f11a5e45a3ada2ccbbaf38a5b76630ceb2f41915
[movit] / effect_chain.cpp
1 #include <epoxy/gl.h>
2 #include <assert.h>
3 #include <math.h>
4 #include <stddef.h>
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <string.h>
8 #include <algorithm>
9 #include <set>
10 #include <stack>
11 #include <utility>
12 #include <vector>
13 #include <Eigen/Core>
14
15 #include "alpha_division_effect.h"
16 #include "alpha_multiplication_effect.h"
17 #include "colorspace_conversion_effect.h"
18 #include "dither_effect.h"
19 #include "effect.h"
20 #include "effect_chain.h"
21 #include "effect_util.h"
22 #include "gamma_compression_effect.h"
23 #include "gamma_expansion_effect.h"
24 #include "init.h"
25 #include "input.h"
26 #include "resource_pool.h"
27 #include "util.h"
28 #include "ycbcr_conversion_effect.h"
29
30 using namespace Eigen;
31 using namespace std;
32
33 namespace movit {
34
35 namespace {
36
37 // An effect whose only purpose is to sit in a phase on its own and take the
38 // texture output from a compute shader and display it to the normal backbuffer
39 // (or any FBO). That phase can be skipped when rendering using render_to_textures().
40 class ComputeShaderOutputDisplayEffect : public Effect {
41 public:
42         ComputeShaderOutputDisplayEffect() {}
43         string effect_type_id() const override { return "ComputeShaderOutputDisplayEffect"; }
44         string output_fragment_shader() override { return read_file("identity.frag"); }
45         bool needs_texture_bounce() const override { return true; }
46 };
47
48 }  // namespace
49
50 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
51         : aspect_nom(aspect_nom),
52           aspect_denom(aspect_denom),
53           output_color_rgba(false),
54           num_output_color_ycbcr(0),
55           dither_effect(nullptr),
56           ycbcr_conversion_effect_node(nullptr),
57           intermediate_format(GL_RGBA16F),
58           intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION),
59           num_dither_bits(0),
60           output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
61           finalized(false),
62           resource_pool(resource_pool),
63           do_phase_timing(false) {
64         if (resource_pool == nullptr) {
65                 this->resource_pool = new ResourcePool();
66                 owns_resource_pool = true;
67         } else {
68                 owns_resource_pool = false;
69         }
70
71         // Generate a VBO with some data in (shared position and texture coordinate data).
72         float vertices[] = {
73                 0.0f, 2.0f,
74                 0.0f, 0.0f,
75                 2.0f, 0.0f
76         };
77         vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
78 }
79
80 EffectChain::~EffectChain()
81 {
82         for (unsigned i = 0; i < nodes.size(); ++i) {
83                 delete nodes[i]->effect;
84                 delete nodes[i];
85         }
86         for (unsigned i = 0; i < phases.size(); ++i) {
87                 resource_pool->release_glsl_program(phases[i]->glsl_program_num);
88                 delete phases[i];
89         }
90         if (owns_resource_pool) {
91                 delete resource_pool;
92         }
93         glDeleteBuffers(1, &vbo);
94         check_error();
95 }
96
97 Input *EffectChain::add_input(Input *input)
98 {
99         assert(!finalized);
100         inputs.push_back(input);
101         add_node(input);
102         return input;
103 }
104
105 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
106 {
107         assert(!finalized);
108         assert(!output_color_rgba);
109         output_format = format;
110         output_alpha_format = alpha_format;
111         output_color_rgba = true;
112 }
113
114 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
115                                    const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting,
116                                    GLenum output_type)
117 {
118         assert(!finalized);
119         assert(num_output_color_ycbcr < 2);
120         output_format = format;
121         output_alpha_format = alpha_format;
122
123         if (num_output_color_ycbcr == 1) {
124                 // Check that the format is the same.
125                 assert(output_ycbcr_format.luma_coefficients == ycbcr_format.luma_coefficients);
126                 assert(output_ycbcr_format.full_range == ycbcr_format.full_range);
127                 assert(output_ycbcr_format.num_levels == ycbcr_format.num_levels);
128                 assert(output_ycbcr_format.chroma_subsampling_x == 1);
129                 assert(output_ycbcr_format.chroma_subsampling_y == 1);
130                 assert(output_ycbcr_type == output_type);
131         } else {
132                 output_ycbcr_format = ycbcr_format;
133                 output_ycbcr_type = output_type;
134         }
135         output_ycbcr_splitting[num_output_color_ycbcr++] = output_splitting;
136
137         assert(ycbcr_format.chroma_subsampling_x == 1);
138         assert(ycbcr_format.chroma_subsampling_y == 1);
139 }
140
141 void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format)
142 {
143         assert(num_output_color_ycbcr > 0);
144         assert(output_ycbcr_format.chroma_subsampling_x == 1);
145         assert(output_ycbcr_format.chroma_subsampling_y == 1);
146
147         output_ycbcr_format = ycbcr_format;
148         if (finalized) {
149                 YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(ycbcr_conversion_effect_node->effect);
150                 effect->change_output_format(ycbcr_format);
151         }
152 }
153
154 Node *EffectChain::add_node(Effect *effect)
155 {
156         for (unsigned i = 0; i < nodes.size(); ++i) {
157                 assert(nodes[i]->effect != effect);
158         }
159
160         Node *node = new Node;
161         node->effect = effect;
162         node->disabled = false;
163         node->output_color_space = COLORSPACE_INVALID;
164         node->output_gamma_curve = GAMMA_INVALID;
165         node->output_alpha_type = ALPHA_INVALID;
166         node->needs_mipmaps = Effect::DOES_NOT_NEED_MIPMAPS;
167         node->one_to_one_sampling = false;
168         node->strong_one_to_one_sampling = false;
169
170         nodes.push_back(node);
171         node_map[effect] = node;
172         effect->inform_added(this);
173         return node;
174 }
175
176 void EffectChain::connect_nodes(Node *sender, Node *receiver)
177 {
178         sender->outgoing_links.push_back(receiver);
179         receiver->incoming_links.push_back(sender);
180 }
181
182 void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
183 {
184         new_receiver->incoming_links = old_receiver->incoming_links;
185         old_receiver->incoming_links.clear();
186         
187         for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
188                 Node *sender = new_receiver->incoming_links[i];
189                 for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
190                         if (sender->outgoing_links[j] == old_receiver) {
191                                 sender->outgoing_links[j] = new_receiver;
192                         }
193                 }
194         }       
195 }
196
197 void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
198 {
199         new_sender->outgoing_links = old_sender->outgoing_links;
200         old_sender->outgoing_links.clear();
201         
202         for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
203                 Node *receiver = new_sender->outgoing_links[i];
204                 for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
205                         if (receiver->incoming_links[j] == old_sender) {
206                                 receiver->incoming_links[j] = new_sender;
207                         }
208                 }
209         }       
210 }
211
212 void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
213 {
214         for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
215                 if (sender->outgoing_links[i] == receiver) {
216                         sender->outgoing_links[i] = middle;
217                         middle->incoming_links.push_back(sender);
218                 }
219         }
220         for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
221                 if (receiver->incoming_links[i] == sender) {
222                         receiver->incoming_links[i] = middle;
223                         middle->outgoing_links.push_back(receiver);
224                 }
225         }
226
227         assert(middle->incoming_links.size() == middle->effect->num_inputs());
228 }
229
230 GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
231 {
232         assert(node->effect->needs_texture_bounce());
233         assert(input_num < node->incoming_links.size());
234         assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
235         assert(node->incoming_links[input_num]->bound_sampler_num < 8);
236         return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
237 }
238
239 GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
240 {
241         assert(input_num < node->incoming_links.size());
242         return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
243                 node->incoming_links[input_num]->bound_sampler_num < 8;
244 }
245
246 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
247 {
248         if (node->output_gamma_curve == GAMMA_LINEAR &&
249             node->effect->effect_type_id() != "GammaCompressionEffect") {
250                 return;
251         }
252         if (node->effect->num_inputs() == 0) {
253                 nonlinear_inputs->push_back(node);
254         } else {
255                 assert(node->effect->num_inputs() == node->incoming_links.size());
256                 for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
257                         find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
258                 }
259         }
260 }
261
262 Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
263 {
264         assert(!finalized);
265         assert(inputs.size() == effect->num_inputs());
266         Node *node = add_node(effect);
267         for (unsigned i = 0; i < inputs.size(); ++i) {
268                 assert(node_map.count(inputs[i]) != 0);
269                 connect_nodes(node_map[inputs[i]], node);
270         }
271         return effect;
272 }
273
274 // ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
275 string replace_prefix(const string &text, const string &prefix)
276 {
277         string output;
278         size_t start = 0;
279
280         while (start < text.size()) {
281                 size_t pos = text.find("PREFIX(", start);
282                 if (pos == string::npos) {
283                         output.append(text.substr(start, string::npos));
284                         break;
285                 }
286
287                 output.append(text.substr(start, pos - start));
288                 output.append(prefix);
289                 output.append("_");
290
291                 pos += strlen("PREFIX(");
292         
293                 // Output stuff until we find the matching ), which we then eat.
294                 int depth = 1;
295                 size_t end_arg_pos = pos;
296                 while (end_arg_pos < text.size()) {
297                         if (text[end_arg_pos] == '(') {
298                                 ++depth;
299                         } else if (text[end_arg_pos] == ')') {
300                                 --depth;
301                                 if (depth == 0) {
302                                         break;
303                                 }
304                         }
305                         ++end_arg_pos;
306                 }
307                 output.append(text.substr(pos, end_arg_pos - pos));
308                 ++end_arg_pos;
309                 assert(depth == 0);
310                 start = end_arg_pos;
311         }
312         return output;
313 }
314
315 namespace {
316
317 template<class T>
318 void extract_uniform_declarations(const vector<Uniform<T>> &effect_uniforms,
319                                   const string &type_specifier,
320                                   const string &effect_id,
321                                   vector<Uniform<T>> *phase_uniforms,
322                                   string *glsl_string)
323 {
324         for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
325                 phase_uniforms->push_back(effect_uniforms[i]);
326                 phase_uniforms->back().prefix = effect_id;
327
328                 *glsl_string += string("uniform ") + type_specifier + " " + effect_id
329                         + "_" + effect_uniforms[i].name + ";\n";
330         }
331 }
332
333 template<class T>
334 void extract_uniform_array_declarations(const vector<Uniform<T>> &effect_uniforms,
335                                         const string &type_specifier,
336                                         const string &effect_id,
337                                         vector<Uniform<T>> *phase_uniforms,
338                                         string *glsl_string)
339 {
340         for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
341                 phase_uniforms->push_back(effect_uniforms[i]);
342                 phase_uniforms->back().prefix = effect_id;
343
344                 char buf[256];
345                 snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
346                         type_specifier.c_str(), effect_id.c_str(),
347                         effect_uniforms[i].name.c_str(),
348                         int(effect_uniforms[i].num_values));
349                 *glsl_string += buf;
350         }
351 }
352
353 template<class T>
354 void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T>> *phase_uniforms)
355 {
356         for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
357                 Uniform<T> &uniform = (*phase_uniforms)[i];
358                 uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
359         }
360 }
361
362 }  // namespace
363
364 void EffectChain::compile_glsl_program(Phase *phase)
365 {
366         string frag_shader_header;
367         if (phase->is_compute_shader) {
368                 frag_shader_header = read_file("header.comp");
369         } else {
370                 frag_shader_header = read_version_dependent_file("header", "frag");
371         }
372         string frag_shader = "";
373
374         // Create functions and uniforms for all the texture inputs that we need.
