X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect.h;h=073780f172bdbc929db64238c1705221264691c4;hp=43849dfa5cdddd61035a58a528061f7045276ead;hb=65c6584f77bff0af0c8e38d1ac90298bcd55e9ac;hpb=5d4c0425579de66b3e2dd7d0095e890278f8efcf diff --git a/effect.h b/effect.h index 43849df..073780f 100644 --- a/effect.h +++ b/effect.h @@ -1,5 +1,5 @@ -#ifndef _EFFECT_H -#define _EFFECT_H 1 +#ifndef _MOVIT_EFFECT_H +#define _MOVIT_EFFECT_H 1 // Effect is the base class for every effect. It basically represents a single // GLSL function, with an optional set of user-settable parameters. @@ -10,22 +10,24 @@ // effect instance; use the macro PREFIX() around your identifiers to // automatically prepend that prefix. +#include +#include +#include #include #include #include - -#include - #include -#include -#include "util.h" +#include "defs.h" + +namespace movit { class EffectChain; class Node; // Can alias on a float[2]. struct Point2D { + Point2D() {} Point2D(float x, float y) : x(x), y(y) {} @@ -34,6 +36,7 @@ struct Point2D { // Can alias on a float[3]. struct RGBTriplet { + RGBTriplet() {} RGBTriplet(float r, float g, float b) : r(r), g(g), b(b) {} @@ -41,22 +44,23 @@ struct RGBTriplet { }; // Can alias on a float[4]. -struct RGBATriplet { - RGBATriplet(float r, float g, float b, float a) +struct RGBATuple { + RGBATuple() {} + RGBATuple(float r, float g, float b, float a) : r(r), g(g), b(b), a(a) {} float r, g, b, a; }; -// Convenience functions that deal with prepending the prefix. -GLint get_uniform_location(GLuint glsl_program_num, const std::string &prefix, const std::string &key); -void set_uniform_int(GLuint glsl_program_num, const std::string &prefix, const std::string &key, int value); -void set_uniform_float(GLuint glsl_program_num, const std::string &prefix, const std::string &key, float value); -void set_uniform_vec2(GLuint glsl_program_num, const std::string &prefix, const std::string &key, const float *values); -void set_uniform_vec3(GLuint glsl_program_num, const std::string &prefix, const std::string &key, const float *values); -void set_uniform_vec4(GLuint glsl_program_num, const std::string &prefix, const std::string &key, const float *values); -void set_uniform_vec4_array(GLuint glsl_program_num, const std::string &prefix, const std::string &key, const float *values, size_t num_values); -void set_uniform_mat3(GLuint glsl_program_num, const std::string &prefix, const std::string &key, const Eigen::Matrix3d &matrix); +// Represents a registered uniform. +template +struct Uniform { + std::string name; // Without prefix. + const T *value; // Owner by the effect. + size_t num_values; // Number of elements; for arrays only. _Not_ the vector length. + std::string prefix; // Filled in only after phases have been constructed. + GLint location; // Filled in only after phases have been constructed. -1 if no location. +}; class Effect { public: @@ -99,7 +103,8 @@ public: // This is the most natural format for processing, and the default in // most of Movit (just like linear light is). // - // If you set INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA, all of your inputs + // If you set INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA or + // INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK, all of your inputs // (if any) are guaranteed to also be in premultiplied alpha. // Otherwise, you can get postmultiplied or premultiplied alpha; // you won't know. If you have multiple inputs, you will get the same @@ -121,9 +126,19 @@ public: // If you set this, you should also set needs_linear_light(). INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA, - // Keeps the type of alpha unchanged from input to output. - // Usually appropriate if you process all color channels - // in a linear fashion, and do not change alpha. + // Like INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA, but also guarantees + // that if you get blank alpha in, you also keep blank alpha out. + // This is a somewhat weaker guarantee than DONT_CARE_ALPHA_TYPE, + // but is still useful in many situations, and appropriate when + // e.g. you don't touch alpha at all. + // + // Does not make sense for inputs. + INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK, + + // Keeps the type of alpha (premultiplied, postmultiplied, blank) + // unchanged from input to output. Usually appropriate if you + // process all color channels in a linear fashion, do not change + // alpha, and do not produce any new pixels that have alpha != 1.0. // // Does not make sense for inputs. DONT_CARE_ALPHA_TYPE, @@ -161,12 +176,61 @@ public: // needs mipmaps, you will also get them). virtual bool needs_mipmaps() const { return false; } + // Whether there is a direct correspondence between input and output + // texels. Specifically, the effect must not: + // + // 1. Try to sample in the border (ie., outside the 0.0 to 1.0 area). + // 2. Try to sample between texels. + // 3. Sample with an x- or y-derivative different from -1 or 1. + // (This also means needs_mipmaps() and one_to_one_sampling() + // together would make no sense.) + // + // The most common case for this would be an effect that has an exact + // 1:1-correspondence between input and output texels, e.g. SaturationEffect. + // However, more creative things, like mirroring/flipping or padding, + // would also be allowed. + // + // The primary gain from setting