Import aom library

This is the reference implementation for the Alliance for Open Media's av1 video code.

The commit used was 4d668d7feb1f8abd809d1bca0418570a7f142a36.

third_party/aom/av1/encoder/aq_complexity.c

@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "av1/encoder/aq_complexity.h"
+#include "av1/encoder/aq_variance.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/common/seg_common.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/system_state.h"
+
+#define AQ_C_SEGMENTS 5
+#define DEFAULT_AQ2_SEG 3  // Neutral Q segment
+#define AQ_C_STRENGTHS 3
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { 1.75, 1.25, 1.05, 1.00, 0.90 },
+  { 2.00, 1.50, 1.15, 1.00, 0.85 },
+  { 2.50, 1.75, 1.25, 1.00, 0.80 }
+};
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { 0.15, 0.30, 0.55, 2.00, 100.0 },
+  { 0.20, 0.40, 0.65, 2.00, 100.0 },
+  { 0.25, 0.50, 0.75, 2.00, 100.0 }
+};
+static const double aq_c_var_thresholds[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { -4.0, -3.0, -2.0, 100.00, 100.0 },
+  { -3.5, -2.5, -1.5, 100.00, 100.0 },
+  { -3.0, -2.0, -1.0, 100.00, 100.0 }
+};
+
+#define DEFAULT_COMPLEXITY 64
+
+static int get_aq_c_strength(int q_index, aom_bit_depth_t bit_depth) {
+  // Approximate base quatizer (truncated to int)
+  const int base_quant = av1_ac_quant(q_index, 0, bit_depth) / 4;
+  return (base_quant > 10) + (base_quant > 25);
+}
+
+void av1_setup_in_frame_q_adj(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  struct segmentation *const seg = &cm->seg;
+
+  // Make SURE use of floating point in this function is safe.
+  aom_clear_system_state();
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    // Clear down the segment map.
+    memset(cpi->segmentation_map, DEFAULT_AQ2_SEG, cm->mi_rows * cm->mi_cols);
+
+    av1_clearall_segfeatures(seg);
+
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      av1_disable_segmentation(seg);
+      return;
+    }
+
+    av1_enable_segmentation(seg);
+
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Default segment "Q" feature is disabled so it defaults to the baseline Q.
+    av1_disable_segfeature(seg, DEFAULT_AQ2_SEG, SEG_LVL_ALT_Q);
+
+    // Use some of the segments for in frame Q adjustment.
+    for (segment = 0; segment < AQ_C_SEGMENTS; ++segment) {
+      int qindex_delta;
+
+      if (segment == DEFAULT_AQ2_SEG) continue;
+
+      qindex_delta = av1_compute_qdelta_by_rate(
+          &cpi->rc, cm->frame_type, cm->base_qindex,
+          aq_c_q_adj_factor[aq_strength][segment], cm->bit_depth);
+
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+      if ((cm->base_qindex + qindex_delta) > 0) {
+        av1_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
+        av1_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+      }
+    }
+  }
+}
+
+#define DEFAULT_LV_THRESH 10.0
+#define MIN_DEFAULT_LV_THRESH 8.0
+#define VAR_STRENGTH_STEP 0.25
+// Select a segment for the current block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average and its spatial complexity.
+void av1_caq_select_segment(const AV1_COMP *cpi, MACROBLOCK *mb, BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int xmis = AOMMIN(cm->mi_cols - mi_col, mi_size_wide[bs]);
+  const int ymis = AOMMIN(cm->mi_rows - mi_row, mi_size_high[bs]);
+  int x, y;
+  int i;
+  unsigned char segment;
+
+  if (0) {
+    segment = DEFAULT_AQ2_SEG;
+  } else {
+    // Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
+    // It is converted to bits * 256 units.
+    const int64_t num = (int64_t)cpi->rc.sb64_target_rate * xmis * ymis * 256;
+    const int denom = cm->mib_size * cm->mib_size;
+    const int target_rate = (int)(num / denom);
+    double logvar;
+    double low_var_thresh;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    aom_clear_system_state();
+    low_var_thresh = (cpi->oxcf.pass == 2) ? AOMMAX(cpi->twopass.mb_av_energy,
+                                                    MIN_DEFAULT_LV_THRESH)
+                                           : DEFAULT_LV_THRESH;
+
+    av1_setup_src_planes(mb, cpi->source, mi_row, mi_col);
+    logvar = av1_log_block_var(cpi, mb, bs);
+
+    segment = AQ_C_SEGMENTS - 1;  // Just in case no break out below.
+    for (i = 0; i < AQ_C_SEGMENTS; ++i) {
+      // Test rate against a threshold value and variance against a threshold.
+      // Increasing segment number (higher variance and complexity) = higher Q.
+      if ((projected_rate < target_rate * aq_c_transitions[aq_strength][i]) &&
+          (logvar < (low_var_thresh + aq_c_var_thresholds[aq_strength][i]))) {
+        segment = i;
+        break;
+      }
+    }
+  }
+
+  // Fill in the entires in the segment map corresponding to this SB64.
+  for (y = 0; y < ymis; y++) {
+    for (x = 0; x < xmis; x++) {
+      cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
+    }
+  }
+}

