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.
33 #include "image_format.h"
43 // For internal use within Node.
51 // Whether you want pre- or postmultiplied alpha in the output
52 // (see effect.h for a discussion of pre- versus postmultiplied alpha).
53 enum OutputAlphaFormat {
54 OUTPUT_ALPHA_FORMAT_PREMULTIPLIED,
55 OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED,
58 // RGBA output is nearly always packed; Y'CbCr, however, is often planar
59 // due to chroma subsampling. This enum controls how add_ycbcr_output()
60 // distributes the color channels between the fragment shader outputs.
61 // Obviously, anything except YCBCR_OUTPUT_INTERLEAVED will be meaningless
62 // unless you use render_to_fbo() and have an FBO with multiple render
63 // targets attached (the other outputs will be discarded).
64 enum YCbCrOutputSplitting {
65 // Only one output: Store Y'CbCr into the first three output channels,
66 // respectively, plus alpha. This is also called “chunked” or
68 YCBCR_OUTPUT_INTERLEAVED,
70 // Store Y' and alpha into the first output (in the red and alpha
71 // channels; effect to the others is undefined), and Cb and Cr into
72 // the first two channels of the second output. This is particularly
73 // useful if you want to end up in a format like NV12, where all the
74 // Y' samples come first and then Cb and Cr come interlevaed afterwards.
75 // You will still need to do the chroma subsampling yourself to actually
76 // get down to NV12, though.
77 YCBCR_OUTPUT_SPLIT_Y_AND_CBCR,
79 // Store Y' and alpha into the first output, Cb into the first channel
80 // of the second output and Cr into the first channel of the third output.
81 // (Effect on the other channels is undefined.) Essentially gives you
82 // 4:4:4 planar, or ”yuv444p”.
86 // A node in the graph; basically an effect and some associated information.
92 // Edges in the graph (forward and backward).
93 std::vector<Node *> outgoing_links;
94 std::vector<Node *> incoming_links;
96 // For unit tests only. Do not use from other code.
97 // Will contain an arbitrary choice if the node is in multiple phases.
98 Phase *containing_phase;
101 // Logical size of the output of this effect, ie. the resolution
102 // you would get if you sampled it as a texture. If it is undefined
103 // (since the inputs differ in resolution), it will be 0x0.
104 // If both this and output_texture_{width,height} are set,
105 // they will be equal.
106 unsigned output_width, output_height;
108 // If the effect has is_single_texture(), or if the output went to RTT
109 // and that texture has been bound to a sampler, the sampler number
110 // will be stored here.
112 // TODO: Can an RTT texture be used as inputs to multiple effects
113 // within the same phase? If so, we have a problem with modifying
114 // sampler state here.
115 int bound_sampler_num;
117 // Used during the building of the effect chain.
118 Colorspace output_color_space;
119 GammaCurve output_gamma_curve;
120 AlphaType output_alpha_type;
121 bool needs_mipmaps; // Directly or indirectly.
123 // Set if this effect, and all effects consuming output from this node
124 // (in the same phase) have one_to_one_sampling() set.
125 bool one_to_one_sampling;
127 friend class EffectChain;
130 // A rendering phase; a single GLSL program rendering a single quad.
134 GLuint glsl_program_num; // Owned by the resource_pool.
135 bool input_needs_mipmaps;
137 // Inputs are only inputs from other phases (ie., those that come from RTT);
138 // input textures are counted as part of <effects>.
139 std::vector<Phase *> inputs;
140 // Bound sampler numbers for each input. Redundant in a sense
141 // (it always corresponds to the index), but we need somewhere
142 // to hold the value for the uniform.
143 std::vector<int> input_samplers;
144 std::vector<Node *> effects; // In order.
145 unsigned output_width, output_height, virtual_output_width, virtual_output_height;
147 // Identifier used to create unique variables in GLSL.
148 // Unique per-phase to increase cacheability of compiled shaders.
149 std::map<Node *, std::string> effect_ids;
151 // Uniforms for this phase; combined from all the effects.
152 std::vector<Uniform<int> > uniforms_sampler2d;
153 std::vector<Uniform<bool> > uniforms_bool;
154 std::vector<Uniform<int> > uniforms_int;
155 std::vector<Uniform<float> > uniforms_float;
156 std::vector<Uniform<float> > uniforms_vec2;
157 std::vector<Uniform<float> > uniforms_vec3;
158 std::vector<Uniform<float> > uniforms_vec4;
159 std::vector<Uniform<Eigen::Matrix3d> > uniforms_mat3;
161 // For measurement of GPU time used.
