1 #ifndef _MOVIT_EFFECT_CHAIN_H
2 #define _MOVIT_EFFECT_CHAIN_H 1
4 // An EffectChain is the largest basic entity in Movit; it contains everything
5 // needed to connects a series of effects, from inputs to outputs, and render
6 // them. Generally you set up your effect chain once and then call its render
7 // functions once per frame; setting one up can be relatively expensive,
8 // but rendering is fast.
10 // Threading considerations: EffectChain is “thread-compatible”; you can use
11 // different EffectChains in multiple threads at the same time (assuming the
12 // threads do not use the same OpenGL context, but this is a good idea anyway),
13 // but you may not use one EffectChain from multiple threads simultaneously.
14 // You _are_ allowed to use one EffectChain from multiple threads as long as
15 // you only use it from one at a time (possibly by doing your own locking),
16 // but if so, the threads' contexts need to be set up to share resources, since
17 // the EffectChain holds textures and other OpenGL objects that are tied to the
20 // Memory management (only relevant if you use multiple contexts):
21 // See corresponding comment in resource_pool.h. This holds even if you don't
22 // allocate your own ResourcePool, but let EffectChain hold its own.
34 #include "image_format.h"
44 // For internal use within Node.
52 // Whether you want pre- or postmultiplied alpha in the output
53 // (see effect.h for a discussion of pre- versus postmultiplied alpha).
54 enum OutputAlphaFormat {
55 OUTPUT_ALPHA_FORMAT_PREMULTIPLIED,
56 OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED,
59 // RGBA output is nearly always packed; Y'CbCr, however, is often planar
60 // due to chroma subsampling. This enum controls how add_ycbcr_output()
61 // distributes the color channels between the fragment shader outputs.
62 // Obviously, anything except YCBCR_OUTPUT_INTERLEAVED will be meaningless
63 // unless you use render_to_fbo() and have an FBO with multiple render
64 // targets attached (the other outputs will be discarded).
65 enum YCbCrOutputSplitting {
66 // Only one output: Store Y'CbCr into the first three output channels,
67 // respectively, plus alpha. This is also called “chunked” or
69 YCBCR_OUTPUT_INTERLEAVED,
71 // Store Y' and alpha into the first output (in the red and alpha
72 // channels; effect to the others is undefined), and Cb and Cr into
73 // the first two channels of the second output. This is particularly
74 // useful if you want to end up in a format like NV12, where all the
75 // Y' samples come first and then Cb and Cr come interlevaed afterwards.
76 // You will still need to do the chroma subsampling yourself to actually
77 // get down to NV12, though.
78 YCBCR_OUTPUT_SPLIT_Y_AND_CBCR,
80 // Store Y' and alpha into the first output, Cb into the first channel
81 // of the second output and Cr into the first channel of the third output.
82 // (Effect on the other channels is undefined.) Essentially gives you
83 // 4:4:4 planar, or ”yuv444p”.
87 // Where (0,0) is taken to be in the output. If you want to render to an
88 // OpenGL screen, you should keep the default of bottom-left, as that is
89 // OpenGL's natural coordinate system. However, there are cases, such as if you
90 // render to an FBO and read the pixels back into some other system, where
91 // you'd want a top-left origin; if so, an additional flip step will be added
92 // at the very end (but done in a vertex shader, so it will have zero extra
95 // Note that Movit's coordinate system in general consistently puts (0,0) in
96 // the top left for _input_, no matter what you set as output origin.
98 OUTPUT_ORIGIN_BOTTOM_LEFT,
99 OUTPUT_ORIGIN_TOP_LEFT,
102 // A node in the graph; basically an effect and some associated information.
108 // Edges in the graph (forward and backward).
109 std::vector<Node *> outgoing_links;
110 std::vector<Node *> incoming_links;
112 // For unit tests only. Do not use from other code.
113 // Will contain an arbitrary choice if the node is in multiple phases.
114 Phase *containing_phase;
117 // Logical size of the output of this effect, ie. the resolution
118 // you would get if you sampled it as a texture. If it is undefined
119 // (since the inputs differ in resolution), it will be 0x0.
120 // If both this and output_texture_{width,height} are set,
121 // they will be equal.
