intern/cycles/kernel/filter/filter_transform.h

   1 /*
   2  * Copyright 2011-2017 Blender Foundation
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  * http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 CCL_NAMESPACE_BEGIN
  18
  19 ccl_device void kernel_filter_construct_transform(float ccl_restrict_ptr buffer,
  20                                                   int x, int y, int4 rect,
  21                                                   int pass_stride,
  22                                                   float *transform, int *rank,
  23                                                   int radius, float pca_threshold)
  24 {
  25         int buffer_w = align_up(rect.z - rect.x, 4);
  26
  27         float features[DENOISE_FEATURES];
  28
  29         /* Temporary storage, used in different steps of the algorithm. */
  30         float tempmatrix[DENOISE_FEATURES*DENOISE_FEATURES];
  31         float tempvector[2*DENOISE_FEATURES];
  32         float ccl_restrict_ptr pixel_buffer;
  33         int2 pixel;
  34
  35
  36
  37
  38         /* === Calculate denoising window. === */
  39         int2 low  = make_int2(max(rect.x, x - radius),
  40                               max(rect.y, y - radius));
  41         int2 high = make_int2(min(rect.z, x + radius + 1),
  42                               min(rect.w, y + radius + 1));
  43
  44
  45
  46
  47         /* === Shift feature passes to have mean 0. === */
  48         float feature_means[DENOISE_FEATURES];
  49         math_vector_zero(feature_means, DENOISE_FEATURES);
  50         FOR_PIXEL_WINDOW {
  51                 filter_get_features(pixel, pixel_buffer, features, NULL, pass_stride);
  52                 math_vector_add(feature_means, features, DENOISE_FEATURES);
  53         } END_FOR_PIXEL_WINDOW
  54
  55         float pixel_scale = 1.0f / ((high.y - low.y) * (high.x - low.x));
  56         math_vector_scale(feature_means, pixel_scale, DENOISE_FEATURES);
  57
  58         /* === Scale the shifted feature passes to a range of [-1; 1], will be baked into the transform later. === */
  59         float *feature_scale = tempvector;
  60         math_vector_zero(feature_scale, DENOISE_FEATURES);
  61
  62         FOR_PIXEL_WINDOW {
  63                 filter_get_feature_scales(pixel, pixel_buffer, features, feature_means, pass_stride);
  64                 math_vector_max(feature_scale, features, DENOISE_FEATURES);
  65         } END_FOR_PIXEL_WINDOW
  66
  67         filter_calculate_scale(feature_scale);
  68
  69
  70         /* === Generate the feature transformation. ===
  71          * This transformation maps the DENOISE_FEATURES-dimentional feature space to a reduced feature (r-feature) space
  72          * which generally has fewer dimensions. This mainly helps to prevent overfitting. */
  73         float* feature_matrix = tempmatrix;
  74         math_matrix_zero(feature_matrix, DENOISE_FEATURES);
  75         FOR_PIXEL_WINDOW {
  76                 filter_get_features(pixel, pixel_buffer, features, feature_means, pass_stride);
  77                 math_vector_mul(features, feature_scale, DENOISE_FEATURES);
  78                 math_matrix_add_gramian(feature_matrix, DENOISE_FEATURES, features, 1.0f);
  79         } END_FOR_PIXEL_WINDOW
  80
  81         math_matrix_jacobi_eigendecomposition(feature_matrix, transform, DENOISE_FEATURES, 1);
  82         *rank = 0;
  83         if(pca_threshold < 0.0f) {
  84                 float threshold_energy = 0.0f;
  85                 for(int i = 0; i < DENOISE_FEATURES; i++) {
  86                         threshold_energy += feature_matrix[i*DENOISE_FEATURES+i];
  87                 }
  88                 threshold_energy *= 1.0f - (-pca_threshold);
  89
  90                 float reduced_energy = 0.0f;
  91                 for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
  92                         if(i >= 2 && reduced_energy >= threshold_energy)
  93                                 break;
  94                         float s = feature_matrix[i*DENOISE_FEATURES+i];
  95                         reduced_energy += s;
  96                 }
  97         }
  98         else {
  99                 for(int i = 0; i < DENOISE_FEATURES; i++, (*rank)++) {
 100                         float s = feature_matrix[i*DENOISE_FEATURES+i];
 101                         if(i >= 2 && sqrtf(s) < pca_threshold)
 102                                 break;
 103                 }
 104         }
 105
 106         /* Bake the feature scaling into the transformation matrix. */
 107         for(int i = 0; i < (*rank); i++) {
 108                 math_vector_mul(transform + i*DENOISE_FEATURES, feature_scale, DENOISE_FEATURES);
 109         }
 110         math_matrix_transpose(transform, DENOISE_FEATURES, 1);
 111 }
 112
 113 CCL_NAMESPACE_END