162 GLuint timer_query_object;
163 uint64_t time_elapsed_ns;
164 uint64_t num_measured_iterations;
169 // Aspect: e.g. 16.0f, 9.0f for 16:9.
170 // resource_pool is a pointer to a ResourcePool with which to share shaders
171 // and other resources (see resource_pool.h). If NULL (the default),
172 // will create its own that is not shared with anything else. Does not take
173 // ownership of the passed-in ResourcePool, but will naturally take ownership
174 // of its own internal one if created.
175 EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool = NULL);
179 // input, effects, output, finalize need to come in that specific order.
181 // EffectChain takes ownership of the given input.
182 // input is returned back for convenience.
183 Input *add_input(Input *input);
185 // EffectChain takes ownership of the given effect.
186 // effect is returned back for convenience.
187 Effect *add_effect(Effect *effect) {
188 return add_effect(effect, last_added_effect());
190 Effect *add_effect(Effect *effect, Effect *input) {
191 std::vector<Effect *> inputs;
192 inputs.push_back(input);
193 return add_effect(effect, inputs);
195 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2) {
196 std::vector<Effect *> inputs;
197 inputs.push_back(input1);
198 inputs.push_back(input2);
199 return add_effect(effect, inputs);
201 Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3) {
202 std::vector<Effect *> inputs;
203 inputs.push_back(input1);
204 inputs.push_back(input2);
205 inputs.push_back(input3);
206 return add_effect(effect, inputs);
208 Effect *add_effect(Effect *effect, const std::vector<Effect *> &inputs);
210 // Adds an RGB output. Note that you can only have one output.
211 void add_output(const ImageFormat &format, OutputAlphaFormat alpha_format);
213 // Adds an YCbCr output. Note that you can only have one output.
214 // Currently, only chunked packed output is supported, and only 4:4:4
215 // (so chroma_subsampling_x and chroma_subsampling_y must both be 1).
216 void add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
217 const YCbCrFormat &ycbcr_format,
218 YCbCrOutputSplitting output_splitting = YCBCR_OUTPUT_INTERLEAVED);
220 // Set number of output bits, to scale the dither.
221 // 8 is the right value for most outputs.
222 // The default, 0, is a special value that means no dither.
223 void set_dither_bits(unsigned num_bits)
225 this->num_dither_bits = num_bits;
230 // Measure the GPU time used for each actual phase during rendering.
231 // Note that this is only available if GL_ARB_timer_query
232 // (or, equivalently, OpenGL 3.3) is available. Also note that measurement
233 // will incur a performance cost, as we wait for the measurements to
234 // complete at the end of rendering.
235 void enable_phase_timing(bool enable);
236 void reset_phase_timing();
237 void print_phase_timing();
239 //void render(unsigned char *src, unsigned char *dst);
240 void render_to_screen()
242 render_to_fbo(0, 0, 0);
245 // Render the effect chain to the given FBO. If width=height=0, keeps
246 // the current viewport.
247 void render_to_fbo(GLuint fbo, unsigned width, unsigned height);
249 Effect *last_added_effect() {
253 return nodes.back()->effect;
257 // API for manipulating the graph directly. Intended to be used from
258 // effects and by EffectChain itself.
260 // Note that for nodes with multiple inputs, the order of calls to
261 // connect_nodes() will matter.
262 Node *add_node(Effect *effect);
263 void connect_nodes(Node *sender, Node *receiver);
264 void replace_receiver(Node *old_receiver, Node *new_receiver);
265 void replace_sender(Node *new_sender, Node *receiver);
266 void insert_node_between(Node *sender, Node *middle, Node *receiver);
267 Node *find_node_for_effect(Effect *effect) { return node_map[effect]; }
269 // Get the OpenGL sampler (GL_TEXTURE0, GL_TEXTURE1, etc.) for the
270 // input of the given node, so that one can modify the sampler state
271 // directly. Only valid to call during set_gl_state().
273 // Also, for this to be allowed, <node>'s effect must have
274 // needs_texture_bounce() set, so that it samples directly from a
275 // single-sampler input, or from an RTT texture.
276 GLenum get_input_sampler(Node *node, unsigned input_num) const;
278 // Get the current resource pool assigned to this EffectChain.
279 // Primarily to let effects allocate textures as needed.
280 // Any resources you get from the pool must be returned to the pool
281 // no later than in the Effect's destructor.