third_party/aom/av1/encoder/aq_complexity.h

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_COMPLEXITY_H_
+#define AV1_ENCODER_AQ_COMPLEXITY_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/common/enums.h"
+
+struct AV1_COMP;
+struct macroblock;
+
+// Select a segment for the current Block.
+void av1_caq_select_segment(const struct AV1_COMP *cpi, struct macroblock *,
+                            BLOCK_SIZE bs, int mi_row, int mi_col,
+                            int projected_rate);
+
+// This function sets up a set of segments with delta Q values around
+// the baseline frame quantizer.
+void av1_setup_in_frame_q_adj(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_COMPLEXITY_H_

third_party/aom/av1/encoder/aq_cyclicrefresh.c

@@ -0,0 +1,566 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "av1/common/seg_common.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/system_state.h"
+
+struct CYCLIC_REFRESH {
+  // Percentage of blocks per frame that are targeted as candidates
+  // for cyclic refresh.
+  int percent_refresh;
+  // Maximum q-delta as percentage of base q.
+  int max_qdelta_perc;
+  // Superblock starting index for cycling through the frame.
+  int sb_index;
+  // Controls how long block will need to wait to be refreshed again, in
+  // excess of the cycle time, i.e., in the case of all zero motion, block
+  // will be refreshed every (100/percent_refresh + time_for_refresh) frames.
+  int time_for_refresh;
+  // Target number of (8x8) blocks that are set for delta-q.
+  int target_num_seg_blocks;
+  // Actual number of (8x8) blocks that were applied delta-q.
+  int actual_num_seg1_blocks;
+  int actual_num_seg2_blocks;
+  // RD mult. parameters for segment 1.
+  int rdmult;
+  // Cyclic refresh map.
+  signed char *map;
+  // Map of the last q a block was coded at.
+  uint8_t *last_coded_q_map;
+  // Thresholds applied to the projected rate/distortion of the coding block,
+  // when deciding whether block should be refreshed.
+  int64_t thresh_rate_sb;
+  int64_t thresh_dist_sb;
+  // Threshold applied to the motion vector (in units of 1/8 pel) of the
+  // coding block, when deciding whether block should be refreshed.
+  int16_t motion_thresh;
+  // Rate target ratio to set q delta.
+  double rate_ratio_qdelta;
+  // Boost factor for rate target ratio, for segment CR_SEGMENT_ID_BOOST2.
+  int rate_boost_fac;
+  double low_content_avg;
+  int qindex_delta[3];
+};
+
+CYCLIC_REFRESH *av1_cyclic_refresh_alloc(int mi_rows, int mi_cols) {
+  size_t last_coded_q_map_size;
+  CYCLIC_REFRESH *const cr = aom_calloc(1, sizeof(*cr));
+  if (cr == NULL) return NULL;
+
+  cr->map = aom_calloc(mi_rows * mi_cols, sizeof(*cr->map));
+  if (cr->map == NULL) {
+    av1_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  last_coded_q_map_size = mi_rows * mi_cols * sizeof(*cr->last_coded_q_map);
+  cr->last_coded_q_map = aom_malloc(last_coded_q_map_size);
+  if (cr->last_coded_q_map == NULL) {
+    av1_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  assert(MAXQ <= 255);
+  memset(cr->last_coded_q_map, MAXQ, last_coded_q_map_size);
+
+  return cr;
+}
+
+void av1_cyclic_refresh_free(CYCLIC_REFRESH *cr) {
+  aom_free(cr->map);
+  aom_free(cr->last_coded_q_map);
+  aom_free(cr);
+}
+
+// Check if we should turn off cyclic refresh based on bitrate condition.
+static int apply_cyclic_refresh_bitrate(const AV1_COMMON *cm,
+                                        const RATE_CONTROL *rc) {
+  // Turn off cyclic refresh if bits available per frame is not sufficiently
+  // larger than bit cost of segmentation. Segment map bit cost should scale
+  // with number of seg blocks, so compare available bits to number of blocks.
+  // Average bits available per frame = avg_frame_bandwidth
+  // Number of (8x8) blocks in frame = mi_rows * mi_cols;
+  const float factor = 0.25;
+  const int number_blocks = cm->mi_rows * cm->mi_cols;
+  // The condition below corresponds to turning off at target bitrates:
+  // (at 30fps), ~12kbps for CIF, 36kbps for VGA, 100kps for HD/720p.
+  // Also turn off at very small frame sizes, to avoid too large fraction of
+  // superblocks to be refreshed per frame. Threshold below is less than QCIF.
+  if (rc->avg_frame_bandwidth < factor * number_blocks ||
+      number_blocks / 64 < 5)
+    return 0;
+  else
+    return 1;
+}
+
+// Check if this coding block, of size bsize, should be considered for refresh
+// (lower-qp coding). Decision can be based on various factors, such as
+// size of the coding block (i.e., below min_block size rejected), coding
+// mode, and rate/distortion.
+static int candidate_refresh_aq(const CYCLIC_REFRESH *cr,
