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

[movit] / util.cpp

diff --git a/util.cpp b/util.cpp
index 54c815cc2be0f7e45967bcfddabea04cd4060639..401664424fe2149cb97c4f9718d341f1f8dfa812 100644 (file)
--- a/util.cpp
+++ b/util.cpp
@@ -219,59 +219,6 @@ string output_glsl_vec3(const string &name, float x, float y, float z)
        return ss.str();
 }
 
-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(movit_initialized);
-       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;
-       }
-}
-
-// Explicit instantiations.
-template
-void combine_two_samples<float>(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels,
-                                float *offset, float *total_weight, float *sum_sq_error);
-
-template
-void combine_two_samples<fp16_int_t>(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels,
-                                     fp16_int_t *offset, fp16_int_t *total_weight, float *sum_sq_error);
-
 GLuint generate_vbo(GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 {
        GLuint vbo;