Fix vp10_accumulate_frame_counts once and for all.

This ensures the multi-threaded and single-threaded encoder/decoder
always uses the same probability contexts.

Change-Id: I6f1e7c6bd8808c390c1dc0a628ae97db3acedf6d

vp10/common/entropymode.h

@@ -116,6 +116,8 @@ typedef struct frame_contexts {
 } FRAME_CONTEXT;
 
 typedef struct FRAME_COUNTS {
+  // Note: This structure should only contain 'unsigned int' fields, or
+  // aggregates built solely from 'unsigned int' fields/elements
   unsigned int kf_y_mode[INTRA_MODES][INTRA_MODES][INTRA_MODES];
   unsigned int y_mode[BLOCK_SIZE_GROUPS][INTRA_MODES];
   unsigned int uv_mode[INTRA_MODES][INTRA_MODES];

vp10/common/thread_common.c

@@ -336,270 +336,15 @@ void vp10_loop_filter_dealloc(VP9LfSync *lf_sync) {
   }
 }
 
-// Accumulate frame counts.
-void vp10_accumulate_frame_counts(VP10_COMMON *cm, FRAME_COUNTS *counts,
-                                 int is_dec) {
-  int i, j, k, l, m;
+// Accumulate frame counts. FRAME_COUNTS consist solely of 'unsigned int'
+// members, so we treat it as an array, and sum over the whole length.
+void vp10_accumulate_frame_counts(VP10_COMMON *cm, FRAME_COUNTS *counts) {
+  unsigned int *const acc = (unsigned int*)&cm->counts;
+  const unsigned int *const cnt = (unsigned int*)counts;
 
-  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
-    for (j = 0; j < INTRA_MODES; j++)
-      cm->counts.y_mode[i][j] += counts->y_mode[i][j];
+  const unsigned int n_counts = sizeof(FRAME_COUNTS)/sizeof(unsigned int);
+  unsigned int i;
 