375         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
376                 Node *input = phase->inputs[i]->output_node;
377                 char effect_id[256];
378                 sprintf(effect_id, "in%u", i);
379                 phase->effect_ids.insert(make_pair(make_pair(input, IN_ANOTHER_PHASE), effect_id));
380         
381                 frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
382                 frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
383                 frag_shader += "\tvec4 tmp = tex2D(tex_" + string(effect_id) + ", tc);\n";
384
385                 if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
386                     phase->inputs[i]->output_node->output_gamma_curve == GAMMA_LINEAR) {
387                         frag_shader += "\ttmp.rgb *= tmp.rgb;\n";
388                 }
389
390                 frag_shader += "\treturn tmp;\n";
391                 frag_shader += "}\n";
392                 frag_shader += "\n";
393
394                 Uniform<int> uniform;
395                 uniform.name = effect_id;
396                 uniform.value = &phase->input_samplers[i];
397                 uniform.prefix = "tex";
398                 uniform.num_values = 1;
399                 uniform.location = -1;
400                 phase->uniforms_sampler2d.push_back(uniform);
401         }
402
403         // Give each effect in the phase its own ID.
404         for (unsigned i = 0; i < phase->effects.size(); ++i) {
405                 Node *node = phase->effects[i];
406                 char effect_id[256];
407                 sprintf(effect_id, "eff%u", i);
408                 bool inserted = phase->effect_ids.insert(make_pair(make_pair(node, IN_SAME_PHASE), effect_id)).second;
409                 assert(inserted);
410         }
411
412         for (unsigned i = 0; i < phase->effects.size(); ++i) {
413                 Node *node = phase->effects[i];
414                 const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
415                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
416                         if (node->incoming_links.size() == 1) {
417                                 frag_shader += "#define INPUT";
418                         } else {
419                                 char buf[256];
420                                 sprintf(buf, "#define INPUT%d", j + 1);
421                                 frag_shader += buf;
422                         }
423
424                         Node *input = node->incoming_links[j];
425                         NodeLinkType link_type = node->incoming_link_type[j];
426                         if (i != 0 &&
427                             input->effect->is_compute_shader() &&
428                             node->incoming_link_type[j] == IN_SAME_PHASE) {
429                                 // First effect after the compute shader reads the value
430                                 // that cs_output() wrote to a global variable,
431                                 // ignoring the tc (since all such effects have to be
432                                 // strong one-to-one).
433                                 frag_shader += "(tc) CS_OUTPUT_VAL\n";
434                         } else {
435                                 assert(phase->effect_ids.count(make_pair(input, link_type)));
436                                 frag_shader += string(" ") + phase->effect_ids[make_pair(input, link_type)] + "\n";
437                         }
438                 }
439         
440                 frag_shader += "\n";
441                 frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
442                 if (node->effect->is_compute_shader()) {
443                         frag_shader += string("#define NORMALIZE_TEXTURE_COORDS(tc) ((tc) * ") + effect_id + "_inv_output_size + " + effect_id + "_output_texcoord_adjust)\n";
444                 }
445                 frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
446                 frag_shader += "#undef FUNCNAME\n";
447                 if (node->incoming_links.size() == 1) {
448                         frag_shader += "#undef INPUT\n";
449                 } else {
450                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
451                                 char buf[256];
452                                 sprintf(buf, "#undef INPUT%d\n", j + 1);
453                                 frag_shader += buf;
454                         }
455                 }
456                 frag_shader += "\n";
457         }
458         if (phase->is_compute_shader) {
459                 assert(phase->effect_ids.count(make_pair(phase->compute_shader_node, IN_SAME_PHASE)));
460                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->compute_shader_node, IN_SAME_PHASE)] + "\n";
461                 if (phase->compute_shader_node == phase->effects.back()) {
462                         // No postprocessing.
463                         frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n";
464                 } else {
465                         frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
466                 }
467         } else {
468                 assert(phase->effect_ids.count(make_pair(phase->effects.back(), IN_SAME_PHASE)));
469                 frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
470         }
471
472         // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
473         vector<string> frag_shader_outputs;  // In order.
474         if (phase->output_node->outgoing_links.empty() && num_output_color_ycbcr > 0) {
475                 switch (output_ycbcr_splitting[0]) {
476                 case YCBCR_OUTPUT_INTERLEAVED:
477                         // No #defines set.
478                         frag_shader_outputs.push_back("FragColor");
479                         break;
480                 case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
481                         frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
482                         frag_shader_outputs.push_back("Y");
483                         frag_shader_outputs.push_back("Chroma");
484                         break;
485                 case YCBCR_OUTPUT_PLANAR:
486                         frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
487                         frag_shader_outputs.push_back("Y");
488                         frag_shader_outputs.push_back("Cb");
489                         frag_shader_outputs.push_back("Cr");
490                         break;
491                 default:
492                         assert(false);
493                 }
494
495                 if (num_output_color_ycbcr > 1) {
496                         switch (output_ycbcr_splitting[1]) {
497                         case YCBCR_OUTPUT_INTERLEAVED:
498                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_INTERLEAVED 1\n";
499                                 frag_shader_outputs.push_back("YCbCr2");
500                                 break;
501                         case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
502                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
503                                 frag_shader_outputs.push_back("Y2");
504                                 frag_shader_outputs.push_back("Chroma2");
505                                 break;
506                         case YCBCR_OUTPUT_PLANAR:
507                                 frag_shader += "#define SECOND_YCBCR_OUTPUT_PLANAR 1\n";
508                                 frag_shader_outputs.push_back("Y2");
509                                 frag_shader_outputs.push_back("Cb2");
510                                 frag_shader_outputs.push_back("Cr2");
511                                 break;
512                         default:
513                                 assert(false);
514                         }
515                 }
516
517                 if (output_color_rgba) {
518                         // Note: Needs to come in the header, because not only the
519                         // output needs to see it (YCbCrConversionEffect and DitherEffect
520                         // do, too).
521                         frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
522                         frag_shader_outputs.push_back("RGBA");
523                 }
524         }
525
526         // If we're bouncing to a temporary texture, signal transformation if desired.
527         if (!phase->output_node->outgoing_links.empty()) {
528                 if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
529                     phase->output_node->output_gamma_curve == GAMMA_LINEAR) {
530                         frag_shader += "#define SQUARE_ROOT_TRANSFORMATION 1\n";
531                 }
532         }
533
534         if (phase->is_compute_shader) {
535                 frag_shader.append(read_file("footer.comp"));
536                 phase->compute_shader_node->effect->register_uniform_ivec2("output_size", phase->uniform_output_size);
537                 phase->compute_shader_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size);
538                 phase->compute_shader_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust);
539         } else {
540                 frag_shader.append(read_file("footer.frag"));
541         }
542
543         // Collect uniforms from all effects and output them. Note that this needs
544         // to happen after output_fragment_shader(), even though the uniforms come
545         // before in the output source, since output_fragment_shader() is allowed
546         // to register new uniforms (e.g. arrays that are of unknown length until
547         // finalization time).
548         // TODO: Make a uniform block for platforms that support it.
549         string frag_shader_uniforms = "";
550         for (unsigned i = 0; i < phase->effects.size(); ++i) {
551                 Node *node = phase->effects[i];
552                 Effect *effect = node->effect;
553                 const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
554                 extract_uniform_declarations(effect->uniforms_image2d, "image2D", effect_id, &phase->uniforms_image2d, &frag_shader_uniforms);
555                 extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
556                 extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
557                 extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
558                 extract_uniform_declarations(effect->uniforms_ivec2, "ivec2", effect_id, &phase->uniforms_ivec2, &frag_shader_uniforms);
559                 extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
560                 extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
561                 extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
562                 extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
563                 extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
564                 extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
565                 extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
566                 extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
567                 extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
568         }
569
570         string vert_shader = read_version_dependent_file("vs", "vert");
571
572         // If we're the last phase and need to flip the picture to compensate for
573         // the origin, tell the vertex or compute shader so.
574         bool is_last_phase;
575         if (has_dummy_effect) {
576                 is_last_phase = (phase->output_node->outgoing_links.size() == 1 &&
577                         phase->output_node->outgoing_links[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
578         } else {
579                 is_last_phase = phase->output_node->outgoing_links.empty();
580         }
581         if (is_last_phase && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
582                 if (phase->is_compute_shader) {
583                         frag_shader_header += "#define FLIP_ORIGIN 1\n";
584                 } else {
585                         const string needle = "#define FLIP_ORIGIN 0";
586                         size_t pos = vert_shader.find(needle);
587                         assert(pos != string::npos);
588
589                         vert_shader[pos + needle.size() - 1] = '1';
590                 }
591         }
592
593         frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
594
595         if (phase->is_compute_shader) {
596                 phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader);
597
598                 Uniform<int> uniform;
599                 uniform.name = "outbuf";
600                 uniform.value = &phase->outbuf_image_unit;
601                 uniform.prefix = "tex";
602                 uniform.num_values = 1;
603                 uniform.location = -1;
604                 phase->uniforms_image2d.push_back(uniform);
605         } else {
606                 phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
607         }
608         GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
609         GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
610         if (position_attribute_index != -1) {
611                 phase->attribute_indexes.insert(position_attribute_index);
612         }
613         if (texcoord_attribute_index != -1) {
614                 phase->attribute_indexes.insert(texcoord_attribute_index);
615         }
616
617         // Collect the resulting location numbers for each uniform.
618         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_image2d);
619         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
620         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
621         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
622         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_ivec2);
623         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
624         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
625         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
626         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
627         collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
628 }
629
630 // Construct GLSL programs, starting at the given effect and following
631 // the chain from there. We end a program every time we come to an effect
632 // marked as "needs texture bounce", one that is used by multiple other
633 // effects, every time we need to bounce due to output size change
634 // (not all size changes require ending), and of course at the end.
635 //
636 // We follow a quite simple depth-first search from the output, although
637 // without recursing explicitly within each phase.