this is that you can sample directly + // from an effect that changes output size (see changes_output_size() below), + // without going through a bounce texture. It won't work for effects that + // set sets_virtual_output_size(), though. + // + // Does not make a lot of sense together with needs_texture_bounce(). + virtual bool one_to_one_sampling() const { return false; } + // Whether this effect wants to output to a different size than - // its input(s) (see inform_input_size(), below). If you set this to - // true, the output will be bounced to a texture (similarly to if the - // next effect set needs_texture_bounce()). + // its input(s) (see inform_input_size(), below). See also + // sets_virtual_output_size() below. virtual bool changes_output_size() const { return false; } + // Whether your get_output_size() function (see below) intends to ever set + // virtual_width different from width, or similar for height. + // It does not make sense to set this to true if changes_output_size() is false. + virtual bool sets_virtual_output_size() const { return changes_output_size(); } + + // Whether this effect is effectively sampling from a a single texture. + // If so, it will override needs_texture_bounce(); however, there are also + // two demands it needs to fulfill: + // + // 1. It needs to be an Input, ie. num_inputs() == 0. + // 2. It needs to allocate exactly one sampler in set_gl_state(), + // and allow dependent effects to change that sampler state. + virtual bool is_single_texture() const { return false; } + + // If set, this effect should never be bounced to an output, even if a + // dependent effect demands texture bounce. + // + // Note that setting this can invoke undefined behavior, up to and including crashing, + // so you should only use it if you have deep understanding of your entire chain + // and Movit's processing of it. The most likely use case is if you have an input + // that's cheap to compute but not a single texture (e.g. YCbCrInput), and want + // to run a ResampleEffect directly from it. Normally, this would require a bounce, + // but it's faster not to. (However, also note that in this case, effective texel + // subpixel precision will be too optimistic, since chroma is already subsampled.) + // + // Has no effect if is_single_texture() is set. + virtual bool override_disable_bounce() const { return false; } + // If changes_output_size() is true, you must implement this to tell // the framework what output size you want. Also, you can set a // virtual width/height, which is the size the next effect (if any) @@ -181,6 +245,45 @@ public: assert(false); } + // Whether this effect uses a compute shader instead of a regular fragment shader. + // Compute shaders are more flexible in that they can have multiple outputs + // for each invocation and also communicate between instances (by using shared + // memory within each group), but are not universally supported. The typical + // pattern would be to check movit_compute_shaders_supported and rewrite the + // graph to use a compute shader effect instead of a regular effect if it is + // available, in order to get better performance. Since compute shaders can reuse + // loads (again typically through shared memory), using needs_texture_bounce() + // is usually not needed, although it is allowed; the best candidates for compute + // shaders are typically those that sample many times from their input + // but can reuse those loads across neighboring instances. + // + // Compute shaders commonly work with unnormalized texture coordinates + // (where coordinates are integers [0..W) and [0..H)), whereas the rest + // of Movit, including any inputs you may want to sample from, works + // with normalized coordinates ([0..1)). Movit gives you uniforms + // PREFIX(inv_output_size) and PREFIX(output_texcoord_adjust) that you + // can use to transform unnormalized to normalized, as well as a macro + // NORMALIZE_TEXTURE_COORDS(vec2) that does it for you. + // + // Since compute shaders have flexible output, it is difficult to chain other + // effects after them in the same phase, and thus, they will always be last. + // (This limitation may be lifted for the special case of one-to-one effects + // in the future.) Furthermore, they cannot write to the framebuffer, just to + // textures, so Movit may have to insert an extra phase just to do the output + // from a texture to the screen in some cases. However, this is transparent + // to both the effect and the user. + virtual bool is_compute_shader() const { return false; } + + // For a compute shader (see the previous member function), what dimensions + // it should be invoked over. Called every frame, before uniforms are set + // (so you are allowed to update uniforms based from this call). + virtual void get_compute_dimensions(unsigned output_width, unsigned output_height, + unsigned *x, unsigned *y, unsigned *z) const { + *x = output_width; + *y = output_height; + *z = 1; + } + // Tells the effect the resolution of each of its input. // This will be called every frame, and always before get_output_size(), // so you can change your output size based on the input if so desired. @@ -198,6 +301,12 @@ public: // if you have several, they will be INPUT1(), INPUT2(), and so on. virtual unsigned num_inputs() const { return 1; } + // Inform the effect that it has been just added to the