In ResampleEffect, precompute the Lanczos function into a table.

[movit] / util.cpp

diff --git a/util.cpp b/util.cpp
index e08a0859a25ef09ae7ddba9945ef648b26934260..59f1dcdd67529ef9b4fd54f959fdc49bdc41d633 100644 (file)
--- a/util.cpp
+++ b/util.cpp
@@ -87,17 +87,41 @@ string read_file(const string &filename)
 {
        const string full_pathname = *movit_data_directory + "/" + filename;
 
-       static char buf[131072];
        FILE *fp = fopen(full_pathname.c_str(), "r");
        if (fp == NULL) {
                perror(full_pathname.c_str());
                exit(1);
        }
 
-       int len = fread(buf, 1, sizeof(buf), fp);
+       int ret = fseek(fp, 0, SEEK_END);
+       if (ret == -1) {
+               perror("fseek(SEEK_END)");
+               exit(1);
+       }
+
+       int size = ftell(fp);
+
+       ret = fseek(fp, 0, SEEK_SET);
+       if (ret == -1) {
+               perror("fseek(SEEK_SET)");
+               exit(1);
+       }
+
+       string str;
+       str.resize(size);
+       ret = fread(&str[0], size, 1, fp);
+       if (ret == -1) {
+               perror("fread");
+               exit(1);
+       }
+       if (ret == 0) {
+               fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
+                       size, full_pathname.c_str());
+               exit(1);
+       }
        fclose(fp);
 
-       return string(buf, len);
+       return str;
 }
 
 string read_version_dependent_file(const string &base, const string &extension)
@@ -132,6 +156,7 @@ GLuint compile_shader(const string &shader_src, GLenum type)
        GLint status;
        glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
        if (status == GL_FALSE) {
+               fprintf(stderr, "Failed to compile shader: %s\n", shader_src.c_str());
                exit(1);
        }
 
@@ -152,6 +177,7 @@ string output_glsl_mat3(const string &name, const Eigen::Matrix3d &m)
        stringstream ss;
        ss.imbue(locale("C"));
        ss.precision(8);
+       ss << scientific;
        ss << "const mat3 " << name << " = mat3(\n";
        ss << "    " << m(0,0) << ", " << m(1,0) << ", " << m(2,0) << ",\n";
        ss << "    " << m(0,1) << ", " << m(1,1) << ", " << m(2,1) << ",\n";
@@ -159,12 +185,24 @@ string output_glsl_mat3(const string &name, const Eigen::Matrix3d &m)
        return ss.str();
 }
 
+string output_glsl_float(const string &name, float x)
+{
+       // Use stringstream to be independent of the current locale in a thread-safe manner.
+       stringstream ss;
+       ss.imbue(locale("C"));
+       ss.precision(8);
+       ss << scientific;
+       ss << "const float " << name << " = " << x << ";\n";
+       return ss.str();
+}
+
 string output_glsl_vec2(const string &name, float x, float y)
 {
        // Use stringstream to be independent of the current locale in a thread-safe manner.
        stringstream ss;
        ss.imbue(locale("C"));
        ss.precision(8);
+       ss << scientific;
        ss << "const vec2 " << name << " = vec2(" << x << ", " << y << ");\n";
        return ss.str();
 }
@@ -175,12 +213,13 @@ string output_glsl_vec3(const string &name, float x, float y, float z)
        stringstream ss;
        ss.imbue(locale("C"));
        ss.precision(8);
+       ss << scientific;
        ss << "const vec3 " << name << " = vec3(" << x << ", " << y << ", " << z << ");\n";
        return ss.str();
 }
 
 template<class DestFloat>
-void combine_two_samples(float w1, float w2, float pos1, float pos2, unsigned size,
+void combine_two_samples(float w1, float w2, float pos1, float pos2, float num_subtexels, float inv_num_subtexels,
                          DestFloat *offset, DestFloat *total_weight, float *sum_sq_error)
 {
        assert(movit_initialized);
@@ -199,9 +238,7 @@ void combine_two_samples(float w1, float w2, float pos1, float pos2, unsigned si
        // 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 *= size;  // Move to pixel coordinates,
-       z = lrintf(z / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;  // Round.
-       z /= size;  // Move back to normalized coordinates.
+       z = lrintf(z * num_subtexels) * inv_num_subtexels;
        
        // Choose total weight w so that we minimize total squared error
        // for the effective weights:
@@ -216,22 +253,22 @@ void combine_two_samples(float w1, float w2, float pos1, float pos2, unsigned si
        //   w = (a(1-z) + bz) / ((1-z)² + z²)
        //
        // 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));
+       *total_weight = from_fp64<DestFloat>((w1 + z * (w2 - w1)) / (z * z + (1 - z) * (1 - z)));
 
        if (sum_sq_error != NULL) {
-               float err1 = *total_weight * (1 - z) - w1;
-               float err2 = *total_weight * z - w2;
+               float err1 = to_fp64(*total_weight) * (1 - z) - w1;
+               float err2 = to_fp64(*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 pos2, unsigned size,
+void combine_two_samples<float>(float w1, float w2, float pos1, float pos2, 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 pos2, unsigned size,
+void combine_two_samples<fp16_int_t>(float w1, float w2, float pos1, float pos2, float num_subtexels, float inv_num_subtexels,
                                      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)