638 Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
639 {
640         if (completed_effects->count(output)) {
641                 return (*completed_effects)[output];
642         }
643
644         Phase *phase = new Phase;
645         phase->output_node = output;
646         phase->is_compute_shader = false;
647         phase->compute_shader_node = nullptr;
648
649         // If the output effect has one-to-one sampling, we try to trace this
650         // status down through the dependency chain. This is important in case
651         // we hit an effect that changes output size (and not sets a virtual
652         // output size); if we have one-to-one sampling, we don't have to break
653         // the phase.
654         output->one_to_one_sampling = output->effect->one_to_one_sampling();
655         output->strong_one_to_one_sampling = output->effect->strong_one_to_one_sampling();
656
657         // Effects that we have yet to calculate, but that we know should
658         // be in the current phase.
659         stack<Node *> effects_todo_this_phase;
660         effects_todo_this_phase.push(output);
661
662         while (!effects_todo_this_phase.empty()) {
663                 Node *node = effects_todo_this_phase.top();
664                 effects_todo_this_phase.pop();
665
666                 assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling());
667
668                 if (node->effect->needs_mipmaps() != Effect::DOES_NOT_NEED_MIPMAPS) {
669                         // Can't have incompatible requirements imposed on us from a dependent effect;
670                         // if so, it should have started a new phase instead.
671                         assert(node->needs_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS ||
672                                node->needs_mipmaps == node->effect->needs_mipmaps());
673                         node->needs_mipmaps = node->effect->needs_mipmaps();
674                 }
675
676                 // This should currently only happen for effects that are inputs
677                 // (either true inputs or phase outputs). We special-case inputs,
678                 // and then deduplicate phase outputs below.
679                 if (node->effect->num_inputs() == 0) {
680                         if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
681                                 continue;
682                         }
683                 } else {
684                         assert(completed_effects->count(node) == 0);
685                 }
686
687                 phase->effects.push_back(node);
688                 if (node->effect->is_compute_shader()) {
689                         assert(phase->compute_shader_node == nullptr ||
690                                phase->compute_shader_node == node);
691                         phase->is_compute_shader = true;
692                         phase->compute_shader_node = node;
693                 }
694
695                 // Find all the dependencies of this effect, and add them to the stack.
696                 assert(node->effect->num_inputs() == node->incoming_links.size());
697                 for (Node *dep : node->incoming_links) {
698                         bool start_new_phase = false;
699
700                         Effect::MipmapRequirements save_needs_mipmaps = dep->needs_mipmaps;
701
702                         if (node->effect->needs_texture_bounce() &&
703                             !dep->effect->is_single_texture() &&
704                             !dep->effect->override_disable_bounce()) {
705                                 start_new_phase = true;
706                         }
707
708                         // Propagate information about needing mipmaps down the chain,
709                         // breaking the phase if we notice an incompatibility.
710                         //
711                         // Note that we cannot do this propagation as a normal pass,
712                         // because it needs information about where the phases end
713                         // (we should not propagate the flag across phases).
714                         if (node->needs_mipmaps != Effect::DOES_NOT_NEED_MIPMAPS) {
715                                 // The node can have a value set (ie. not DOES_NOT_NEED_MIPMAPS)
716                                 // if we have diamonds in the graph; if so, choose that.
717                                 // If not, the effect on the node can also decide (this is the
718                                 // more common case).
719                                 Effect::MipmapRequirements dep_mipmaps = dep->needs_mipmaps;
720                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
721                                         if (dep->effect->num_inputs() == 0) {
722                                                 Input *input = static_cast<Input *>(dep->effect);
723                                                 dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS;
724                                         } else {
725                                                 dep_mipmaps = dep->effect->needs_mipmaps();
726                                         }
727                                 }
728                                 if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
729                                         dep->needs_mipmaps = node->needs_mipmaps;
730                                 } else if (dep_mipmaps != node->needs_mipmaps) {
731                                         // The dependency cannot supply our mipmap demands
732                                         // (either because it's an input that can't do mipmaps,
733                                         // or because there's a conflict between mipmap-needing
734                                         // and mipmap-refusing effects somewhere in the graph),
735                                         // so they cannot be in the same phase.
736                                         start_new_phase = true;
737                                 }
738                         }
739
740                         if (dep->outgoing_links.size() > 1) {
741                                 if (!dep->effect->is_single_texture()) {
742                                         // More than one effect uses this as the input,
743                                         // and it is not a texture itself.
744                                         // The easiest thing to do (and probably also the safest
745                                         // performance-wise in most cases) is to bounce it to a texture
746                                         // and then let the next passes read from that.
747                                         start_new_phase = true;
748                                 } else {
749                                         assert(dep->effect->num_inputs() == 0);
750
751                                         // For textures, we try to be slightly more clever;
752                                         // if none of our outputs need a bounce, we don't bounce
753                                         // but instead simply use the effect many times.
754                                         //
755                                         // Strictly speaking, we could bounce it for some outputs
756                                         // and use it directly for others, but the processing becomes
757                                         // somewhat simpler if the effect is only used in one such way.
758                                         for (unsigned j = 0; j < dep->outgoing_links.size(); ++j) {
759                                                 Node *rdep = dep->outgoing_links[j];
760                                                 start_new_phase |= rdep->effect->needs_texture_bounce();
761                                         }
762                                 }
763                         }
764
765                         if (dep->effect->is_compute_shader()) {
766                                 if (phase->is_compute_shader) {
767                                         // Only one compute shader per phase.
768                                         start_new_phase = true;
769                                 } else if (!node->strong_one_to_one_sampling) {
770                                         // If all nodes so far are strong one-to-one, we can put them after
771                                         // the compute shader (ie., process them on the output).
772                                         start_new_phase = true;
773                                 } else if (!start_new_phase) {
774                                         phase->is_compute_shader = true;
775                                         phase->compute_shader_node = dep;
776                                 }
777                         } else if (dep->effect->sets_virtual_output_size()) {
778                                 assert(dep->effect->changes_output_size());
779                                 // If the next effect sets a virtual size to rely on OpenGL's
780                                 // bilinear sampling, we'll really need to break the phase here.
781                                 start_new_phase = true;
782                         } else if (dep->effect->changes_output_size() && !node->one_to_one_sampling) {
783                                 // If the next effect changes size and we don't have one-to-one sampling,
784                                 // we also need to break here.
785                                 start_new_phase = true;
786                         }
787
788                         if (start_new_phase) {
789                                 // Since we're starting a new phase here, we don't need to impose any
790                                 // new demands on this effect. Restore the status we had before we
791                                 // started looking at it.
792                                 dep->needs_mipmaps = save_needs_mipmaps;
793
794                                 phase->inputs.push_back(construct_phase(dep, completed_effects));
795                         } else {
796                                 effects_todo_this_phase.push(dep);
797
798                                 // Propagate the one-to-one status down through the dependency.
799                                 dep->one_to_one_sampling = node->one_to_one_sampling &&
800                                         dep->effect->one_to_one_sampling();
801                                 dep->strong_one_to_one_sampling = node->strong_one_to_one_sampling &&
802                                         dep->effect->strong_one_to_one_sampling();
803                         }
804
805                         node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE);
806                 }
807         }
808
809         // No more effects to do this phase. Take all the ones we have,
810         // and create a GLSL program for it.
811         assert(!phase->effects.empty());
812
813         // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
814         // that would be nondeterministic and thus reduce cacheability.
815         // TODO: Make this even more deterministic.
816         vector<Phase *> dedup_inputs;
817         set<Phase *> seen_inputs;
818         for (size_t i = 0; i < phase->inputs.size(); ++i) {
819                 if (seen_inputs.insert(phase->inputs[i]).second) {
820                         dedup_inputs.push_back(phase->inputs[i]);
821                 }
822         }
823         swap(phase->inputs, dedup_inputs);
824
825         // Allocate samplers for each input.
826         phase->input_samplers.resize(phase->inputs.size());
827
828         // We added the effects from the output and back, but we need to output
829         // them in topological sort order in the shader.
830         phase->effects = topological_sort(phase->effects);
831
832         // Figure out if we need mipmaps or not, and if so, tell the inputs that.
833         // (RTT inputs have different logic, which is checked in execute_phase().)
834         for (unsigned i = 0; i < phase->effects.size(); ++i) {
835                 Node *node = phase->effects[i];
836                 if (node->effect->num_inputs() == 0) {
837                         Input *input = static_cast<Input *>(node->effect);
838                         assert(node->needs_mipmaps != Effect::NEEDS_MIPMAPS || input->can_supply_mipmaps());
839                         CHECK(input->set_int("needs_mipmaps", node->needs_mipmaps == Effect::NEEDS_MIPMAPS));
840                 }
841         }
842
843         // Tell each node which phase it ended up in, so that the unit test
844         // can check that the phases were split in the right place.
845         // Note that this ignores that effects may be part of multiple phases;
846         // if the unit tests need to test such cases, we'll reconsider.
847         for (unsigned i = 0; i < phase->effects.size(); ++i) {
848                 phase->effects[i]->containing_phase = phase;
849         }
850
851         // Actually make the shader for this phase.
852         compile_glsl_program(phase);
853
854         // Initialize timers.
855         if (movit_timer_queries_supported) {
856                 phase->time_elapsed_ns = 0;
857                 phase->num_measured_iterations = 0;
858         }
859
860         assert(completed_effects->count(output) == 0);
861         completed_effects->insert(make_pair(output, phase));
862         phases.push_back(phase);
863         return phase;
864 }
865
866 void EffectChain::output_dot(const char *filename)
867 {
868         if (movit_debug_level != MOVIT_DEBUG_ON) {
869                 return;
870         }
871
872         FILE *fp = fopen(filename, "w");
873         if (fp == nullptr) {
874                 perror(filename);
875                 exit(1);
876         }
877
878         fprintf(fp, "digraph G {\n");
879         fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
880         for (unsigned i = 0; i < nodes.size(); ++i) {
881                 // Find out which phase this event belongs to.
882                 vector<int> in_phases;
883                 for (unsigned j = 0; j < phases.size(); ++j) {
884                         const Phase* p = phases[j];
885                         if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
886                                 in_phases.push_back(j);
887                         }
888                 }
889
890                 if (in_phases.empty()) {
891                         fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
892                 } else if (in_phases.size() == 1) {
893                         fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
894                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
895                                 (in_phases[0] % 8) + 1);
896                 } else {
897                         // If we had new enough Graphviz, style="wedged" would probably be ideal here.
898                         // But alas.
899                         fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
900                                 (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
901                                 (in_phases[0] % 8) + 1);
902                 }
903
904                 char from_node_id[256];
905                 snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
906
907                 for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
908                         char to_node_id[256];
909                         snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
910
911                         vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
912                         output_dot_edge(fp, from_node_id, to_node_id, labels);
913                 }
914
915                 if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
916                         // Output node.
