Prune BAWP by GOP usage at speed >= 4

av2/encoder/encodeframe.c

@@ -2038,6 +2038,22 @@ static void get_histogram(int32_t *hist, const YV12_BUFFER_CONFIG *buf,
   }
 }
 
+// GOP-level BAWP pruning (sf->inter_sf.prune_bawp_by_gop_usage). Returns 1 if
+// BAWP is to be turned off for the current frame: a frame at pyramid level
+// >= 3 whose GOP's level-1 and level-2 frames used BAWP on average on at most
+// half of their BAWP-eligible area. Intra and level-1 frames start a new GOP.
+static int should_prune_bawp_by_gop_usage(AV2_COMP *const cpi) {
+  const AV2_COMMON *const cm = &cpi->common;
+  const unsigned int level = cm->current_frame.pyramid_level;
+  if (frame_is_intra_only(cm) || level == 1) {
+    cpi->bawp_gop_usage[0] = cpi->bawp_gop_usage[1] = -1;
+    return 0;
+  }
+  if (level < 3) return 0;
+  if (cpi->bawp_gop_usage[0] < 0 || cpi->bawp_gop_usage[1] < 0) return 0;
+  return cpi->bawp_gop_usage[0] + cpi->bawp_gop_usage[1] <= 1.0;
+}
+
 static int av2_set_on_bawp_picture_level(AV2_COMP *const cpi) {
   int enable_curr_pic_bawp = 0;
   assert(cpi->sf.inter_sf.enable_fast_bawp == 1);
@@ -2214,6 +2230,8 @@ static AVM_INLINE void encode_frame_internal(AV2_COMP *cpi) {
   av2_zero(rdc->comp_pred_diff);
   av2_zero(rdc->tx_type_used);
   av2_zero(rdc->warped_used);
+  rdc->bawp_eligible_area = 0;
+  rdc->bawp_used_area = 0;
 
   for (i = 0; i < CCSO_NUM_COMPONENTS; ++i) {
     cm->ccso_info.ccso_frame_flag = 0;
@@ -2372,6 +2390,10 @@ static AVM_INLINE void encode_frame_internal(AV2_COMP *cpi) {
 #endif
   av2_compute_global_motion_facade(cpi);
 
+  // Called for every frame, so that the GOP state follows the coded frames.
+  if (cpi->sf.inter_sf.prune_bawp_by_gop_usage &&
+      should_prune_bawp_by_gop_usage(cpi))
+    features->enable_bawp = 0;
   if (features->enable_bawp && cpi->sf.inter_sf.enable_fast_bawp)
     features->enable_bawp = av2_set_on_bawp_picture_level(cpi);
 
@@ -2439,6 +2461,18 @@ static AVM_INLINE void encode_frame_internal(AV2_COMP *cpi) {
     }
   }
 
+  // BAWP usage of the GOP's level-1 and level-2 frames, for its later frames.
+  // A frame coded with BAWP off has no eligible area and counts as 0.
+  if (cpi->sf.inter_sf.prune_bawp_by_gop_usage && !frame_is_intra_only(cm)) {
+    const unsigned int level = cm->current_frame.pyramid_level;
+    if (level == 1 || level == 2) {
+      cpi->bawp_gop_usage[level - 1] =
+          rdc->bawp_eligible_area > 0
+              ? (double)rdc->bawp_used_area / (double)rdc->bawp_eligible_area
+              : 0.0;
+    }
+  }
+
   // If intrabc is allowed but never selected, reset the allow_intrabc flag.
   if (features->allow_intrabc && !cpi->intrabc_used) {
     features->allow_intrabc = 0;

av2/encoder/encoder.c

@@ -1538,6 +1538,9 @@ AV2_COMP *av2_create_compressor(AV2EncoderConfig *oxcf, BufferPool *const pool,
   if (!cm) return NULL;
 
   av2_zero(*cpi);
+  // Unknown until the first level-1 or level-2 frame is coded
+  // (sf->inter_sf.prune_bawp_by_gop_usage).
+  cpi->bawp_gop_usage[0] = cpi->bawp_gop_usage[1] = -1.0;
 