122 unsigned output_width, output_height;
124 // If the effect has is_single_texture(), or if the output went to RTT
125 // and that texture has been bound to a sampler, the sampler number
126 // will be stored here.
128 // TODO: Can an RTT texture be used as inputs to multiple effects
129 // within the same phase? If so, we have a problem with modifying
130 // sampler state here.
131 int bound_sampler_num;
133 // Used during the building of the effect chain.
134 Colorspace output_color_space;
135 GammaCurve output_gamma_curve;
136 AlphaType output_alpha_type;
137 bool needs_mipmaps; // Directly or indirectly.
139 // Set if this effect, and all effects consuming output from this node
140 // (in the same phase) have one_to_one_sampling() set.
141 bool one_to_one_sampling;
143 friend class EffectChain;
146 // A rendering phase; a single GLSL program rendering a single quad.
150 GLuint glsl_program_num; // Owned by the resource_pool.
152 // Position and texcoord attribute indexes, although it doesn't matter
153 // which is which, because they contain the same data.
154 std::set<GLint> attribute_indexes;
156 bool input_needs_mipmaps;
158 // Inputs are only inputs from other phases (ie., those that come from RTT);
159 // input textures are counted as part of <effects>.
160 std::vector<Phase *> inputs;
161 // Bound sampler numbers for each input. Redundant in a sense
162 // (it always corresponds to the index), but we need somewhere
163 // to hold the value for the uniform.
164 std::vector<int> input_samplers;
165 std::vector<Node *> effects; // In order.
166 unsigned output_width, output_height, virtual_output_width, virtual_output_height;
168 // Identifier used to create unique variables in GLSL.
169 // Unique per-phase to increase cacheability of compiled shaders.
170 std::map<Node *, std::string> effect_ids;
172 // Uniforms for this phase; combined from all the effects.
173 std::vector<Uniform<int> > uniforms_sampler2d;
174 std::vector<Uniform<bool> > uniforms_bool;
175 std::vector<Uniform<int> > uniforms_int;
176 std::vector<Uniform<float> > uniforms_float;
177 std::vector<Uniform<float> > uniforms_vec2;
178 std::vector<Uniform<float> > uniforms_vec3;
179 std::vector<Uniform<float> > uniforms_vec4;
180 std::vector<Uniform<Eigen::Matrix3d> > uniforms_mat3;
182 // For measurement of GPU time used.
183 std::list<GLuint> timer_query_objects_running;
184 std::list<GLuint> timer_query_objects_free;
185 uint64_t time_elapsed_ns;
186 uint64_t num_measured_iterations;
191 // Aspect: e.g. 16.0f, 9.0f for 16:9.
192 // resource_pool is a pointer to a ResourcePool with which to share shaders
193 // and other resources (see resource_pool.h). If NULL (the default),
194 // will create its own that is not shared with anything else. Does not take
195 // ownership of the passed-in ResourcePool, but will naturally take ownership
196 // of its own internal one if created.
197 EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool = NULL, GLenum intermediate_format = GL_RGBA16F);
201 // input, effects, output, finalize need to come in that specific order.
203 // EffectChain takes ownership of the given input.
204 // input is returned back for convenience.
205 Input *add_input(Input *input);
207 // EffectChain takes ownership of the given effect.
208 // effect is returned back for convenience.
209 Effect *add_effect(Effect *effect) {
210 return add_effect(effect, last_added_effect());
212 Effect *add_effect(Effect *effect, Effect *input) {
213 std::vector<Effect *> inputs;
214 inputs.push_back(input);
215 return add_effect(effect, inputs);
217 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2) {
218 std::vector<Effect *> inputs;
219 inputs.push_back(input1);
220 inputs.push_back(input2);
221 return add_effect(effect, inputs);
223 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3) {
224 std::vector<Effect *> inputs;
225 inputs.push_back(input1);
226 inputs.push_back(input2);
227 inputs.push_back(input3);
228 return add_effect(effect, inputs);
230 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3, Effect *input4) {
231 std::vector<Effect *> inputs;
232 inputs.push_back(input1);
233 inputs.push_back(input2);
234 inputs.push_back(input3);
235 inputs.push_back(input4);
236 return add_effect(effect, inputs);
238 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3, Effect *input4, Effect *input5) {
239 std::vector<Effect *> inputs;
240 inputs.push_back(input1);
241 inputs.push_back(input2);
242 inputs.push_back(input3);
243 inputs.push_back(input4);
244 inputs.push_back(input5);
245 return add_effect(effect, inputs);
247 Effect *add_effect(Effect *effect, const std::vector<Effect *> &inputs);
249 // Adds an RGBA output. Note that you can have at most one RGBA output and one
250 // Y'CbCr output (see below for details).
