X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.h;h=2e89c3bcc38717473c10f8b05cfffe80a68b9689;hp=7731b54ad5458db41988f0f1d39e2a8b47ae5802;hb=6f1efa8348a90a393187c12d70fd10d81bbd2c99;hpb=1ca4785183b5af9a2a255bba0a28dfdae156470a diff --git a/effect_chain.h b/effect_chain.h index 7731b54..2e89c3b 100644 --- a/effect_chain.h +++ b/effect_chain.h @@ -1,15 +1,102 @@ -#ifndef _EFFECT_CHAIN_H -#define _EFFECT_CHAIN_H 1 - +#ifndef _MOVIT_EFFECT_CHAIN_H +#define _MOVIT_EFFECT_CHAIN_H 1 + +// An EffectChain is the largest basic entity in Movit; it contains everything +// needed to connects a series of effects, from inputs to outputs, and render +// them. Generally you set up your effect chain once and then call its render +// functions once per frame; setting one up can be relatively expensive, +// but rendering is fast. +// +// Threading considerations: EffectChain is “thread-compatible”; you can use +// different EffectChains in multiple threads at the same time (assuming the +// threads do not use the same OpenGL context, but this is a good idea anyway), +// but you may not use one EffectChain from multiple threads simultaneously. +// You _are_ allowed to use one EffectChain from multiple threads as long as +// you only use it from one at a time (possibly by doing your own locking), +// but if so, the threads' contexts need to be set up to share resources, since +// the EffectChain holds textures and other OpenGL objects that are tied to the +// context. +// +// Memory management (only relevant if you use multiple contexts): +// See corresponding comment in resource_pool.h. This holds even if you don't +// allocate your own ResourcePool, but let EffectChain hold its own. + +#include +#include +#include #include +#include #include +#include #include "effect.h" #include "image_format.h" -#include "input.h" +#include "ycbcr.h" + +namespace movit { + +class Effect; +class Input; +struct Phase; +class ResourcePool; + +// For internal use within Node. +enum AlphaType { + ALPHA_INVALID = -1, + ALPHA_BLANK, + ALPHA_PREMULTIPLIED, + ALPHA_POSTMULTIPLIED, +}; + +// Whether you want pre- or postmultiplied alpha in the output +// (see effect.h for a discussion of pre- versus postmultiplied alpha). +enum OutputAlphaFormat { + OUTPUT_ALPHA_FORMAT_PREMULTIPLIED, + OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, +}; + +// RGBA output is nearly always packed; Y'CbCr, however, is often planar +// due to chroma subsampling. This enum controls how add_ycbcr_output() +// distributes the color channels between the fragment shader outputs. +// Obviously, anything except YCBCR_OUTPUT_INTERLEAVED will be meaningless +// unless you use render_to_fbo() and have an FBO with multiple render +// targets attached (the other outputs will be discarded). +enum YCbCrOutputSplitting { + // Only one output: Store Y'CbCr into the first three output channels, + // respectively, plus alpha. This is also called “chunked” or + // ”packed” mode. + YCBCR_OUTPUT_INTERLEAVED, + + // Store Y' and alpha into the first output (in the red and alpha + // channels; effect to the others is undefined), and Cb and Cr into + // the first two channels of the second output. This is particularly + // useful if you want to end up in a format like NV12, where all the + // Y' samples come first and then Cb and Cr come interlevaed afterwards. + // You will still need to do the chroma subsampling yourself to actually + // get down to NV12, though. + YCBCR_OUTPUT_SPLIT_Y_AND_CBCR, + + // Store Y' and alpha into the first output, Cb into the first channel + // of the second output and Cr into the first channel of the third output. + // (Effect on the other channels is undefined.) Essentially gives you + // 4:4:4 planar, or ”yuv444p”. + YCBCR_OUTPUT_PLANAR, +}; -class EffectChain; -class Phase; +// Where (0,0) is taken to be in the output. If you want to render to an +// OpenGL screen, you should keep the default of bottom-left, as that is +// OpenGL's natural coordinate system. However, there are cases, such as if you +// render to an FBO and read the pixels back into some other system, where +// you'd want a top-left origin; if so, an additional flip step will be added +// at the very end (but done in a vertex shader, so it will have zero extra +// cost). +// +// Note that Movit's coordinate system in general consistently puts (0,0) in +// the top left for _input_, no matter what you set as output origin. +enum OutputOrigin { + OUTPUT_ORIGIN_BOTTOM_LEFT, + OUTPUT_ORIGIN_TOP_LEFT, +}; // A node in the graph; basically an effect and some associated information. class Node { @@ -21,10 +108,11 @@ public: std::vector outgoing_links; std::vector incoming_links; -private: - // Identifier used to create unique variables in GLSL. - std::string effect_id; + // For unit tests only. Do not use from other code. + // Will contain an arbitrary choice if the node is in multiple phases. + Phase *containing_phase; +private: // Logical size of the output of this effect, ie. the resolution // you would get if you sampled it as a texture. If it is undefined // (since the inputs differ in resolution), it will be 0x0. @@ -32,36 +120,74 @@ private: // they will be equal. unsigned output_width, output_height; - // If output goes to RTT (otherwise, none of these are set). - // The Phase pointer