libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_firstpass.c

 /*
  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 20 matching lines @@
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_mcomp.h"
 #include "vp9/encoder/vp9_quantize.h"
 #include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_variance.h"
 
 #define OUTPUT_FPF          0
 #define ARF_STATS_OUTPUT    0
 
+#define GROUP_ADAPTIVE_MAXQ 0
+
 #define BOOST_BREAKOUT      12.5
 #define BOOST_FACTOR        12.5
 #define ERR_DIVISOR         128.0
 #define FACTOR_PT_LOW       0.70
 #define FACTOR_PT_HIGH      0.90
 #define FIRST_PASS_Q        10.0
 #define GF_MAX_BOOST        96.0
 #define INTRA_MODE_PENALTY  1024
 #define KF_MAX_BOOST        128.0
 #define MIN_ARF_GF_BOOST    240
 #define MIN_DECAY_FACTOR    0.01
 #define MIN_GF_INTERVAL     4
 #define MIN_KF_BOOST        300
 #define NEW_MV_MODE_PENALTY 32
 #define SVC_FACTOR_PT_LOW   0.45
 #define DARK_THRESH         64
+#define DEFAULT_GRP_WEIGHT  1.0
 
 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
 
 #if ARF_STATS_OUTPUT
 unsigned int arf_count = 0;
 #endif
 
 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
   YV12_BUFFER_CONFIG temp = *a;
   *a = *b;
@@ skipping 458 matching lines @@
     vp9_scale_references(cpi);
 
     // Use either last frame or alt frame for motion search.
     if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
       first_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
       if (first_ref_buf == NULL)
         first_ref_buf = get_ref_frame_buffer(cpi, LAST_FRAME);
     }
 
     if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
+      BufferPool *const pool = cm->buffer_pool;
       const int ref_idx =
           cm->ref_frame_map[get_ref_frame_idx(cpi, GOLDEN_FRAME)];
       const int scaled_idx = cpi->scaled_ref_idx[GOLDEN_FRAME - 1];
 
-      gld_yv12 = (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf :
+      gld_yv12 = (scaled_idx != ref_idx) ? &pool->frame_bufs[scaled_idx].buf :
                  get_ref_frame_buffer(cpi, GOLDEN_FRAME);
     } else {
       gld_yv12 = NULL;
     }
 
     recon_y_stride = new_yv12->y_stride;
     recon_uv_stride = new_yv12->uv_stride;
     uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
 
     set_ref_ptrs(cm, xd,
@@ skipping 525 matching lines @@
 
   return fclamp(pow(error_term, power_term), 0.05, 5.0);
 }
 
 // Larger image formats are expected to be a little harder to code relatively
 // given the same prediction error score. This in part at least relates to the
 // increased size and hence coding cost of motion vectors.
 #define EDIV_SIZE_FACTOR 800
 
 static int get_twopass_worst_quality(const VP9_COMP *cpi,
-                                     const FIRSTPASS_STATS *stats,
-                                     int section_target_bandwidth) {
+                                     const double section_err,
+                                     int section_target_bandwidth,
+                                     double group_weight_factor) {
   const RATE_CONTROL *const rc = &cpi->rc;
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
 
   if (section_target_bandwidth <= 0) {
     return rc->worst_quality;  // Highest value allowed
   } else {
     const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
                         ? cpi->initial_mbs : cpi->common.MBs;
-    const double section_err = stats->coded_error / stats->count;
     const double err_per_mb = section_err / num_mbs;
     const double speed_term = 1.0 + 0.04 * oxcf->speed;
     const double ediv_size_correction = num_mbs / EDIV_SIZE_FACTOR;
     const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
                                          BPER_MB_NORMBITS) / num_mbs;
 
     int q;
     int is_svc_upper_layer = 0;
+
     if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)
       is_svc_upper_layer = 1;
 
