         assert(node->incoming_links.size() == 1);
1318                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1319                                 node->output_alpha_type = ALPHA_BLANK;
1320                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1321                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1322                         } else {
1323                                 node->output_alpha_type = ALPHA_INVALID;
1324                         }
1325                         continue;
1326                 }
1327
1328                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1329                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1330                 // taken care of above. Rationale: Even if you could imagine
1331                 // e.g. an effect that took in an image and set alpha=1.0
1332                 // unconditionally, it wouldn't make any sense to have it as
1333                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1334                 // got its input pre- or postmultiplied, so it wouldn't know
1335                 // whether to divide away the old alpha or not.
1336                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1337                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1338                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1339                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1340
1341                 // If the node has multiple inputs, check that they are all valid and
1342                 // the same.
1343                 bool any_invalid = false;
1344                 bool any_premultiplied = false;
1345                 bool any_postmultiplied = false;
1346
1347                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1348                         switch (node->incoming_links[j]->output_alpha_type) {
1349                         case ALPHA_INVALID:
1350                                 any_invalid = true;
1351                                 break;
1352                         case ALPHA_BLANK:
1353                                 // Blank is good as both pre- and postmultiplied alpha,
1354                                 // so just ignore it.
1355                                 break;
1356                         case ALPHA_PREMULTIPLIED:
1357                                 any_premultiplied = true;
1358                                 break;
1359                         case ALPHA_POSTMULTIPLIED:
1360                                 any_postmultiplied = true;
1361                                 break;
1362                         default:
1363                                 assert(false);
1364                         }
1365                 }
1366
1367                 if (any_invalid) {
1368                         node->output_alpha_type = ALPHA_INVALID;
1369                         continue;
1370                 }
1371
1372                 // Inputs must be of the same type.
1373                 if (any_premultiplied && any_postmultiplied) {
1374                         node->output_alpha_type = ALPHA_INVALID;
1375                         continue;
1376                 }
1377
1378                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1379                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1380                         // This combination (requiring premultiplied alpha, but _not_ requiring
1381                         // linear light) is illegal, since the combination of premultiplied alpha
1382                         // and nonlinear inputs is meaningless.
1383                         assert(node->effect->needs_linear_light());
1384
1385                         // If the effect has asked for premultiplied alpha, check that it has got it.
1386                         if (any_postmultiplied) {
1387                                 node->output_alpha_type = ALPHA_INVALID;
1388                         } else if (!any_premultiplied &&
1389                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1390                                 // Blank input alpha, and the effect preserves blank alpha.
1391                                 node->output_alpha_type = ALPHA_BLANK;
1392                         } else {
1393                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1394                         }
1395                 } else {
1396                         // OK, all inputs are the same, and this effect is not going
1397                         // to change it.
1398                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1399                         if (any_premultiplied) {
1400                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1401                         } else if (any_postmultiplied) {
1402                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1403                         } else {
1404                                 node->output_alpha_type = ALPHA_BLANK;
1405                         }
1406                 }
1407         }
1408 }
1409
1410 bool EffectChain::node_needs_colorspace_fix(Node *node)
1411 {
1412         if (node->disabled) {
1413                 return false;
1414         }
1415         if (node->effect->num_inputs() == 0) {
1416                 return false;
1417         }
1418
1419         // propagate_gamma_and_color_space() has already set our output
1420         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1421         if (node->output_color_space == COLORSPACE_INVALID) {
1422                 return true;
1423         }
1424         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1425 }
1426
1427 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1428 // the graph. Our strategy is not always optimal, but quite simple:
1429 // Find an effect that's as early as possible where the inputs are of
1430 // unacceptable colorspaces (that is, either different, or, if the effect only
1431 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1432 // propagate the information anew, and repeat until there are no more such
1433 // effects.
1434 void EffectChain::fix_internal_color_spaces()
1435 {
1436         unsigned colorspace_propagation_pass = 0;
1437         bool found_any;
1438         do {
1439                 found_any = false;
1440                 for (unsigned i = 0; i < nodes.size(); ++i) {
1441                         Node *node = nodes[i];
1442                         if (!node_needs_colorspace_fix(node)) {
1443                                 continue;
1444                         }
1445
1446                         // Go through each input that is not sRGB, and insert
1447                         // a colorspace conversion after it.
1448                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1449                                 Node *input = node->incoming_links[j];
1450                                 assert(input->output_color_space != COLORSPACE_INVALID);
1451                                 if (input->output_color_space == COLORSPACE_sRGB) {
1452                                         continue;
1453                                 }
1454                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1455                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1456                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1457                                 conversion->output_color_space = COLORSPACE_sRGB;
1458                                 replace_sender(input, conversion);
1459                                 connect_nodes(input, conversion);
1460                         }
1461
1462                         // Re-sort topologically, and propagate the new information.
