Make the row based multi-threaded encoder deterministic

This patch followed allow_exhaustive_searches feature modification and
continued to modify the encoder to achieve the determinism in the row
based multi-threaded encoding. While row-mt = 1 and using multiple
threads, the adaptive feature in encoder was disabled, which gave
BDRate gain(at speed 1, -0.6% ~ -0.7%; at speed 2, -0.46% ~ -0.59%),
but some encoder speed losses(7% ~ 10% at speed 1 and 3% ~ 6% at
speed 2). These speed losses were acceptable considering the speed
gains obtained from row-mt.

Change-Id: I60d87a25346ebc487a864b57d559f560b7e398bb

11 files changed, 27 insertions, 131 deletions

diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 12dc226da..e2adf731f 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -4341,7 +4341,6 @@ void vp9_init_tile_data(VP9_COMP *cpi) {
           }
         }
 #if CONFIG_MULTITHREAD
-        tile_data->enc_row_mt_mutex = NULL;
         tile_data->row_base_thresh_freq_fact = NULL;
 #endif
       }
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index d82b706be..c26e5cf47 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -5281,4 +5281,9 @@ void vp9_set_row_mt(VP9_COMP *cpi) {
   if (cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5 && cpi->oxcf.row_mt) {
     cpi->row_mt = 1;
   }
+
+  if (cpi->row_mt && cpi->oxcf.max_threads > 1)
+    cpi->row_mt_bit_exact = 1;
+  else
+    cpi->row_mt_bit_exact = 0;
 }
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index 94ebc3b6f..d31b67138 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -268,7 +268,6 @@ typedef struct VP9EncoderConfig {
   VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
 
   int row_mt;
-  unsigned int row_mt_bit_exact;
   unsigned int motion_vector_unit_test;
 } VP9EncoderConfig;
 
@@ -286,9 +285,6 @@ typedef struct TileDataEnc {
 
   // Used for adaptive_rd_thresh with row multithreading
   int *row_base_thresh_freq_fact;
-#if CONFIG_MULTITHREAD
-  pthread_mutex_t *enc_row_mt_mutex;
-#endif
 } TileDataEnc;
 
 typedef struct RowMTInfo {
@@ -692,7 +688,9 @@ typedef struct VP9_COMP {
   void (*row_mt_sync_read_ptr)(VP9RowMTSync *const, int, int);
   void (*row_mt_sync_write_ptr)(VP9RowMTSync *const, int, int, const int);
   ARNRFilterData arnr_filter_data;
+
   int row_mt;
+  unsigned int row_mt_bit_exact;
 
   // Previous Partition Info
   BLOCK_SIZE *prev_partition;
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index e7639a7c5..84a7ce6e1 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -979,12 +979,12 @@ void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
     if (log_intra < 10.0) {
       mb_intra_factor = 1.0 + ((10.0 - log_intra) * 0.05);
       fp_acc_data->intra_factor += mb_intra_factor;
-      if (cpi->oxcf.row_mt_bit_exact)
+      if (cpi->row_mt_bit_exact)
         cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor =
             mb_intra_factor;
     } else {
       fp_acc_data->intra_factor += 1.0;
-      if (cpi->oxcf.row_mt_bit_exact)
+      if (cpi->row_mt_bit_exact)
         cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor = 1.0;
     }
 
@@ -999,12 +999,12 @@ void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
     if ((level_sample < DARK_THRESH) && (log_intra < 9.0)) {
       mb_brightness_factor = 1.0 + (0.01 * (DARK_THRESH - level_sample));
       fp_acc_data->brightness_factor += mb_brightness_factor;
-      if (cpi->oxcf.row_mt_bit_exact)
+      if (cpi->row_mt_bit_exact)
         cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
             mb_brightness_factor;
     } else {
       fp_acc_data->brightness_factor += 1.0;
-      if (cpi->oxcf.row_mt_bit_exact)
+      if (cpi->row_mt_bit_exact)
         cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
             1.0;
     }
@@ -1166,7 +1166,7 @@ void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
         if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
             (this_error < (2 * intrapenalty))) {
           fp_acc_data->neutral_count += 1.0;
-          if (cpi->oxcf.row_mt_bit_exact)
+          if (cpi->row_mt_bit_exact)
             cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
                 1.0;
           // Also track cases where the intra is not much worse than the inter
@@ -1176,7 +1176,7 @@ void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
           mb_neutral_count =
               (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_error);
           fp_acc_data->neutral_count += mb_neutral_count;
-          if (cpi->oxcf.row_mt_bit_exact)
+          if (cpi->row_mt_bit_exact)
             cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
                 mb_neutral_count;
         }
@@ -1424,7 +1424,7 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
 
   cm->log2_tile_rows = 0;
 