+
     // Try and pick a max Q that will be high enough to encode the
     // content at the given rate.
     for (q = rc->best_quality; q < rc->worst_quality; ++q) {
       const double factor =
           calc_correction_factor(err_per_mb,
                                  ERR_DIVISOR - ediv_size_correction,
                                  is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
                                  FACTOR_PT_LOW, FACTOR_PT_HIGH, q,
                                  cpi->common.bit_depth);
-      const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
-                                                 factor * speed_term,
-                                                 cpi->common.bit_depth);
+      const int bits_per_mb =
+        vp9_rc_bits_per_mb(INTER_FRAME, q,
+                           factor * speed_term * group_weight_factor,
+                           cpi->common.bit_depth);
       if (bits_per_mb <= target_norm_bits_per_mb)
         break;
     }
 
     // Restriction on active max q for constrained quality mode.
     if (cpi->oxcf.rc_mode == VPX_CQ)
       q = MAX(q, oxcf->cq_level);
     return q;
   }
 }
@@ skipping 563 matching lines @@
   RATE_CONTROL *const rc = &cpi->rc;
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   TWO_PASS *const twopass = &cpi->twopass;
   FIRSTPASS_STATS next_frame;
   const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
   int i;
 
   double boost_score = 0.0;
   double old_boost_score = 0.0;
   double gf_group_err = 0.0;
+#if GROUP_ADAPTIVE_MAXQ
+  double gf_group_raw_error = 0.0;
+#endif
   double gf_first_frame_err = 0.0;
   double mod_frame_err = 0.0;
 
   double mv_ratio_accumulator = 0.0;
   double decay_accumulator = 1.0;
   double zero_motion_accumulator = 1.0;
 
   double loop_decay_rate = 1.00;
   double last_loop_decay_rate = 1.00;
 
@@ skipping 23 matching lines @@
 
   // Load stats for the current frame.
   mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
 
   // Note the error of the frame at the start of the group. This will be
   // the GF frame error if we code a normal gf.
   gf_first_frame_err = mod_frame_err;
 
   // If this is a key frame or the overlay from a previous arf then
   // the error score / cost of this frame has already been accounted for.
-  if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
+  if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active) {
     gf_group_err -= gf_first_frame_err;
+#if GROUP_ADAPTIVE_MAXQ
+    gf_group_raw_error -= this_frame->coded_error;
+#endif
+  }
 
   // Motion breakout threshold for loop below depends on image size.
   mv_ratio_accumulator_thresh =
       (cpi->initial_height + cpi->initial_width) / 4.0;
 
   // Set a maximum and minimum interval for the GF group.
   // If the image appears almost completely static we can extend beyond this.
   {
     int int_max_q =
       (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
@@ skipping 18 matching lines @@
     }
   }
 
   i = 0;
   while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
     ++i;
 
     // Accumulate error score of frames in this gf group.
     mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
     gf_group_err += mod_frame_err;
+#if GROUP_ADAPTIVE_MAXQ
+    gf_group_raw_error += this_frame->coded_error;
+#endif
 
     if (EOF == input_stats(twopass, &next_frame))
       break;
 
     // Test for the case where there is a brief flash but the prediction
     // quality back to an earlier frame is then restored.
     flash_detected = detect_flash(twopass, 0);
 
     // Update the motion related elements to the boost calculation.
     accumulate_frame_motion_stats(&next_frame,
@@ skipping 41 matching lines @@
       boost_score = old_boost_score;
       break;
     }
 
     *this_frame = next_frame;
     old_boost_score = boost_score;
   }
 
   twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
 
+  // Was the group length constrained by the requirement for a new KF?
+  rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
+
   // Set the interval until the next gf.
   if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
     rc->baseline_gf_interval = i - 1;
   else
     rc->baseline_gf_interval = i;
 
   // Only encode alt reference frame in temporal base layer. So
   // baseline_gf_interval should be multiple of a temporal layer group
   // (typically the frame distance between two base layer frames)
   if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
     int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
     int new_gf_interval = (rc->baseline_gf_interval + count) & (~count);
     int j;
     for (j = 0; j < new_gf_interval - rc->baseline_gf_interval; ++j) {
       if (EOF == input_stats(twopass, this_frame))
         break;
       gf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
+#if GROUP_ADAPTIVE_MAXQ
+      gf_group_raw_error += this_frame->coded_error;
+#endif
     }
     rc->baseline_gf_interval = new_gf_interval;
   }
 
   rc->frames_till_gf_update_due = rc->baseline_gf_interval;
 
