Merge branch 'me_early_terminate'

Conflicts:
	configure.ac
	src/cfg.c
	src/cli.c
	src/kvazaar.h
	src/search_inter.c

configure.ac

@@ -23,7 +23,7 @@ AC_CONFIG_SRCDIR([src/encmain.c])
 #
 # Here is a somewhat sane guide to lib versioning: http://apr.apache.org/versioning.html
 ver_major=3
-ver_minor=7
+ver_minor=8
 ver_release=0
 
 # Prevents configure from adding a lot of defines to the CFLAGS

src/cfg.c

@@ -109,6 +109,8 @@ int kvz_config_init(kvz_config *cfg)
   cfg->mv_constraint = KVZ_MV_CONSTRAIN_NONE;
   cfg->crypto_features = KVZ_CRYPTO_OFF;
 
+  cfg->me_early_termination = 0;
+
   return 1;
 }
 
@@ -298,6 +300,8 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
   static const char * const crypto_toggle_names[] = { "off", "on", NULL };
   static const char * const crypto_feature_names[] = { "mvs", "mv_signs", "trans_coeffs", "trans_coeff_signs", NULL };
 
+  static const char * const me_early_termination_names[] = { "off", "on", "sensitive", NULL };
+
   static const char * const preset_values[11][32] = {
       { 
         "ultrafast", 
@@ -897,6 +901,12 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
 
     return 1;
   }
+  else if OPT("me-early-termination"){
+    int8_t mode = 0;
+    int result = parse_enum(value, me_early_termination_names, &mode);
+    cfg->me_early_termination = mode;
+    return result;
+  }
   else
     return 0;
 #undef OPT

src/cli.c

@@ -107,6 +107,7 @@ static const struct option long_options[] = {
   { "hash",               required_argument, NULL, 0 },
   {"cu-split-termination",required_argument, NULL, 0 },
   { "crypto",             required_argument, NULL, 0 },
+  { "me-early-termination",required_argument, NULL, 0 },
   {0, 0, 0, 0}
 };
 
@@ -380,6 +381,10 @@ void print_help(void)
     "                                     \"zero\": Terminate when splitting gives little\n"
     "                                               improvement.\n"
     "                                     \"off\": Don't terminate splitting early\n"
+    "          --me-early-termitanion : Specify the me early termination behaviour\n"
+    "                                     \"off\": Early termination is off\n"
+    "                                     \"on\": Early termination is on\n"
+    "                                     \"sensitive\": Sensitive early termination is on\n"
     "\n"
     "  Video Usability Information:\n"
     "          --sar <width:height>   : Specify Sample Aspect Ratio\n"

src/kvazaar.h

@@ -131,6 +131,7 @@ enum kvz_hash
 
 /**
 * \brief cu split termination mode
+* \since since 3.8.0
 */
 enum kvz_cu_split_termination
 {
@@ -151,6 +152,17 @@ enum kvz_crypto_features {
   KVZ_CRYPTO_ON = (1 << 4) - 1,
 };
 
+/**
+* \brief me early termination mode
+* \since since 3.8.0
+*/
+enum kvz_me_early_termination
+{
+  KVZ_ME_EARLY_TERMINATION_OFF = 0,
+  KVZ_ME_EARLY_TERMINATION_ON = 1,
+  KVZ_ME_EARLY_TERMINATION_SENSITIVE = 2
+};
+
 /**
  * \brief GoP picture configuration.
  */
@@ -258,9 +270,12 @@ typedef struct kvz_config
   enum kvz_mv_constraint mv_constraint;  /*!< \since 3.3.0 \brief Constrain movement vectors. */
   enum kvz_hash hash;  /*!< \since 3.5.0 \brief What hash algorithm to use. */
 
-  enum kvz_cu_split_termination cu_split_termination; /*!< \brief Mode of cu split termination. */
+  enum kvz_cu_split_termination cu_split_termination; /*!< \since 3.8.0 \brief Mode of cu split termination. */
 
   enum kvz_crypto_features crypto_features; /*!< \since 3.7.0 */
+
+  enum kvz_me_early_termination me_early_termination; /*!< \since 3.8.0 \brief Mode of me early termination. */
+
 } kvz_config;
 
 /**

src/search_inter.c

@@ -257,6 +257,61 @@ static int calc_mvd_cost(encoder_state_t * const state, int x, int y, int mv_shi
 }
 