917                         vector<string> labels = get_labels_for_edge(nodes[i], nullptr);
918                         output_dot_edge(fp, from_node_id, "output", labels);
919                 }
920         }
921         fprintf(fp, "}\n");
922
923         fclose(fp);
924 }
925
926 vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
927 {
928         vector<string> labels;
929
930         if (to != nullptr && to->effect->needs_texture_bounce()) {
931                 labels.push_back("needs_bounce");
932         }
933         if (from->effect->changes_output_size()) {
934                 labels.push_back("resize");
935         }
936
937         switch (from->output_color_space) {
938         case COLORSPACE_INVALID:
939                 labels.push_back("spc[invalid]");
940                 break;
941         case COLORSPACE_REC_601_525:
942                 labels.push_back("spc[rec601-525]");
943                 break;
944         case COLORSPACE_REC_601_625:
945                 labels.push_back("spc[rec601-625]");
946                 break;
947         default:
948                 break;
949         }
950
951         switch (from->output_gamma_curve) {
952         case GAMMA_INVALID:
953                 labels.push_back("gamma[invalid]");
954                 break;
955         case GAMMA_sRGB:
956                 labels.push_back("gamma[sRGB]");
957                 break;
958         case GAMMA_REC_601:  // and GAMMA_REC_709
959                 labels.push_back("gamma[rec601/709]");
960                 break;
961         default:
962                 break;
963         }
964
965         switch (from->output_alpha_type) {
966         case ALPHA_INVALID:
967                 labels.push_back("alpha[invalid]");
968                 break;
969         case ALPHA_BLANK:
970                 labels.push_back("alpha[blank]");
971                 break;
972         case ALPHA_POSTMULTIPLIED:
973                 labels.push_back("alpha[postmult]");
974                 break;
975         default:
976                 break;
977         }
978
979         return labels;
980 }
981
982 void EffectChain::output_dot_edge(FILE *fp,
983                                   const string &from_node_id,
984                                   const string &to_node_id,
985                                   const vector<string> &labels)
986 {
987         if (labels.empty()) {
988                 fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
989         } else {
990                 string label = labels[0];
991                 for (unsigned k = 1; k < labels.size(); ++k) {
992                         label += ", " + labels[k];
993                 }
994                 fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
995         }
996 }
997
998 void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
999 {
1000         unsigned scaled_width, scaled_height;
1001
1002         if (float(width) * aspect_denom >= float(height) * aspect_nom) {
1003                 // Same aspect, or W/H > aspect (image is wider than the frame).
1004                 // In either case, keep width, and adjust height.
1005                 scaled_width = width;
1006                 scaled_height = lrintf(width * aspect_denom / aspect_nom);
1007         } else {
1008                 // W/H < aspect (image is taller than the frame), so keep height,
1009                 // and adjust width.
1010                 scaled_width = lrintf(height * aspect_nom / aspect_denom);
1011                 scaled_height = height;
1012         }
1013
1014         // We should be consistently larger or smaller then the existing choice,
1015         // since we have the same aspect.
1016         assert(!(scaled_width < *output_width && scaled_height > *output_height));
1017         assert(!(scaled_height < *output_height && scaled_width > *output_width));
1018
1019         if (scaled_width >= *output_width && scaled_height >= *output_height) {
1020                 *output_width = scaled_width;
1021                 *output_height = scaled_height;
1022         }
1023 }
1024
1025 // Propagate input texture sizes throughout, and inform effects downstream.
1026 // (Like a lot of other code, we depend on effects being in topological order.)
1027 void EffectChain::inform_input_sizes(Phase *phase)
1028 {
1029         // All effects that have a defined size (inputs and RTT inputs)
1030         // get that. Reset all others.
1031         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1032                 Node *node = phase->effects[i];
1033                 if (node->effect->num_inputs() == 0) {
1034                         Input *input = static_cast<Input *>(node->effect);
1035                         node->output_width = input->get_width();
1036                         node->output_height = input->get_height();
1037                         assert(node->output_width != 0);
1038                         assert(node->output_height != 0);
1039                 } else {
1040                         node->output_width = node->output_height = 0;
1041                 }
1042         }
1043         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1044                 Phase *input = phase->inputs[i];
1045                 input->output_node->output_width = input->virtual_output_width;
1046                 input->output_node->output_height = input->virtual_output_height;
1047                 assert(input->output_node->output_width != 0);
1048                 assert(input->output_node->output_height != 0);
1049         }
1050
1051         // Now propagate from the inputs towards the end, and inform as we go.
1052         // The rules are simple:
1053         //
1054         //   1. Don't touch effects that already have given sizes (ie., inputs
1055         //      or effects that change the output size).
1056         //   2. If all of your inputs have the same size, that will be your output size.
1057         //   3. Otherwise, your output size is 0x0.
1058         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1059                 Node *node = phase->effects[i];
1060                 if (node->effect->num_inputs() == 0) {
1061                         continue;
1062                 }
1063                 unsigned this_output_width = 0;
1064                 unsigned this_output_height = 0;
1065                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1066                         Node *input = node->incoming_links[j];
1067                         node->effect->inform_input_size(j, input->output_width, input->output_height);
1068                         if (j == 0) {
1069                                 this_output_width = input->output_width;
1070                                 this_output_height = input->output_height;
1071                         } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
1072                                 // Inputs disagree.
1073                                 this_output_width = 0;
1074                                 this_output_height = 0;
1075                         }
1076                 }
1077                 if (node->effect->changes_output_size()) {
1078                         // We cannot call get_output_size() before we've done inform_input_size()
1079                         // on all inputs.
1080                         unsigned real_width, real_height;
1081                         node->effect->get_output_size(&real_width, &real_height,
1082                                                       &node->output_width, &node->output_height);
1083                         assert(node->effect->sets_virtual_output_size() ||
1084                                (real_width == node->output_width &&
1085                                 real_height == node->output_height));
1086                 } else {
1087                         node->output_width = this_output_width;
1088                         node->output_height = this_output_height;
1089                 }
1090         }
1091 }
1092
1093 // Note: You should call inform_input_sizes() before this, as the last effect's
1094 // desired output size might change based on the inputs.
1095 void EffectChain::find_output_size(Phase *phase)
1096 {
1097         Node *output_node = phase->is_compute_shader ? phase->compute_shader_node : phase->effects.back();
1098
1099         // If the last effect explicitly sets an output size, use that.
1100         if (output_node->effect->changes_output_size()) {
1101                 output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
1102                                                      &phase->virtual_output_width, &phase->virtual_output_height);
1103                 assert(output_node->effect->sets_virtual_output_size() ||
1104                        (phase->output_width == phase->virtual_output_width &&
1105                         phase->output_height == phase->virtual_output_height));
1106                 return;
1107         }
1108
1109         // If all effects have the same size, use that.
1110         unsigned output_width = 0, output_height = 0;
1111         bool all_inputs_same_size = true;
1112
1113         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1114                 Phase *input = phase->inputs[i];
1115                 assert(input->output_width != 0);
1116                 assert(input->output_height != 0);
1117                 if (output_width == 0 && output_height == 0) {
1118                         output_width = input->virtual_output_width;
1119                         output_height = input->virtual_output_height;
1120                 } else if (output_width != input->virtual_output_width ||
1121                            output_height != input->virtual_output_height) {
1122                         all_inputs_same_size = false;
1123                 }
1124         }
1125         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1126                 Effect *effect = phase->effects[i]->effect;
1127                 if (effect->num_inputs() != 0) {
1128                         continue;
1129                 }
1130
1131                 Input *input = static_cast<Input *>(effect);
1132                 if (output_width == 0 && output_height == 0) {
1133                         output_width = input->get_width();
1134                         output_height = input->get_height();
1135                 } else if (output_width != input->get_width() ||
1136                            output_height != input->get_height()) {
1137                         all_inputs_same_size = false;
1138                 }
1139         }
1140
1141         if (all_inputs_same_size) {
1142                 assert(output_width != 0);
1143                 assert(output_height != 0);
1144                 phase->virtual_output_width = phase->output_width = output_width;
1145                 phase->virtual_output_height = phase->output_height = output_height;
1146                 return;
1147         }
1148
1149         // If not, fit all the inputs into the current aspect, and select the largest one. 
1150         output_width = 0;
1151         output_height = 0;
1152         for (unsigned i = 0; i < phase->inputs.size(); ++i) {
1153                 Phase *input = phase->inputs[i];
1154                 assert(input->output_width != 0);
1155                 assert(input->output_height != 0);
1156                 size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height);
1157         }
1158         for (unsigned i = 0; i < phase->effects.size(); ++i) {
1159                 Effect *effect = phase->effects[i]->effect;
1160                 if (effect->num_inputs() != 0) {
1161                         continue;
1162                 }
1163
1164                 Input *input = static_cast<Input *>(effect);
1165                 size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
1166         }
1167         assert(output_width != 0);
1168         assert(output_height != 0);
1169         phase->virtual_output_width = phase->output_width = output_width;
1170         phase->virtual_output_height = phase->output_height = output_height;
1171 }
1172
1173 void EffectChain::sort_all_nodes_topologically()
1174 {
1175         nodes = topological_sort(nodes);
1176 }
1177
1178 vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
1179 {
1180         set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
1181         vector<Node *> sorted_list;
1182         for (unsigned i = 0; i < nodes.size(); ++i) {
1183                 topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
1184         }
1185         reverse(sorted_list.begin(), sorted_list.end());
1186         return sorted_list;
1187 }
1188
1189 void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
1190 {
1191         if (nodes_left_to_visit->count(node) == 0) {
1192                 return;
1193         }
1194         nodes_left_to_visit->erase(node);
1195         for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
1196                 topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
1197         }
1198         sorted_list->push_back(node);
1199 }
1200
1201 void EffectChain::find_color_spaces_for_inputs()
1202 {
1203         for (unsigned i = 0; i < nodes.size(); ++i) {
1204                 Node *node = nodes[i];
1205                 if (node->disabled) {
1206                         continue;
1207                 }
1208                 if (node->incoming_links.size() == 0) {
1209                         Input *input = static_cast<Input *>(node->effect);
1210                         node->output_color_space = input->get_color_space();
1211                         node->output_gamma_curve = input->get_gamma_curve();
1212
1213                         Effect::AlphaHandling alpha_handling = input->alpha_handling();
1214                         switch (alpha_handling) {
1215                         case Effect::OUTPUT_BLANK_ALPHA:
1216                                 node->output_alpha_type = ALPHA_BLANK;
1217                                 break;
1218                         case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
1219                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1220                                 break;
1221                         case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
1222                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1223                                 break;
1224                         case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
1225                         case Effect::DONT_CARE_ALPHA_TYPE:
1226                         default:
1227                                 assert(false);
1228                         }
1229
1230                         if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
1231                                 assert(node->output_gamma_curve == GAMMA_LINEAR);
1232                         }
1233                 }
1234         }
1235 }
1236
1237 // Propagate gamma and color space information as far as we can in the graph.
