X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.h;h=918769b3872451576f329a5252568fc214889a03;hp=70dc256f41f19616d4b2fce2875ec98d049e7847;hb=f216b7bef5a968c89f6fc78e83cc26a91e504a8a;hpb=2fd06b9c44225d1e740cb2de08a9dfa5c9cd0031

diff --git a/effect_chain.h b/effect_chain.h
index 70dc256..918769b 100644
--- a/effect_chain.h
+++ b/effect_chain.h
@@ -23,6 +23,7 @@
 
 #include <epoxy/gl.h>
 #include <stdio.h>
+#include <list>
 #include <map>
 #include <set>
 #include <string>
@@ -98,6 +99,19 @@ enum OutputOrigin {
 	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:
@@ -179,7 +193,8 @@ struct Phase {
 	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;
 };
@@ -275,6 +290,46 @@ public:
 		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.
@@ -286,7 +341,6 @@ public:
 	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);
@@ -436,6 +490,8 @@ private:
 	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;