Revert the optimization of the bilinear weights.

[movit] / util.cpp

diff --git a/util.cpp b/util.cpp
index 579efb35409d4a17ddc22838ce3dbcec4c82fc92..b12ee80712ba83e541f28472760167e5e8fb1d22 100644 (file)
--- a/util.cpp
+++ b/util.cpp
@@ -6,6 +6,7 @@
 #include <string.h>
 #include <Eigen/Core>
 
+#include "fp16.h"
 #include "init.h"
 #include "util.h"
 
@@ -157,21 +158,29 @@ string output_glsl_mat3(const string &name, const Eigen::Matrix3d &m)
        return buf;
 }
 
-void combine_two_samples(float w1, float w2, float *offset, float *total_weight, float *sum_sq_error)
+template<class DestFloat>
+void combine_two_samples(float w1, float w2, float pos1, float pos2, unsigned size,
+                         DestFloat *offset, DestFloat *total_weight, float *sum_sq_error)
 {
        assert(movit_initialized);
        assert(w1 * w2 >= 0.0f);  // Should not have differing signs.
-       float z;  // Just a shorter name for offset.
+       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_fp64<DestFloat>(pos1 + z * (pos2 - pos1));
+       z = (to_fp64(*offset) - pos1) / (pos2 - pos1);
+
        // 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 / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;
+       z *= size;  // Move to pixel coordinates,
+       z = lrintf(z / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;  // Round.
+       z /= size;  // Move back to normalized coordinates.
        
        // Choose total weight w so that we minimize total squared error
        // for the effective weights:
@@ -187,18 +196,23 @@ void combine_two_samples(float w1, float w2, float *offset, float *total_weight,
        //
        // If z had infinite precision, this would simply reduce to w = w1 + w2.
        *total_weight = (w1 * (1 - z) + w2 * z) / (z * z + (1 - z) * (1 - z));
-       *offset = z;
 
        if (sum_sq_error != NULL) {
                float err1 = *total_weight * (1 - z) - w1;
                float err2 = *total_weight * z - w2;
                *sum_sq_error = err1 * err1 + err2 * err2;
        }
-
-       assert(*offset >= 0.0f);
-       assert(*offset <= 1.0f);
 }
 
+// Explicit instantiations.
+template
+void combine_two_samples<float>(float w1, float w2, float pos1, float pos2, unsigned size,
+                                float *offset, float *total_weight, float *sum_sq_error);
+
+template
+void combine_two_samples<fp16_int_t>(float w1, float w2, float pos1, float pos2, unsigned size,
+                                     fp16_int_t *offset, fp16_int_t *total_weight, float *sum_sq_error);
+
 GLuint fill_vertex_attribute(GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 {
        int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());