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