1238 // The rules are simple: Anything where all the inputs agree, get that as
1239 // output as well. Anything else keeps having *_INVALID.
1240 void EffectChain::propagate_gamma_and_color_space()
1241 {
1242         // We depend on going through the nodes in order.
1243         sort_all_nodes_topologically();
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->incoming_links.size() == node->effect->num_inputs());
1251                 if (node->incoming_links.size() == 0) {
1252                         assert(node->output_color_space != COLORSPACE_INVALID);
1253                         assert(node->output_gamma_curve != GAMMA_INVALID);
1254                         continue;
1255                 }
1256
1257                 Colorspace color_space = node->incoming_links[0]->output_color_space;
1258                 GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
1259                 for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
1260                         if (node->incoming_links[j]->output_color_space != color_space) {
1261                                 color_space = COLORSPACE_INVALID;
1262                         }
1263                         if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
1264                                 gamma_curve = GAMMA_INVALID;
1265                         }
1266                 }
1267
1268                 // The conversion effects already have their outputs set correctly,
1269                 // so leave them alone.
1270                 if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
1271                         node->output_color_space = color_space;
1272                 }               
1273                 if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
1274                     node->effect->effect_type_id() != "GammaExpansionEffect") {
1275                         node->output_gamma_curve = gamma_curve;
1276                 }               
1277         }
1278 }
1279
1280 // Propagate alpha information as far as we can in the graph.
1281 // Similar to propagate_gamma_and_color_space().
1282 void EffectChain::propagate_alpha()
1283 {
1284         // We depend on going through the nodes in order.
1285         sort_all_nodes_topologically();
1286
1287         for (unsigned i = 0; i < nodes.size(); ++i) {
1288                 Node *node = nodes[i];
1289                 if (node->disabled) {
1290                         continue;
1291                 }
1292                 assert(node->incoming_links.size() == node->effect->num_inputs());
1293                 if (node->incoming_links.size() == 0) {
1294                         assert(node->output_alpha_type != ALPHA_INVALID);
1295                         continue;
1296                 }
1297
1298                 // The alpha multiplication/division effects are special cases.
1299                 if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
1300                         assert(node->incoming_links.size() == 1);
1301                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
1302                         node->output_alpha_type = ALPHA_PREMULTIPLIED;
1303                         continue;
1304                 }
1305                 if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
1306                         assert(node->incoming_links.size() == 1);
1307                         assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1308                         node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1309                         continue;
1310                 }
1311
1312                 // GammaCompressionEffect and GammaExpansionEffect are also a special case,
1313                 // because they are the only one that _need_ postmultiplied alpha.
1314                 if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
1315                     node->effect->effect_type_id() == "GammaExpansionEffect") {
1316                         assert(node->incoming_links.size() == 1);
1317                         if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
1318                                 node->output_alpha_type = ALPHA_BLANK;
1319                         } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
1320                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1321                         } else {
1322                                 node->output_alpha_type = ALPHA_INVALID;
1323                         }
1324                         continue;
1325                 }
1326
1327                 // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
1328                 // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
1329                 // taken care of above. Rationale: Even if you could imagine
1330                 // e.g. an effect that took in an image and set alpha=1.0
1331                 // unconditionally, it wouldn't make any sense to have it as
1332                 // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
1333                 // got its input pre- or postmultiplied, so it wouldn't know
1334                 // whether to divide away the old alpha or not.
1335                 Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
1336                 assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1337                        alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
1338                        alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1339
1340                 // If the node has multiple inputs, check that they are all valid and
1341                 // the same.
1342                 bool any_invalid = false;
1343                 bool any_premultiplied = false;
1344                 bool any_postmultiplied = false;
1345
1346                 for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1347                         switch (node->incoming_links[j]->output_alpha_type) {
1348                         case ALPHA_INVALID:
1349                                 any_invalid = true;
1350                                 break;
1351                         case ALPHA_BLANK:
1352                                 // Blank is good as both pre- and postmultiplied alpha,
1353                                 // so just ignore it.
1354                                 break;
1355                         case ALPHA_PREMULTIPLIED:
1356                                 any_premultiplied = true;
1357                                 break;
1358                         case ALPHA_POSTMULTIPLIED:
1359                                 any_postmultiplied = true;
1360                                 break;
1361                         default:
1362                                 assert(false);
1363                         }
1364                 }
1365
1366                 if (any_invalid) {
1367                         node->output_alpha_type = ALPHA_INVALID;
1368                         continue;
1369                 }
1370
1371                 // Inputs must be of the same type.
1372                 if (any_premultiplied && any_postmultiplied) {
1373                         node->output_alpha_type = ALPHA_INVALID;
1374                         continue;
1375                 }
1376
1377                 if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
1378                     alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1379                         // This combination (requiring premultiplied alpha, but _not_ requiring
1380                         // linear light) is illegal, since the combination of premultiplied alpha
1381                         // and nonlinear inputs is meaningless.
1382                         assert(node->effect->needs_linear_light());
1383
1384                         // If the effect has asked for premultiplied alpha, check that it has got it.
1385                         if (any_postmultiplied) {
1386                                 node->output_alpha_type = ALPHA_INVALID;
1387                         } else if (!any_premultiplied &&
1388                                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
1389                                 // Blank input alpha, and the effect preserves blank alpha.
1390                                 node->output_alpha_type = ALPHA_BLANK;
1391                         } else {
1392                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1393                         }
1394                 } else {
1395                         // OK, all inputs are the same, and this effect is not going
1396                         // to change it.
1397                         assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
1398                         if (any_premultiplied) {
1399                                 node->output_alpha_type = ALPHA_PREMULTIPLIED;
1400                         } else if (any_postmultiplied) {
1401                                 node->output_alpha_type = ALPHA_POSTMULTIPLIED;
1402                         } else {
1403                                 node->output_alpha_type = ALPHA_BLANK;
1404                         }
1405                 }
1406         }
1407 }
1408
1409 bool EffectChain::node_needs_colorspace_fix(Node *node)
1410 {
1411         if (node->disabled) {
1412                 return false;
1413         }
1414         if (node->effect->num_inputs() == 0) {
1415                 return false;
1416         }
1417
1418         // propagate_gamma_and_color_space() has already set our output
1419         // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
1420         if (node->output_color_space == COLORSPACE_INVALID) {
1421                 return true;
1422         }
1423         return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
1424 }
1425
1426 // Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
1427 // the graph. Our strategy is not always optimal, but quite simple:
1428 // Find an effect that's as early as possible where the inputs are of
1429 // unacceptable colorspaces (that is, either different, or, if the effect only
1430 // wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
1431 // propagate the information anew, and repeat until there are no more such
1432 // effects.
1433 void EffectChain::fix_internal_color_spaces()
1434 {
1435         unsigned colorspace_propagation_pass = 0;
1436         bool found_any;
1437         do {
1438                 found_any = false;
1439                 for (unsigned i = 0; i < nodes.size(); ++i) {
1440                         Node *node = nodes[i];
1441                         if (!node_needs_colorspace_fix(node)) {
1442                                 continue;
1443                         }
1444
1445                         // Go through each input that is not sRGB, and insert
1446                         // a colorspace conversion after it.
1447                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1448                                 Node *input = node->incoming_links[j];
1449                                 assert(input->output_color_space != COLORSPACE_INVALID);
1450                                 if (input->output_color_space == COLORSPACE_sRGB) {
1451                                         continue;
1452                                 }
1453                                 Node *conversion = add_node(new ColorspaceConversionEffect());
1454                                 CHECK(conversion->effect->set_int("source_space", input->output_color_space));
1455                                 CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
1456                                 conversion->output_color_space = COLORSPACE_sRGB;
1457                                 replace_sender(input, conversion);
1458                                 connect_nodes(input, conversion);
1459                         }
1460
1461                         // Re-sort topologically, and propagate the new information.
1462                         propagate_gamma_and_color_space();
1463                         
1464                         found_any = true;
1465                         break;
1466                 }
1467         
1468                 char filename[256];
1469                 sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
1470                 output_dot(filename);
1471                 assert(colorspace_propagation_pass < 100);
1472         } while (found_any);
1473
1474         for (unsigned i = 0; i < nodes.size(); ++i) {
1475                 Node *node = nodes[i];
1476                 if (node->disabled) {
1477                         continue;
1478                 }
1479                 assert(node->output_color_space != COLORSPACE_INVALID);
1480         }
1481 }
1482
1483 bool EffectChain::node_needs_alpha_fix(Node *node)
1484 {
1485         if (node->disabled) {
1486                 return false;
1487         }
1488
1489         // propagate_alpha() has already set our output to ALPHA_INVALID if the
1490         // inputs differ or we are otherwise in mismatch, so we can rely on that.
1491         return (node->output_alpha_type == ALPHA_INVALID);
1492 }
1493
1494 // Fix up alpha so that there are no ALPHA_INVALID nodes left in
1495 // the graph. Similar to fix_internal_color_spaces().
1496 void EffectChain::fix_internal_alpha(unsigned step)
1497 {
1498         unsigned alpha_propagation_pass = 0;
1499         bool found_any;
1500         do {
1501                 found_any = false;
1502                 for (unsigned i = 0; i < nodes.size(); ++i) {
1503                         Node *node = nodes[i];
1504                         if (!node_needs_alpha_fix(node)) {
1505                                 continue;
1506                         }
1507
1508                         // If we need to fix up GammaExpansionEffect, then clearly something
1509                         // is wrong, since the combination of premultiplied alpha and nonlinear inputs
1510                         // is meaningless.
1511                         assert(node->effect->effect_type_id() != "GammaExpansionEffect");
1512
1513                         AlphaType desired_type = ALPHA_PREMULTIPLIED;
1514
1515                         // GammaCompressionEffect is special; it needs postmultiplied alpha.
1516                         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1517                                 assert(node->incoming_links.size() == 1);
1518                                 assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
1519                                 desired_type = ALPHA_POSTMULTIPLIED;
1520                         }
1521
1522                         // Go through each input that is not premultiplied alpha, and insert
1523                         // a conversion before it.
1524                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1525                                 Node *input = node->incoming_links[j];
1526                                 assert(input->output_alpha_type != ALPHA_INVALID);
1527                                 if (input->output_alpha_type == desired_type ||
1528                                     input->output_alpha_type == ALPHA_BLANK) {
1529                                         continue;
1530                                 }
1531                                 Node *conversion;
1532                                 if (desired_type == ALPHA_PREMULTIPLIED) {
1533                                         conversion = add_node(new AlphaMultiplicationEffect());
1534                                 } else {
1535                                         conversion = add_node(new AlphaDivisionEffect());
1536                                 }
1537                                 conversion->output_alpha_type = desired_type;
1538                                 replace_sender(input, conversion);
1539                                 connect_nodes(input, conversion);
1540                         }
1541
1542                         // Re-sort topologically, and propagate the new information.
