Adaptive mode search scheduling

This commit enables an adaptive mode search order scheduling scheme
in the rate-distortion optimization. It changes the compression
performance by -0.433% and -0.420% for derf and stdhd respectively.
It provides speed improvement for speed 3:

bus CIF 1000 kbps
24590 b/f, 35.513 dB, 7864 ms ->
24696 b/f, 35.491 dB, 7408 ms (6% speed-up)

stockholm 720p 1000 kbps
8983 b/f, 35.078 dB, 65698 ms ->
8962 b/f, 35.054 dB, 60298 ms (8%)

old_town_cross 720p 1000 kbps
11804 b/f, 35.666 dB, 62492 ms ->
11778 b/f, 35.609 dB, 56040 ms (10%)

blue_sky 1080p 1500 kbps
57173 b/f, 36.179 dB, 77879 ms ->
57199 b/f, 36.131 dB, 69821 ms (10%)

pedestrian_area 1080p 2000 kbps
74241 b/f, 41.105 dB, 144031 ms ->
74271 b/f, 41.091 dB, 133614 ms (8%)

Change-Id: Iaad28cbc99399030fc5f9951eb5aa7fa633f320e

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 7b7a6e9..107bed5 100644 (file)
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -991,8 +991,10 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
 
   // Default rd threshold factors for mode selection
   for (i = 0; i < BLOCK_SIZES; ++i) {
-    for (j = 0; j < MAX_MODES; ++j)
+    for (j = 0; j < MAX_MODES; ++j) {
       cpi->rd.thresh_freq_fact[i][j] = 32;
+      cpi->rd.mode_map[i][j] = j;
+    }
   }
 
 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\

diff --git a/vp9/encoder/vp9_rd.h b/vp9/encoder/vp9_rd.h
index eeb5e0f..5dcb2f8 100644 (file)
--- a/vp9/encoder/vp9_rd.h
+++ b/vp9/encoder/vp9_rd.h
@@ -51,6 +51,12 @@ typedef enum {
 
   THR_NEARMV,
   THR_NEARA,
+  THR_NEARG,
+
+  THR_ZEROMV,
+  THR_ZEROG,
+  THR_ZEROA,
+
   THR_COMP_NEARESTLA,
   THR_COMP_NEARESTGA,
 
@@ -58,13 +64,9 @@ typedef enum {
 
   THR_COMP_NEARLA,
   THR_COMP_NEWLA,
-  THR_NEARG,
   THR_COMP_NEARGA,
   THR_COMP_NEWGA,
 
-  THR_ZEROMV,
-  THR_ZEROG,
-  THR_ZEROA,
   THR_COMP_ZEROLA,
   THR_COMP_ZEROGA,
 
@@ -98,6 +100,8 @@ typedef struct RD_OPT {
   int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
   int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
 
+  int mode_map[BLOCK_SIZES][MAX_MODES];
+
   int64_t comp_pred_diff[REFERENCE_MODES];
   int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
   int64_t tx_select_diff[TX_MODES];

diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 37832b7..81f6878 100644 (file)
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -91,6 +91,12 @@ static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
 
   {NEARMV,    {LAST_FRAME,   NONE}},
   {NEARMV,    {ALTREF_FRAME, NONE}},
+  {NEARMV,    {GOLDEN_FRAME, NONE}},
+
+  {ZEROMV,    {LAST_FRAME,   NONE}},
+  {ZEROMV,    {GOLDEN_FRAME, NONE}},
+  {ZEROMV,    {ALTREF_FRAME, NONE}},
+
   {NEARESTMV, {LAST_FRAME,   ALTREF_FRAME}},
   {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
 
@@ -98,13 +104,9 @@ static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
 
   {NEARMV,    {LAST_FRAME,   ALTREF_FRAME}},
   {NEWMV,     {LAST_FRAME,   ALTREF_FRAME}},
-  {NEARMV,    {GOLDEN_FRAME, NONE}},
   {NEARMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
   {NEWMV,     {GOLDEN_FRAME, ALTREF_FRAME}},
 
-  {ZEROMV,    {LAST_FRAME,   NONE}},
-  {ZEROMV,    {GOLDEN_FRAME, NONE}},
-  {ZEROMV,    {ALTREF_FRAME, NONE}},
   {ZEROMV,    {LAST_FRAME,   ALTREF_FRAME}},
   {ZEROMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
 
