Rework coeff probability model update for rtc coding

This commit reworks the ONE_LOOP_REDUCED coefficient probability
model update process. It allows model update for every coefficient
across the spectrum at a coarser resolution, instead of performing
precise update only for certain subset of probability models.

The overall runtime remains nearly same (<1% change) for speed -6.
The compression performance is improved by 7.5% in PSNR for speed
-5 and 4.57% for speed -6, respectively.

Change-Id: Ifb17136382ee7e39a9f34ff4a4f09a753125c8d1

vp9/encoder/vp9_bitstream.c

@@ -535,6 +535,8 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
   const vp9_prob upd = DIFF_UPDATE_PROB;
   const int entropy_nodes_update = UNCONSTRAINED_NODES;
   int i, j, k, l, t;
+  int stepsize = cpi->sf.coeff_prob_appx_step;
+
   switch (cpi->sf.use_fast_coef_updates) {
     case TWO_LOOP: {
       /* dry run to see if there is any update at all needed */
@@ -552,7 +554,7 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd, stepsize);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
@@ -590,7 +592,7 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd, stepsize);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t],
@@ -613,14 +615,14 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
 
     case ONE_LOOP:
     case ONE_LOOP_REDUCED: {
-      const int prev_coef_contexts_to_update =
-          cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED ?
-              COEFF_CONTEXTS >> 1 : COEFF_CONTEXTS;
-      const int coef_band_to_update =
-          cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED ?
-              COEF_BANDS >> 1 : COEF_BANDS;
       int updates = 0;
       int noupdates_before_first = 0;
+
+      if (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8) {
+        vp9_write_bit(bc, 0);
+        return;
+      }
+
       for (i = 0; i < PLANE_TYPES; ++i) {
         for (j = 0; j < REF_TYPES; ++j) {
           for (k = 0; k < COEF_BANDS; ++k) {
@@ -631,21 +633,19 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                 vp9_prob *oldp = old_coef_probs[i][j][k][l] + t;
                 int s;
                 int u = 0;
-                if (l >= prev_coef_contexts_to_update ||
-                    k >= coef_band_to_update) {
-                  u = 0;
-                } else {
-                  if (t == PIVOT_NODE)
+
+                if (t == PIVOT_NODE) {
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                        old_coef_probs[i][j][k][l], &newp, upd);
-                  else
+                      old_coef_probs[i][j][k][l], &newp, upd, stepsize);
+                } else {
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t],
                       *oldp, &newp, upd);
+                }
+
                 if (s > 0 && newp != *oldp)
                   u = 1;
-                }
                 updates += u;
                 if (u == 0 && updates == 0) {
                   noupdates_before_first++;

vp9/encoder/vp9_speed_features.c

@@ -249,7 +249,6 @@ static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf,
     sf->use_uv_intra_rd_estimate = 1;
     sf->skip_encode_sb = 1;
     sf->mv.subpel_iters_per_step = 1;
-    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
     sf->adaptive_rd_thresh = 4;
     sf->mode_skip_start = 6;
     sf->allow_skip_recode = 0;
@@ -304,6 +303,9 @@ static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf,
     // This feature is only enabled when partition search is disabled.
     sf->reuse_inter_pred_sby = 1;
     sf->partition_search_breakout_rate_thr = 200;
+    sf->coeff_prob_appx_step = 4;
+    sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED;
+
     if (!is_keyframe) {
       int i;
       if (content == VP9E_CONTENT_SCREEN) {
@@ -394,6 +396,7 @@ void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
   sf->mv.subpel_force_stop = 0;
   sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
   sf->mv.reduce_first_step_size = 0;
+  sf->coeff_prob_appx_step = 1;
   sf->mv.auto_mv_step_size = 0;
   sf->mv.fullpel_search_step_param = 6;
   sf->comp_inter_joint_search_thresh = BLOCK_4X4;

vp9/encoder/vp9_speed_features.h

@@ -236,6 +236,9 @@ typedef struct SPEED_FEATURES {
   // level within a frame.
   int allow_skip_recode;
 
+  // Coefficient probability model approximation step size
+  int coeff_prob_appx_step;
+
   // The threshold is to determine how slow the motino is, it is used when
   // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
   MOTION_THRESHOLD lf_motion_threshold;

vp9/encoder/vp9_subexp.c

@@ -140,7 +140,8 @@ int vp9_prob_diff_update_savings_search(const unsigned int *ct,
 int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
                                               const vp9_prob *oldp,
                                               vp9_prob *bestp,
-                                              vp9_prob upd) {
+                                              vp9_prob upd,
+                                              int stepsize) {
   int i, old_b, new_b, update_b, savings, bestsavings, step;
   int newp;
   vp9_prob bestnewp, newplist[ENTROPY_NODES], oldplist[ENTROPY_NODES];
@@ -153,9 +154,9 @@ int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
   bestsavings = 0;
   bestnewp = oldp[PIVOT_NODE];
 
-  step = (*bestp > oldp[PIVOT_NODE] ? -1 : 1);
-
-  for (newp = *bestp; newp != oldp[PIVOT_NODE]; newp += step) {
+  if (*bestp > oldp[PIVOT_NODE]) {
+    step = -stepsize;
+    for (newp = *bestp; newp > oldp[PIVOT_NODE]; newp += step) {
       if (newp < 1 || newp > 255)
         continue;
       newplist[PIVOT_NODE] = newp;
@@ -171,6 +172,26 @@ int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
         bestnewp = newp;
       }
     }
+  } else {
+    step = stepsize;
+    for (newp = *bestp; newp < oldp[PIVOT_NODE]; newp += step) {
+      if (newp < 1 || newp > 255)
+        continue;
+      newplist[PIVOT_NODE] = newp;
+      vp9_model_to_full_probs(newplist, newplist);
+      for (i = UNCONSTRAINED_NODES, new_b = 0; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i]);
+      new_b += cost_branch256(ct + 2 * PIVOT_NODE, newplist[PIVOT_NODE]);
+      update_b = prob_diff_update_cost(newp, oldp[PIVOT_NODE]) +
+          vp9_cost_upd256;
+      savings = old_b - new_b - update_b;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+
   *bestp = bestnewp;
   return bestsavings;
 }

vp9/encoder/vp9_subexp.h

@@ -30,7 +30,8 @@ int vp9_prob_diff_update_savings_search(const unsigned int *ct,
 int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
                                               const vp9_prob *oldp,
                                               vp9_prob *bestp,
-                                              vp9_prob upd);
+                                              vp9_prob upd,
+                                              int stepsize);
 
 #ifdef __cplusplus
 }  // extern "C"