1543                         propagate_gamma_and_color_space();
1544                         propagate_alpha();
1545                         
1546                         found_any = true;
1547                         break;
1548                 }
1549         
1550                 char filename[256];
1551                 sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
1552                 output_dot(filename);
1553                 assert(alpha_propagation_pass < 100);
1554         } while (found_any);
1555
1556         for (unsigned i = 0; i < nodes.size(); ++i) {
1557                 Node *node = nodes[i];
1558                 if (node->disabled) {
1559                         continue;
1560                 }
1561                 assert(node->output_alpha_type != ALPHA_INVALID);
1562         }
1563 }
1564
1565 // Make so that the output is in the desired color space.
1566 void EffectChain::fix_output_color_space()
1567 {
1568         Node *output = find_output_node();
1569         if (output->output_color_space != output_format.color_space) {
1570                 Node *conversion = add_node(new ColorspaceConversionEffect());
1571                 CHECK(conversion->effect->set_int("source_space", output->output_color_space));
1572                 CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
1573                 conversion->output_color_space = output_format.color_space;
1574                 connect_nodes(output, conversion);
1575                 propagate_alpha();
1576                 propagate_gamma_and_color_space();
1577         }
1578 }
1579
1580 // Make so that the output is in the desired pre-/postmultiplication alpha state.
1581 void EffectChain::fix_output_alpha()
1582 {
1583         Node *output = find_output_node();
1584         assert(output->output_alpha_type != ALPHA_INVALID);
1585         if (output->output_alpha_type == ALPHA_BLANK) {
1586                 // No alpha output, so we don't care.
1587                 return;
1588         }
1589         if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
1590             output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
1591                 Node *conversion = add_node(new AlphaDivisionEffect());
1592                 connect_nodes(output, conversion);
1593                 propagate_alpha();
1594                 propagate_gamma_and_color_space();
1595         }
1596         if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
1597             output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
1598                 Node *conversion = add_node(new AlphaMultiplicationEffect());
1599                 connect_nodes(output, conversion);
1600                 propagate_alpha();
1601                 propagate_gamma_and_color_space();
1602         }
1603 }
1604
1605 bool EffectChain::node_needs_gamma_fix(Node *node)
1606 {
1607         if (node->disabled) {
1608                 return false;
1609         }
1610
1611         // Small hack since the output is not an explicit node:
1612         // If we are the last node and our output is in the wrong
1613         // space compared to EffectChain's output, we need to fix it.
1614         // This will only take us to linear, but fix_output_gamma()
1615         // will come and take us to the desired output gamma
1616         // if it is needed.
1617         //
1618         // This needs to be before everything else, since it could
1619         // even apply to inputs (if they are the only effect).
1620         if (node->outgoing_links.empty() &&
1621             node->output_gamma_curve != output_format.gamma_curve &&
1622             node->output_gamma_curve != GAMMA_LINEAR) {
1623                 return true;
1624         }
1625
1626         if (node->effect->num_inputs() == 0) {
1627                 return false;
1628         }
1629
1630         // propagate_gamma_and_color_space() has already set our output
1631         // to GAMMA_INVALID if the inputs differ, so we can rely on that,
1632         // except for GammaCompressionEffect.
1633         if (node->output_gamma_curve == GAMMA_INVALID) {
1634                 return true;
1635         }
1636         if (node->effect->effect_type_id() == "GammaCompressionEffect") {
1637                 assert(node->incoming_links.size() == 1);
1638                 return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
1639         }
1640
1641         return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
1642 }
1643
1644 // Very similar to fix_internal_color_spaces(), but for gamma.
1645 // There is one difference, though; before we start adding conversion nodes,
1646 // we see if we can get anything out of asking the sources to deliver
1647 // linear gamma directly. fix_internal_gamma_by_asking_inputs()
1648 // does that part, while fix_internal_gamma_by_inserting_nodes()
1649 // inserts nodes as needed afterwards.
1650 void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
1651 {
1652         unsigned gamma_propagation_pass = 0;
1653         bool found_any;
1654         do {
1655                 found_any = false;
1656                 for (unsigned i = 0; i < nodes.size(); ++i) {
1657                         Node *node = nodes[i];
1658                         if (!node_needs_gamma_fix(node)) {
1659                                 continue;
1660                         }
1661
1662                         // See if all inputs can give us linear gamma. If not, leave it.
1663                         vector<Node *> nonlinear_inputs;
1664                         find_all_nonlinear_inputs(node, &nonlinear_inputs);
1665                         assert(!nonlinear_inputs.empty());
1666
1667                         bool all_ok = true;
1668                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1669                                 Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
1670                                 all_ok &= input->can_output_linear_gamma();
1671                         }
1672
1673                         if (!all_ok) {
1674                                 continue;
1675                         }
1676
1677                         for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
1678                                 CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
1679                                 nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
1680                         }
1681
1682                         // Re-sort topologically, and propagate the new information.
1683                         propagate_gamma_and_color_space();
1684                         
1685                         found_any = true;
1686                         break;
1687                 }
1688         
1689                 char filename[256];
1690                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1691                 output_dot(filename);
1692                 assert(gamma_propagation_pass < 100);
1693         } while (found_any);
1694 }
1695
1696 void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
1697 {
1698         unsigned gamma_propagation_pass = 0;
1699         bool found_any;
1700         do {
1701                 found_any = false;
1702                 for (unsigned i = 0; i < nodes.size(); ++i) {
1703                         Node *node = nodes[i];
1704                         if (!node_needs_gamma_fix(node)) {
1705                                 continue;
1706                         }
1707
1708                         // Special case: We could be an input and still be asked to
1709                         // fix our gamma; if so, we should be the only node
1710                         // (as node_needs_gamma_fix() would only return true in
1711                         // for an input in that case). That means we should insert
1712                         // a conversion node _after_ ourselves.
1713                         if (node->incoming_links.empty()) {
1714                                 assert(node->outgoing_links.empty());
1715                                 Node *conversion = add_node(new GammaExpansionEffect());
1716                                 CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
1717                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1718                                 connect_nodes(node, conversion);
1719                         }
1720
1721                         // If not, go through each input that is not linear gamma,
1722                         // and insert a gamma conversion after it.
1723                         for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
1724                                 Node *input = node->incoming_links[j];
1725                                 assert(input->output_gamma_curve != GAMMA_INVALID);
1726                                 if (input->output_gamma_curve == GAMMA_LINEAR) {
1727                                         continue;
1728                                 }
1729                                 Node *conversion = add_node(new GammaExpansionEffect());
1730                                 CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
1731                                 conversion->output_gamma_curve = GAMMA_LINEAR;
1732                                 replace_sender(input, conversion);
1733                                 connect_nodes(input, conversion);
1734                         }
1735
1736                         // Re-sort topologically, and propagate the new information.
1737                         propagate_alpha();
1738                         propagate_gamma_and_color_space();
1739                         
1740                         found_any = true;
1741                         break;
1742                 }
1743         
1744                 char filename[256];
1745                 sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
1746                 output_dot(filename);
1747                 assert(gamma_propagation_pass < 100);
1748         } while (found_any);
1749
1750         for (unsigned i = 0; i < nodes.size(); ++i) {
1751                 Node *node = nodes[i];
1752                 if (node->disabled) {
1753                         continue;
1754                 }
1755                 assert(node->output_gamma_curve != GAMMA_INVALID);
1756         }
1757 }
1758
1759 // Make so that the output is in the desired gamma.
1760 // Note that this assumes linear input gamma, so it might create the need
1761 // for another pass of fix_internal_gamma().
1762 void EffectChain::fix_output_gamma()
1763 {
1764         Node *output = find_output_node();
1765         if (output->output_gamma_curve != output_format.gamma_curve) {
1766                 Node *conversion = add_node(new GammaCompressionEffect());
1767                 CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
1768                 conversion->output_gamma_curve = output_format.gamma_curve;
1769                 connect_nodes(output, conversion);
1770         }
1771 }
1772
1773 // If the user has requested Y'CbCr output, we need to do this conversion
1774 // _after_ GammaCompressionEffect etc., but before dither (see below).
1775 // This is because Y'CbCr, with the exception of a special optional mode
1776 // in Rec. 2020 (which we currently don't support), is defined to work on
1777 // gamma-encoded data.
1778 void EffectChain::add_ycbcr_conversion_if_needed()
1779 {
1780         assert(output_color_rgba || num_output_color_ycbcr > 0);
1781         if (num_output_color_ycbcr == 0) {
1782                 return;
1783         }
1784         Node *output = find_output_node();
1785         ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type));
1786         connect_nodes(output, ycbcr_conversion_effect_node);
1787 }
1788         
1789 // If the user has requested dither, add a DitherEffect right at the end
1790 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
1791 // since dither is about the only effect that can _not_ be done in linear space.
1792 void EffectChain::add_dither_if_needed()
1793 {
1794         if (num_dither_bits == 0) {
1795                 return;
1796         }
1797         Node *output = find_output_node();
1798         Node *dither = add_node(new DitherEffect());
1799         CHECK(dither->effect->set_int("num_bits", num_dither_bits));
1800         connect_nodes(output, dither);
1801
1802         dither_effect = dither->effect;
1803 }
1804
1805 namespace {
1806
1807 // Whether this effect will cause the phase it is in to become a compute shader phase.
1808 bool induces_compute_shader(Node *node)
1809 {
1810         if (node->effect->is_compute_shader()) {
1811                 return true;
1812         }
1813         if (!node->effect->strong_one_to_one_sampling()) {
1814                 // This effect can't be chained after a compute shader.
1815                 return false;
1816         }
1817         // If at least one of the effects we depend on is a compute shader,
1818         // one of them will be put in the same phase as us (the other ones,
1819         // if any, will be bounced).
1820         for (Node *dep : node->incoming_links) {
1821                 if (induces_compute_shader(dep)) {
1822                         return true;
1823                 }
1824         }
1825         return false;
1826 }
1827
1828 }  // namespace
1829
1830 // Compute shaders can't output to the framebuffer, so if the last
1831 // phase ends in a compute shader, add a dummy phase at the end that
1832 // only blits directly from the temporary texture.
1833 void EffectChain::add_dummy_effect_if_needed()
1834 {
1835         Node *output = find_output_node();
1836         if (induces_compute_shader(output)) {
1837                 Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
1838                 connect_nodes(output, dummy);
1839                 has_dummy_effect = true;
1840         }
1841 }
1842
1843 // Find the output node. This is, simply, one that has no outgoing links.
1844 // If there are multiple ones, the graph is malformed (we do not support
1845 // multiple outputs right now).
