X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=util.h;h=5e8cc6d721da56616728bbffdafd599baefc5554;hp=a89d3a2adf4be608e296b2357e0b79f6104b3151;hb=b4ec29a251e118f463ce940ffaf0945188bf6894;hpb=42f0fd5ccbb3560a76d55f3e725416a5e0f93523

diff --git a/util.h b/util.h
index a89d3a2..5e8cc6d 100644
--- a/util.h
+++ b/util.h
@@ -8,8 +8,10 @@
 #include <stdlib.h>
 #include <Eigen/Core>
 #include <string>
+#include "defs.h"
+#include "fp16.h"
 
-#define BUFFER_OFFSET(i) ((char *)NULL + (i))
+#define BUFFER_OFFSET(i) ((char *)nullptr + (i))
 
 namespace movit {
 
@@ -38,6 +40,11 @@ void print_3x3_matrix(const Eigen::Matrix3d &m);
 // Output a GLSL 3x3 matrix declaration.
 std::string output_glsl_mat3(const std::string &name, const Eigen::Matrix3d &m);
 
+// Output GLSL scalar, 2-length and 3-length vector declarations.
+std::string output_glsl_float(const std::string &name, float x);
+std::string output_glsl_vec2(const std::string &name, float x, float y);
+std::string output_glsl_vec3(const std::string &name, float x, float y, float z);
+
 // Calculate a / b, rounding up. Does not handle overflow correctly.
 unsigned div_round_up(unsigned a, unsigned b);
 
@@ -49,12 +56,16 @@ enum CombineRoundingBehavior {
 // Calculate where to sample, and with what weight, if one wants to use
 // the GPU's bilinear hardware to sample w1 * x[pos1] + w2 * x[pos2],
 // where pos1 and pos2 must be normalized coordinates describing neighboring
-// pixels in the mipmap level at which you sample, and the total number of
-// pixels (in given mipmap level) is <size>.
+// texels in the mipmap level at which you sample. <num_subtexels> is the
+// number of distinct accessible subtexels in the given mipmap level,
+// calculated by num_texels / movit_texel_subpixel_precision. It is a float
+// for performance reasons, even though it is expected to be a whole number.
+// <inv_num_subtexels> is simply its inverse (1/x). <pos1_pos2_diff> is
+// (pos2-pos1) and <inv_pos1_pos2_diff> is 1/(pos2-pos1).
 //
 // Note that since the GPU might have limited precision in its linear
 // interpolation, the effective weights might be different from the ones you
-// asked for. sum_sq_error, if not NULL, will contain the sum of the
+// asked for. sum_sq_error, if not nullptr, will contain the sum of the
 // (estimated) squared errors of the two weights.
 //
 // The answer, in "offset", comes as a normalized coordinate,
@@ -62,8 +73,51 @@ enum CombineRoundingBehavior {
 // is COMBINE_ROUND_TO_FP16, the coordinate is assumed to be stored as a
 // rounded fp16 value. This enables more precise calculation of total_weight
 // and sum_sq_error.
-void combine_two_samples(float w1, float w2, float pos1, float pos2, unsigned size, CombineRoundingBehavior rounding_behavior,
-                         float *offset, float *total_weight, float *sum_sq_error);
+template<class DestFloat>
+void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels,
+                         DestFloat *offset, DestFloat *total_weight, float *sum_sq_error)
+{
+	assert(w1 * w2 >= 0.0f);  // Should not have differing signs.
+	float z;  // Normalized 0..1 between pos1 and pos2.
+	if (fabs(w1 + w2) < 1e-6) {
+		z = 0.5f;
+	} else {
+		z = w2 / (w1 + w2);
+	}
+
+	// Round to the desired precision. Note that this might take z outside the 0..1 range.
+	*offset = from_fp32<DestFloat>(pos1 + z * pos1_pos2_diff);
+	z = (to_fp32(*offset) - pos1) * inv_pos1_pos2_diff;
+
+	// Round to the minimum number of bits we have measured earlier.
+	// The card will do this for us anyway, but if we know what the real z
+	// is, we can pick a better total_weight below.
+	z = lrintf(z * num_subtexels) * inv_num_subtexels;
+
+	// Choose total weight w so that we minimize total squared error
+	// for the effective weights:
+	//
+	//   e = (w(1-z) - a)² + (wz - b)²
+	//
+	// Differentiating by w and setting equal to zero:
+	//
+	//   2(w(1-z) - a)(1-z) + 2(wz - b)z = 0
+	//   w(1-z)² - a(1-z) + wz² - bz = 0
+	//   w((1-z)² + z²) = a(1-z) + bz
+	//   w = (a(1-z) + bz) / ((1-z)² + z²)
+	//
+	// If z had infinite precision, this would simply reduce to w = w1 + w2.
+	*total_weight = from_fp32<DestFloat>((w1 + z * (w2 - w1)) / (z * z + (1 - z) * (1 - z)));
+
+	if (sum_sq_error != nullptr) {
+		float err1 = to_fp32(*total_weight) * (1 - z) - w1;
+		float err2 = to_fp32(*total_weight) * z - w2;
+		*sum_sq_error = err1 * err1 + err2 * err2;
+	}
+}
+
+// Create a VBO with the given data. Returns the VBO number.
+GLuint generate_vbo(GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data);
 
 // Create a VBO with the given data, and bind it to the vertex attribute
 // with name <attribute_name>. Returns the VBO number.
@@ -81,12 +135,15 @@ unsigned next_power_of_two(unsigned v);
 // back into anything you intend to pass into OpenGL.
 void *get_gl_context_identifier();
 
+// Used in the check_error() macro, below.
+void abort_gl_error(GLenum err, const char *filename, int line) DOES_NOT_RETURN;
+
 }  // namespace movit
 
 #ifdef NDEBUG
 #define check_error()
 #else
-#define check_error() { int err = glGetError(); if (err != GL_NO_ERROR) { printf("GL error 0x%x at %s:%d\n", err, __FILE__, __LINE__); abort(); } }
+#define check_error() { GLenum err = glGetError(); if (err != GL_NO_ERROR) { movit::abort_gl_error(err, __FILE__, __LINE__); } }
 #endif
 
 // CHECK() is like assert(), but retains any side effects no matter the compilation mode.