Merge "vp9: Add feature to use block source_sad for realtime mode."

vp9/encoder/vp9_block.h

@@ -151,6 +151,8 @@ struct macroblock {
 
   uint8_t sb_is_skin;
 
+  uint8_t skip_low_source_sad;
+
   // Used to save the status of whether a block has a low variance in
   // choose_partitioning. 0 for 64x64, 1~2 for 64x32, 3~4 for 32x64, 5~8 for
   // 32x32, 9~24 for 16x16.

vp9/encoder/vp9_encodeframe.c

@@ -925,7 +925,12 @@ static int choose_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
   int variance4x4downsample[16];
   int segment_id;
   int sb_offset = (cm->mi_stride >> 3) * (mi_row >> 3) + (mi_col >> 3);
-
+  if (cpi->sf.use_source_sad && !is_key_frame) {
+    // The sb_offset2 is to make it consistent with the index in the function
+    // vp9_avg_source_sad() in vp9_ratectrl.c.
+    int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
+    x->skip_low_source_sad = cpi->avg_source_sad_sb[sb_offset2] == 1 ? 1 : 0;
+  }
   set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
   segment_id = xd->mi[0]->segment_id;
   if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
@@ -3857,6 +3862,7 @@ static void encode_nonrd_sb_row(VP9_COMP *cpi, ThreadData *td,
     x->color_sensitivity[0] = 0;
     x->color_sensitivity[1] = 0;
     x->sb_is_skin = 0;
+    x->skip_low_source_sad = 0;
 
     if (seg->enabled) {
       const uint8_t *const map =

vp9/encoder/vp9_encoder.c

@@ -462,6 +462,9 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
   vpx_free(cpi->copied_frame_cnt);
   cpi->copied_frame_cnt = NULL;
 
+  vpx_free(cpi->avg_source_sad_sb);
+  cpi->avg_source_sad_sb = NULL;
+
   vp9_cyclic_refresh_free(cpi->cyclic_refresh);
   cpi->cyclic_refresh = NULL;
 
@@ -3156,7 +3159,7 @@ static void encode_without_recode_loop(VP9_COMP *cpi, size_t *size,
        (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
         cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
        cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
-       cpi->noise_estimate.enabled))
+       cpi->noise_estimate.enabled || cpi->sf.use_source_sad))
     cpi->Last_Source =
         vp9_scale_if_required(cm, cpi->unscaled_last_source,
                               &cpi->scaled_last_source, (cpi->oxcf.pass == 0));

vp9/encoder/vp9_encoder.h

@@ -642,6 +642,8 @@ typedef struct VP9_COMP {
   uint8_t *copied_frame_cnt;
   uint8_t max_copied_frame;
 
+  uint8_t *avg_source_sad_sb;
+
   LevelConstraint level_constraint;
 } VP9_COMP;
 

vp9/encoder/vp9_pickmode.c

@@ -1953,7 +1953,8 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
   if ((!force_skip_low_temp_var || bsize < BLOCK_32X32) && perform_intra_pred &&
       (best_rdc.rdcost == INT64_MAX ||
        (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
-        bsize <= cpi->sf.max_intra_bsize))) {
+        bsize <= cpi->sf.max_intra_bsize)) &&
+      !x->skip_low_source_sad) {
     struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
     int i;
     TX_SIZE best_intra_tx_size = TX_SIZES;

vp9/encoder/vp9_ratectrl.c

@@ -2242,6 +2242,7 @@ void vp9_avg_source_sad(VP9_COMP *cpi) {
         const BLOCK_SIZE bsize = BLOCK_64X64;
         // Loop over sub-sample of frame, compute average sad over 64x64 blocks.
         uint64_t avg_sad = 0;
+        uint64_t tmp_sad = 0;
         int num_samples = 0;
         int sb_cols = (cm->mi_cols + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
         int sb_rows = (cm->mi_rows + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
@@ -2260,9 +2261,12 @@ void vp9_avg_source_sad(VP9_COMP *cpi) {
                  (sbi_row < sb_rows - 1 && sbi_col < sb_cols - 1) &&
                  ((sbi_row % 2 == 0 && sbi_col % 2 == 0) ||
                   (sbi_row % 2 != 0 && sbi_col % 2 != 0)))) {
+              tmp_sad = cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y,
+                                               last_src_ystride);
+              if (cpi->sf.use_source_sad)
+                cpi->avg_source_sad_sb[num_samples] = tmp_sad < 5000 ? 1 : 0;
+              avg_sad += tmp_sad;
               num_samples++;
-              avg_sad += cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y,
-                                                last_src_ystride);
             }
             src_y += 64;
             last_src_y += 64;

vp9/encoder/vp9_speed_features.c

@@ -494,6 +494,15 @@ static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf, int speed,
       sf->mv.search_method = NSTEP;
       sf->mv.fullpel_search_step_param = 6;
     }
+    if (!cpi->use_svc && !cpi->resize_pending && !cpi->resize_state &&
+        !cpi->external_resize && cpi->oxcf.resize_mode == RESIZE_NONE)
+      sf->use_source_sad = 1;
+    if (sf->use_source_sad) {
+      if (cpi->avg_source_sad_sb == NULL) {
+        cpi->avg_source_sad_sb = (uint8_t *)vpx_calloc(
+            (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), sizeof(uint8_t));
+      }
+    }
   }
 
   if (speed >= 8) {