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