251 void add_output(const ImageFormat &format, OutputAlphaFormat alpha_format);
253 // Adds an YCbCr output. Note that you can only have one output.
254 // Currently, only chunked packed output is supported, and only 4:4:4
255 // (so chroma_subsampling_x and chroma_subsampling_y must both be 1).
257 // If you have both RGBA and Y'CbCr output, the RGBA output will come
258 // in the last draw buffer. Also, <format> and <alpha_format> must be
259 // identical between the two.
260 void add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
261 const YCbCrFormat &ycbcr_format,
262 YCbCrOutputSplitting output_splitting = YCBCR_OUTPUT_INTERLEAVED);
264 // Set number of output bits, to scale the dither.
265 // 8 is the right value for most outputs.
266 // The default, 0, is a special value that means no dither.
267 void set_dither_bits(unsigned num_bits)
269 this->num_dither_bits = num_bits;
272 // Set where (0,0) is taken to be in the output. The default is
273 // OUTPUT_ORIGIN_BOTTOM_LEFT, which is usually what you want
274 // (see OutputOrigin above for more details).
275 void set_output_origin(OutputOrigin output_origin)
277 this->output_origin = output_origin;
282 // Measure the GPU time used for each actual phase during rendering.
283 // Note that this is only available if GL_ARB_timer_query
284 // (or, equivalently, OpenGL 3.3) is available. Also note that measurement
285 // will incur a performance cost, as we wait for the measurements to
286 // complete at the end of rendering.
287 void enable_phase_timing(bool enable);
288 void reset_phase_timing();
289 void print_phase_timing();
291 //void render(unsigned char *src, unsigned char *dst);
292 void render_to_screen()
294 render_to_fbo(0, 0, 0);
297 // Render the effect chain to the given FBO. If width=height=0, keeps
298 // the current viewport.
299 void render_to_fbo(GLuint fbo, unsigned width, unsigned height);
301 Effect *last_added_effect() {
305 return nodes.back()->effect;
309 // API for manipulating the graph directly. Intended to be used from
310 // effects and by EffectChain itself.
312 // Note that for nodes with multiple inputs, the order of calls to
313 // connect_nodes() will matter.
314 Node *add_node(Effect *effect);
315 void connect_nodes(Node *sender, Node *receiver);
316 void replace_receiver(Node *old_receiver, Node *new_receiver);
317 void replace_sender(Node *new_sender, Node *receiver);
318 void insert_node_between(Node *sender, Node *middle, Node *receiver);
319 Node *find_node_for_effect(Effect *effect) { return node_map[effect]; }
321 // Get the OpenGL sampler (GL_TEXTURE0, GL_TEXTURE1, etc.) for the
322 // input of the given node, so that one can modify the sampler state
323 // directly. Only valid to call during set_gl_state().
325 // Also, for this to be allowed, <node>'s effect must have
326 // needs_texture_bounce() set, so that it samples directly from a
327 // single-sampler input, or from an RTT texture.
328 GLenum get_input_sampler(Node *node, unsigned input_num) const;
330 // Whether input <input_num> of <node> corresponds to a single sampler
331 // (see get_input_sampler()). Normally, you should not need to call this;
332 // however, if the input Effect has set override_texture_bounce(),
333 // this will return false, and you could be flexible and check it first
335 GLenum has_input_sampler(Node *node, unsigned input_num) const;
337 // Get the current resource pool assigned to this EffectChain.
338 // Primarily to let effects allocate textures as needed.
339 // Any resources you get from the pool must be returned to the pool
340 // no later than in the Effect's destructor.
341 ResourcePool *get_resource_pool() { return resource_pool; }
344 // Make sure the output rectangle is at least large enough to hold
345 // the given input rectangle in both dimensions, and is of the
346 // current aspect ratio (aspect_nom/aspect_denom).
347 void size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height);
349 // Compute the input sizes for all inputs for all effects in a given phase,
350 // and inform the effects about the results.