 
+static bool early_terminate(int16_t num_cand, inter_merge_cand_t *merge_cand, vector2d_t *mv_in_out, vector2d_t *mv, const encoder_state_t *const state,
+  const vector2d_t *orig, unsigned width, unsigned height, int wpp_limit, const kvz_picture *pic, const kvz_picture *ref,
+  int16_t mv_cand[2][2], int32_t ref_idx, unsigned *best_cost, uint32_t *bitcost_out, uint32_t *best_bitcost,
+  int(*calc_mvd)(const encoder_state_t * const, int, int, int, int16_t[2][2], inter_merge_cand_t[MRG_MAX_NUM_CANDS],
+  int16_t, int32_t, uint32_t *))
+{
+  static const vector2d_t small_hexbs[5] = {
+      { 0, 0 },
+      { 0, -1 }, { -1, 0 }, { 1, 0 }, { 0, 1 },
+  };
+  double multiplier = 1;
+  // If early termination is set to fast set multiplier to 0.9
+  if (state->encoder_control->cfg->me_early_termination == KVZ_ME_EARLY_TERMINATION_SENSITIVE){
+    multiplier = 0.95;
+  }
+  const vector2d_t *offset;
+  for (int k = 0; k < 2; ++k){
+    unsigned best_index = 0;
+    for (int i = 1; i < 5; ++i) {
+      offset = &small_hexbs[i];
+      if (!intmv_within_tile(state, orig, mv->x + offset->x, mv->y + offset->y, width, height, wpp_limit)) {
+        continue;
+      }
+
+      unsigned cost = kvz_image_calc_sad(pic, ref, orig->x, orig->y,
+        (state->tile->lcu_offset_x * LCU_WIDTH) + orig->x + mv->x + offset->x,
+        (state->tile->lcu_offset_y * LCU_WIDTH) + orig->y + mv->y + offset->y,
+        width, height, -1);
+      unsigned bitcost;
+      cost += calc_mvd(state, mv->x + offset->x, mv->y + offset->y, 2, mv_cand, merge_cand, num_cand, ref_idx, &bitcost);
+
+      if (cost < multiplier * *best_cost ) {
+        *best_cost = cost;
+        best_index = i;
+        *best_bitcost = bitcost;
+      }
+    }
+    // Adjust the movement vector
+    mv->x += small_hexbs[best_index].x;
+    mv->y += small_hexbs[best_index].y;
+
+    // if best match is at center we stop the search
+    if (best_index == 0){
+      // Return final movement vector in quarter-pixel precision.
+      mv_in_out->x = mv->x << 2;
+      mv_in_out->y = mv->y << 2;
+
+      *bitcost_out = *best_bitcost;
+      return true;
+    }
+  }
+  return false;
+}
+
+
 unsigned kvz_tz_pattern_search(encoder_state_t * const state, const kvz_picture *pic, const kvz_picture *ref, unsigned pattern_type,
                            const vector2d_t *orig, const int iDist, vector2d_t *mv, unsigned best_cost, int *best_dist,
                            int16_t mv_cand[2][2], inter_merge_cand_t merge_cand[MRG_MAX_NUM_CANDS], int16_t num_cand, int32_t ref_idx, uint32_t *best_bitcost,
@@ -536,6 +591,12 @@ static unsigned tz_search(encoder_state_t * const state,
   best_cost = select_starting_point(num_cand, merge_cand, mv_in_out, &mv, state, orig, width, height, wpp_limit,
                                 pic, ref, mv_cand, ref_idx, best_cost, &best_index, &best_bitcost, calc_mvd);
 
+  // Check if we should stop search
+  if (state->encoder_control->cfg->me_early_termination){
+    if (early_terminate(num_cand, merge_cand, mv_in_out, &mv, state, orig, width, height, wpp_limit,
+      pic, ref, mv_cand, ref_idx, &best_cost, bitcost_out, &best_bitcost, calc_mvd)) return best_cost;
+  }
+
   //step 2, grid search
   for (iDist = 1; iDist <= iSearchRange; iDist *= 2)
   {
@@ -685,6 +746,12 @@ static unsigned hexagon_search(encoder_state_t * const state,
   best_cost = select_starting_point(num_cand, merge_cand, mv_in_out, &mv, state, orig, width, height, wpp_limit,
                                 pic, ref, mv_cand, ref_idx, best_cost, &best_index, &best_bitcost, calc_mvd);
 
+  // Check if we should stop search
+  if (state->encoder_control->cfg->me_early_termination){
+    if (early_terminate(num_cand, merge_cand, mv_in_out, &mv, state, orig, width, height, wpp_limit,
+      pic, ref, mv_cand, ref_idx, &best_cost, bitcost_out, &best_bitcost, calc_mvd)) return best_cost;
+  }
+
   // Search the initial 7 points of the hexagon.
   best_index = 0;
   for (i = 0; i < 7; ++i) {