Another benchmarking data set fix.

[movit] / effect_chain.h

#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 <epoxy/gl.h>
#include <stdio.h>
#include <list>
#include <map>
#include <set>
#include <string>
#include <vector>
#include <Eigen/Core>

#include "effect.h"
#include "image_format.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,
};

// 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,
};

// Transformation to apply (if any) to pixel data in temporary buffers.
// See set_intermediate_format() below for more information.
enum FramebufferTransformation {
        // The default; just store the value. This is what you usually want.
        NO_FRAMEBUFFER_TRANSFORMATION,

        // If the values are in linear light, store sqrt(x) to the framebuffer
        // instead of x itself, of course undoing it with x² on read. Useful as
        // a rough approximation to the sRGB curve. (If the values are not in
        // linear light, just store them as-is.)
        SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION,
};

// A node in the graph; basically an effect and some associated information.
class Node {
public:
        Effect *effect;
        bool disabled;

        // Edges in the graph (forward and backward).
        std::vector<Node *> outgoing_links;
        std::vector<Node *> incoming_links;

        // 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.
        // If both this and output_texture_{width,height} are set,
        // they will be equal.
        unsigned output_width, output_height;

        // 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;

        // Same, for strong_one_to_one_sampling().
        bool strong_one_to_one_sampling;

        friend class EffectChain;
};

// A rendering phase; a single GLSL program rendering a single quad.
struct Phase {
        Node *output_node;

        GLuint glsl_program_num;  // Owned by the resource_pool.

        // Position and texcoord attribute indexes, although it doesn't matter
        // which is which, because they contain the same data.
        std::set<GLint> attribute_indexes;

        bool input_needs_mipmaps;

        // Inputs are only inputs from other phases (ie., those that come from RTT);
        // input textures are counted as part of <effects>.
        std::vector<Phase *> 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<int> input_samplers;
        std::vector<Node *> effects;  // In order.
        unsigned output_width, output_height, virtual_output_width, virtual_output_height;

        // Whether this phase is compiled as a compute shader, ie., the last effect is
        // marked as one.
        bool is_compute_shader;
        Node *compute_shader_node;

        // If <is_compute_shader>, which image unit the output buffer is bound to.
        // This is used as source for a Uniform<int> below.
        int outbuf_image_unit;

        // These are used in transforming from unnormalized to normalized coordinates
        // in compute shaders.
        int uniform_output_size[2];
        Point2D inv_output_size, output_texcoord_adjust;

        // Identifier used to create unique variables in GLSL.
        // Unique per-phase to increase cacheability of compiled shaders.
        std::map<Node *, std::string> effect_ids;

        // Uniforms for this phase; combined from all the effects.
        std::vector<Uniform<int>> uniforms_image2d;
        std::vector<Uniform<int>> uniforms_sampler2d;
        std::vector<Uniform<bool>> uniforms_bool;
        std::vector<Uniform<int>> uniforms_int;
        std::vector<Uniform<int>> uniforms_ivec2;
        std::vector<Uniform<float>> uniforms_float;
        std::vector<Uniform<float>> uniforms_vec2;
        std::vector<Uniform<float>> uniforms_vec3;
        std::vector<Uniform<float>> uniforms_vec4;
        std::vector<Uniform<Eigen::Matrix3d>> uniforms_mat3;

        // For measurement of GPU time used.
        std::list<GLuint> timer_query_objects_running;
        std::list<GLuint> timer_query_objects_free;
        uint64_t time_elapsed_ns;
        uint64_t num_measured_iterations;
};

class EffectChain {
public:
        // 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 nullptr (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 = nullptr);
        ~EffectChain();

        // User API:
        // input, effects, output, finalize need to come in that specific order.

        // EffectChain takes ownership of the given input.
        // input is returned back for convenience.
        Input *add_input(Input *input);

        // EffectChain takes ownership of the given effect.
        // effect is returned back for convenience.
        Effect *add_effect(Effect *effect) {
                return add_effect(effect, last_added_effect());
        }
        Effect *add_effect(Effect *effect, Effect *input) {
                std::vector<Effect *> inputs;
                inputs.push_back(input);
                return add_effect(effect, inputs);
        }
        Effect *add_effect(Effect *effect, Effect *input1, Effect *input2) {
                std::vector<Effect *> inputs;
                inputs.push_back(input1);
                inputs.push_back(input2);
                return add_effect(effect, inputs);
        }
        Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3) {
                std::vector<Effect *> inputs;
                inputs.push_back(input1);
                inputs.push_back(input2);
                inputs.push_back(input3);
                return add_effect(effect, inputs);
        }
        Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3, Effect *input4) {
                std::vector<Effect *> inputs;
                inputs.push_back(input1);
                inputs.push_back(input2);
                inputs.push_back(input3);
                inputs.push_back(input4);
                return add_effect(effect, inputs);
        }
        Effect *add_effect(Effect *effect, Effect *input1, Effect *input2, Effect *input3, Effect *input4, Effect *input5) {
                std::vector<Effect *> inputs;
                inputs.push_back(input1);
                inputs.push_back(input2);
                inputs.push_back(input3);
                inputs.push_back(input4);
                inputs.push_back(input5);
                return add_effect(effect, inputs);
        }
        Effect *add_effect(Effect *effect, const std::vector<Effect *> &inputs);