+                                const MB_MODE_INFO *mbmi, int64_t rate,
+                                int64_t dist, int bsize) {
+  MV mv = mbmi->mv[0].as_mv;
+  // Reject the block for lower-qp coding if projected distortion
+  // is above the threshold, and any of the following is true:
+  // 1) mode uses large mv
+  // 2) mode is an intra-mode
+  // Otherwise accept for refresh.
+  if (dist > cr->thresh_dist_sb &&
+      (mv.row > cr->motion_thresh || mv.row < -cr->motion_thresh ||
+       mv.col > cr->motion_thresh || mv.col < -cr->motion_thresh ||
+       !is_inter_block(mbmi)))
+    return CR_SEGMENT_ID_BASE;
+  else if (bsize >= BLOCK_16X16 && rate < cr->thresh_rate_sb &&
+           is_inter_block(mbmi) && mbmi->mv[0].as_int == 0 &&
+           cr->rate_boost_fac > 10)
+    // More aggressive delta-q for bigger blocks with zero motion.
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BOOST1;
+}
+
+// Compute delta-q for the segment.
+static int compute_deltaq(const AV1_COMP *cpi, int q, double rate_factor) {
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int deltaq = av1_compute_qdelta_by_rate(rc, cpi->common.frame_type, q,
+                                          rate_factor, cpi->common.bit_depth);
+  if ((-deltaq) > cr->max_qdelta_perc * q / 100) {
+    deltaq = -cr->max_qdelta_perc * q / 100;
+  }
+  return deltaq;
+}
+
+// For the just encoded frame, estimate the bits, incorporating the delta-q
+// from non-base segment. For now ignore effect of multiple segments
+// (with different delta-q). Note this function is called in the postencode
+// (called from rc_update_rate_correction_factors()).
+int av1_cyclic_refresh_estimate_bits_at_q(const AV1_COMP *cpi,
+                                          double correction_factor) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int estimated_bits;
+  int mbs = cm->MBs;
+  int num8x8bl = mbs << 2;
+  // Weight for non-base segments: use actual number of blocks refreshed in
+  // previous/just encoded frame. Note number of blocks here is in 8x8 units.
+  double weight_segment1 = (double)cr->actual_num_seg1_blocks / num8x8bl;
+  double weight_segment2 = (double)cr->actual_num_seg2_blocks / num8x8bl;
+  // Take segment weighted average for estimated bits.
+  estimated_bits =
+      (int)((1.0 - weight_segment1 - weight_segment2) *
+                av1_estimate_bits_at_q(cm->frame_type, cm->base_qindex, mbs,
+                                       correction_factor, cm->bit_depth) +
+            weight_segment1 *
+                av1_estimate_bits_at_q(cm->frame_type,
+                                       cm->base_qindex + cr->qindex_delta[1],
+                                       mbs, correction_factor, cm->bit_depth) +
+            weight_segment2 *
+                av1_estimate_bits_at_q(cm->frame_type,
+                                       cm->base_qindex + cr->qindex_delta[2],
+                                       mbs, correction_factor, cm->bit_depth));
+  return estimated_bits;
+}
+
+// Prior to encoding the frame, estimate the bits per mb, for a given q = i and
+// a corresponding delta-q (for segment 1). This function is called in the
+// rc_regulate_q() to set the base qp index.
+// Note: the segment map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or
+// to 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock, prior to encoding.
+int av1_cyclic_refresh_rc_bits_per_mb(const AV1_COMP *cpi, int i,
+                                      double correction_factor) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int bits_per_mb;
+  int num8x8bl = cm->MBs << 2;
+  // Weight for segment prior to encoding: take the average of the target
+  // number for the frame to be encoded and the actual from the previous frame.
+  double weight_segment =
+      (double)((cr->target_num_seg_blocks + cr->actual_num_seg1_blocks +
+                cr->actual_num_seg2_blocks) >>
+               1) /
+      num8x8bl;
+  // Compute delta-q corresponding to qindex i.
+  int deltaq = compute_deltaq(cpi, i, cr->rate_ratio_qdelta);
+  // Take segment weighted average for bits per mb.
+  bits_per_mb = (int)((1.0 - weight_segment) *
+                          av1_rc_bits_per_mb(cm->frame_type, i,
+                                             correction_factor, cm->bit_depth) +
+                      weight_segment *
+                          av1_rc_bits_per_mb(cm->frame_type, i + deltaq,
+                                             correction_factor, cm->bit_depth));
+  return bits_per_mb;
+}
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void av1_cyclic_refresh_update_segment(const AV1_COMP *cpi,