   // The jmp_buf is valid only for the duration of the function that calls
   // setjmp(). Therefore, this function must reset the 'setjmp' field to 0

av2/encoder/encoder.h

@@ -1705,6 +1705,10 @@ typedef struct RD_COUNTS {
   int skip_mode_used_flag;
   int tx_type_used[TX_SIZES_ALL][PRIMARY_TX_TYPES];
   int warped_used[2];
+  // Luma area (in 4x4 units) of the final coded blocks that may use BAWP, and
+  // of those that use it. Used by sf->inter_sf.prune_bawp_by_gop_usage.
+  int64_t bawp_eligible_area;
+  int64_t bawp_used_area;
 } RD_COUNTS;
 
 typedef struct ThreadData {
@@ -2513,6 +2517,13 @@ typedef struct AV2_COMP {
    */
   YV12_BUFFER_CONFIG *unfiltered_source;
 
+  /*!
+   * BAWP share of the BAWP-eligible area on the current GOP's level-1 and
+   * level-2 frames, or -1 while unknown. Used by
+   * sf->inter_sf.prune_bawp_by_gop_usage.
+   */
+  double bawp_gop_usage[2];
+
   /*!
    * Parameters related to tpl.
    */

av2/encoder/ethread.c

@@ -40,6 +40,8 @@ static AVM_INLINE void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
   for (int i = 0; i < 2; i++) {
     td->rd_counts.warped_used[i] += td_t->rd_counts.warped_used[i];
   }
+  td->rd_counts.bawp_eligible_area += td_t->rd_counts.bawp_eligible_area;
+  td->rd_counts.bawp_used_area += td_t->rd_counts.bawp_used_area;
 }
 
 void av2_row_mt_sync_read_dummy(AV2EncRowMultiThreadSync *row_mt_sync, int r,

av2/encoder/partition_search.c

@@ -1813,6 +1813,15 @@ static void encode_b(const AV2_COMP *const cpi, TileDataEnc *tile_data,
     }
 
     if (tile_data->allow_update_cdf) update_stats(&cpi->common, td);
+    if (cpi->sf.inter_sf.prune_bawp_by_gop_usage && !frame_is_intra_only(cm) &&
+        xd->tree_type != CHROMA_PART && is_inter_block(mbmi, xd->tree_type) &&
+        !is_intrabc_block(mbmi, xd->tree_type) && cm->features.enable_bawp &&
+        av2_allow_bawp(cm, mbmi, xd->mi_row, xd->mi_col)) {
+      const BLOCK_SIZE bs = mbmi->sb_type[PLANE_TYPE_Y];
+      const int64_t area = (block_size_wide[bs] * block_size_high[bs]) >> 4;
+      td->rd_counts.bawp_eligible_area += area;
+      if (mbmi->bawp_flag[0]) td->rd_counts.bawp_used_area += area;
+    }
 
     // Gather warped motion count to update the probability.
     if (cpi->sf.inter_sf.prune_warped_prob_thresh > 0 ||

av2/encoder/speed_features.c

@@ -623,6 +623,7 @@ static void set_good_speed_features_framesize_independent(
 
     sf->inter_sf.alt_ref_search_fp = 1;
     sf->inter_sf.txfm_rd_gate_level = boosted ? 0 : 4;
+    sf->inter_sf.prune_bawp_by_gop_usage = 1;
     sf->inter_sf.fast_warp_delta_decoupled_search = 1;
 
     sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 3;
@@ -996,6 +997,7 @@ static AVM_INLINE void init_inter_sf(INTER_MODE_SPEED_FEATURES *inter_sf) {
   inter_sf->prune_comp_mode_eval_using_est_rd = false;
   inter_sf->prune_warp_newmv_ref_mv_idx = false;
   inter_sf->enable_fast_bawp = 0;
+  inter_sf->prune_bawp_by_gop_usage = 0;
   inter_sf->limit_max_ref_mv_idx = false;
 }
 

av2/encoder/speed_features.h

@@ -904,6 +904,13 @@ typedef struct INTER_MODE_SPEED_FEATURES {
   // 1: enable histogram based frame level on/off
   int enable_fast_bawp;
 
+  // GOP-level BAWP pruning: turn the frame-level BAWP flag off on frames at
+  // pyramid level >= 3 when, on the GOP's level-1 and level-2 frames, BAWP
+  // covered on average at most half of the BAWP-eligible area.
+  // 0: off
+  // 1: on
+  int prune_bawp_by_gop_usage;
+
   // Limit the maximum number of reference MVs searched for single-DRL compound
   // prediction modes. The cap is applicable for larger blocks and is not
   // applied on key/golden/arf frames.