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