+                                       MB_MODE_INFO *const mbmi, int mi_row,
+                                       int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int xmis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = AOMMIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  const int refresh_this_block =
+      candidate_refresh_aq(cr, mbmi, rate, dist, bsize);
+  // Default is to not update the refresh map.
+  int new_map_value = cr->map[block_index];
+  int x = 0;
+  int y = 0;
+
+  // If this block is labeled for refresh, check if we should reset the
+  // segment_id.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    mbmi->segment_id = refresh_this_block;
+    // Reset segment_id if will be skipped.
+    if (skip) mbmi->segment_id = CR_SEGMENT_ID_BASE;
+  }
+
+  // Update the cyclic refresh map, to be used for setting segmentation map
+  // for the next frame. If the block  will be refreshed this frame, mark it
+  // as clean. The magnitude of the -ve influences how long before we consider
+  // it for refresh again.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    new_map_value = -cr->time_for_refresh;
+  } else if (refresh_this_block) {
+    // Else if it is accepted as candidate for refresh, and has not already
+    // been refreshed (marked as 1) then mark it as a candidate for cleanup
+    // for future time (marked as 0), otherwise don't update it.
+    if (cr->map[block_index] == 1) new_map_value = 0;
+  } else {
+    // Leave it marked as block that is not candidate for refresh.
+    new_map_value = 1;
+  }
+
+  // Update entries in the cyclic refresh map with new_map_value, and
+  // copy mbmi->segment_id into global segmentation map.
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      int map_offset = block_index + y * cm->mi_cols + x;
+      cr->map[map_offset] = new_map_value;
+      cpi->segmentation_map[map_offset] = mbmi->segment_id;
+      // Inter skip blocks were clearly not coded at the current qindex, so
+      // don't update the map for them. For cases where motion is non-zero or
+      // the reference frame isn't the previous frame, the previous value in
+      // the map for this spatial location is not entirely correct.
+      if ((!is_inter_block(mbmi) || !skip) &&
+          mbmi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] = clamp(
+            cm->base_qindex + cr->qindex_delta[mbmi->segment_id], 0, MAXQ);
+      } else if (is_inter_block(mbmi) && skip &&
+                 mbmi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] =
+            AOMMIN(clamp(cm->base_qindex + cr->qindex_delta[mbmi->segment_id],
+                         0, MAXQ),
+                   cr->last_coded_q_map[map_offset]);
+      }
+    }
+}
+
+// Update the actual number of blocks that were applied the segment delta q.
+void av1_cyclic_refresh_postencode(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int mi_row, mi_col;
+  cr->actual_num_seg1_blocks = 0;
+  cr->actual_num_seg2_blocks = 0;
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++)
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      if (cyclic_refresh_segment_id(seg_map[mi_row * cm->mi_cols + mi_col]) ==
+          CR_SEGMENT_ID_BOOST1)
+        cr->actual_num_seg1_blocks++;
+      else if (cyclic_refresh_segment_id(
+                   seg_map[mi_row * cm->mi_cols + mi_col]) ==
+               CR_SEGMENT_ID_BOOST2)
+        cr->actual_num_seg2_blocks++;
+    }
+}
+
+// Set golden frame update interval, for 1 pass CBR mode.
+void av1_cyclic_refresh_set_golden_update(AV1_COMP *const cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  // Set minimum gf_interval for GF update to a multiple (== 2) of refresh
+  // period. Depending on past encoding stats, GF flag may be reset and update
+  // may not occur until next baseline_gf_interval.
+  if (cr->percent_refresh > 0)
+    rc->baseline_gf_interval = 4 * (100 / cr->percent_refresh);
+  else
+    rc->baseline_gf_interval = 40;
+}
+
+// Update some encoding stats (from the just encoded frame). If this frame's
+// background has high motion, refresh the golden frame. Otherwise, if the
+// golden reference is to be updated check if we should NOT update the golden
+// ref.
+void av1_cyclic_refresh_check_golden_update(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int mi_row, mi_col;
+  double fraction_low = 0.0;
+  int low_content_frame = 0;
+
+  MODE_INFO **mi;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int rows = cm->mi_rows, cols = cm->mi_cols;
+  int cnt1 = 0, cnt2 = 0;
+  int force_gf_refresh = 0;
+
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      int16_t abs_mvr = mi[0]->mbmi.mv[0].as_mv.row >= 0
