intern/cycles/kernel/filter/filter_nlm_gpu.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_inline void kernel_filter_nlm_calc_difference(int x, int y,
  20                                                          int dx, int dy,
  21                                                          const ccl_global float *ccl_restrict weight_image,
  22                                                          const ccl_global float *ccl_restrict variance_image,
  23                                                          ccl_global float *difference_image,
  24                                                          int4 rect, int w,
  25                                                          int channel_offset,
  26                                                          float a, float k_2)
  27 {
  28         float diff = 0.0f;
  29         int numChannels = channel_offset? 3 : 1;
  30         for(int c = 0; c < numChannels; c++) {
  31                 float cdiff = weight_image[c*channel_offset + y*w+x] - weight_image[c*channel_offset + (y+dy)*w+(x+dx)];
  32                 float pvar = variance_image[c*channel_offset + y*w+x];
  33                 float qvar = variance_image[c*channel_offset + (y+dy)*w+(x+dx)];
  34                 diff += (cdiff*cdiff - a*(pvar + min(pvar, qvar))) / (1e-8f + k_2*(pvar+qvar));
  35         }
  36         if(numChannels > 1) {
  37                 diff *= 1.0f/numChannels;
  38         }
  39         difference_image[y*w+x] = diff;
  40 }
  41
  42 ccl_device_inline void kernel_filter_nlm_blur(int x, int y,
  43                                               const ccl_global float *ccl_restrict difference_image,
  44                                               ccl_global float *out_image,
  45                                               int4 rect, int w, int f)
  46 {
  47         float sum = 0.0f;
  48         const int low = max(rect.y, y-f);
  49         const int high = min(rect.w, y+f+1);
  50         for(int y1 = low; y1 < high; y1++) {
  51                 sum += difference_image[y1*w+x];
  52         }
  53         sum *= 1.0f/(high-low);
  54         out_image[y*w+x] = sum;
  55 }
  56
  57 ccl_device_inline void kernel_filter_nlm_calc_weight(int x, int y,
  58                                                      const ccl_global float *ccl_restrict difference_image,
  59                                                      ccl_global float *out_image,
  60                                                      int4 rect, int w, int f)
  61 {
  62         float sum = 0.0f;
  63         const int low = max(rect.x, x-f);
  64         const int high = min(rect.z, x+f+1);
  65         for(int x1 = low; x1 < high; x1++) {
  66                 sum += difference_image[y*w+x1];
  67         }
  68         sum *= 1.0f/(high-low);
  69         out_image[y*w+x] = fast_expf(-max(sum, 0.0f));
  70 }
  71
  72 ccl_device_inline void kernel_filter_nlm_update_output(int x, int y,
  73                                                        int dx, int dy,
  74                                                        const ccl_global float *ccl_restrict difference_image,
  75                                                        const ccl_global float *ccl_restrict image,
  76                                                        ccl_global float *out_image,
  77                                                        ccl_global float *accum_image,
  78                                                        int4 rect, int w, int f)
  79 {
  80         float sum = 0.0f;
  81         const int low = max(rect.x, x-f);
  82         const int high = min(rect.z, x+f+1);
  83         for(int x1 = low; x1 < high; x1++) {
  84                 sum += difference_image[y*w+x1];
  85         }
  86         sum *= 1.0f/(high-low);
  87         if(out_image) {
  88                 accum_image[y*w+x] += sum;
  89                 out_image[y*w+x] += sum*image[(y+dy)*w+(x+dx)];
  90         }
  91         else {
  92                 accum_image[y*w+x] = sum;
  93         }
  94 }
  95
  96 ccl_device_inline void kernel_filter_nlm_construct_gramian(int fx, int fy,
  97                                                            int dx, int dy,
  98                                                            const ccl_global float *ccl_restrict difference_image,
  99                                                            const ccl_global float *ccl_restrict buffer,
 100                                                            const ccl_global float *ccl_restrict transform,
 101                                                            ccl_global int *rank,
 102                                                            ccl_global float *XtWX,
 103                                                            ccl_global float3 *XtWY,
 104                                                            int4 rect,
 105                                                            int4 filter_rect,
 106                                                            int w, int h, int f,
 107                                                            int pass_stride,
 108                                                            int localIdx)
 109 {
 110         int y = fy + filter_rect.y;
 111         int x = fx + filter_rect.x;
 112         const int low = max(rect.x, x-f);
 113         const int high = min(rect.z, x+f+1);
 114         float sum = 0.0f;
 115         for(int x1 = low; x1 < high; x1++) {
 116                 sum += difference_image[y*w+x1];
 117         }
 118         float weight = sum * (1.0f/(high - low));
 119
 120         int storage_ofs = fy*filter_rect.z + fx;
 121         transform += storage_ofs;
 122         rank += storage_ofs;
 123         XtWX += storage_ofs;
 124         XtWY += storage_ofs;
 125
 126         kernel_filter_construct_gramian(x, y,
 127                                         filter_rect.z*filter_rect.w,
 128                                         dx, dy, w, h,
 129                                         pass_stride,
 130                                         buffer,
 131                                         transform, rank,
 132                                         weight, XtWX, XtWY,
 133                                         localIdx);
 134 }
 135
 136 ccl_device_inline void kernel_filter_nlm_normalize(int x, int y,
 137                                                    ccl_global float *out_image,
 138                                                    const ccl_global float *ccl_restrict accum_image,
 139                                                    int4 rect, int w)
 140 {
 141         out_image[y*w+x] /= accum_image[y*w+x];
 142 }
 143
 144 CCL_NAMESPACE_END