        // Adds an RGBA output. Note that you can have at most one RGBA output and two
        // Y'CbCr outputs (see below for details).
        void add_output(const ImageFormat &format, OutputAlphaFormat alpha_format);

        // Adds an YCbCr output. Note that you can only have at most two Y'CbCr
        // outputs, and they must have the same <ycbcr_format> and <type>.
        // (This limitation may be lifted in the future, to allow e.g. simultaneous
        // 8- and 10-bit output. Currently, multiple Y'CbCr outputs are only
        // useful in some very limited circumstances, like if one texture goes
        // to some place you cannot easily read from later.)
        //
        // Only 4:4:4 output is supported due to fragment shader limitations,
        // so chroma_subsampling_x and chroma_subsampling_y must both be 1.
        // <type> should match the data type of the FBO you are rendering to,
        // so that if you use 16-bit output (GL_UNSIGNED_SHORT), you will get
        // 8-, 10- or 12-bit output correctly as determined by <ycbcr_format.num_levels>.
        // Using e.g. ycbcr_format.num_levels == 1024 with GL_UNSIGNED_BYTE is
        // nonsensical and invokes undefined behavior.
        //
        // If you have both RGBA and Y'CbCr output(s), the RGBA output will come
        // in the last draw buffer. Also, <format> and <alpha_format> must be
        // identical between the two.
        void add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
                              const YCbCrFormat &ycbcr_format,
                              YCbCrOutputSplitting output_splitting = YCBCR_OUTPUT_INTERLEAVED,
                              GLenum output_type = GL_UNSIGNED_BYTE);

        // Change Y'CbCr output format. (This can be done also after finalize()).
        // Note that you are not allowed to change subsampling parameters;
        // however, you can change the color space parameters, ie.,
        // luma_coefficients, full_range and num_levels.
        void change_ycbcr_output_format(const YCbCrFormat &ycbcr_format);

        // Set number of output bits, to scale the dither.
        // 8 is the right value for most outputs.
        //
        // Special note for 10- and 12-bit Y'CbCr packed into GL_UNSIGNED_SHORT:
        // This is relative to the actual output, not the logical one, so you should
        // specify 16 here, not 10 or 12.
        //
        // 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;
        }

        // Set intermediate format for framebuffers used when we need to bounce
        // to a temporary texture. The default, GL_RGBA16F, is good for most uses;
        // it is precise, has good range, and is relatively efficient. However,
        // if you need even more speed and your chain can do with some loss of
        // accuracy, you can change the format here (before calling finalize).
        // Calculations between bounce buffers are still in 32-bit floating-point
        // no matter what you specify.
        //
        // Of special interest is GL_SRGB8_ALPHA8, which stores sRGB-encoded RGB
        // and linear alpha; this is half the memory bandwidth of GL_RGBA16F,
        // while retaining reasonable precision for typical image data. It will,
        // however, cause some gamut clipping if your colorspace is far from sRGB,
        // as it cannot represent values outside [0,1]. NOTE: If you construct
        // a chain where you end up bouncing pixels in non-linear light
        // (gamma different from GAMMA_LINEAR), this will be the wrong thing.
        // However, it's hard to see how this could happen in a non-contrived
        // chain; few effects ever need texture bounce or resizing without also
        // combining multiple pixels, which really needs linear light and thus
        // triggers a conversion before the bounce.
        //
        // If you don't need alpha (or can do with very little of it), GL_RGB10_A2
        // is even better, as it has two more bits for each color component. There
        // is no GL_SRGB10, unfortunately, so on its own, it is somewhat worse than
        // GL_SRGB8, but you can set <transformation> to SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION,
        // and sqrt(x) will be stored instead of x. This is a rough approximation to
        // the sRGB curve, and reduces maximum error (in sRGB distance) by almost an
        // order of magnitude, well below what you can get from 8-bit true sRGB.
        // (Note that this strategy avoids the problem with bounced non-linear data
        // above, since the square root is turned off in that case.) However, texture
        // filtering will happen on the transformed values, so if you have heavy
        // downscaling or the likes (e.g. mipmaps), you could get subtly bad results.
        // You'll need to see which of the two that works the best for you in practice.
        void set_intermediate_format(
                GLenum intermediate_format,
                FramebufferTransformation transformation = NO_FRAMEBUFFER_TRANSFORMATION)
        {
                this->intermediate_format = intermediate_format;
                this->intermediate_transformation = transformation;
        }

        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();

        // Note: If you already know the width and height of the viewport,
        // calling render_to_fbo() directly will be slightly more efficient,
        // as it saves it from getting it from OpenGL.
        void render_to_screen()
        {
                render_to_fbo(0, 0, 0);
        }

        // Render the effect chain to the given FBO. If width=height=0, keeps
        // the current viewport.
        void render_to_fbo(GLuint fbo, unsigned width, unsigned height);