+                            ? mi[0]->mbmi.mv[0].as_mv.row
+                            : -1 * mi[0]->mbmi.mv[0].as_mv.row;
+      int16_t abs_mvc = mi[0]->mbmi.mv[0].as_mv.col >= 0
+                            ? mi[0]->mbmi.mv[0].as_mv.col
+                            : -1 * mi[0]->mbmi.mv[0].as_mv.col;
+
+      // Calculate the motion of the background.
+      if (abs_mvr <= 16 && abs_mvc <= 16) {
+        cnt1++;
+        if (abs_mvr == 0 && abs_mvc == 0) cnt2++;
+      }
+      mi++;
+
+      // Accumulate low_content_frame.
+      if (cr->map[mi_row * cols + mi_col] < 1) low_content_frame++;
+    }
+  }
+
+  // For video conference clips, if the background has high motion in current
+  // frame because of the camera movement, set this frame as the golden frame.
+  // Use 70% and 5% as the thresholds for golden frame refreshing.
+  // Also, force this frame as a golden update frame if this frame will change
+  // the resolution (resize_pending != 0).
+  if (cpi->resize_pending != 0 ||
+      (cnt1 * 10 > (70 * rows * cols) && cnt2 * 20 < cnt1)) {
+    av1_cyclic_refresh_set_golden_update(cpi);
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    force_gf_refresh = 1;
+  }
+
+  fraction_low = (double)low_content_frame / (rows * cols);
+  // Update average.
+  cr->low_content_avg = (fraction_low + 3 * cr->low_content_avg) / 4;
+  if (!force_gf_refresh && cpi->refresh_golden_frame == 1) {
+    // Don't update golden reference if the amount of low_content for the
+    // current encoded frame is small, or if the recursive average of the
+    // low_content over the update interval window falls below threshold.
+    if (fraction_low < 0.8 || cr->low_content_avg < 0.7)
+      cpi->refresh_golden_frame = 0;
+    // Reset for next internal.
+    cr->low_content_avg = fraction_low;
+  }
+}
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+// The map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or to
+// 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock.
+// Blocks labeled as BOOST1 may later get set to BOOST2 (during the
+// encoding of the superblock).
+static void cyclic_refresh_update_map(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int i, block_count, bl_index, sb_rows, sb_cols, sbs_in_frame;
+  int xmis, ymis, x, y;
+  memset(seg_map, CR_SEGMENT_ID_BASE, cm->mi_rows * cm->mi_cols);
+  sb_cols = (cm->mi_cols + cm->mib_size - 1) / cm->mib_size;
+  sb_rows = (cm->mi_rows + cm->mib_size - 1) / cm->mib_size;
+  sbs_in_frame = sb_cols * sb_rows;
+  // Number of target blocks to get the q delta (segment 1).
+  block_count = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
+  // Set the segmentation map: cycle through the superblocks, starting at
+  // cr->mb_index, and stopping when either block_count blocks have been found
+  // to be refreshed, or we have passed through whole frame.
+  assert(cr->sb_index < sbs_in_frame);
+  i = cr->sb_index;
+  cr->target_num_seg_blocks = 0;
+  do {
+    int sum_map = 0;
+    // Get the mi_row/mi_col corresponding to superblock index i.
+    int sb_row_index = (i / sb_cols);
+    int sb_col_index = i - sb_row_index * sb_cols;
+    int mi_row = sb_row_index * cm->mib_size;
+    int mi_col = sb_col_index * cm->mib_size;
+    int qindex_thresh =
+        cpi->oxcf.content == AOM_CONTENT_SCREEN
+            ? av1_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST2, cm->base_qindex)
+            : 0;
+    assert(mi_row >= 0 && mi_row < cm->mi_rows);
+    assert(mi_col >= 0 && mi_col < cm->mi_cols);
+    bl_index = mi_row * cm->mi_cols + mi_col;
+    // Loop through all MI blocks in superblock and update map.
+    xmis = AOMMIN(cm->mi_cols - mi_col, cm->mib_size);
+    ymis = AOMMIN(cm->mi_rows - mi_row, cm->mib_size);
+    for (y = 0; y < ymis; y++) {
+      for (x = 0; x < xmis; x++) {
+        const int bl_index2 = bl_index + y * cm->mi_cols + x;
+        // If the block is as a candidate for clean up then mark it
+        // for possible boost/refresh (segment 1). The segment id may get
+        // reset to 0 later if block gets coded anything other than ZEROMV.
+        if (cr->map[bl_index2] == 0) {
+          if (cr->last_coded_q_map[bl_index2] > qindex_thresh) sum_map++;
+        } else if (cr->map[bl_index2] < 0) {
+          cr->map[bl_index2]++;
+        }
+      }
+    }
+    // Enforce constant segment over superblock.
+    // If segment is at least half of superblock, set to 1.
+    if (sum_map >= xmis * ymis / 2) {
+      for (y = 0; y < ymis; y++)
+        for (x = 0; x < xmis; x++) {
+          seg_map[bl_index + y * cm->mi_cols + x] = CR_SEGMENT_ID_BOOST1;
+        }
+      cr->target_num_seg_blocks += xmis * ymis;