1846 Node *EffectChain::find_output_node()
1847 {
1848         vector<Node *> output_nodes;
1849         for (unsigned i = 0; i < nodes.size(); ++i) {
1850                 Node *node = nodes[i];
1851                 if (node->disabled) {
1852                         continue;
1853                 }
1854                 if (node->outgoing_links.empty()) {
1855                         output_nodes.push_back(node);
1856                 }
1857         }
1858         assert(output_nodes.size() == 1);
1859         return output_nodes[0];
1860 }
1861
1862 void EffectChain::finalize()
1863 {
1864         // Output the graph as it is before we do any conversions on it.
1865         output_dot("step0-start.dot");
1866
1867         // Give each effect in turn a chance to rewrite its own part of the graph.
1868         // Note that if more effects are added as part of this, they will be
1869         // picked up as part of the same for loop, since they are added at the end.
1870         for (unsigned i = 0; i < nodes.size(); ++i) {
1871                 nodes[i]->effect->rewrite_graph(this, nodes[i]);
1872         }
1873         output_dot("step1-rewritten.dot");
1874
1875         find_color_spaces_for_inputs();
1876         output_dot("step2-input-colorspace.dot");
1877
1878         propagate_alpha();
1879         output_dot("step3-propagated-alpha.dot");
1880
1881         propagate_gamma_and_color_space();
1882         output_dot("step4-propagated-all.dot");
1883
1884         fix_internal_color_spaces();
1885         fix_internal_alpha(6);
1886         fix_output_color_space();
1887         output_dot("step7-output-colorspacefix.dot");
1888         fix_output_alpha();
1889         output_dot("step8-output-alphafix.dot");
1890
1891         // Note that we need to fix gamma after colorspace conversion,
1892         // because colorspace conversions might create needs for gamma conversions.
1893         // Also, we need to run an extra pass of fix_internal_gamma() after 
1894         // fixing the output gamma, as we only have conversions to/from linear,
1895         // and fix_internal_alpha() since GammaCompressionEffect needs
1896         // postmultiplied input.
1897         fix_internal_gamma_by_asking_inputs(9);
1898         fix_internal_gamma_by_inserting_nodes(10);
1899         fix_output_gamma();
1900         output_dot("step11-output-gammafix.dot");
1901         propagate_alpha();
1902         output_dot("step12-output-alpha-propagated.dot");
1903         fix_internal_alpha(13);
1904         output_dot("step14-output-alpha-fixed.dot");
1905         fix_internal_gamma_by_asking_inputs(15);
1906         fix_internal_gamma_by_inserting_nodes(16);
1907
1908         output_dot("step17-before-ycbcr.dot");
1909         add_ycbcr_conversion_if_needed();
1910
1911         output_dot("step18-before-dither.dot");
1912         add_dither_if_needed();
1913
1914         output_dot("step19-before-dummy-effect.dot");
1915         add_dummy_effect_if_needed();
1916
1917         output_dot("step20-final.dot");
1918         
1919         // Construct all needed GLSL programs, starting at the output.
1920         // We need to keep track of which effects have already been computed,
1921         // as an effect with multiple users could otherwise be calculated
1922         // multiple times.
1923         map<Node *, Phase *> completed_effects;
1924         construct_phase(find_output_node(), &completed_effects);
1925
1926         output_dot("step21-split-to-phases.dot");
1927
1928         // There are some corner cases where we thought we needed to add a dummy
1929         // effect, but then it turned out later we didn't (e.g. induces_compute_shader()
1930         // didn't see a mipmap conflict coming, which would cause the compute shader
1931         // to be split off from the inal phase); if so, remove the extra phase
1932         // at the end, since it will give us some trouble during execution.
1933         //
1934         // TODO: Remove induces_compute_shader() and replace it with precise tracking.
1935         if (has_dummy_effect && !phases[phases.size() - 2]->is_compute_shader) {
1936                 resource_pool->release_glsl_program(phases.back()->glsl_program_num);
1937                 delete phases.back();
1938                 phases.pop_back();
1939                 has_dummy_effect = false;
1940         }
1941
1942         output_dot("step22-dummy-phase-removal.dot");
1943
1944         assert(phases[0]->inputs.empty());
1945         
1946         finalized = true;
1947 }
1948
1949 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
1950 {
1951         // Save original viewport.
1952         GLuint x = 0, y = 0;
1953
1954         if (width == 0 && height == 0) {
1955                 GLint viewport[4];
1956                 glGetIntegerv(GL_VIEWPORT, viewport);
1957                 x = viewport[0];
1958                 y = viewport[1];
1959                 width = viewport[2];
1960                 height = viewport[3];
1961         }
1962
1963         render(dest_fbo, {}, x, y, width, height);
1964 }
1965
1966 void EffectChain::render_to_texture(const vector<DestinationTexture> &destinations, unsigned width, unsigned height)
1967 {
1968         assert(finalized);
1969         assert(!destinations.empty());
1970
1971         if (!has_dummy_effect) {
1972                 // We don't end in a compute shader, so there's nothing specific for us to do.
1973                 // Create an FBO for this set of textures, and just render to that.
1974                 GLuint texnums[4] = { 0, 0, 0, 0 };
1975                 for (unsigned i = 0; i < destinations.size() && i < 4; ++i) {
1976                         texnums[i] = destinations[i].texnum;
1977                 }
1978                 GLuint dest_fbo = resource_pool->create_fbo(texnums[0], texnums[1], texnums[2], texnums[3]);
1979                 render(dest_fbo, {}, 0, 0, width, height);
1980                 resource_pool->release_fbo(dest_fbo);
1981         } else {
1982                 render((GLuint)-1, destinations, 0, 0, width, height);
1983         }
1984 }
1985
1986 void EffectChain::render(GLuint dest_fbo, const vector<DestinationTexture> &destinations, unsigned x, unsigned y, unsigned width, unsigned height)
1987 {
1988         assert(finalized);
1989         assert(destinations.size() <= 1);
1990
1991         // This needs to be set anew, in case we are coming from a different context
1992         // from when we initialized.
1993         check_error();
1994         glDisable(GL_DITHER);
1995         check_error();
1996
1997         const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB);
1998         check_error();
1999         bool current_srgb = final_srgb;
2000
2001         // Basic state.
2002         check_error();
2003         glDisable(GL_BLEND);
2004         check_error();
2005         glDisable(GL_DEPTH_TEST);
2006         check_error();
2007         glDepthMask(GL_FALSE);
2008         check_error();
2009
2010         set<Phase *> generated_mipmaps;
2011
2012         // We keep one texture per output, but only for as long as we actually have any
2013         // phases that need it as an input. (We don't make any effort to reorder phases
2014         // to minimize the number of textures in play, as register allocation can be
2015         // complicated and we rarely have much to gain, since our graphs are typically
2016         // pretty linear.)
2017         map<Phase *, GLuint> output_textures;
2018         map<Phase *, int> ref_counts;
2019         for (Phase *phase : phases) {
2020                 for (Phase *input : phase->inputs) {
2021                         ++ref_counts[input];
2022                 }
2023         }
2024
2025         size_t num_phases = phases.size();
2026         if (destinations.empty()) {
2027                 assert(dest_fbo != (GLuint)-1);
2028         } else {
2029                 assert(has_dummy_effect);
2030                 assert(x == 0);
2031                 assert(y == 0);
2032                 assert(num_phases >= 2);
2033                 assert(!phases.back()->is_compute_shader);
2034                 assert(phases[phases.size() - 2]->is_compute_shader);
2035                 assert(phases.back()->effects.size() == 1);
2036                 assert(phases.back()->effects[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
2037
2038                 // We are rendering to a set of textures, so we can run the compute shader
2039                 // directly and skip the dummy phase.
2040                 --num_phases;
2041         }
2042
2043         for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) {
2044                 Phase *phase = phases[phase_num];
2045
2046                 if (do_phase_timing) {
2047                         GLuint timer_query_object;
2048                         if (phase->timer_query_objects_free.empty()) {
2049                                 glGenQueries(1, &timer_query_object);
2050                         } else {
2051                                 timer_query_object = phase->timer_query_objects_free.front();
2052                                 phase->timer_query_objects_free.pop_front();
2053                         }
2054                         glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
2055                         phase->timer_query_objects_running.push_back(timer_query_object);
2056                 }
2057                 bool last_phase = (phase_num == num_phases - 1);
2058                 if (last_phase) {
2059                         // Last phase goes to the output the user specified.
2060                         if (!phase->is_compute_shader) {
2061                                 assert(dest_fbo != (GLuint)-1);
2062                                 glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
2063                                 check_error();
2064                                 GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
2065                                 assert(status == GL_FRAMEBUFFER_COMPLETE);
2066                                 glViewport(x, y, width, height);
2067                         }
2068                         if (dither_effect != nullptr) {
2069                                 CHECK(dither_effect->set_int("output_width", width));
2070                                 CHECK(dither_effect->set_int("output_height", height));
2071                         }
2072                 }
2073
2074                 // Enable sRGB rendering for intermediates in case we are
2075                 // rendering to an sRGB format.
2076                 // TODO: Support this for compute shaders.
2077                 bool needs_srgb = last_phase ? final_srgb : true;
2078                 if (needs_srgb && !current_srgb) {
2079                         glEnable(GL_FRAMEBUFFER_SRGB);
2080                         check_error();
2081                         current_srgb = true;
2082                 } else if (!needs_srgb && current_srgb) {
2083                         glDisable(GL_FRAMEBUFFER_SRGB);
2084                         check_error();
2085                         current_srgb = true;
2086                 }
2087
2088                 // Find a texture for this phase.
2089                 inform_input_sizes(phase);
2090                 find_output_size(phase);
2091                 vector<DestinationTexture> phase_destinations;
2092                 if (!last_phase) {
2093                         GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
2094                         output_textures.insert(make_pair(phase, tex_num));
2095                         phase_destinations.push_back(DestinationTexture{ tex_num, intermediate_format });
2096
2097                         // The output texture needs to have valid state to be written to by a compute shader.
2098                         glActiveTexture(GL_TEXTURE0);
2099                         check_error();
2100                         glBindTexture(GL_TEXTURE_2D, tex_num);
2101                         check_error();
2102                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2103                         check_error();
2104                 } else if (phase->is_compute_shader) {
2105                         assert(!destinations.empty());
2106                         phase_destinations = destinations;
2107                 }
2108
2109                 execute_phase(phase, output_textures, phase_destinations, &generated_mipmaps);
2110                 if (do_phase_timing) {
2111                         glEndQuery(GL_TIME_ELAPSED);
2112                 }
2113
2114                 // Drop any input textures we don't need anymore.
2115                 for (Phase *input : phase->inputs) {
2116                         assert(ref_counts[input] > 0);
2117                         if (--ref_counts[input] == 0) {
2118                                 resource_pool->release_2d_texture(output_textures[input]);
2119                                 output_textures.erase(input);
2120                         }
2121                 }
2122         }
2123
2124         for (const auto &phase_and_texnum : output_textures) {
2125                 resource_pool->release_2d_texture(phase_and_texnum.second);
2126         }
2127
2128         glBindFramebuffer(GL_FRAMEBUFFER, 0);
2129         check_error();
2130         glUseProgram(0);
2131         check_error();
2132
2133         glBindBuffer(GL_ARRAY_BUFFER, 0);
2134         check_error();
2135         glBindVertexArray(0);
2136         check_error();
2137
2138         if (do_phase_timing) {
2139                 // Get back the timer queries.