        // Render the effect chain to the given set of textures. This is equivalent
        // to render_to_fbo() with a freshly created FBO bound to the given textures,
        // except that it is more efficient if the last phase contains a compute shader.
        // Thus, prefer this to render_to_fbo() where possible.
        //
        // Only one destination texture is supported. This restriction will be lifted
        // in the future.
        //
        // All destination textures must be exactly of size <width> x <height>,
        // and must either come from the same ResourcePool the effect uses, or outlive
        // the EffectChain (otherwise, we could be allocating FBOs that end up being
        // stale). Textures must also have valid state; in particular, they must either
        // be mipmap complete or have a non-mipmapped minification mode.
        //
        // width and height can not be zero.
        struct DestinationTexture {
                GLuint texnum;
                GLenum format;
        };
        void render_to_texture(const std::vector<DestinationTexture> &destinations, unsigned width, unsigned height);

        Effect *last_added_effect() {
                if (nodes.empty()) {
                        return nullptr;
                } else {
                        return nodes.back()->effect;
                }       
        }

        // API for manipulating the graph directly. Intended to be used from
        // effects and by EffectChain itself.
        //
        // Note that for nodes with multiple inputs, the order of calls to
        // connect_nodes() will matter.
        Node *add_node(Effect *effect);
        void connect_nodes(Node *sender, Node *receiver);
        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, <node>'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;

        // Whether input <input_num> of <node> corresponds to a single sampler
        // (see get_input_sampler()). Normally, you should not need to call this;
        // however, if the input Effect has set override_texture_bounce(),
        // this will return false, and you could be flexible and check it first
        // if you want.
        GLenum has_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:
        // 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.    
        void inform_input_sizes(Phase *phase);

        // Determine the preferred output size of a given phase.
        // Requires that all input phases (if any) already have output sizes set.
        void find_output_size(Phase *phase);

        // Find all inputs eventually feeding into this effect that have
        // output gamma different from GAMMA_LINEAR.
        void find_all_nonlinear_inputs(Node *effect, std::vector<Node *> *nonlinear_inputs);

        // 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 depend on it (whenever possible). Returns the phase that has <output>
        // as the last effect. Also pushes all phases in order onto <phases>.
        Phase *construct_phase(Node *output, std::map<Node *, Phase *> *completed_effects);

        // Do the actual rendering of the chain. If <dest_fbo> is not (GLuint)-1,
        // renders to that FBO. If <destinations> is non-empty, render to that set
        // of textures (last phase, save for the dummy phase, must be a compute shader),
        // with x/y ignored. Having both set is an error.
        void render(GLuint dest_fbo, const std::vector<DestinationTexture> &destinations,
                    unsigned x, unsigned y, unsigned width, unsigned height);

        // Execute one phase, ie. set up all inputs, effects and outputs, and render the quad.
        // If <destinations> is empty, uses whatever output is current (and the phase must not be
        // a compute shader).
        void execute_phase(Phase *phase,
                           const std::map<Phase *, GLuint> &output_textures,
                           const std::vector<DestinationTexture> &destinations,
                           std::set<Phase *> *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<std::string> 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<std::string> &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_all_nodes_topologically();

        // Do the actual topological sort. <nodes> must be a connected, acyclic subgraph;
        // links that go to nodes not in the set will be ignored.
        std::vector<Node *> topological_sort(const std::vector<Node *> &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<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list);

        // Used during finalize().
        void find_color_spaces_for_inputs();
        void propagate_alpha();
        void propagate_gamma_and_color_space();
        Node *find_output_node();

        bool node_needs_colorspace_fix(Node *node);
        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();
        void add_dummy_effect_if_needed();

        float aspect_nom, aspect_denom;
        ImageFormat output_format;
        OutputAlphaFormat output_alpha_format;

        bool output_color_rgba;
        int num_output_color_ycbcr;                      // Max 2.
        YCbCrFormat output_ycbcr_format;                 // If num_output_color_ycbcr is > 0.
        GLenum output_ycbcr_type;                        // If num_output_color_ycbcr is > 0.
        YCbCrOutputSplitting output_ycbcr_splitting[2];  // If num_output_color_ycbcr is > N.

        std::vector<Node *> nodes;
        std::map<Effect *, Node *> node_map;
        Effect *dither_effect;
        Node *ycbcr_conversion_effect_node;

        std::vector<Input *> inputs;  // Also contained in nodes.
        std::vector<Phase *> phases;

        GLenum intermediate_format;
        FramebufferTransformation intermediate_transformation;
        unsigned num_dither_bits;
        OutputOrigin output_origin;
        bool finalized;
        GLuint vbo;  // Contains vertex and texture coordinate data.

        // Whether the last effect (which will then be in a phase all by itself)
        // is a dummy effect that is only added because the last phase uses a compute
        // shader, which cannot output directly to the backbuffer. This means that
        // the phase can be skipped if we are _not_ rendering to the backbuffer.
        bool has_dummy_effect = false;

        ResourcePool *resource_pool;
        bool owns_resource_pool;

        bool do_phase_timing;
};

}  // namespace movit

#endif // !defined(_MOVIT_EFFECT_CHAIN_H)