+    }
+    i++;
+    if (i == sbs_in_frame) {
+      i = 0;
+    }
+  } while (cr->target_num_seg_blocks < block_count && i != cr->sb_index);
+  cr->sb_index = i;
+}
+
+// Set cyclic refresh parameters.
+void av1_cyclic_refresh_update_parameters(AV1_COMP *const cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  cr->percent_refresh = 10;
+  cr->max_qdelta_perc = 50;
+  cr->time_for_refresh = 0;
+  // Use larger delta-qp (increase rate_ratio_qdelta) for first few (~4)
+  // periods of the refresh cycle, after a key frame.
+  if (rc->frames_since_key < 4 * cr->percent_refresh)
+    cr->rate_ratio_qdelta = 3.0;
+  else
+    cr->rate_ratio_qdelta = 2.0;
+  // Adjust some parameters for low resolutions at low bitrates.
+  if (cm->width <= 352 && cm->height <= 288 && rc->avg_frame_bandwidth < 3400) {
+    cr->motion_thresh = 4;
+    cr->rate_boost_fac = 10;
+  } else {
+    cr->motion_thresh = 32;
+    cr->rate_boost_fac = 17;
+  }
+}
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void av1_cyclic_refresh_setup(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  struct segmentation *const seg = &cm->seg;
+  const int apply_cyclic_refresh = apply_cyclic_refresh_bitrate(cm, rc);
+  if (cm->current_video_frame == 0) cr->low_content_avg = 0.0;
+  // Don't apply refresh on key frame or enhancement layer frames.
+  if (!apply_cyclic_refresh || cm->frame_type == KEY_FRAME) {
+    // Set segmentation map to 0 and disable.
+    unsigned char *const seg_map = cpi->segmentation_map;
+    memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
+    av1_disable_segmentation(&cm->seg);
+    if (cm->frame_type == KEY_FRAME) {
+      memset(cr->last_coded_q_map, MAXQ,
+             cm->mi_rows * cm->mi_cols * sizeof(*cr->last_coded_q_map));
+      cr->sb_index = 0;
+    }
+    return;
+  } else {
+    int qindex_delta = 0;
+    int qindex2;
+    const double q = av1_convert_qindex_to_q(cm->base_qindex, cm->bit_depth);
+    aom_clear_system_state();
+    // Set rate threshold to some multiple (set to 2 for now) of the target
+    // rate (target is given by sb64_target_rate and scaled by 256).
+    cr->thresh_rate_sb = ((int64_t)(rc->sb64_target_rate) << 8) << 2;
+    // Distortion threshold, quadratic in Q, scale factor to be adjusted.
+    // q will not exceed 457, so (q * q) is within 32bit; see:
+    // av1_convert_qindex_to_q(), av1_ac_quant(), ac_qlookup*[].
+    cr->thresh_dist_sb = ((int64_t)(q * q)) << 2;
+
+    // Set up segmentation.
+    // Clear down the segment map.
+    av1_enable_segmentation(&cm->seg);
+    av1_clearall_segfeatures(seg);
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Note: setting temporal_update has no effect, as the seg-map coding method
+    // (temporal or spatial) is determined in
+    // av1_choose_segmap_coding_method(),
+    // based on the coding cost of each method. For error_resilient mode on the
+    // last_frame_seg_map is set to 0, so if temporal coding is used, it is
+    // relative to 0 previous map.
+    // seg->temporal_update = 0;
+
+    // Segment BASE "Q" feature is disabled so it defaults to the baseline Q.
+    av1_disable_segfeature(seg, CR_SEGMENT_ID_BASE, SEG_LVL_ALT_Q);
+    // Use segment BOOST1 for in-frame Q adjustment.
+    av1_enable_segfeature(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q);
+    // Use segment BOOST2 for more aggressive in-frame Q adjustment.
+    av1_enable_segfeature(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q);
+
+    // Set the q delta for segment BOOST1.
+    qindex_delta = compute_deltaq(cpi, cm->base_qindex, cr->rate_ratio_qdelta);
+    cr->qindex_delta[1] = qindex_delta;
+
+    // Compute rd-mult for segment BOOST1.
+    qindex2 = clamp(cm->base_qindex + cm->y_dc_delta_q + qindex_delta, 0, MAXQ);
+
+    cr->rdmult = av1_compute_rd_mult(cpi, qindex2);
+
+    av1_set_segdata(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Set a more aggressive (higher) q delta for segment BOOST2.
+    qindex_delta = compute_deltaq(
+        cpi, cm->base_qindex,
+        AOMMIN(CR_MAX_RATE_TARGET_RATIO,
+               0.1 * cr->rate_boost_fac * cr->rate_ratio_qdelta));
+    cr->qindex_delta[2] = qindex_delta;
+    av1_set_segdata(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Update the segmentation and refresh map.
+    cyclic_refresh_update_map(cpi);
+  }
+}
+
+int av1_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr) {
+  return cr->rdmult;
+}
+
+void av1_cyclic_refresh_reset_resize(AV1_COMP *const cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  memset(cr->map, 0, cm->mi_rows * cm->mi_cols);
+  cr->sb_index = 0;
+  cpi->refresh_golden_frame = 1;
+}