   // Should we use the alternate reference frame.
   if (allow_alt_ref &&
       (i < cpi->oxcf.lag_in_frames) &&
       (i >= MIN_GF_INTERVAL)) {
@@ skipping 10 matching lines @@
     rc->gfu_boost = MAX((int)boost_score, MIN_ARF_GF_BOOST);
     rc->source_alt_ref_pending = 0;
   }
 
   // Reset the file position.
   reset_fpf_position(twopass, start_pos);
 
   // Calculate the bits to be allocated to the gf/arf group as a whole
   gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
 
+#if GROUP_ADAPTIVE_MAXQ
+  // Calculate an estimate of the maxq needed for the group.
+  // We are more agressive about correcting for sections
+  // where there could be significant overshoot than for easier
+  // sections where we do not wish to risk creating an overshoot
+  // of the allocated bit budget.
+  if ((cpi->oxcf.rc_mode != VPX_Q) && (rc->baseline_gf_interval > 1)) {
+    const int vbr_group_bits_per_frame =
+      (int)(gf_group_bits / rc->baseline_gf_interval);
+    const double group_av_err = gf_group_raw_error  / rc->baseline_gf_interval;
+    const int tmp_q =
+      get_twopass_worst_quality(cpi, group_av_err, vbr_group_bits_per_frame,
+                                twopass->kfgroup_inter_fraction);
+
+    if (tmp_q < twopass->baseline_worst_quality) {
+      twopass->active_worst_quality =
+        (tmp_q + twopass->baseline_worst_quality + 1) / 2;
+    } else {
+      twopass->active_worst_quality = tmp_q;
+    }
+  }
+#endif
+
   // Calculate the extra bits to be used for boosted frame(s)
   gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
                                      rc->gfu_boost, gf_group_bits);
 
   // Adjust KF group bits and error remaining.
   twopass->kf_group_error_left -= (int64_t)gf_group_err;
 
   // If this is an arf update we want to remove the score for the overlay
   // frame at the end which will usually be very cheap to code.
   // The overlay frame has already, in effect, been coded so we want to spread
@@ skipping 203 matching lines @@
   // is between 1x and 2x.
   if (cpi->oxcf.auto_key &&
       rc->frames_to_key > cpi->oxcf.key_freq) {
     FIRSTPASS_STATS tmp_frame = first_frame;
 
     rc->frames_to_key /= 2;
 
     // Reset to the start of the group.
     reset_fpf_position(twopass, start_position);
 
-    kf_group_err = 0;
+    kf_group_err = 0.0;
 
     // Rescan to get the correct error data for the forced kf group.
     for (i = 0; i < rc->frames_to_key; ++i) {
       kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
       input_stats(twopass, &tmp_frame);
     }
     rc->next_key_frame_forced = 1;
   } else if (twopass->stats_in == twopass->stats_in_end ||
              rc->frames_to_key >= cpi->oxcf.key_freq) {
     rc->next_key_frame_forced = 1;
@@ skipping 89 matching lines @@
 
   // Apply various clamps for min and max boost
   rc->kf_boost = (int)(av_decay_accumulator * boost_score);
   rc->kf_boost = MAX(rc->kf_boost, (rc->frames_to_key * 3));
   rc->kf_boost = MAX(rc->kf_boost, MIN_KF_BOOST);
 
   // Work out how many bits to allocate for the key frame itself.
   kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
                                   rc->kf_boost, twopass->kf_group_bits);
 
+  // Work out the fraction of the kf group bits reserved for the inter frames
+  // within the group after discounting the bits for the kf itself.
+  if (twopass->kf_group_bits) {
+    twopass->kfgroup_inter_fraction =
+      (double)(twopass->kf_group_bits - kf_bits) /
+      (double)twopass->kf_group_bits;
+  } else {
+    twopass->kfgroup_inter_fraction = 1.0;
+  }
+
   twopass->kf_group_bits -= kf_bits;
 
   // Save the bits to spend on the key frame.
   gf_group->bit_allocation[0] = kf_bits;
   gf_group->update_type[0] = KF_UPDATE;
   gf_group->rf_level[0] = KF_STD;
 