EffectChain. + // The primary use for this is to store the ResourcePool uesd by + // the chain; for modifications to it, rewrite_graph() below + // is probably a better fit. + virtual void inform_added(EffectChain *chain) {} + // Let the effect rewrite the effect chain as it sees fit. // Most effects won't need to do this, but this is very useful // if you have an effect that consists of multiple sub-effects @@ -212,11 +321,6 @@ public: // itself from all other effects. virtual void rewrite_graph(EffectChain *graph, Node *self) {} - // Outputs one GLSL uniform declaration for each registered parameter - // (see below), with the right prefix prepended to each uniform name. - // If you do not want this behavior, you can override this function. - virtual std::string output_convenience_uniforms() const; - // Returns the GLSL fragment shader string for this effect. virtual std::string output_fragment_shader() = 0; @@ -245,45 +349,76 @@ public: protected: // Register a parameter. Whenever set_*() is called with the same key, // it will update the value in the given pointer (typically a pointer - // to some private member variable in your effect). + // to some private member variable in your effect). It will also + // register a uniform of the same name (plus an arbitrary prefix + // which you can access using the PREFIX macro) that you can access. // // Neither of these take ownership of the pointer. - // int is special since GLSL pre-1.30 doesn't have integer uniforms. - // Thus, ints that you register will _not_ be converted to GLSL uniforms. + // These correspond directly to int/float/vec2/vec3/vec4 in GLSL. void register_int(const std::string &key, int *value); - - // These correspond directly to float/vec2/vec3/vec4 in GLSL. void register_float(const std::string &key, float *value); void register_vec2(const std::string &key, float *values); void register_vec3(const std::string &key, float *values); void register_vec4(const std::string &key, float *values); - // This will register a 1D texture, which will be bound to a sampler - // when your GLSL code runs (so it corresponds 1:1 to a sampler2D uniform - // in GLSL). + // Register uniforms, such that they will automatically be set + // before the shader runs. This is more efficient than set_uniform_* + // in effect_util.h, because it doesn't need to do name lookups + // every time. Also, in the future, it will use uniform buffer objects + // (UBOs) if available to reduce the number of calls into the driver. // - // Note that if you change the contents of , you will need to - // call invalidate_1d_texture() to have the picture re-uploaded on the - // next frame. This is in contrast to all the other parameters, which are - // set anew every frame. - void register_1d_texture(const std::string &key, float *values, size_t size); - void invalidate_1d_texture(const std::string &key); - -private: - struct Texture1D { - float *values; - size_t size; - bool needs_update; - GLuint texture_num; - }; + // May not be called after output_fragment_shader() has returned. + // The pointer must be valid for the entire lifetime of the Effect, + // since the value is pulled from it each execution. The value is + // guaranteed to be read after set_gl_state() for the effect has + // returned, so you can safely update its value from there. + // + // Note that this will also declare the uniform in the shader for you, + // so you should not do that yourself. (This is so it can be part of + // the right uniform block.) However, it is probably a good idea to + // have a commented-out declaration so that it is easier to see the + // type and thus understand the shader on its own. + // + // Calling register_* will automatically imply register_uniform_*, + // except for register_int as noted above. + void register_uniform_sampler2d(const std::string &key, const int *value); + void register_uniform_bool(const std::string &key, const bool *value); + void register_uniform_int(const std::string &key, const int *value); // Note: Requires GLSL 1.30 or newer. + void register_uniform_float(const std::string &key, const float *value); + void register_uniform_vec2(const std::string &key, const float *values); + void register_uniform_vec3(const std::string &key, const float *values); + void register_uniform_vec4(const std::string &key, const float *values); + void register_uniform_float_array(const std::string &key, const float *values, size_t num_values); + void register_uniform_vec2_array(const std::string &key, const float *values, size_t num_values); + void register_uniform_vec3_array(const std::string &key, const float *values, size_t num_values); + void register_uniform_vec4_array(const std::string &key, const float *values, size_t num_values); + void register_uniform_mat3(const std::string &key, const Eigen::Matrix3d *matrix); +private: std::map params_int; std::map params_float; std::map params_vec2; std::map params_vec3; std::map params_vec4; - std::map params_tex_1d; + + // Picked out by EffectChain during finalization. + std::vector > uniforms_image2d; + 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_float_array; + std::vector > uniforms_vec2_array; + std::vector > uniforms_vec3_array; + std::vector > uniforms_vec4_array; + std::vector > uniforms_mat3; + friend class EffectChain; }; -#endif // !defined(_EFFECT_H) +} // namespace movit + +#endif // !defined(_MOVIT_EFFECT_H)