third_party/aom/av1/encoder/aq_cyclicrefresh.h

@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_CYCLICREFRESH_H_
+#define AV1_ENCODER_AQ_CYCLICREFRESH_H_
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The segment ids used in cyclic refresh: from base (no boost) to increasing
+// boost (higher delta-qp).
+#define CR_SEGMENT_ID_BASE 0
+#define CR_SEGMENT_ID_BOOST1 1
+#define CR_SEGMENT_ID_BOOST2 2
+
+// Maximum rate target ratio for setting segment delta-qp.
+#define CR_MAX_RATE_TARGET_RATIO 4.0
+
+struct AV1_COMP;
+
+struct CYCLIC_REFRESH;
+typedef struct CYCLIC_REFRESH CYCLIC_REFRESH;
+
+CYCLIC_REFRESH *av1_cyclic_refresh_alloc(int mi_rows, int mi_cols);
+
+void av1_cyclic_refresh_free(CYCLIC_REFRESH *cr);
+
+// Estimate the bits, incorporating the delta-q from segment 1, after encoding
+// the frame.
+int av1_cyclic_refresh_estimate_bits_at_q(const struct AV1_COMP *cpi,
+                                          double correction_factor);
+
+// Estimate the bits per mb, for a given q = i and a corresponding delta-q
+// (for segment 1), prior to encoding the frame.
+int av1_cyclic_refresh_rc_bits_per_mb(const struct AV1_COMP *cpi, int i,
+                                      double correction_factor);
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void av1_cyclic_refresh_update_segment(const struct AV1_COMP *cpi,
+                                       MB_MODE_INFO *const mbmi, int mi_row,
+                                       int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip);
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+void av1_cyclic_refresh_update__map(struct AV1_COMP *const cpi);
+
+// Update the actual number of blocks that were applied the segment delta q.
+void av1_cyclic_refresh_postencode(struct AV1_COMP *const cpi);
+
+// Set golden frame update interval, for 1 pass CBR mode.
+void av1_cyclic_refresh_set_golden_update(struct AV1_COMP *const cpi);
+
+// Check if we should not update golden reference, based on past refresh stats.
+void av1_cyclic_refresh_check_golden_update(struct AV1_COMP *const cpi);
+
+// Set/update global/frame level refresh parameters.
+void av1_cyclic_refresh_update_parameters(struct AV1_COMP *const cpi);
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void av1_cyclic_refresh_setup(struct AV1_COMP *const cpi);
+
+int av1_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr);
+
+void av1_cyclic_refresh_reset_resize(struct AV1_COMP *const cpi);
+
+static INLINE int cyclic_refresh_segment_id_boosted(int segment_id) {
+  return segment_id == CR_SEGMENT_ID_BOOST1 ||
+         segment_id == CR_SEGMENT_ID_BOOST2;
+}
+
+static INLINE int cyclic_refresh_segment_id(int segment_id) {
+  if (segment_id == CR_SEGMENT_ID_BOOST1)
+    return CR_SEGMENT_ID_BOOST1;
+  else if (segment_id == CR_SEGMENT_ID_BOOST2)
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BASE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_CYCLICREFRESH_H_