@@ -2572,7 +2574,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
   MB_MODE_INFO best_mbmode;
   int best_mode_skippable = 0;
-  int mode_index, best_mode_index = -1;
+  int midx, best_mode_index = -1;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
   int64_t best_intra_rd = INT64_MAX;
@@ -2590,8 +2592,11 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   int mode_skip_start = cpi->sf.mode_skip_start + 1;
   const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
   const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize];
+  int mode_threshold[MAX_MODES];
+  int *mode_map = rd_opt->mode_map[bsize];
   const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags;
   vp9_zero(best_mbmode);
+
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
@@ -2686,7 +2691,25 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   mode_skip_mask[INTRA_FRAME] |=
       ~(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]]);
 
-  for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
+  for (i = 0; i < MAX_MODES; ++i)
+    mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
+
+  midx =  cpi->sf.schedule_mode_search ? mode_skip_start : 0;
+  while (midx > 4) {
+    uint8_t end_pos = 0;
+    for (i = 5; i < midx; ++i) {
+      if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
+        uint8_t tmp = mode_map[i];
+        mode_map[i] = mode_map[i - 1];
+        mode_map[i - 1] = tmp;
+        end_pos = i;
+      }
+    }
+    midx = end_pos;
+  }
+
+  for (midx = 0; midx < MAX_MODES; ++midx) {
+    int mode_index = mode_map[midx];
     int mode_excluded = 0;
     int64_t this_rd = INT64_MAX;
     int disable_skip = 0;
@@ -2706,7 +2729,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
 
     // Look at the reference frame of the best mode so far and set the
     // skip mask to look at a subset of the remaining modes.
-    if (mode_index == mode_skip_start && best_mode_index >= 0) {
+    if (midx == mode_skip_start && best_mode_index >= 0) {
       switch (best_mbmode.ref_frame[0]) {
         case INTRA_FRAME:
           break;
@@ -2736,8 +2759,10 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
       continue;
 
     // Test best rd so far against threshold for trying this mode.
-    if (rd_less_than_thresh(best_rd, rd_threshes[mode_index],
-                            rd_thresh_freq_fact[mode_index]))
+    if (best_mode_skippable && cpi->sf.schedule_mode_search)
+      mode_threshold[mode_index] <<= 1;
+
+    if (best_rd < mode_threshold[mode_index])
       continue;
 
     if (cpi->sf.motion_field_mode_search) {
@@ -3129,7 +3154,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
          (cm->interp_filter == best_mbmode.interp_filter) ||
          !is_inter_block(&best_mbmode));
 
-  update_rd_thresh_fact(cpi, bsize, best_mode_index);
+  if (!cpi->rc.is_src_frame_alt_ref)
+    update_rd_thresh_fact(cpi, bsize, best_mode_index);
 
   // macroblock modes
   *mbmi = best_mbmode;

diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 7ddeac7..52e9a8e 100644 (file)
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -95,9 +95,11 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
                                                         : USE_LARGESTALL;
     if (MIN(cm->width, cm->height) >= 720) {
       sf->disable_split_mask = DISABLE_ALL_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
     } else {
       sf->max_intra_bsize = BLOCK_32X32;
       sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
     }
     sf->adaptive_pred_interp_filter = 0;
     sf->adaptive_mode_search = 1;
@@ -376,6 +378,7 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->use_fast_coef_updates = TWO_LOOP;
   sf->use_fast_coef_costing = 0;
   sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
+  sf->schedule_mode_search = 0;
   sf->use_nonrd_pick_mode = 0;
   for (i = 0; i < BLOCK_SIZES; ++i)
     sf->inter_mode_mask[i] = INTER_ALL;

diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index a56d7de..ed84008 100644 (file)
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -318,6 +318,8 @@ typedef struct SPEED_FEATURES {
   // point for this motion search and limits the search range around it.
   int adaptive_motion_search;
 
+  int schedule_mode_search;
+
   // Allows sub 8x8 modes to use the prediction filter that was determined
   // best for 8x8 mode. If set to 0 we always re check all the filters for
   // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter