Inline combine_two_samples (and remove an obsolete assert). Helps 13–14% on ResampleE...

[movit] / util.h

diff --git a/util.h b/util.h
index 577140b50f3ba7a05796a61f27a6683606c6b227..feefea5fa9f1fa1b992570b69791ed7e308cdce8 100644 (file)
--- a/util.h
+++ b/util.h
@@ -9,6 +9,7 @@
 #include <Eigen/Core>
 #include <string>
 #include "defs.h"
+#include "fp16.h"
 
 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
 
@@ -74,7 +75,46 @@ enum CombineRoundingBehavior {
 // and 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);
+                         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 != NULL) {
+               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);