-  for (i = 0; i < INTRA_MODES; i++)
-    for (j = 0; j < INTRA_MODES; j++)
-      cm->counts.uv_mode[i][j] += counts->uv_mode[i][j];
-
-  for (i = 0; i < PARTITION_CONTEXTS; i++)
-#if CONFIG_EXT_PARTITION_TYPES
-    for (j = 0; j < (i ? EXT_PARTITION_TYPES : PARTITION_TYPES); j++)
-#else
-    for (j = 0; j < PARTITION_TYPES; j++)
-#endif
-      cm->counts.partition[i][j] += counts->partition[i][j];
-
-  if (is_dec) {
-    int n;
-    for (i = 0; i < TX_SIZES; i++)
-      for (j = 0; j < PLANE_TYPES; j++)
-        for (k = 0; k < REF_TYPES; k++)
-          for (l = 0; l < COEF_BANDS; l++)
-            for (m = 0; m < COEFF_CONTEXTS; m++) {
-              cm->counts.eob_branch[i][j][k][l][m] +=
-                  counts->eob_branch[i][j][k][l][m];
-              for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
-                cm->counts.coef[i][j][k][l][m][n] +=
-                    counts->coef[i][j][k][l][m][n];
-            }
-  } else {
-    for (i = 0; i < TX_SIZES; i++)
-      for (j = 0; j < PLANE_TYPES; j++)
-        for (k = 0; k < REF_TYPES; k++)
-          for (l = 0; l < COEF_BANDS; l++)
-            for (m = 0; m < COEFF_CONTEXTS; m++)
-              cm->counts.eob_branch[i][j][k][l][m] +=
-                  counts->eob_branch[i][j][k][l][m];
-                // In the encoder, cm->counts.coef is only updated at frame
-                // level, so not need to accumulate it here.
-                // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
-                //   cm->counts.coef[i][j][k][l][m][n] +=
-                //       counts->coef[i][j][k][l][m][n];
-  }
-
-  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
-    for (j = 0; j < SWITCHABLE_FILTERS; j++)
-      cm->counts.switchable_interp[i][j] += counts->switchable_interp[i][j];
-#if CONFIG_OBMC
-  for (i = 0; i < BLOCK_SIZES; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.obmc[i][j] += counts->obmc[i][j];
-#endif  // CONFIG_OBMC
-
-#if CONFIG_REF_MV
-  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.newmv_mode[i][j] += counts->newmv_mode[i][j];
-
-  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.zeromv_mode[i][j] += counts->zeromv_mode[i][j];
-
-  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.refmv_mode[i][j] += counts->refmv_mode[i][j];
-
-
-  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.drl_mode[i][j] += counts->drl_mode[i][j];
-
-#if CONFIG_EXT_INTER
-  for (j = 0; j < 2; ++j)
-    cm->counts.new2mv_mode[j] += counts->new2mv_mode[j];
-#endif  // CONFIG_EXT_INTER
-#endif
-
-  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
-    for (j = 0; j < INTER_MODES; j++)
-      cm->counts.inter_mode[i][j] += counts->inter_mode[i][j];
-
-#if CONFIG_EXT_INTER
-  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
-    for (j = 0; j < INTER_COMPOUND_MODES; j++)
-      cm->counts.inter_compound_mode[i][j] +=
-          counts->inter_compound_mode[i][j];
-
-  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.interintra[i][j] += counts->interintra[i][j];
-
-  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
-    for (j = 0; j < INTERINTRA_MODES; j++)
-      cm->counts.interintra_mode[i][j] += counts->interintra_mode[i][j];
-
-  for (i = 0; i < BLOCK_SIZES; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.wedge_interintra[i][j] += counts->wedge_interintra[i][j];
-  for (i = 0; i < BLOCK_SIZES; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.wedge_interinter[i][j] += counts->wedge_interinter[i][j];
-#endif  // CONFIG_EXT_INTER
-
-  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.intra_inter[i][j] += counts->intra_inter[i][j];
-
-  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.comp_inter[i][j] += counts->comp_inter[i][j];
-
-  for (i = 0; i < REF_CONTEXTS; i++)
-    for (j = 0; j < (SINGLE_REFS - 1); j++)
-      for (k = 0; k < 2; k++)
-        cm->counts.single_ref[i][j][k] += counts->single_ref[i][j][k];
-
-  for (i = 0; i < REF_CONTEXTS; i++)
-    for (j = 0; j < (COMP_REFS - 1); j++)
-      for (k = 0; k < 2; k++)
-        cm->counts.comp_ref[i][j][k] += counts->comp_ref[i][j][k];
-
-  for (i = 0; i < TX_SIZES - 1; ++i)
-    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
-      for (k = 0; k < i + 2; ++k)
-        cm->counts.tx_size[i][j][k] += counts->tx_size[i][j][k];
-
-  for (i = 0; i < TX_SIZES; ++i)
-    cm->counts.tx_size_totals[i] += counts->tx_size_totals[i];
-
-#if CONFIG_VAR_TX
-  for (i = 0; i < TXFM_PARTITION_CONTEXTS; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.txfm_partition[i][j] += counts->txfm_partition[i][j];
-#endif
-
-  for (i = 0; i < SKIP_CONTEXTS; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.skip[i][j] += counts->skip[i][j];
-
-#if CONFIG_REF_MV
-  for (m = 0; m < NMV_CONTEXTS; ++m) {
-    for (i = 0; i < MV_JOINTS; i++)
-      cm->counts.mv[m].joints[i] += counts->mv[m].joints[i];
-
-    for (k = 0; k < 2; k++) {
-      nmv_component_counts *comps = &cm->counts.mv[m].comps[k];
-      nmv_component_counts *comps_t = &counts->mv[m].comps[k];
-
-      for (i = 0; i < 2; i++) {
-        comps->sign[i] += comps_t->sign[i];
-        comps->class0_hp[i] += comps_t->class0_hp[i];
-        comps->hp[i] += comps_t->hp[i];
-      }
-
-      for (i = 0; i < MV_CLASSES; i++)
-        comps->classes[i] += comps_t->classes[i];
-
-      for (i = 0; i < CLASS0_SIZE; i++) {
-        comps->class0[i] += comps_t->class0[i];
-        for (j = 0; j < MV_FP_SIZE; j++)
-          comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
-      }
-
-      for (i = 0; i < MV_OFFSET_BITS; i++)
-        for (j = 0; j < 2; j++)
-          comps->bits[i][j] += comps_t->bits[i][j];
-
-      for (i = 0; i < MV_FP_SIZE; i++)
-        comps->fp[i] += comps_t->fp[i];
-    }
-  }
-#else
-  for (i = 0; i < MV_JOINTS; i++)
-    cm->counts.mv.joints[i] += counts->mv.joints[i];
-
-  for (k = 0; k < 2; k++) {
-    nmv_component_counts *comps = &cm->counts.mv.comps[k];
-    nmv_component_counts *comps_t = &counts->mv.comps[k];
-
-    for (i = 0; i < 2; i++) {
-      comps->sign[i] += comps_t->sign[i];
-      comps->class0_hp[i] += comps_t->class0_hp[i];
-      comps->hp[i] += comps_t->hp[i];
-    }
-
-    for (i = 0; i < MV_CLASSES; i++)
-      comps->classes[i] += comps_t->classes[i];
-
-    for (i = 0; i < CLASS0_SIZE; i++) {
-      comps->class0[i] += comps_t->class0[i];
-      for (j = 0; j < MV_FP_SIZE; j++)
-        comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
-    }
-
-    for (i = 0; i < MV_OFFSET_BITS; i++)
-      for (j = 0; j < 2; j++)
-        comps->bits[i][j] += comps_t->bits[i][j];
-
-    for (i = 0; i < MV_FP_SIZE; i++)
-      comps->fp[i] += comps_t->fp[i];
-  }
-#endif
-
-#if CONFIG_EXT_TX
-  for (i = 0; i < EXT_TX_SIZES; i++) {
-    int s, k;
-    for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
-      if (use_inter_ext_tx_for_txsize[s][i]) {
-        for (k = 0; k < TX_TYPES; k++)
-          cm->counts.inter_ext_tx[s][i][k] += counts->inter_ext_tx[s][i][k];
-      }
-    }
-    for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
-      if (use_intra_ext_tx_for_txsize[s][i]) {
-        int j;
-        for (j = 0; j < INTRA_MODES; ++j)
-          for (k = 0; k < TX_TYPES; k++)
-            cm->counts.intra_ext_tx[s][i][j][k] +=
-                counts->intra_ext_tx[s][i][j][k];
-      }
-    }
-  }
-#else
-  for (i = 0; i < EXT_TX_SIZES; i++) {
-    int j;
-    for (j = 0; j < TX_TYPES; ++j)
-      for (k = 0; k < TX_TYPES; k++)
-        cm->counts.intra_ext_tx[i][j][k] += counts->intra_ext_tx[i][j][k];
-  }
-  for (i = 0; i < EXT_TX_SIZES; i++) {
-    for (k = 0; k < TX_TYPES; k++)
-      cm->counts.inter_ext_tx[i][k] += counts->inter_ext_tx[i][k];
-  }
-#endif  // CONFIG_EXT_TX
-
-#if CONFIG_SUPERTX
-  for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; i++)
-    for (j = 0; j < TX_SIZES; j++)
-      for (k = 0; k < 2; k++)
-        cm->counts.supertx[i][j][k] += counts->supertx[i][j][k];
-  for (i = 0; i < TX_SIZES; i++)
-    cm->counts.supertx_size[i] += counts->supertx_size[i];
-#endif  // CONFIG_SUPERTX
-
-  for (i = 0; i < PREDICTION_PROBS; i++)
-    for (j = 0; j < 2; j++)
-      cm->counts.seg.pred[i][j] += counts->seg.pred[i][j];
-
-  for (i = 0; i < MAX_SEGMENTS; i++) {
-    cm->counts.seg.tree_total[i] += counts->seg.tree_total[i];
-    cm->counts.seg.tree_mispred[i] += counts->seg.tree_mispred[i];
-  }
-
-#if CONFIG_EXT_INTRA
-  for (i = 0; i < PLANE_TYPES; ++i)
-    for (j = 0; j < 2; ++j)
-      cm->counts.ext_intra[i][j] += counts->ext_intra[i][j];
-  for (i = 0; i < INTRA_FILTERS + 1; ++i)
-    for (j = 0; j < INTRA_FILTERS; ++j)
-      cm->counts.intra_filter[i][j] += counts->intra_filter[i][j];
-#endif  // CONFIG_EXT_INTRA
+  for (i = 0; i < n_counts; i++)
+    acc[i] += cnt[i];
 }

