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