1463                         propagate_gamma_and_color_space();
1464
1465                         found_any = true;
1466                         break;
1467                 }
1468
1469                 char filename[256];
1470                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1471                 output_dot(filename);
1472                 assert(colorspace_propagation_pass < 100);
1473         } while (found_any);
1474
1475         for (unsigned i = 0; i < nodes.size(); ++i) {
1476                 Node *node = nodes[i];
1477                 if (node->disabled) {
1478                         continue;
1479                 }
1480                 assert(node->output_color_space != COLORSPACE_INVALID);
1481         }
1482 }
1483
1484 bool EffectChain::node_needs_alpha_fix(Node *node)
1485 {
1486         if (node->disabled) {
1487                 return false;
1488         }
1489
1490         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1491         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1492         return (node->output_alpha_type == ALPHA_INVALID);
1493 }
1494
1495 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1496 // the graph. Similar to fix_internal_color_spaces().
1497 void EffectChain::fix_internal_alpha(unsigned step)
1498 {
1499         unsigned alpha_propagation_pass = 0;
1500         bool found_any;
1501         do {
1502                 found_any = false;
1503                 for (unsigned i = 0; i < nodes.size(); ++i) {
1504                         Node *node = nodes[i];
1505                         if (!node_needs_alpha_fix(node)) {
1506                                 continue;
1507                         }
1508
1509                         // If we need to fix up GammaExpansionEffect, then clearly something
1510                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1511                         // is meaningless.
1512                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1513
1514                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1515
1516                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1517                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1518                                 assert(node->incoming_links.size() == 1);
1519                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1520                                 desired_type = ALPHA_POSTMULTIPLIED;
1521                         }
1522
1523                         // Go through each input that is not premultiplied alpha, and insert
1524                         // a conversion before it.
1525                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1526                                 Node *input = node->incoming_links[j];
1527                                 assert(input->output_alpha_type != ALPHA_INVALID);
1528                                 if (input->output_alpha_type == desired_type ||
1529                                     input->output_alpha_type == ALPHA_BLANK) {
1530                                         continue;
1531                                 }
1532                                 Node *conversion;
1533                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1534                                         conversion = add_node(new AlphaMultiplicationEffect());
1535                                 } else {
1536                                         conversion = add_node(new AlphaDivisionEffect());
1537                                 }
1538                                 conversion->output_alpha_type = desired_type;
1539                                 replace_sender(input, conversion);
1540                                 connect_nodes(input, conversion);
1541                         }
1542
1543                         // Re-sort topologically, and propagate the new information.
1544                         propagate_gamma_and_color_space();
1545                         propagate_alpha();
1546
1547                         found_any = true;
1548                         break;
1549                 }
1550
1551                 char filename[256];
1552                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1553                 output_dot(filename);
1554                 assert(alpha_propagation_pass < 100);
1555         } while (found_any);
1556
1557         for (unsigned i = 0; i < nodes.size(); ++i) {
1558                 Node *node = nodes[i];
1559                 if (node->disabled) {
1560                         continue;
1561                 }
1562                 assert(node->output_alpha_type != ALPHA_INVALID);
1563         }
1564 }
1565
1566 // Make so that the output is in the desired color space.
1567 void EffectChain::fix_output_color_space()
1568 {
1569         Node *output = find_output_node();
1570         if (output->output_color_space != output_format.color_space) {
1571                 Node *conversion = add_node(new ColorspaceConversionEffect());
1572                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1573                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1574                 conversion->output_color_space = output_format.color_space;
1575                 connect_nodes(output, conversion);
1576                 propagate_alpha();
1577                 propagate_gamma_and_color_space();
1578         }
1579 }
1580
1581 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1582 void EffectChain::fix_output_alpha()
1583 {
1584         Node *output = find_output_node();
1585         assert(output->output_alpha_type != ALPHA_INVALID);
1586         if (output->output_alpha_type == ALPHA_BLANK) {
1587                 // No alpha output, so we don't care.
1588                 return;
1589         }
1590         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1591             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1592                 Node *conversion = add_node(new AlphaDivisionEffect());
1593                 connect_nodes(output, conversion);
1594                 propagate_alpha();
1595                 propagate_gamma_and_color_space();
1596         }
1597         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1598             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1599                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1600                 connect_nodes(output, conversion);
1601                 propagate_alpha();
1602                 propagate_gamma_and_color_space();
1603         }
1604 }
1605
1606 bool EffectChain::node_needs_gamma_fix(Node *node)
1607 {
1608         if (node->disabled) {
1609                 return false;
1610         }
1611
1612         // Small hack since the output is not an explicit node:
1613         // If we are the last node and our output is in the wrong
1614         // space compared to EffectChain's output, we need to fix it.
1615         // This will only take us to linear, but fix_output_gamma()
1616         // will come and take us to the desired output gamma
1617         // if it is needed.
1618         //
1619         // This needs to be before everything else, since it could
1620         // even apply to inputs (if they are the only effect).
1621         if (node->outgoing_links.empty() &&
1622             node->output_gamma_curve != output_format.gamma_curve &&
1623             node->output_gamma_curve != GAMMA_LINEAR) {
1624                 return true;
1625         }
1626
1627         if (node->effect->num_inputs() == 0) {
1628                 return false;
1629         }
1630
1631         // propagate_gamma_and_color_space() has already set our output
1632         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1633         // except for GammaCompressionEffect.
1634         if (node->output_gamma_curve == GAMMA_INVALID) {
1635                 return true;
1636         }
1637         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1638                 assert(node->incoming_links.size() == 1);
1639                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1640         }
1641
1642         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1643 }
1644
1645 // Very similar to fix_internal_color_spaces(), but for gamma.
1646 // There is one difference, though; before we start adding conversion nodes,
1647 // we see if we can get anything out of asking the sources to deliver
1648 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1649 // does that part, while fix_internal_gamma_by_inserting_nodes()
1650 // inserts nodes as needed afterwards.
1651 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1652 {
1653         unsigned gamma_propagation_pass = 0;
1654         bool found_any;
1655         do {
1656                 found_any = false;
1657                 for (unsigned i = 0; i < nodes.size(); ++i) {
1658                         Node *node = nodes[i];
1659                         if (!node_needs_gamma_fix(node)) {
1660                                 continue;
1661                         }
1662
1663                         // See if all inputs can give us linear gamma. If not, leave it.
1664                         vector<Node *> nonlinear_inputs;
1665                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1666                         assert(!nonlinear_inputs.empty());
1667
1668                         bool all_ok = true;
1669                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1670                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1671                                 all_ok &= input->can_output_linear_gamma();
1672                         }
1673
1674                         if (!all_ok) {
1675                                 continue;
1676                         }
1677
1678                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1679                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1680                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1681                         }
1682
1683                         // Re-sort topologically, and propagate the new information.
1684                         propagate_gamma_and_color_space();
1685
1686                         found_any = true;
1687                         break;
1688                 }
1689
1690                 char filename[256];
1691                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1692                 output_dot(filename);
1693                 assert(gamma_propagation_pass < 100);
1694         } while (found_any);
1695 }
1696
1697 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1698 {
1699         unsigned gamma_propagation_pass = 0;
1700         bool found_any;
1701         do {
1702                 found_any = false;
1703                 for (unsigned i = 0; i < nodes.size(); ++i) {
1704                         Node *node = nodes[i];
1705                         if (!node_needs_gamma_fix(node)) {
1706                                 continue;
1707                         }
1708
1709                         // Special case: We could be an input and still be asked to
1710                         // fix our gamma; if so, we should be the only node
1711                         // (as node_needs_gamma_fix() would only return true in
1712                         // for an input in that case). That means we should insert
1713                         // a conversion node _after_ ourselves.
1714                         if (node->incoming_links.empty()) {
1715                                 assert(node->outgoing_links.empty());
1716                                 Node *conversion = add_node(new GammaExpansionEffect());
1717                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1718                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1719                                 connect_nodes(node, conversion);
1720                         }
1721
1722                         // If not, go through each input that is not linear gamma,
1723                         // and insert a gamma conversion after it.
1724                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1725                                 Node *input = node->incoming_links[j];
1726                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1727                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1728                                         continue;
1729                                 }
1730                                 Node *conversion = add_node(new GammaExpansionEffect());
1731                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1732                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1733                                 replace_sender(input, conversion);
1734                                 connect_nodes(input, conversion);
1735                         }
1736
1737                         // Re-sort topologically, and propagate the new information.
1738                         propagate_alpha();
1739                         propagate_gamma_and_color_space();
1740
1741                         found_any = true;
1742                         break;
1743                 }
1744
1745                 char filename[256];
1746                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1747                 output_dot(filename);
1748                 assert(gamma_propagation_pass < 100);
1749         } while (found_any);
1750
1751         for (unsigned i = 0; i < nodes.size(); ++i) {
1752                 Node *node = nodes[i];
1753                 if (node->disabled) {
1754                         continue;
1755                 }
1756                 assert(node->output_gamma_curve != GAMMA_INVALID);
1757         }
1758 }
1759
1760 // Make so that the output is in the desired gamma.
1761 // Note that this assumes linear input gamma, so it might create the need
1762 // for another pass of fix_internal_gamma().
1763 void EffectChain::fix_output_gamma()
1764 {
1765         Node *output = find_output_node();
1766         if (output->output_gamma_curve != output_format.gamma_curve) {
1767                 Node *conversion = add_node(new GammaCompressionEffect());
1768                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1769                 conversion->output_gamma_curve = output_format.gamma_curve;
1770                 connect_nodes(output, conversion);
1771         }
1772 }
1773
1774 // If the user has requested Y'CbCr output, we need to do this conversion
1775 // _after_ GammaCompressionEffect etc., but before dither (see below).