282 ResourcePool *get_resource_pool() { return resource_pool; }
285 // Make sure the output rectangle is at least large enough to hold
286 // the given input rectangle in both dimensions, and is of the
287 // current aspect ratio (aspect_nom/aspect_denom).
288 void size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height);
290 // Compute the input sizes for all inputs for all effects in a given phase,
291 // and inform the effects about the results.
292 void inform_input_sizes(Phase *phase);
294 // Determine the preferred output size of a given phase.
295 // Requires that all input phases (if any) already have output sizes set.
296 void find_output_size(Phase *phase);
298 // Find all inputs eventually feeding into this effect that have
299 // output gamma different from GAMMA_LINEAR.
300 void find_all_nonlinear_inputs(Node *effect, std::vector<Node *> *nonlinear_inputs);
302 // Create a GLSL program computing the effects for this phase in order.
303 void compile_glsl_program(Phase *phase);
305 // Create all GLSL programs needed to compute the given effect, and all outputs
306 // that depend on it (whenever possible). Returns the phase that has <output>
307 // as the last effect. Also pushes all phases in order onto <phases>.
308 Phase *construct_phase(Node *output, std::map<Node *, Phase *> *completed_effects);
310 // Execute one phase, ie. set up all inputs, effects and outputs, and render the quad.
311 void execute_phase(Phase *phase, bool last_phase, std::map<Phase *, GLuint> *output_textures, std::set<Phase *> *generated_mipmaps);
313 // Set up uniforms for one phase. The program must already be bound.
314 void setup_uniforms(Phase *phase);
316 // Set up the given sampler number for sampling from an RTT texture.
317 void setup_rtt_sampler(int sampler_num, bool use_mipmaps);
319 // Output the current graph to the given file in a Graphviz-compatible format;
320 // only useful for debugging.
321 void output_dot(const char *filename);
322 std::vector<std::string> get_labels_for_edge(const Node *from, const Node *to);
323 void output_dot_edge(FILE *fp,
324 const std::string &from_node_id,
325 const std::string &to_node_id,
326 const std::vector<std::string> &labels);
328 // Some of the graph algorithms assume that the nodes array is sorted
329 // topologically (inputs are always before outputs), but some operations
330 // (like graph rewriting) can change that. This function restores that order.
331 void sort_all_nodes_topologically();
333 // Do the actual topological sort. <nodes> must be a connected, acyclic subgraph;
334 // links that go to nodes not in the set will be ignored.
335 std::vector<Node *> topological_sort(const std::vector<Node *> &nodes);
337 // Utility function used by topological_sort() to do a depth-first search.
338 // The reason why we store nodes left to visit instead of a more conventional
339 // list of nodes to visit is that we want to be able to limit ourselves to
340 // a subgraph instead of all nodes. The set thus serves a dual purpose.
341 void topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list);
343 // Used during finalize().
344 void find_color_spaces_for_inputs();
345 void propagate_alpha();
346 void propagate_gamma_and_color_space();
347 Node *find_output_node();
349 bool node_needs_colorspace_fix(Node *node);
350 void fix_internal_color_spaces();
351 void fix_output_color_space();
353 bool node_needs_alpha_fix(Node *node);
354 void fix_internal_alpha(unsigned step);
355 void fix_output_alpha();
357 bool node_needs_gamma_fix(Node *node);
358 void fix_internal_gamma_by_asking_inputs(unsigned step);
359 void fix_internal_gamma_by_inserting_nodes(unsigned step);
360 void fix_output_gamma();
361 void add_ycbcr_conversion_if_needed();
362 void add_dither_if_needed();
364 float aspect_nom, aspect_denom;
365 ImageFormat output_format;
366 OutputAlphaFormat output_alpha_format;
368 enum OutputColorType { OUTPUT_COLOR_RGB, OUTPUT_COLOR_YCBCR };
369 OutputColorType output_color_type;
370 YCbCrFormat output_ycbcr_format; // If output_color_type == OUTPUT_COLOR_YCBCR.
371 YCbCrOutputSplitting output_ycbcr_splitting; // If output_color_type == OUTPUT_COLOR_YCBCR.
373 std::vector<Node *> nodes;
374 std::map<Effect *, Node *> node_map;
375 Effect *dither_effect;
377 std::vector<Input *> inputs; // Also contained in nodes.
378 std::vector<Phase *> phases;
380 unsigned num_dither_bits;
383 ResourcePool *resource_pool;
384 bool owns_resource_pool;
386 bool do_phase_timing;
391 #endif // !defined(_MOVIT_EFFECT_CHAIN_H)