vp10/common/thread_common.h

@@ -56,7 +56,7 @@ void vp10_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                               VP9LfSync *lf_sync);
 
 void vp10_accumulate_frame_counts(struct VP10Common *cm,
-                                 struct FRAME_COUNTS *counts, int is_dec);
+                                 struct FRAME_COUNTS *counts);
 
 #ifdef __cplusplus
 }  // extern "C"

vp10/decoder/decodeframe.c

@@ -3491,7 +3491,7 @@ static const uint8_t *decode_tiles_mt(VP10Decoder *pbi,
   if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
     for (i = 0; i < num_workers; ++i) {
       TileWorkerData *const twd = (TileWorkerData*)pbi->tile_workers[i].data1;
-      vp10_accumulate_frame_counts(cm, &twd->counts, 1);
+      vp10_accumulate_frame_counts(cm, &twd->counts);
     }
   }
 

vp10/encoder/ethread.c

@@ -167,7 +167,7 @@ void vp10_encode_tiles_mt(VP10_COMP *cpi) {
 
     // Accumulate counters.
     if (i < cpi->num_workers - 1) {
-      vp10_accumulate_frame_counts(cm, thread_data->td->counts, 0);
+      vp10_accumulate_frame_counts(cm, thread_data->td->counts);
       accumulate_rd_opt(&cpi->td, thread_data->td);
     }
   }