   // Note the total error score of the kf group minus the key frame itself.
   twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
 
@@ skipping 70 matching lines @@
     twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
 }
 
 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
   TWO_PASS *const twopass = &cpi->twopass;
   GF_GROUP *const gf_group = &twopass->gf_group;
   int frames_left;
   FIRSTPASS_STATS this_frame;
-  FIRSTPASS_STATS this_frame_copy;
 
   int target_rate;
   LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?
         &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : 0;
 
   if (lc != NULL) {
     frames_left = (int)(twopass->total_stats.count -
                   lc->current_video_frame_in_layer);
   } else {
     frames_left = (int)(twopass->total_stats.count -
@@ skipping 37 matching lines @@
 
   vp9_clear_system_state();
 
   if (cpi->oxcf.rc_mode == VPX_Q) {
     twopass->active_worst_quality = cpi->oxcf.cq_level;
   } else if (cm->current_video_frame == 0 ||
              (lc != NULL && lc->current_video_frame_in_layer == 0)) {
     // Special case code for first frame.
     const int section_target_bandwidth = (int)(twopass->bits_left /
                                                frames_left);
-    const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
-                                                section_target_bandwidth);
+    const double section_error =
+      twopass->total_left_stats.coded_error / twopass->total_left_stats.count;
+    const int tmp_q =
+      get_twopass_worst_quality(cpi, section_error,
+                                section_target_bandwidth, DEFAULT_GRP_WEIGHT);
+
     twopass->active_worst_quality = tmp_q;
+    twopass->baseline_worst_quality = tmp_q;
     rc->ni_av_qi = tmp_q;
     rc->last_q[INTER_FRAME] = tmp_q;
     rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
     rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
     rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
     rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
   }
   vp9_zero(this_frame);
   if (EOF == input_stats(twopass, &this_frame))
     return;
 
-  // Local copy of the current frame's first pass stats.
-  this_frame_copy = this_frame;
-
   // Keyframe and section processing.
   if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
+    FIRSTPASS_STATS this_frame_copy;
+    this_frame_copy = this_frame;
     // Define next KF group and assign bits to it.
-    find_next_key_frame(cpi, &this_frame_copy);
+    find_next_key_frame(cpi, &this_frame);
+    this_frame = this_frame_copy;
   } else {
     cm->frame_type = INTER_FRAME;
   }
 
   if (lc != NULL) {
     if (cpi->svc.spatial_layer_id == 0) {
       lc->is_key_frame = (cm->frame_type == KEY_FRAME);
       if (lc->is_key_frame) {
         cpi->ref_frame_flags &=
             (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
         lc->frames_from_key_frame = 0;
         // Reset the empty frame resolution since we have a key frame.
         cpi->svc.empty_frame_width = cm->width;
         cpi->svc.empty_frame_height = cm->height;
       }
     } else {
       cm->frame_type = INTER_FRAME;
       lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
 
       if (lc->is_key_frame) {
         cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
         lc->frames_from_key_frame = 0;
       }
     }
   }
 
   // Define a new GF/ARF group. (Should always enter here for key frames).
   if (rc->frames_till_gf_update_due == 0) {
-    define_gf_group(cpi, &this_frame_copy);
+    define_gf_group(cpi, &this_frame);
 
     rc->frames_till_gf_update_due = rc->baseline_gf_interval;
     if (lc != NULL)
       cpi->refresh_golden_frame = 1;
 
 #if ARF_STATS_OUTPUT
     {
       FILE *fpfile;
       fpfile = fopen("arf.stt", "a");
       ++arf_count;
@@ skipping 96 matching lines @@
       // Unwind undershoot or overshoot adjustment.
       if (rc->rolling_target_bits < rc->rolling_actual_bits)
         --twopass->extend_minq;
       else if (rc->rolling_target_bits > rc->rolling_actual_bits)
         --twopass->extend_maxq;
     }
     twopass->extend_minq = clamp(twopass->extend_minq, 0, MINQ_ADJ_LIMIT);
     twopass->extend_maxq = clamp(twopass->extend_maxq, 0, maxq_adj_limit);
   }
 }