351 void inform_input_sizes(Phase *phase);
353 // Determine the preferred output size of a given phase.
354 // Requires that all input phases (if any) already have output sizes set.
355 void find_output_size(Phase *phase);
357 // Find all inputs eventually feeding into this effect that have
358 // output gamma different from GAMMA_LINEAR.
359 void find_all_nonlinear_inputs(Node *effect, std::vector<Node *> *nonlinear_inputs);
361 // Create a GLSL program computing the effects for this phase in order.
362 void compile_glsl_program(Phase *phase);
364 // Create all GLSL programs needed to compute the given effect, and all outputs
365 // that depend on it (whenever possible). Returns the phase that has <output>
366 // as the last effect. Also pushes all phases in order onto <phases>.
367 Phase *construct_phase(Node *output, std::map<Node *, Phase *> *completed_effects);
369 // Execute one phase, ie. set up all inputs, effects and outputs, and render the quad.
370 void execute_phase(Phase *phase, bool last_phase,
371 std::set<GLint> *bound__attribute_indices,
372 std::map<Phase *, GLuint> *output_textures,
373 std::set<Phase *> *generated_mipmaps);
375 // Set up uniforms for one phase. The program must already be bound.
376 void setup_uniforms(Phase *phase);
378 // Set up the given sampler number for sampling from an RTT texture.
379 void setup_rtt_sampler(int sampler_num, bool use_mipmaps);
381 // Output the current graph to the given file in a Graphviz-compatible format;
382 // only useful for debugging.
383 void output_dot(const char *filename);
384 std::vector<std::string> get_labels_for_edge(const Node *from, const Node *to);
385 void output_dot_edge(FILE *fp,
386 const std::string &from_node_id,
387 const std::string &to_node_id,
388 const std::vector<std::string> &labels);
390 // Some of the graph algorithms assume that the nodes array is sorted
391 // topologically (inputs are always before outputs), but some operations
392 // (like graph rewriting) can change that. This function restores that order.
393 void sort_all_nodes_topologically();
395 // Do the actual topological sort. <nodes> must be a connected, acyclic subgraph;
396 // links that go to nodes not in the set will be ignored.
397 std::vector<Node *> topological_sort(const std::vector<Node *> &nodes);
399 // Utility function used by topological_sort() to do a depth-first search.
400 // The reason why we store nodes left to visit instead of a more conventional
401 // list of nodes to visit is that we want to be able to limit ourselves to
402 // a subgraph instead of all nodes. The set thus serves a dual purpose.
403 void topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list);
405 // Used during finalize().
406 void find_color_spaces_for_inputs();
407 void propagate_alpha();
408 void propagate_gamma_and_color_space();
409 Node *find_output_node();
411 bool node_needs_colorspace_fix(Node *node);
412 void fix_internal_color_spaces();
413 void fix_output_color_space();
415 bool node_needs_alpha_fix(Node *node);
416 void fix_internal_alpha(unsigned step);
417 void fix_output_alpha();
419 bool node_needs_gamma_fix(Node *node);
420 void fix_internal_gamma_by_asking_inputs(unsigned step);
421 void fix_internal_gamma_by_inserting_nodes(unsigned step);
422 void fix_output_gamma();
423 void add_ycbcr_conversion_if_needed();
424 void add_dither_if_needed();
426 float aspect_nom, aspect_denom;
427 ImageFormat output_format;
428 OutputAlphaFormat output_alpha_format;
430 bool output_color_rgba, output_color_ycbcr;
431 YCbCrFormat output_ycbcr_format; // If output_color_ycbcr is true.
432 YCbCrOutputSplitting output_ycbcr_splitting; // If output_color_ycbcr is true.
434 std::vector<Node *> nodes;
435 std::map<Effect *, Node *> node_map;
436 Effect *dither_effect;
438 std::vector<Input *> inputs; // Also contained in nodes.
439 std::vector<Phase *> phases;
441 GLenum intermediate_format;
442 unsigned num_dither_bits;
443 OutputOrigin output_origin;
445 GLuint vbo; // Contains vertex and texture coordinate data.
447 ResourcePool *resource_pool;
448 bool owns_resource_pool;
450 bool do_phase_timing;
455 #endif // !defined(_MOVIT_EFFECT_CHAIN_H)