-  if (cpi->oxcf.row_mt_bit_exact && cpi->twopass.fp_mb_float_stats == NULL)
+  if (cpi->row_mt_bit_exact && cpi->twopass.fp_mb_float_stats == NULL)
     CHECK_MEM_ERROR(
         cm, cpi->twopass.fp_mb_float_stats,
         vpx_calloc(cm->MBs * sizeof(*cpi->twopass.fp_mb_float_stats), 1));
@@ -1441,13 +1441,13 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
     } else {
       cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
       cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
-      if (cpi->oxcf.row_mt_bit_exact) {
+      if (cpi->row_mt_bit_exact) {
         cm->log2_tile_cols = 0;
         vp9_zero_array(cpi->twopass.fp_mb_float_stats, cm->MBs);
       }
       vp9_encode_fp_row_mt(cpi);
       first_tile_col = &cpi->tile_data[0];
-      if (cpi->oxcf.row_mt_bit_exact)
+      if (cpi->row_mt_bit_exact)
         accumulate_floating_point_stats(cpi, first_tile_col);
       first_pass_stat_calc(cpi, &fps, &(first_tile_col->fp_data));
     }
diff --git a/vp9/encoder/vp9_multi_thread.c b/vp9/encoder/vp9_multi_thread.c
index 0ad5dcc48..da06fb151 100644
--- a/vp9/encoder/vp9_multi_thread.c
+++ b/vp9/encoder/vp9_multi_thread.c
@@ -110,19 +110,6 @@ void vp9_row_mt_mem_alloc(VP9_COMP *cpi) {
     multi_thread_ctxt->num_tile_vert_sbs[tile_row] =
         get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
   }
-
-#if CONFIG_MULTITHREAD
-  for (tile_row = 0; tile_row < tile_rows; tile_row++) {
-    for (tile_col = 0; tile_col < tile_cols; tile_col++) {
-      TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
-
-      CHECK_MEM_ERROR(cm, this_tile->enc_row_mt_mutex,
-                      vpx_malloc(sizeof(*this_tile->enc_row_mt_mutex)));
-
-      pthread_mutex_init(this_tile->enc_row_mt_mutex, NULL);
-    }
-  }
-#endif
 }
 
 void vp9_row_mt_mem_dealloc(VP9_COMP *cpi) {
@@ -165,9 +152,6 @@ void vp9_row_mt_mem_dealloc(VP9_COMP *cpi) {
           this_tile->row_base_thresh_freq_fact = NULL;
         }
       }
-      pthread_mutex_destroy(this_tile->enc_row_mt_mutex);
-      vpx_free(this_tile->enc_row_mt_mutex);
-      this_tile->enc_row_mt_mutex = NULL;
     }
   }
 #endif
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 9c6138f6d..501cd32cf 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -1720,15 +1720,8 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
          rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
                                     &rd_thresh_freq_fact[mode_index])) ||
         (!cpi->sf.adaptive_rd_thresh_row_mt &&
-         rd_less_than_thresh(
-             best_rdc.rdcost, mode_rd_thresh,
-#if CONFIG_MULTITHREAD
-             // Synchronization of this function
-             // is only necessary when
-             // adaptive_rd_thresh is > 0.
-             cpi->sf.adaptive_rd_thresh ? tile_data->enc_row_mt_mutex : NULL,
-#endif
-             &rd_thresh_freq_fact[mode_index])))
+         rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
+                             &rd_thresh_freq_fact[mode_index])))
       continue;
 
     if (this_mode == NEWMV) {
@@ -2098,15 +2091,8 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
            rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
                                       &rd_thresh_freq_fact[mode_index])) ||
           (!cpi->sf.adaptive_rd_thresh_row_mt &&
-           rd_less_than_thresh(
-               best_rdc.rdcost, mode_rd_thresh,
-#if CONFIG_MULTITHREAD
-               // Synchronization of this function
-               // is only necessary when
-               // adaptive_rd_thresh is > 0.
-               cpi->sf.adaptive_rd_thresh ? tile_data->enc_row_mt_mutex : NULL,
-#endif
-               &rd_thresh_freq_fact[mode_index])))
+           rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
+                               &rd_thresh_freq_fact[mode_index])))
         continue;
 
       mi->mode = this_mode;
diff --git a/vp9/encoder/vp9_rd.c b/vp9/encoder/vp9_rd.c
index 3c49fe665..39a7742f0 100644
--- a/vp9/encoder/vp9_rd.c
+++ b/vp9/encoder/vp9_rd.c
@@ -650,15 +650,7 @@ void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
 }
 