is a but ugly; we should probably fix so - // that Phase takes other phases as inputs, instead of Node. - GLuint output_texture; - unsigned output_texture_width, output_texture_height; - Phase *phase; + // If the effect has is_single_texture(), or if the output went to RTT + // and that texture has been bound to a sampler, the sampler number + // will be stored here. + // + // TODO: Can an RTT texture be used as inputs to multiple effects + // within the same phase? If so, we have a problem with modifying + // sampler state here. + int bound_sampler_num; // Used during the building of the effect chain. Colorspace output_color_space; GammaCurve output_gamma_curve; + AlphaType output_alpha_type; + bool needs_mipmaps; // Directly or indirectly. + + // Set if this effect, and all effects consuming output from this node + // (in the same phase) have one_to_one_sampling() set. + bool one_to_one_sampling; friend class EffectChain; }; // A rendering phase; a single GLSL program rendering a single quad. struct Phase { - GLint glsl_program_num, vertex_shader, fragment_shader; + Node *output_node; + + GLuint glsl_program_num; // Owned by the resource_pool. bool input_needs_mipmaps; // Inputs are only inputs from other phases (ie., those that come from RTT); - // input textures are not counted here. - std::vector inputs; - + // input textures are counted as part of . + std::vector inputs; + // Bound sampler numbers for each input. Redundant in a sense + // (it always corresponds to the index), but we need somewhere + // to hold the value for the uniform. + std::vector input_samplers; std::vector effects; // In order. - unsigned output_width, output_height; + unsigned output_width, output_height, virtual_output_width, virtual_output_height; + + // Identifier used to create unique variables in GLSL. + // Unique per-phase to increase cacheability of compiled shaders. + std::map effect_ids; + + // Uniforms for this phase; combined from all the effects. + std::vector > uniforms_sampler2d; + std::vector > uniforms_bool; + std::vector > uniforms_int; + std::vector > uniforms_float; + std::vector > uniforms_vec2; + std::vector > uniforms_vec3; + std::vector > uniforms_vec4; + std::vector > uniforms_mat3; + + // For measurement of GPU time used. + GLuint timer_query_object; + uint64_t time_elapsed_ns; + uint64_t num_measured_iterations; }; class EffectChain { public: - EffectChain(float aspect_nom, float aspect_denom); // E.g., 16.0f, 9.0f for 16:9. + // Aspect: e.g. 16.0f, 9.0f for 16:9. + // resource_pool is a pointer to a ResourcePool with which to share shaders + // and other resources (see resource_pool.h). If NULL (the default), + // will create its own that is not shared with anything else. Does not take + // ownership of the passed-in ResourcePool, but will naturally take ownership + // of its own internal one if created. + EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool = NULL); ~EffectChain(); // User API: @@ -87,11 +213,52 @@ public: inputs.push_back(input2); return add_effect(effect, inputs); } + Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3) { + std::vector inputs; + inputs.push_back(input1); + inputs.push_back(input2); + inputs.push_back(input3); + return add_effect(effect, inputs); + } Effect *add_effect(Effect *effect, const std::vector &inputs); - void add_output(const ImageFormat &format); + // Adds an RGB output. Note that you can only have one output. + void add_output(const ImageFormat &format, OutputAlphaFormat alpha_format); + + // Adds an YCbCr output. Note that you can only have one output. + // Currently, only chunked packed output is supported, and only 4:4:4 + // (so chroma_subsampling_x and chroma_subsampling_y must both be 1). + void add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format, + const YCbCrFormat &ycbcr_format, + YCbCrOutputSplitting output_splitting = YCBCR_OUTPUT_INTERLEAVED); + + // Set number of output bits, to scale the dither. + // 8 is the right value for most outputs. + // The default, 0, is a special value that means no dither. + void set_dither_bits(unsigned num_bits) + { + this->num_dither_bits = num_bits; + } + + // Set where (0,0) is taken to be in the output. The default is + // OUTPUT_ORIGIN_BOTTOM_LEFT, which is usually what you want + // (see OutputOrigin above for more details). + void set_output_origin(OutputOrigin output_origin) + { + this->output_origin = output_origin; + } + void finalize(); + // Measure the GPU time used for each actual phase during rendering. + // Note that this is only available if GL_ARB_timer_query + // (or, equivalently, OpenGL 3.3) is available. Also note that measurement + // will incur a performance cost, as we wait for the measurements to + // complete at the end of rendering. + void enable_phase_timing(bool enable); + void reset_phase_timing(); + void print_phase_timing(); + //void render(unsigned char *src, unsigned char *dst); void render_to_screen() { @@ -120,11 +287,28 @@ public: void replace_receiver(Node *old_receiver, Node *new_receiver); void replace_sender(Node *new_sender, Node *receiver); void insert_node_between(Node *sender, Node *middle, Node *receiver); + Node *find_node_for_effect(Effect *effect) { return node_map[effect]; } + + // Get the OpenGL sampler (GL_TEXTURE0, GL_TEXTURE1, etc.) for the + // input of the given node, so that one can modify the sampler state + // directly. Only valid to call during set_gl_state(). + // + // Also, for this to be allowed, 's effect must have + // needs_texture_bounce() set, so that it samples directly from a + // single-sampler input, or from an RTT texture. + GLenum get_input_sampler(Node *node, unsigned input_num) const; + + // Get the current resource pool assigned to this EffectChain. + // Primarily to let effects allocate textures as needed. + // Any resources you get from the pool must be returned to the pool + // no later than in the Effect's destructor. + ResourcePool *get_resource_pool() { return resource_pool; } private: - // Fits a rectangle of the given size to the current aspect ratio - // (aspect_nom/aspect_denom) and returns the new width and height. - unsigned fit_rectangle_to_aspect(unsigned width, unsigned height); + // Make sure the output rectangle is at least large enough to hold + // the given input rectangle in both dimensions, and is of the + // current aspect ratio (aspect_nom/aspect_denom). + void size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height); // Compute the input sizes for all inputs for all effects in a given phase, // and inform the effects about the results. @@ -138,25 +322,50 @@ private: // output gamma different from GAMMA_LINEAR. void find_all_nonlinear_inputs(Node *effect, std::vector *nonlinear_inputs); - // Create a GLSL program computing the given effects in order. - Phase *compile_glsl_program(const std::vector &inputs, - const std::vector &effects); + // Create a GLSL program computing the effects for this phase in order. + void compile_glsl_program(Phase *phase); // Create all GLSL programs needed to compute the given effect, and all outputs - // that depends on it (whenever possible). - void construct_glsl_programs(Node *output); + // that depend on it (whenever possible). Returns the phase that has + // as the last effect. Also pushes all phases in order onto . + Phase *construct_phase(Node *output, std::map *completed_effects); + + // Execute one phase, ie. set up all inputs, effects and outputs, and render the quad. + void execute_phase(Phase *phase, bool last_phase, std::map *output_textures, std::set *generated_mipmaps); + + // Set up uniforms for one phase. The program must already be bound. + void setup_uniforms(Phase *phase); + + // Set up the given sampler number for sampling from an RTT texture. + void setup_rtt_sampler(int sampler_num, bool use_mipmaps); // Output the current graph to the given file in a Graphviz-compatible format; // only useful for debugging. void output_dot(const char *filename); + std::vector get_labels_for_edge(const Node *from, const Node *to); + void output_dot_edge(FILE *fp, + const std::string &from_node_id, + const std::string &to_node_id, + const std::vector &labels); // Some of the graph algorithms assume that the nodes array is sorted // topologically (inputs are always before outputs), but some operations // (like graph rewriting) can change that. This function restores that order. - void sort_nodes_topologically(); - void topological_sort_visit_node(Node *node, std::set *visited_nodes, std::vector *sorted_list); + void sort_all_nodes_topologically(); + + // Do the actual topological sort. must be a connected, acyclic subgraph; + // links that go to nodes not in the set will be ignored. + std::vector topological_sort(const std::vector &nodes); + + // Utility function used by topological_sort() to do a depth-first search. + // The reason why we store nodes left to visit instead of a more conventional + // list of nodes to visit is that we want to be able to limit ourselves to + // a subgraph instead of all nodes. The set thus serves a dual purpose. + void topological_sort_visit_node(Node *node, std::set *nodes_left_to_visit, std::vector *sorted_list); // Used during finalize(). + void find_color_spaces_for_inputs(); + void propagate_alpha(); void propagate_gamma_and_color_space(); Node *find_output_node(); @@ -164,24 +373,43 @@ private: void fix_internal_color_spaces(); void fix_output_color_space(); + bool node_needs_alpha_fix(Node *node); + void fix_internal_alpha(unsigned step); + void fix_output_alpha(); + bool node_needs_gamma_fix(Node *node); void fix_internal_gamma_by_asking_inputs(unsigned step); void fix_internal_gamma_by_inserting_nodes(unsigned step); void fix_output_gamma(); + void add_ycbcr_conversion_if_needed(); + void add_dither_if_needed(); float aspect_nom, aspect_denom; ImageFormat output_format; + OutputAlphaFormat output_alpha_format; + + enum OutputColorType { OUTPUT_COLOR_RGB, OUTPUT_COLOR_YCBCR }; + OutputColorType output_color_type; + YCbCrFormat output_ycbcr_format; // If output_color_type == OUTPUT_COLOR_YCBCR. + YCbCrOutputSplitting output_ycbcr_splitting; // If output_color_type == OUTPUT_COLOR_YCBCR. std::vector nodes; std::map node_map; + Effect *dither_effect; std::vector inputs; // Also contained in nodes. - - GLuint fbo; std::vector phases; - GLenum format, bytes_per_pixel; + unsigned num_dither_bits; + OutputOrigin output_origin; bool finalized; + + ResourcePool *resource_pool; + bool owns_resource_pool; + + bool do_phase_timing; }; -#endif // !defined(_EFFECT_CHAIN_H) +} // namespace movit + +#endif // !defined(_MOVIT_EFFECT_CHAIN_H)