2140                 for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2141                         Phase *phase = phases[phase_num];
2142                         for (auto timer_it = phase->timer_query_objects_running.cbegin();
2143                              timer_it != phase->timer_query_objects_running.cend(); ) {
2144                                 GLint timer_query_object = *timer_it;
2145                                 GLint available;
2146                                 glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
2147                                 if (available) {
2148                                         GLuint64 time_elapsed;
2149                                         glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
2150                                         phase->time_elapsed_ns += time_elapsed;
2151                                         ++phase->num_measured_iterations;
2152                                         phase->timer_query_objects_free.push_back(timer_query_object);
2153                                         phase->timer_query_objects_running.erase(timer_it++);
2154                                 } else {
2155                                         ++timer_it;
2156                                 }
2157                         }
2158                 }
2159         }
2160 }
2161
2162 void EffectChain::enable_phase_timing(bool enable)
2163 {
2164         if (enable) {
2165                 assert(movit_timer_queries_supported);
2166         }
2167         this->do_phase_timing = enable;
2168 }
2169
2170 void EffectChain::reset_phase_timing()
2171 {
2172         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2173                 Phase *phase = phases[phase_num];
2174                 phase->time_elapsed_ns = 0;
2175                 phase->num_measured_iterations = 0;
2176         }
2177 }
2178
2179 void EffectChain::print_phase_timing()
2180 {
2181         double total_time_ms = 0.0;
2182         for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
2183                 Phase *phase = phases[phase_num];
2184                 double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
2185                 printf("Phase %d: %5.1f ms  [", phase_num, avg_time_ms);
2186                 for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
2187                         if (effect_num != 0) {
2188                                 printf(", ");
2189                         }
2190                         printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
2191                 }
2192                 printf("]\n");
2193                 total_time_ms += avg_time_ms;
2194         }
2195         printf("Total:   %5.1f ms\n", total_time_ms);
2196 }
2197
2198 void EffectChain::execute_phase(Phase *phase,
2199                                 const map<Phase *, GLuint> &output_textures,
2200                                 const vector<DestinationTexture> &destinations,
2201                                 set<Phase *> *generated_mipmaps)
2202 {
2203         // Set up RTT inputs for this phase.
2204         for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
2205                 glActiveTexture(GL_TEXTURE0 + sampler);
2206                 Phase *input = phase->inputs[sampler];
2207                 input->output_node->bound_sampler_num = sampler;
2208                 const auto it = output_textures.find(input);
2209                 assert(it != output_textures.end());
2210                 glBindTexture(GL_TEXTURE_2D, it->second);
2211                 check_error();
2212
2213                 // See if anything using this RTT input (in this phase) needs mipmaps.
2214                 // TODO: It could be that we get conflicting logic here, if we have
2215                 // multiple effects with incompatible mipmaps using the same
2216                 // RTT input. However, that is obscure enough that we can deal
2217                 // with it at some future point (preferably when we have
2218                 // universal support for separate sampler objects!). For now,
2219                 // an assert is good enough. See also the TODO at bound_sampler_num.
2220                 bool any_needs_mipmaps = false, any_refuses_mipmaps = false;
2221                 for (Node *node : phase->effects) {
2222                         assert(node->incoming_links.size() == node->incoming_link_type.size());
2223                         for (size_t i = 0; i < node->incoming_links.size(); ++i) {
2224                                 if (node->incoming_links[i] == input->output_node &&
2225                                     node->incoming_link_type[i] == IN_ANOTHER_PHASE) {
2226                                         if (node->needs_mipmaps == Effect::NEEDS_MIPMAPS) {
2227                                                 any_needs_mipmaps = true;
2228                                         } else if (node->needs_mipmaps == Effect::CANNOT_ACCEPT_MIPMAPS) {
2229                                                 any_refuses_mipmaps = true;
2230                                         }
2231                                 }
2232                         }
2233                 }
2234                 assert(!(any_needs_mipmaps && any_refuses_mipmaps));
2235
2236                 if (any_needs_mipmaps && generated_mipmaps->count(input) == 0) {
2237                         glGenerateMipmap(GL_TEXTURE_2D);
2238                         check_error();
2239                         generated_mipmaps->insert(input);
2240                 }
2241                 setup_rtt_sampler(sampler, any_needs_mipmaps);
2242                 phase->input_samplers[sampler] = sampler;  // Bind the sampler to the right uniform.
2243         }
2244
2245         GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
2246         check_error();
2247
2248         // And now the output.
2249         GLuint fbo = 0;
2250         if (phase->is_compute_shader) {
2251                 assert(!destinations.empty());
2252
2253                 // This is currently the only place where we use image units,
2254                 // so we can always start at 0. TODO: Support multiple destinations.
2255                 phase->outbuf_image_unit = 0;
2256                 glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format);
2257                 check_error();
2258                 phase->uniform_output_size[0] = phase->output_width;
2259                 phase->uniform_output_size[1] = phase->output_height;
2260                 phase->inv_output_size.x = 1.0f / phase->output_width;
2261                 phase->inv_output_size.y = 1.0f / phase->output_height;
2262                 phase->output_texcoord_adjust.x = 0.5f / phase->output_width;
2263                 phase->output_texcoord_adjust.y = 0.5f / phase->output_height;
2264         } else if (!destinations.empty()) {
2265                 assert(destinations.size() == 1);
2266                 fbo = resource_pool->create_fbo(destinations[0].texnum);
2267                 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
2268                 glViewport(0, 0, phase->output_width, phase->output_height);
2269         }
2270
2271         // Give the required parameters to all the effects.
2272         unsigned sampler_num = phase->inputs.size();
2273         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2274                 Node *node = phase->effects[i];
2275                 unsigned old_sampler_num = sampler_num;
2276                 node->effect->set_gl_state(instance_program_num, phase->effect_ids[make_pair(node, IN_SAME_PHASE)], &sampler_num);
2277                 check_error();
2278
2279                 if (node->effect->is_single_texture()) {
2280                         assert(sampler_num - old_sampler_num == 1);
2281                         node->bound_sampler_num = old_sampler_num;
2282                 } else {
2283                         node->bound_sampler_num = -1;
2284                 }
2285         }
2286
2287         if (phase->is_compute_shader) {
2288                 unsigned x, y, z;
2289                 phase->compute_shader_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z);
2290
2291                 // Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(),
2292                 // since they can be updated from there.
2293                 setup_uniforms(phase);
2294                 glDispatchCompute(x, y, z);
2295                 check_error();
2296                 glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT | GL_TEXTURE_UPDATE_BARRIER_BIT);
2297                 check_error();
2298         } else {
2299                 // Uniforms need to come after set_gl_state(), since they can be updated
2300                 // from there.
2301                 setup_uniforms(phase);
2302
2303                 // Bind the vertex data.
2304                 GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
2305                 glBindVertexArray(vao);
2306
2307                 glDrawArrays(GL_TRIANGLES, 0, 3);
2308                 check_error();
2309
2310                 resource_pool->release_vec2_vao(vao);
2311         }
2312         
2313         for (unsigned i = 0; i < phase->effects.size(); ++i) {
2314                 Node *node = phase->effects[i];
2315                 node->effect->clear_gl_state();
2316         }
2317
2318         resource_pool->unuse_glsl_program(instance_program_num);
2319
2320         if (fbo != 0) {
2321                 resource_pool->release_fbo(fbo);
2322         }
2323 }
2324
2325 void EffectChain::setup_uniforms(Phase *phase)
2326 {
2327         // TODO: Use UBO blocks.
2328         for (size_t i = 0; i < phase->uniforms_image2d.size(); ++i) {
2329                 const Uniform<int> &uniform = phase->uniforms_image2d[i];
2330                 if (uniform.location != -1) {
2331                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2332                 }
2333         }
2334         for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
2335                 const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
2336                 if (uniform.location != -1) {
2337                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2338                 }
2339         }
2340         for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
2341                 const Uniform<bool> &uniform = phase->uniforms_bool[i];
2342                 assert(uniform.num_values == 1);
2343                 if (uniform.location != -1) {
2344                         glUniform1i(uniform.location, *uniform.value);
2345                 }
2346         }
2347         for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
2348                 const Uniform<int> &uniform = phase->uniforms_int[i];
2349                 if (uniform.location != -1) {
2350                         glUniform1iv(uniform.location, uniform.num_values, uniform.value);
2351                 }
2352         }
2353         for (size_t i = 0; i < phase->uniforms_ivec2.size(); ++i) {
2354                 const Uniform<int> &uniform = phase->uniforms_ivec2[i];
2355                 if (uniform.location != -1) {
2356                         glUniform2iv(uniform.location, uniform.num_values, uniform.value);
2357                 }
2358         }
2359         for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
2360                 const Uniform<float> &uniform = phase->uniforms_float[i];
2361                 if (uniform.location != -1) {
2362                         glUniform1fv(uniform.location, uniform.num_values, uniform.value);
2363                 }
2364         }
2365         for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
2366                 const Uniform<float> &uniform = phase->uniforms_vec2[i];
2367                 if (uniform.location != -1) {
2368                         glUniform2fv(uniform.location, uniform.num_values, uniform.value);
2369                 }
2370         }
2371         for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
2372                 const Uniform<float> &uniform = phase->uniforms_vec3[i];
2373                 if (uniform.location != -1) {
2374                         glUniform3fv(uniform.location, uniform.num_values, uniform.value);
2375                 }
2376         }
2377         for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
2378                 const Uniform<float> &uniform = phase->uniforms_vec4[i];
2379                 if (uniform.location != -1) {
2380                         glUniform4fv(uniform.location, uniform.num_values, uniform.value);
2381                 }
2382         }
2383         for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
2384                 const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
2385                 assert(uniform.num_values == 1);
2386                 if (uniform.location != -1) {
2387                         // Convert to float (GLSL has no double matrices).
2388                         float matrixf[9];
2389                         for (unsigned y = 0; y < 3; ++y) {
2390                                 for (unsigned x = 0; x < 3; ++x) {
2391                                         matrixf[y + x * 3] = (*uniform.value)(y, x);
2392                                 }
2393                         }
2394                         glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
2395                 }
2396         }
2397 }
2398
2399 void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
2400 {
2401         glActiveTexture(GL_TEXTURE0 + sampler_num);
2402         check_error();
2403         if (use_mipmaps) {
2404                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
2405                 check_error();
2406         } else {
2407                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
2408                 check_error();
2409         }
2410         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2411         check_error();
2412         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2413         check_error();
2414 }
2415
2416 }  // namespace movit