 void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
-                               int bsize,
-#if CONFIG_MULTITHREAD
-                               pthread_mutex_t *enc_row_mt_mutex,
-#endif
-                               int best_mode_index) {
-#if CONFIG_MULTITHREAD
-  if (NULL != enc_row_mt_mutex) pthread_mutex_lock(enc_row_mt_mutex);
-#endif
-
+                               int bsize, int best_mode_index) {
   if (rd_thresh > 0) {
     const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
     int mode;
@@ -676,10 +668,6 @@ void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
       }
     }
   }
-
-#if CONFIG_MULTITHREAD
-  if (NULL != enc_row_mt_mutex) pthread_mutex_unlock(enc_row_mt_mutex);
-#endif
 }
 
 int vp9_get_intra_cost_penalty(int qindex, int qdelta,
diff --git a/vp9/encoder/vp9_rd.h b/vp9/encoder/vp9_rd.h
index aae47dcdd..1e1176866 100644
--- a/vp9/encoder/vp9_rd.h
+++ b/vp9/encoder/vp9_rd.h
@@ -170,32 +170,11 @@ void vp9_set_rd_speed_thresholds(struct VP9_COMP *cpi);
 void vp9_set_rd_speed_thresholds_sub8x8(struct VP9_COMP *cpi);
 
 void vp9_update_rd_thresh_fact(int (*fact)[MAX_MODES], int rd_thresh, int bsize,
-#if CONFIG_MULTITHREAD
-                               pthread_mutex_t *enc_row_mt_mutex,
-#endif
                                int best_mode_index);
 
 static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
-#if CONFIG_MULTITHREAD
-                                      pthread_mutex_t *enc_row_mt_mutex,
-#endif
                                       const int *const thresh_fact) {
-  int is_rd_less_than_thresh;
-
-#if CONFIG_MULTITHREAD
-  // Synchronize to ensure data coherency as thresh_freq_fact is maintained at
-  // tile level and not thread-safe with row based multi-threading
-  if (NULL != enc_row_mt_mutex) pthread_mutex_lock(enc_row_mt_mutex);
-#endif
-
-  is_rd_less_than_thresh =
-      best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
-
-#if CONFIG_MULTITHREAD
-  if (NULL != enc_row_mt_mutex) pthread_mutex_unlock(enc_row_mt_mutex);
-#endif
-
-  return is_rd_less_than_thresh;
+  return best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
 }
 
 static INLINE void set_error_per_bit(MACROBLOCK *x, int rdmult) {
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 2278ddc0f..cf46905c9 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -3161,11 +3161,6 @@ void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
 
   for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
 
-#if CONFIG_MULTITHREAD
-  if (NULL != tile_data->enc_row_mt_mutex)
-    pthread_mutex_lock(tile_data->enc_row_mt_mutex);
-#endif
-
   for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
     mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
 
@@ -3187,11 +3182,6 @@ void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
 
   memcpy(mode_map, tile_mode_map, sizeof(mode_map));
 
-#if CONFIG_MULTITHREAD
-  if (NULL != tile_data->enc_row_mt_mutex)
-    pthread_mutex_unlock(tile_data->enc_row_mt_mutex);
-#endif
-
   for (midx = 0; midx < MAX_MODES; ++midx) {
     int mode_index = mode_map[midx];
     int mode_excluded = 0;
@@ -3628,11 +3618,7 @@ void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
 
   if (!cpi->rc.is_src_frame_alt_ref)
     vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
-                              sf->adaptive_rd_thresh, bsize,
-#if CONFIG_MULTITHREAD
-                              tile_data->enc_row_mt_mutex,
-#endif
-                              best_mode_index);
+                              sf->adaptive_rd_thresh, bsize, best_mode_index);
 
   // macroblock modes
   *mi = best_mbmode;
@@ -3772,11 +3758,7 @@ void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, TileDataEnc *tile_data,
          (cm->interp_filter == mi->interp_filter));
 
   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
-                            cpi->sf.adaptive_rd_thresh, bsize,
-#if CONFIG_MULTITHREAD
-                            tile_data->enc_row_mt_mutex,
-#endif
-                            THR_ZEROMV);
+                            cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
 
   vp9_zero(best_pred_diff);
   vp9_zero(best_filter_diff);
@@ -3922,9 +3904,6 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
     if (!internal_active_edge &&
         rd_less_than_thresh(best_rd,
                             rd_opt->threshes[segment_id][bsize][ref_index],
-#if CONFIG_MULTITHREAD
-                            tile_data->enc_row_mt_mutex,
-#endif
                             &rd_thresh_freq_fact[ref_index]))
       continue;
 