1776 // This is because Y'CbCr, with the exception of a special optional mode
1777 // in Rec. 2020 (which we currently don't support), is defined to work on
1778 // gamma-encoded data.
1779 void EffectChain::add_ycbcr_conversion_if_needed()
1780 {
1781         assert(output_color_rgba || num_output_color_ycbcr > 0);
1782         if (num_output_color_ycbcr == 0) {
1783                 return;
1784         }
1785         Node *output = find_output_node();
1786         ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type));
1787         connect_nodes(output, ycbcr_conversion_effect_node);
1788 }
1789
1790 // If the user has requested dither, add a DitherEffect right at the end
1791 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1792 // since dither is about the only effect that can _not_ be done in linear space.
1793 void EffectChain::add_dither_if_needed()
1794 {
1795         if (num_dither_bits == 0) {
1796                 return;
1797         }
1798         Node *output = find_output_node();
1799         Node *dither = add_node(new DitherEffect());
1800         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1801         connect_nodes(output, dither);
1802
1803         dither_effect = dither->effect;
1804 }
1805
1806 namespace {
1807
1808 // Whether this effect will cause the phase it is in to become a compute shader phase.
1809 bool induces_compute_shader(Node *node)
1810 {
1811         if (node->effect->is_compute_shader()) {
1812                 return true;
1813         }
1814         if (!node->effect->strong_one_to_one_sampling()) {
1815                 // This effect can't be chained after a compute shader.
1816                 return false;
1817         }
1818         // If at least one of the effects we depend on is a compute shader,
1819         // one of them will be put in the same phase as us (the other ones,
1820         // if any, will be bounced).
1821         for (Node *dep : node->incoming_links) {
1822                 if (induces_compute_shader(dep)) {
1823                         return true;
1824                 }
1825         }
1826         return false;
1827 }
1828
1829 }  // namespace
1830
1831 // Compute shaders can't output to the framebuffer, so if the last
1832 // phase ends in a compute shader, add a dummy phase at the end that
1833 // only blits directly from the temporary texture.
1834 void EffectChain::add_dummy_effect_if_needed()
1835 {
1836         Node *output = find_output_node();
1837         if (induces_compute_shader(output)) {
1838                 Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
1839                 connect_nodes(output, dummy);
1840                 has_dummy_effect = true;
1841         }
1842 }
1843
1844 // Find the output node. This is, simply, one that has no outgoing links.
1845 // If there are multiple ones, the graph is malformed (we do not support
1846 // multiple outputs right now).
1847 Node *EffectChain::find_output_node()
1848 {
1849         vector<Node *> output_nodes;
1850         for (unsigned i = 0; i < nodes.size(); ++i) {
1851                 Node *node = nodes[i];
1852                 if (node->disabled) {
1853                         continue;
1854                 }
1855                 if (node->outgoing_links.empty()) {
1856                         output_nodes.push_back(node);
1857                 }
1858         }
1859         assert(output_nodes.size() == 1);
1860         return output_nodes[0];
1861 }
1862
1863 void EffectChain::finalize()
1864 {
1865         // Output the graph as it is before we do any conversions on it.
1866         output_dot("step0-start.dot");
1867
1868         // Give each effect in turn a chance to rewrite its own part of the graph.
1869         // Note that if more effects are added as part of this, they will be
1870         // picked up as part of the same for loop, since they are added at the end.
1871         for (unsigned i = 0; i < nodes.size(); ++i) {
1872                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1873         }
1874         output_dot("step1-rewritten.dot");
1875
1876         find_color_spaces_for_inputs();
1877         output_dot("step2-input-colorspace.dot");
1878
1879         propagate_alpha();
1880         output_dot("step3-propagated-alpha.dot");
1881
1882         propagate_gamma_and_color_space();
1883         output_dot("step4-propagated-all.dot");
1884
1885         fix_internal_color_spaces();
1886         fix_internal_alpha(6);
1887         fix_output_color_space();
1888         output_dot("step7-output-colorspacefix.dot");
1889         fix_output_alpha();
1890         output_dot("step8-output-alphafix.dot");
1891
1892         // Note that we need to fix gamma after colorspace conversion,
1893         // because colorspace conversions might create needs for gamma conversions.
1894         // Also, we need to run an extra pass of fix_internal_gamma() after
1895         // fixing the output gamma, as we only have conversions to/from linear,
1896         // and fix_internal_alpha() since GammaCompressionEffect needs
1897         // postmultiplied input.
1898         fix_internal_gamma_by_asking_inputs(9);
1899         fix_internal_gamma_by_inserting_nodes(10);
1900         fix_output_gamma();
1901         output_dot("step11-output-gammafix.dot");
1902         propagate_alpha();
1903         output_dot("step12-output-alpha-propagated.dot");
1904         fix_internal_alpha(13);
1905         output_dot("step14-output-alpha-fixed.dot");
1906         fix_internal_gamma_by_asking_inputs(15);
1907         fix_internal_gamma_by_inserting_nodes(16);
1908
1909         output_dot("step17-before-ycbcr.dot");
1910         add_ycbcr_conversion_if_needed();
1911
1912         output_dot("step18-before-dither.dot");
1913         add_dither_if_needed();
1914
1915         output_dot("step19-before-dummy-effect.dot");
1916         add_dummy_effect_if_needed();
1917
1918         output_dot("step20-final.dot");
1919
1920         // Construct all needed GLSL programs, starting at the output.
