Add the weight combining back, now that we have proper control over the interpolation...
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 28 Oct 2012 17:14:36 +0000 (18:14 +0100)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 28 Oct 2012 17:14:36 +0000 (18:14 +0100)
resample_effect.cpp
resample_effect.h

index 5b01686..c0afe0b 100644 (file)
@@ -40,6 +40,58 @@ unsigned gcd(unsigned a, unsigned b)
        return a;
 }
 
        return a;
 }
 
+unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved)
+{
+       unsigned num_samples_saved = 0;
+       for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
+               // Copy the sample directly; it will be overwritten later if we can combine.
+               if (dst != NULL) {
+                       dst[j * 2 + 0] = src[i * 2 + 0];
+                       dst[j * 2 + 1] = src[i * 2 + 1];
+               }
+
+               if (i == num_src_samples - 1) {
+                       // Last sample; cannot combine.
+                       continue;
+               }
+               assert(num_samples_saved <= max_samples_saved);
+               if (num_samples_saved == max_samples_saved) {
+                       // We could maybe save more here, but other rows can't, so don't bother.
+                       continue;
+               }
+
+               float w1 = src[i * 2 + 0];
+               float w2 = src[(i + 1) * 2 + 0];
+               if (w1 * w2 < 0.0f) {
+                       // Differing signs; cannot combine.
+                       continue;
+               }
+
+               float pos1 = src[i * 2 + 1];
+               float pos2 = src[(i + 1) * 2 + 1];
+               assert(pos2 > pos1);
+
+               float offset, total_weight, sum_sq_error;
+               combine_two_samples(w1, w2, &offset, &total_weight, &sum_sq_error);
+
+               // If the interpolation error is larger than that of one level
+               // at 8-bit precision, don't combine.
+               if (sum_sq_error > 1.0f / (256.0f * 256.0f)) {
+                       continue;
+               }
+
+               // OK, we can combine this and the next sample.
+               if (dst != NULL) {
+                       dst[j * 2 + 0] = total_weight;
+                       dst[j * 2 + 1] = pos1 + offset * (pos2 - pos1);
+               }
+
+               ++i;  // Skip the next sample.
+               ++num_samples_saved;
+       }
+       return num_samples_saved;
+}
+
 }  // namespace
 
 ResampleEffect::ResampleEffect()
 }  // namespace
 
 ResampleEffect::ResampleEffect()
@@ -153,8 +205,6 @@ std::string SingleResamplePassEffect::output_fragment_shader()
 //
 // For horizontal scaling, we fill in the exact same texture;
 // the shader just interprets is differently.
 //
 // For horizontal scaling, we fill in the exact same texture;
 // the shader just interprets is differently.
-//
-// TODO: Support optimization using free linear sampling, like in BlurEffect.
 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
 {
        unsigned src_size, dst_size;
 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
 {
        unsigned src_size, dst_size;
@@ -231,7 +281,7 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std
        // Thus, the kernel width will vary with how much we scale.
        float radius_scaling_factor = std::min(float(dst_size) / float(src_size), 1.0f);
        int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
        // Thus, the kernel width will vary with how much we scale.
        float radius_scaling_factor = std::min(float(dst_size) / float(src_size), 1.0f);
        int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
-       src_samples = int_radius * 2 + 1;
+       int src_samples = int_radius * 2 + 1;
        float *weights = new float[dst_samples * src_samples * 2];
        for (unsigned y = 0; y < dst_samples; ++y) {
                // Find the point around which we want to sample the source image,
        float *weights = new float[dst_samples * src_samples * 2];
        for (unsigned y = 0; y < dst_samples; ++y) {
                // Find the point around which we want to sample the source image,
@@ -248,6 +298,31 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std
                }
        }
 
                }
        }
 
+       // Now make use of the bilinear filtering in the GPU to reduce the number of samples
+       // we need to make. This is a bit more complex than BlurEffect since we cannot combine
+       // two neighboring samples if their weights have differing signs, so we first need to
+       // figure out the maximum number of samples. Then, we downconvert all the weights to
+       // that number -- we could have gone for a variable-length system, but this is simpler,
+       // and the gains would probably be offset by the extra cost of checking when to stop.
+       //
+       // The greedy strategy for combining samples is optimal.
+       src_bilinear_samples = 0;
+       for (unsigned y = 0; y < dst_samples; ++y) {
+               unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX);
+               src_bilinear_samples = std::max<int>(src_bilinear_samples, src_samples - num_samples_saved);
+       }
+
+       // Now that we know the right width, actually combine the samples.
+       float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
+       for (unsigned y = 0; y < dst_samples; ++y) {
+               unsigned num_samples_saved = combine_samples(
+                       weights + (y * src_samples) * 2,
+                       bilinear_weights + (y * src_bilinear_samples) * 2,
+                       src_samples,
+                       src_samples - src_bilinear_samples);
+               assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
+       }       
+
        // Encode as a two-component texture. Note the GL_REPEAT.
        glActiveTexture(GL_TEXTURE0 + *sampler_num);
        check_error();
        // Encode as a two-component texture. Note the GL_REPEAT.
        glActiveTexture(GL_TEXTURE0 + *sampler_num);
        check_error();
@@ -259,10 +334,11 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
        check_error();
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
        check_error();
-       glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_samples, dst_samples, 0, GL_RG, GL_FLOAT, weights);
+       glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights);
        check_error();
 
        delete[] weights;
        check_error();
 
        delete[] weights;
+       delete[] bilinear_weights;
 }
 
 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
 }
 
 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
@@ -287,13 +363,13 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std::
 
        set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
        ++sampler_num;
 
        set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
        ++sampler_num;
-       set_uniform_int(glsl_program_num, prefix, "num_samples", src_samples);
+       set_uniform_int(glsl_program_num, prefix, "num_samples", src_bilinear_samples);
        set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops);
        set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);
 
        // Instructions for how to convert integer sample numbers to positions in the weight texture.
        set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops);
        set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);
 
        // Instructions for how to convert integer sample numbers to positions in the weight texture.
-       set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_samples);
-       set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_samples);
+       set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_bilinear_samples);
+       set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_bilinear_samples);
 
        // We specifically do not want mipmaps on the input texture;
        // they break minification.
 
        // We specifically do not want mipmaps on the input texture;
        // they break minification.
index 1424bbf..5867f25 100644 (file)
@@ -87,7 +87,7 @@ private:
        GLuint texnum;
        int input_width, input_height, output_width, output_height;
        int last_input_width, last_input_height, last_output_width, last_output_height;
        GLuint texnum;
        int input_width, input_height, output_width, output_height;
        int last_input_width, last_input_height, last_output_width, last_output_height;
-       int src_samples, num_loops;
+       int src_bilinear_samples, num_loops;
        float slice_height;
 };
 
        float slice_height;
 };