Merge "Add Tune for SSIM"

diff --cc vp9/encoder/vp9_encoder.c
index cc7d1bb,8cae609..f92c8a4
--- 1/vp9/encoder/vp9_encoder.c
--- 2/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@@ -4718,15 -4731,73 +4732,81 @@@ static void set_frame_index(VP9_COMP *c
    }
  }
  
+ // Implementation and modifications of C. Yeo, H. L. Tan, and Y. H. Tan, "On
+ // rate distortion optimization using SSIM," Circuits and Systems for Video
+ // Technology, IEEE Transactions on, vol. 23, no. 7, pp. 1170-1181, 2013.
+ // SSIM_VAR_SCALE defines the strength of the bias towards SSIM in RDO.
+ // Some sample values are:
+ // SSIM_VAR_SCALE  avg_psnr   ssim   ms_ssim  (for midres test set)
+ //     16.0          2.312   -3.062  -3.882
+ //     32.0          0.852   -2.260  -2.821
+ //     64.0          0.294   -1.606  -1.925
+ #define SSIM_VAR_SCALE 16.0
+ static void set_mb_ssim_rdmult_scaling(VP9_COMP *cpi) {
+   const double c2 = 0.03 * 0.03 * 255 * 255;
+   VP9_COMMON *cm = &cpi->common;
+   ThreadData *td = &cpi->td;
+   MACROBLOCK *x = &td->mb;
+   MACROBLOCKD *xd = &x->e_mbd;
+   uint8_t *y_buffer = cpi->Source->y_buffer;
+   const int y_stride = cpi->Source->y_stride;
+   const int block_size = BLOCK_64X64;
+ 
+   const int num_8x8_w = num_8x8_blocks_wide_lookup[block_size];
+   const int num_8x8_h = num_8x8_blocks_high_lookup[block_size];
+   const int num_cols = (cm->mi_cols + num_8x8_w - 1) / num_8x8_w;
+   const int num_rows = (cm->mi_rows + num_8x8_h - 1) / num_8x8_h;
+   double log_sum = 0.0;
+   int row, col;
+ 
+   // Loop through each 64x64 block.
+   for (row = 0; row < num_rows; ++row) {
+     for (col = 0; col < num_cols; ++col) {
+       int mi_row, mi_col;
+       double var = 0.0, num_of_var = 0.0;
+       const int index = row * num_cols + col;
+ 
+       for (mi_row = row * num_8x8_h;
+            mi_row < cm->mi_rows && mi_row < (row + 1) * num_8x8_h; ++mi_row) {
+         for (mi_col = col * num_8x8_w;
+              mi_col < cm->mi_cols && mi_col < (col + 1) * num_8x8_w; ++mi_col) {
+           struct buf_2d buf;
+           const int row_offset_y = mi_row << 3;
+           const int col_offset_y = mi_col << 3;
+ 
+           buf.buf = y_buffer + row_offset_y * y_stride + col_offset_y;
+           buf.stride = y_stride;
+           var += vp9_get_sby_variance(cpi, &buf, BLOCK_8X8) / 64.0;
+           num_of_var += 1.0;
+         }
+       }
+       var = var / num_of_var / SSIM_VAR_SCALE;
+       var = 2.0 * var + c2;
+       cpi->mi_ssim_rdmult_scaling_factors[index] = var;
+       log_sum += log(var);
+     }
+   }
+   log_sum = exp(log_sum / (double)(num_rows * num_cols));
+ 
+   for (row = 0; row < num_rows; ++row) {
+     for (col = 0; col < num_cols; ++col) {
+       const int index = row * num_cols + col;
+       cpi->mi_ssim_rdmult_scaling_factors[index] /= log_sum;
+     }
+   }
+ 
+   (void)xd;
+ }
+ 
  // Process the wiener variance in 16x16 block basis.
 +static int qsort_comp(const void *elem1, const void *elem2) {
 +  int a = *((const int *)elem1);
 +  int b = *((const int *)elem2);
 +  if (a > b) return 1;
 +  if (a < b) return -1;
 +  return 0;
 +}
 +
  static void set_mb_wiener_variance(VP9_COMP *cpi) {
    VP9_COMMON *cm = &cpi->common;
    uint8_t *buffer = cpi->Source->y_buffer;