1921         // We need to keep track of which effects have already been computed,
1922         // as an effect with multiple users could otherwise be calculated
1923         // multiple times.
1924         map<Node *, Phase *> completed_effects;
1925         construct_phase(find_output_node(), &completed_effects);
1926
1927         output_dot("step21-split-to-phases.dot");
1928
1929         // There are some corner cases where we thought we needed to add a dummy
1930         // effect, but then it turned out later we didn't (e.g. induces_compute_shader()
1931         // didn't see a mipmap conflict coming, which would cause the compute shader
1932         // to be split off from the inal phase); if so, remove the extra phase
1933         // at the end, since it will give us some trouble during execution.
1934         //
1935         // TODO: Remove induces_compute_shader() and replace it with precise tracking.
1936         if (has_dummy_effect && !phases[phases.size() - 2]->is_compute_shader) {
1937                 resource_pool->release_glsl_program(phases.back()->glsl_program_num);
1938                 delete phases.back();
1939                 phases.pop_back();
1940                 has_dummy_effect = false;
1941         }
1942
1943         output_dot("step22-dummy-phase-removal.dot");
1944
1945         assert(phases[0]->inputs.empty());
1946
1947         finalized = true;
1948 }
1949
1950 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1951 {
1952         // Save original viewport.
1953         GLuint x = 0, y = 0;
1954
1955         if (width == 0 && height == 0) {
1956                 GLint viewport[4];
1957                 glGetIntegerv(GL_VIEWPORT, viewport);
1958                 x = viewport[0];
1959                 y = viewport[1];
1960                 width = viewport[2];
1961                 height = viewport[3];
1962         }
1963
1964         render(dest_fbo, {}, x, y, width, height);
1965 }
1966
1967 void EffectChain::render_to_texture(const vector<DestinationTexture> &destinations, unsigned width, unsigned height)
1968 {
1969         assert(finalized);
1970         assert(!destinations.empty());
1971
1972         if (!has_dummy_effect) {
1973                 // We don't end in a compute shader, so there's nothing specific for us to do.
1974                 // Create an FBO for this set of textures, and just render to that.
1975                 GLuint texnums[4] = { 0, 0, 0, 0 };
1976                 for (unsigned i = 0; i < destinations.size() && i < 4; ++i) {
1977                         texnums[i] = destinations[i].texnum;
1978                 }
1979                 GLuint dest_fbo = resource_pool->create_fbo(texnums[0], texnums[1], texnums[2], texnums[3]);
1980                 render(dest_fbo, {}, 0, 0, width, height);
1981                 resource_pool->release_fbo(dest_fbo);
1982         } else {
1983                 render((GLuint)-1, destinations, 0, 0, width, height);
1984         }
1985 }
1986
1987 void EffectChain::render(GLuint dest_fbo, const vector<DestinationTexture> &destinations, unsigned x, unsigned y, unsigned width, unsigned height)
1988 {
1989         assert(finalized);
1990         assert(destinations.size() <= 1);
1991
1992         // This needs to be set anew, in case we are coming from a different context
1993         // from when we initialized.
1994         check_error();
1995         glDisable(GL_DITHER);
1996         check_error();
1997
1998         const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB);
1999         check_error();
2000         bool current_srgb = final_srgb;
2001
2002         // Basic state.
2003         check_error();
2004         glDisable(GL_BLEND);
2005         check_error();
2006         glDisable(GL_DEPTH_TEST);
2007         check_error();
2008         glDepthMask(GL_FALSE);
2009         check_error();
2010
2011         set<Phase *> generated_mipmaps;
2012
2013         // We keep one texture per output, but only for as long as we actually have any
2014         // phases that need it as an input. (We don't make any effort to reorder phases
2015         // to minimize the number of textures in play, as register allocation can be
2016         // complicated and we rarely have much to gain, since our graphs are typically
2017         // pretty linear.)
2018         map<Phase *, GLuint> output_textures;
2019         map<Phase *, int> ref_counts;
2020         for (Phase *phase : phases) {
2021                 for (Phase *input : phase->inputs) {
2022                         ++ref_counts[input];
2023                 }
2024         }
2025
2026         size_t num_phases = phases.size();
2027         if (destinations.empty()) {
2028                 assert(dest_fbo != (GLuint)-1);
2029         } else {
2030                 assert(has_dummy_effect);
2031                 assert(x == 0);
2032                 assert(y == 0);
2033                 assert(num_phases >= 2);
2034                 assert(!phases.back()->is_compute_shader);
2035                 assert(phases[phases.size() - 2]->is_compute_shader);
2036                 assert(phases.back()->effects.size() == 1);
2037                 assert(phases.back()->effects[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
2038
2039                 // We are rendering to a set of textures, so we can run the compute shader
2040                 // directly and skip the dummy phase.
