Add shift stage in FASTSSIM computation

Ported from libvpx/nextgenv2: d1c5cd4

Change-Id: If6cb82d38d2bc63fc61d577f81de939c11f73c5a

aom_dsp/fastssim.c

@@ -138,8 +138,8 @@ static void fs_downsample_level(fs_ctx *_ctx, int _l) {
 
 static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
                                  int _s1ystride, const uint8_t *_src2,
-                                 int _s2ystride, int _w, int _h,
+                                 int _s2ystride, int _w, int _h, uint32_t bd,
-                                 uint32_t bit_depth) {
+                                 uint32_t shift) {
   uint32_t *dst1;
   uint32_t *dst2;
   int w;
@@ -160,7 +160,7 @@ static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
       int i1;
       i0 = 2 * i;
       i1 = FS_MINI(i0 + 1, _w);
-      if (bit_depth == 8) {
+      if (bd == 8 && shift == 0) {
         dst1[j * w + i] =
             _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
             _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
@@ -170,12 +170,14 @@ static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
       } else {
         uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
         uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
-        dst1[j * w + i] =
+        dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
-            src1s[j0 * _s1ystride + i0] + src1s[j0 * _s1ystride + i1] +
+                          (src1s[j0 * _s1ystride + i1] >> shift) +
-            src1s[j1 * _s1ystride + i0] + src1s[j1 * _s1ystride + i1];
+                          (src1s[j1 * _s1ystride + i0] >> shift) +
-        dst2[j * w + i] =
+                          (src1s[j1 * _s1ystride + i1] >> shift);
-            src2s[j0 * _s2ystride + i0] + src2s[j0 * _s2ystride + i1] +
+        dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
-            src2s[j1 * _s2ystride + i0] + src2s[j1 * _s2ystride + i1];
+                          (src2s[j0 * _s2ystride + i1] >> shift) +
+                          (src2s[j1 * _s2ystride + i0] >> shift) +
+                          (src2s[j1 * _s2ystride + i1] >> shift);
       }
     }
   }
@@ -442,21 +444,22 @@ static double convert_ssim_db(double _ssim, double _weight) {
 }
 
 static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
-                        int _dystride, int _w, int _h, uint32_t bit_depth) {
+                        int _dystride, int _w, int _h, uint32_t _bd,
+                        uint32_t _shift) {
   fs_ctx ctx;
   double ret;
   int l;
   ret = 1;
   fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
-  fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h,
+  fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _bd,
-                       bit_depth);
+                       _shift);
   for (l = 0; l < FS_NLEVELS - 1; l++) {
-    fs_calc_structure(&ctx, l, bit_depth);
+    fs_calc_structure(&ctx, l, _bd);
     ret *= fs_average(&ctx, l);
     fs_downsample_level(&ctx, l + 1);
   }
-  fs_calc_structure(&ctx, l, bit_depth);
+  fs_calc_structure(&ctx, l, _bd);
-  fs_apply_luminance(&ctx, l, bit_depth);
+  fs_apply_luminance(&ctx, l, _bd);
   ret *= fs_average(&ctx, l);
   fs_ctx_clear(&ctx);
   return ret;
@@ -464,19 +467,24 @@ static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
 
 double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
                          const YV12_BUFFER_CONFIG *dest, double *ssim_y,
-                         double *ssim_u, double *ssim_v, uint32_t bit_depth) {
+                         double *ssim_u, double *ssim_v, uint32_t bd,
+                         uint32_t in_bd) {
   double ssimv;
+  uint32_t bd_shift = 0;
   aom_clear_system_state();
+  assert(bd >= in_bd);
+
+  bd_shift = bd - in_bd;
 
   *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
                       dest->y_stride, source->y_crop_width,
-                      source->y_crop_height, bit_depth);
+                      source->y_crop_height, in_bd, bd_shift);
   *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
                       dest->uv_stride, source->uv_crop_width,
-                      source->uv_crop_height, bit_depth);
+                      source->uv_crop_height, in_bd, bd_shift);
   *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
                       dest->uv_stride, source->uv_crop_width,
-                      source->uv_crop_height, bit_depth);
+                      source->uv_crop_height, in_bd, bd_shift);
   ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
   return convert_ssim_db(ssimv, 1.0);
 }

aom_dsp/ssim.h

@@ -70,13 +70,10 @@ double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
 double aom_calc_ssim(const YV12_BUFFER_CONFIG *source,
                      const YV12_BUFFER_CONFIG *dest, double *weight);
 
-double aom_calc_ssimg(const YV12_BUFFER_CONFIG *source,
-                      const YV12_BUFFER_CONFIG *dest, double *ssim_y,
-                      double *ssim_u, double *ssim_v);
-
 double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
                          const YV12_BUFFER_CONFIG *dest, double *ssim_y,
-                         double *ssim_u, double *ssim_v, uint32_t bit_depth);
+                         double *ssim_u, double *ssim_v, uint32_t bd,
+                         uint32_t in_bd);
 
 #if CONFIG_AOM_HIGHBITDEPTH
 double aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,

av1/encoder/encoder.c

@@ -3604,7 +3604,8 @@ static void compute_internal_stats(AV1_COMP *cpi) {
       }
     }
 
-    frame_all = aom_calc_fastssim(orig, recon, &y, &u, &v, bit_depth);
+    frame_all =
+        aom_calc_fastssim(orig, recon, &y, &u, &v, bit_depth, in_bit_depth);
     adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
     frame_all = aom_psnrhvs(orig, recon, &y, &u, &v, bit_depth, in_bit_depth);
     adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);

test/hbd_metrics_test.cc

@@ -65,13 +65,13 @@ double compute_hbd_fastssim(const YV12_BUFFER_CONFIG *source,
                             const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
                             uint32_t bd) {
   double tempy, tempu, tempv;
-  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, bd);
+  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, bd, in_bd);
 }
 
 double compute_fastssim(const YV12_BUFFER_CONFIG *source,
                         const YV12_BUFFER_CONFIG *dest) {
   double tempy, tempu, tempv;
-  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, 8);
+  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, 8, 8);
 }
 
 double compute_hbd_aomssim(const YV12_BUFFER_CONFIG *source,