@@ -4374,11 +4353,7 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
          !is_inter_block(&best_mbmode));
 
   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact, sf->adaptive_rd_thresh,
-                            bsize,
-#if CONFIG_MULTITHREAD
-                            tile_data->enc_row_mt_mutex,
-#endif
-                            best_ref_index);
+                            bsize, best_ref_index);
 
   // macroblock modes
   *mi = best_mbmode;
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index fbf13969e..0c200f9be 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -639,10 +639,7 @@ void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi) {
   // With row based multi-threading, the following speed features
   // have to be disabled to guarantee that bitstreams encoded with single thread
   // and multiple threads match
-  if (cpi->oxcf.row_mt_bit_exact) {
-    sf->adaptive_rd_thresh = 0;
-    sf->adaptive_pred_interp_filter = 0;
-  }
+  if (cpi->row_mt_bit_exact) sf->adaptive_rd_thresh = 0;
 
   // This is only used in motion vector unit test.
   if (cpi->oxcf.motion_vector_unit_test == 1)
@@ -796,10 +793,7 @@ void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
   // With row based multi-threading, the following speed features
   // have to be disabled to guarantee that bitstreams encoded with single thread
   // and multiple threads match
-  if (cpi->oxcf.row_mt_bit_exact) {
-    sf->adaptive_rd_thresh = 0;
-    sf->adaptive_pred_interp_filter = 0;
-  }
+  if (cpi->row_mt_bit_exact) sf->adaptive_rd_thresh = 0;
 
   // This is only used in motion vector unit test.
   if (cpi->oxcf.motion_vector_unit_test == 1)
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index a335a4ab5..25fc80a9a 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -52,7 +52,6 @@ struct vp9_extracfg {
   int render_width;
   int render_height;
   unsigned int row_mt;
-  unsigned int row_mt_bit_exact;
   unsigned int motion_vector_unit_test;
 };
 
@@ -86,7 +85,6 @@ static struct vp9_extracfg default_extra_cfg = {
   0,                     // render width
   0,                     // render height
   0,                     // row_mt
-  0,                     // row_mt_bit_exact
   0,                     // motion_vector_unit_test
 };
 
@@ -252,7 +250,6 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
         "or kf_max_dist instead.");
 
   RANGE_CHECK(extra_cfg, row_mt, 0, 1);
-  RANGE_CHECK(extra_cfg, row_mt_bit_exact, 0, 1);
   RANGE_CHECK(extra_cfg, motion_vector_unit_test, 0, 2);
   RANGE_CHECK(extra_cfg, enable_auto_alt_ref, 0, 2);
   RANGE_CHECK(extra_cfg, cpu_used, -8, 8);
@@ -564,7 +561,6 @@ static vpx_codec_err_t set_encoder_config(
   oxcf->target_level = extra_cfg->target_level;
 
   oxcf->row_mt = extra_cfg->row_mt;
-  oxcf->row_mt_bit_exact = extra_cfg->row_mt_bit_exact;
   oxcf->motion_vector_unit_test = extra_cfg->motion_vector_unit_test;
 
   for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
@@ -862,13 +858,6 @@ static vpx_codec_err_t ctrl_set_row_mt(vpx_codec_alg_priv_t *ctx,
   return update_extra_cfg(ctx, &extra_cfg);
 }
 
-static vpx_codec_err_t ctrl_enable_row_mt_bit_exact(vpx_codec_alg_priv_t *ctx,
-                                                    va_list args) {
-  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
-  extra_cfg.row_mt_bit_exact = CAST(VP9E_ENABLE_ROW_MT_BIT_EXACT, args);
-  return update_extra_cfg(ctx, &extra_cfg);
-}
-
 static vpx_codec_err_t ctrl_enable_motion_vector_unit_test(
     vpx_codec_alg_priv_t *ctx, va_list args) {
   struct vp9_extracfg extra_cfg = ctx->extra_cfg;
@@ -1633,7 +1622,6 @@ static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
   { VP9E_SET_RENDER_SIZE, ctrl_set_render_size },
   { VP9E_SET_TARGET_LEVEL, ctrl_set_target_level },
   { VP9E_SET_ROW_MT, ctrl_set_row_mt },
-  { VP9E_ENABLE_ROW_MT_BIT_EXACT, ctrl_enable_row_mt_bit_exact },
   { VP9E_ENABLE_MOTION_VECTOR_UNIT_TEST, ctrl_enable_motion_vector_unit_test },
 
   // Getters