2041                 --num_phases;
2042         }
2043
2044         for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) {
2045                 Phase *phase = phases[phase_num];
2046
2047                 if (do_phase_timing) {
2048                         GLuint timer_query_object;
2049                         if (phase->timer_query_objects_free.empty()) {
2050                                 glGenQueries(1, &timer_query_object);
2051                         } else {
2052                                 timer_query_object = phase->timer_query_objects_free.front();
2053                                 phase->timer_query_objects_free.pop_front();
2054                         }
2055                         glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
2056                         phase->timer_query_objects_running.push_back(timer_query_object);
2057                 }
2058                 bool last_phase = (phase_num == num_phases - 1);
2059                 if (last_phase) {
2060                         // Last phase goes to the output the user specified.
2061                         if (!phase->is_compute_shader) {
2062                                 assert(dest_fbo != (GLuint)-1);
2063                                 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
2064                                 check_error();
2065                                 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
2066                                 assert(status == GL_FRAMEBUFFER_COMPLETE);
2067                                 glViewport(x, y, width, height);
2068                         }
2069                         if (dither_effect != nullptr) {
2070                                 CHECK(dither_effect->set_int("output_width", width));
2071                                 CHECK(dither_effect->set_int("output_height", height));
2072                         }
2073                 }
2074
2075                 // Enable sRGB rendering for intermediates in case we are
2076                 // rendering to an sRGB format.
2077                 // TODO: Support this for compute shaders.
2078                 bool needs_srgb = last_phase ? final_srgb : true;
2079                 if (needs_srgb && !current_srgb) {
2080                         glEnable(GL_FRAMEBUFFER_SRGB);
2081                         check_error();
2082                         current_srgb = true;
2083                 } else if (!needs_srgb && current_srgb) {
2084                         glDisable(GL_FRAMEBUFFER_SRGB);
2085                         check_error();
2086                         current_srgb = true;
2087                 }
2088
2089                 // Find a texture for this phase.
2090                 inform_input_sizes(phase);
2091                 find_output_size(phase);
2092                 vector<DestinationTexture> phase_destinations;
2093                 if (!last_phase) {
2094                         GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
2095                         output_textures.insert(make_pair(phase, tex_num));
2096                         phase_destinations.push_back(DestinationTexture{ tex_num, intermediate_format });
2097
2098                         // The output texture needs to have valid state to be written to by a compute shader.
2099                         glActiveTexture(GL_TEXTURE0);
2100                         check_error();
2101                         glBindTexture(GL_TEXTURE_2D, tex_num);
2102                         check_error();
2103                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2104                         check_error();
2105                 } else if (phase->is_compute_shader) {
2106                         assert(!destinations.empty());
2107                         phase_destinations = destinations;
2108                 }
2109
2110                 execute_phase(phase, output_textures, phase_destinations, &generated_mipmaps);
2111                 if (do_phase_timing) {
2112                         glEndQuery(GL_TIME_ELAPSED);
2113                 }
2114
2115                 // Drop any input textures we don't need anymore.
2116                 for (Phase *input : phase->inputs) {
2117                         assert(ref_counts[input] > 0);
2118                         if (--ref_counts[input] == 0) {
2119                                 resource_pool->release_2d_texture(output_textures[input]);
2120                                 output_textures.erase(input);
2121                         }
2122                 }
2123         }
2124
2125         for (const auto &phase_and_texnum : output_textures) {
2126                 resource_pool->release_2d_texture(phase_and_texnum.second);
2127         }
2128
2129         glBindFramebuffer(GL_FRAMEBUFFER, 0);
2130         check_error();
2131         glUseProgram(0);
2132         check_error();
2133
2134         glBindBuffer(GL_ARRAY_BUFFER, 0);
2135         check_error();
2136         glBindVertexArray(0);
2137         check_error();
2138
2139         if (do_phase_timing) {
2140                 // Get back the timer queries.
