#include <epoxy/gl.h>
#include <stdio.h>
+#include <list>
#include <map>
#include <set>
#include <string>
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:
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);
std::vector<Uniform<Eigen::Matrix3d> > uniforms_mat3;
// For measurement of GPU time used.
- GLuint timer_query_object;
+ std::list<GLuint> timer_query_objects_running;
+ std::list<GLuint> timer_query_objects_free;
uint64_t time_elapsed_ns;
uint64_t num_measured_iterations;
};
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 RGB output. Note that you can only have one output.
+ // Adds an RGBA output. Note that you can have at most one RGBA output and one
+ // Y'CbCr output (see below for details).
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).
+ //
+ // If you have both RGBA and Y'CbCr output, 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);
+ 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.
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 og 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.
void reset_phase_timing();
void print_phase_timing();
- //void render(unsigned char *src, unsigned char *dst);
void render_to_screen()
{
render_to_fbo(0, 0, 0);
// 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
Phase *construct_phase(Node *output, std::map<Node *, Phase *> *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<Phase *, GLuint> *output_textures, std::set<Phase *> *generated_mipmaps);
+ void execute_phase(Phase *phase, bool last_phase,
+ std::set<GLint> *bound__attribute_indices,
+ std::map<Phase *, GLuint> *output_textures,
+ std::set<Phase *> *generated_mipmaps);
// Set up uniforms for one phase. The program must already be bound.
void setup_uniforms(Phase *phase);
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.
+ bool output_color_rgba, output_color_ycbcr;
+ YCbCrFormat output_ycbcr_format; // If output_color_ycbcr is true.
+ YCbCrOutputSplitting output_ycbcr_splitting; // If output_color_ycbcr is true.
std::vector<Node *> nodes;
std::map<Effect *, Node *> node_map;
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.
ResourcePool *resource_pool;
bool owns_resource_pool;
projects / movit / blobdiff