2141                 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2142                         Phase *phase = phases[phase_num];
2143                         for (auto timer_it = phase->timer_query_objects_running.cbegin();
2144                              timer_it != phase->timer_query_objects_running.cend(); ) {
2145                                 GLint timer_query_object = *timer_it;
2146                                 GLint available;
2147                                 glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
2148                                 if (available) {
2149                                         GLuint64 time_elapsed;
2150                                         glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
2151                                         phase->time_elapsed_ns += time_elapsed;
2152                                         ++phase->num_measured_iterations;
2153                                         phase->timer_query_objects_free.push_back(timer_query_object);
2154                                         phase->timer_query_objects_running.erase(timer_it++);
2155                                 } else {
2156                                         ++timer_it;
2157                                 }
2158                         }
2159                 }
2160         }
2161 }
2162
2163 void EffectChain::enable_phase_timing(bool enable)
2164 {
2165         if (enable) {
2166                 assert(movit_timer_queries_supported);
2167         }
2168         this->do_phase_timing = enable;
2169 }
2170
2171 void EffectChain::reset_phase_timing()
2172 {
2173         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2174                 Phase *phase = phases[phase_num];
2175                 phase->time_elapsed_ns = 0;
2176                 phase->num_measured_iterations = 0;
2177         }
2178 }
2179
2180 void EffectChain::print_phase_timing()
2181 {
2182         double total_time_ms = 0.0;
2183         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2184                 Phase *phase = phases[phase_num];
2185                 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
2186                 printf("Phase %d: %5.1f ms  [", phase_num, avg_time_ms);
2187                 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
2188                         if (effect_num != 0) {
2189                                 printf(", ");
2190                         }
2191                         printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
2192                 }
2193                 printf("]\n");
2194                 total_time_ms += avg_time_ms;
2195         }
2196         printf("Total:   %5.1f ms\n", total_time_ms);
2197 }
2198
2199 void EffectChain::execute_phase(Phase *phase,
2200                                 const map<Phase *, GLuint> &output_textures,
2201                                 const vector<DestinationTexture> &destinations,
2202                                 set<Phase *> *generated_mipmaps)
2203 {
2204         // Set up RTT inputs for this phase.
2205         for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
2206                 glActiveTexture(GL_TEXTURE0 + sampler);
2207                 Phase *input = phase->inputs[sampler];
2208                 input->output_node->bound_sampler_num = sampler;
2209                 const auto it = output_textures.find(input);
2210                 assert(it != output_textures.end());
2211                 glBindTexture(GL_TEXTURE_2D, it->second);
2212                 check_error();
2213
2214                 // See if anything using this RTT input (in this phase) needs mipmaps.
2215                 // TODO: It could be that we get conflicting logic here, if we have
2216                 // multiple effects with incompatible mipmaps using the same
2217                 // RTT input. However, that is obscure enough that we can deal
2218                 // with it at some future point (preferably when we have
2219                 // universal support for separate sampler objects!). For now,
2220                 // an assert is good enough. See also the TODO at bound_sampler_num.
2221                 bool any_needs_mipmaps = false, any_refuses_mipmaps = false;
2222                 for (Node *node : phase->effects) {
2223                         assert(node->incoming_links.size() == node->incoming_link_type.size());
2224                         for (size_t i = 0; i < node->incoming_links.size(); ++i) {
2225                                 if (node->incoming_links[i] == input->output_node &&
2226                                     node->incoming_link_type[i] == IN_ANOTHER_PHASE) {
2227                                         if (node->needs_mipmaps == Effect::NEEDS_MIPMAPS) {
2228                                                 any_needs_mipmaps = true;
2229                                         } else if (node->needs_mipmaps == Effect::CANNOT_ACCEPT_MIPMAPS) {
2230                                                 any_refuses_mipmaps = true;
2231                                         }
2232                                 }
2233                         }
2234                 }
2235                 assert(!(any_needs_mipmaps && any_refuses_mipmaps));
2236
2237                 if (any_needs_mipmaps && generated_mipmaps->count(input) == 0) {
2238                         glGenerateMipmap(GL_TEXTURE_2D);
2239                         check_error();
2240                         generated_mipmaps->insert(input);
2241                 }
2242                 setup_rtt_sampler(sampler, any_needs_mipmaps);
2243                 phase->input_samplers[sampler] = sampler;  // Bind the sampler to the right uniform.
2244         }
2245
2246         GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
2247         check_error();
2248
2249         // And now the output.
2250         GLuint fbo = 0;
2251         if (phase->is_compute_shader) {
2252                 assert(!destinations.empty());
2253
2254                 // This is currently the only place where we use image units,
2255                 // so we can always start at 0. TODO: Support multiple destinations.
2256                 phase->outbuf_image_unit = 0;
2257                 glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format);
2258                 check_error();
2259                 phase->uniform_output_size[0] = phase->output_width;
2260                 phase->uniform_output_size[1] = phase->output_height;
2261                 phase->inv_output_size.x = 1.0f / phase->output_width;
2262                 phase->inv_output_size.y = 1.0f / phase->output_height;
2263                 phase->output_texcoord_adjust.x = 0.5f / phase->output_width;
2264                 phase->output_texcoord_adjust.y = 0.5f / phase->output_height;
2265         } else if (!destinations.empty()) {
2266                 assert(destinations.size() == 1);
2267                 fbo = resource_pool->create_fbo(destinations[0].texnum);
2268                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
2269                 glViewport(0, 0, phase->output_width, phase->output_height);
2270         }
2271
2272         // Give the required parameters to all the effects.
2273         unsigned sampler_num = phase->inputs.size();
2274         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2275                 Node *node = phase->effects[i];
2276                 unsigned old_sampler_num = sampler_num;
2277                 node->effect->set_gl_state(instance_program_num, phase->effect_ids[make_pair(node, IN_SAME_PHASE)], &sampler_num);
2278                 check_error();
2279
2280                 if (node->effect->is_single_texture()) {
2281                         assert(sampler_num - old_sampler_num == 1);
2282                         node->bound_sampler_num = old_sampler_num;
2283                 } else {
2284                         node->bound_sampler_num = -1;
2285                 }
2286         }
2287
2288         if (phase->is_compute_shader) {
2289                 unsigned x, y, z;
2290                 phase->compute_shader_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z);
2291
2292                 // Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(),
2293                 // since they can be updated from there.
2294                 setup_uniforms(phase);
2295                 glDispatchCompute(x, y, z);
2296                 check_error();
2297                 glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT | GL_TEXTURE_UPDATE_BARRIER_BIT);
2298                 check_error();
2299         } else {
2300                 // Uniforms need to come after set_gl_state(), since they can be updated
2301                 // from there.
2302                 setup_uniforms(phase);
2303
2304                 // Bind the vertex data.
2305                 GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
2306                 glBindVertexArray(vao);
2307
2308                 glDrawArrays(GL_TRIANGLES, 0, 3);
2309                 check_error();
2310
2311                 resource_pool->release_vec2_vao(vao);
2312         }
2313
2314         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2315                 Node *node = phase->effects[i];
2316                 node->effect->clear_gl_state();
2317         }
2318
2319         resource_pool->unuse_glsl_program(instance_program_num);
2320
2321         if (fbo != 0) {
2322                 resource_pool->release_fbo(fbo);
2323         }
2324 }
2325
2326 void EffectChain::setup_uniforms(Phase *phase)
2327 {
2328         // TODO: Use UBO blocks.
2329         for (size_t i = 0; i < phase->uniforms_image2d.size(); ++i) {
2330                 const Uniform<int> &uniform = phase->uniforms_image2d[i];
2331                 if (uniform.location != -1) {
2332                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2333                 }
2334         }
2335         for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
2336                 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
2337                 if (uniform.location != -1) {
2338                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2339                 }
2340         }
2341         for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
2342                 const Uniform<bool> &uniform = phase->uniforms_bool[i];
2343                 assert(uniform.num_values == 1);
2344                 if (uniform.location != -1) {
2345                         glUniform1i(uniform.location, *uniform.value);
2346                 }
2347         }
2348         for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
2349                 const Uniform<int> &uniform = phase->uniforms_int[i];
2350                 if (uniform.location != -1) {
2351                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2352                 }
2353         }
2354         for (size_t i = 0; i < phase->uniforms_ivec2.size(); ++i) {
2355                 const Uniform<int> &uniform = phase->uniforms_ivec2[i];
2356                 if (uniform.location != -1) {
2357                         glUniform2iv(uniform.location, uniform.num_values, uniform.value);
2358                 }
2359         }
2360         for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
2361                 const Uniform<float> &uniform = phase->uniforms_float[i];
2362                 if (uniform.location != -1) {
2363                         glUniform1fv(uniform.location, uniform.num_values, uniform.value);
2364                 }
2365         }
2366         for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
2367                 const Uniform<float> &uniform = phase->uniforms_vec2[i];
2368                 if (uniform.location != -1) {
2369                         glUniform2fv(uniform.location, uniform.num_values, uniform.value);
2370                 }
2371         }
2372         for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
2373                 const Uniform<float> &uniform = phase->uniforms_vec3[i];
2374                 if (uniform.location != -1) {
2375                         glUniform3fv(uniform.location, uniform.num_values, uniform.value);
2376                 }
2377         }
2378         for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
2379                 const Uniform<float> &uniform = phase->uniforms_vec4[i];
2380                 if (uniform.location != -1) {
2381                         glUniform4fv(uniform.location, uniform.num_values, uniform.value);
2382                 }
2383         }
2384         for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
2385                 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
2386                 assert(uniform.num_values == 1);
2387                 if (uniform.location != -1) {
2388                         // Convert to float (GLSL has no double matrices).
2389                         float matrixf[9];
2390                         for (unsigned y = 0; y < 3; ++y) {
2391                                 for (unsigned x = 0; x < 3; ++x) {
2392                                         matrixf[y + x * 3] = (*uniform.value)(y, x);
2393                                 }
2394                         }
2395                         glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
2396                 }
2397         }
2398 }
2399
2400 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
2401 {
2402         glActiveTexture(GL_TEXTURE0 + sampler_num);
2403         check_error();
2404         if (use_mipmaps) {
2405                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2406                 check_error();
2407         } else {
2408                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2409                 check_error();
2410         }
2411         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2412         check_error();
2413         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2414         check_error();
2415 }
2416
2417 }  // namespace movit