third_party/aom/av1/encoder/aq_variance.c

@@ -0,0 +1,207 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "aom_ports/mem.h"
+
+#include "av1/encoder/aq_variance.h"
+
+#include "av1/common/seg_common.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_ports/system_state.h"
+
+#define ENERGY_MIN (-4)
+#define ENERGY_MAX (1)
+#define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN + 1)
+#define ENERGY_IN_BOUNDS(energy) \
+  assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
+
+static const double rate_ratio[MAX_SEGMENTS] = { 2.5,  2.0, 1.5, 1.0,
+                                                 0.75, 1.0, 1.0, 1.0 };
+static const int segment_id[ENERGY_SPAN] = { 0, 1, 1, 2, 3, 4 };
+
+#define SEGMENT_ID(i) segment_id[(i)-ENERGY_MIN]
+
+DECLARE_ALIGNED(16, static const uint8_t, av1_all_zeros[MAX_SB_SIZE]) = { 0 };
+#if CONFIG_HIGHBITDEPTH
+DECLARE_ALIGNED(16, static const uint16_t,
+                av1_highbd_all_zeros[MAX_SB_SIZE]) = { 0 };
+#endif
+
+unsigned int av1_vaq_segment_id(int energy) {
+  ENERGY_IN_BOUNDS(energy);
+  return SEGMENT_ID(energy);
+}
+
+void av1_vaq_frame_setup(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  struct segmentation *seg = &cm->seg;
+  int i;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    cpi->vaq_refresh = 1;
+
+    av1_enable_segmentation(seg);
+    av1_clearall_segfeatures(seg);
+
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    aom_clear_system_state();
+
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      int qindex_delta =
+          av1_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                     rate_ratio[i], cm->bit_depth);
+
+      // We don't allow qindex 0 in a segment if the base value is not 0.
+      // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+
+      // No need to enable SEG_LVL_ALT_Q for this segment.
+      if (rate_ratio[i] == 1.0) {
+        continue;
+      }
+
+      av1_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
+      av1_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
+    }
+  }
+}
+
+/* TODO(agrange, paulwilkins): The block_variance calls the unoptimized versions
+ * of variance() and highbd_8_variance(). It should not.
+ */
+static void aq_variance(const uint8_t *a, int a_stride, const uint8_t *b,
+                        int b_stride, int w, int h, unsigned int *sse,
+                        int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void aq_highbd_variance64(const uint8_t *a8, int a_stride,
+                                 const uint8_t *b8, int b_stride, int w, int h,
+                                 uint64_t *sse, uint64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void aq_highbd_8_variance(const uint8_t *a8, int a_stride,
+                                 const uint8_t *b8, int b_stride, int w, int h,
+                                 unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  uint64_t sum_long = 0;
+  aq_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static unsigned int block_variance(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                   BLOCK_SIZE bs) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var, sse;
+  int right_overflow =
+      (xd->mb_to_right_edge < 0) ? ((-xd->mb_to_right_edge) >> 3) : 0;
+  int bottom_overflow =
+      (xd->mb_to_bottom_edge < 0) ? ((-xd->mb_to_bottom_edge) >> 3) : 0;
+
+  if (right_overflow || bottom_overflow) {
+    const int bw = 8 * mi_size_wide[bs] - right_overflow;
+    const int bh = 8 * mi_size_high[bs] - bottom_overflow;
+    int avg;
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      aq_highbd_8_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                           CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, bw, bh,
+                           &sse, &avg);
+      sse >>= 2 * (xd->bd - 8);
+      avg >>= (xd->bd - 8);
+    } else {
+      aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
+                  bw, bh, &sse, &avg);
+    }
+#else
+    aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
+                bw, bh, &sse, &avg);
+#endif  // CONFIG_HIGHBITDEPTH
+    var = sse - (unsigned int)(((int64_t)avg * avg) / (bw * bh));
+    return (unsigned int)((uint64_t)var * 256) / (bw * bh);
+  } else {
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      var =
+          cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                             CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, &sse);
+    } else {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                               av1_all_zeros, 0, &sse);
+    }
+#else
+    var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                             av1_all_zeros, 0, &sse);
+#endif  // CONFIG_HIGHBITDEPTH
+    return (unsigned int)((uint64_t)var * 256) >> num_pels_log2_lookup[bs];
+  }
+}
+
+double av1_log_block_var(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  unsigned int var = block_variance(cpi, x, bs);
+  aom_clear_system_state();
+  return log(var + 1.0);
+}
+
+#define DEFAULT_E_MIDPOINT 10.0
+int av1_block_energy(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  double energy;
+  double energy_midpoint;
+  aom_clear_system_state();
+  energy_midpoint =
+      (cpi->oxcf.pass == 2) ? cpi->twopass.mb_av_energy : DEFAULT_E_MIDPOINT;
+  energy = av1_log_block_var(cpi, x, bs) - energy_midpoint;
+  return clamp((int)round(energy), ENERGY_MIN, ENERGY_MAX);
+}

third_party/aom/av1/encoder/aq_variance.h

@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_VARIANCE_H_
+#define AV1_ENCODER_AQ_VARIANCE_H_
+
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int av1_vaq_segment_id(int energy);
+void av1_vaq_frame_setup(AV1_COMP *cpi);
+
+int av1_block_energy(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+double av1_log_block_var(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_VARIANCE_H_

third_party/aom/av1/encoder/arm/neon/dct_neon.c

@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "av1/common/blockd.h"
+#include "aom_dsp/txfm_common.h"
+
+void av1_fdct8x8_quant_neon(const int16_t *input, int stride,
+                            int16_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t *zbin_ptr,
+                            const int16_t *round_ptr, const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                            int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                            uint16_t *eob_ptr, const int16_t *scan_ptr,
+                            const int16_t *iscan_ptr) {
+  int16_t temp_buffer[64];
+  (void)coeff_ptr;
+
+  aom_fdct8x8_neon(input, temp_buffer, stride);
+  av1_quantize_fp_neon(temp_buffer, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
+}

third_party/aom/av1/encoder/arm/neon/error_neon.c

@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./av1_rtcd.h"
+
+int64_t av1_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff,
+                                int block_size) {
+  int64x2_t error = vdupq_n_s64(0);
+
+  assert(block_size >= 8);
+  assert((block_size % 8) == 0);
+
+  do {
+    const int16x8_t c = vld1q_s16(coeff);
+    const int16x8_t d = vld1q_s16(dqcoeff);
+    const int16x8_t diff = vsubq_s16(c, d);
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    // diff is 15-bits, the squares 30, so we can store 2 in 31-bits before
+    // accumulating them in 64-bits.
+    const int32x4_t err0 = vmull_s16(diff_lo, diff_lo);
+    const int32x4_t err1 = vmlal_s16(err0, diff_hi, diff_hi);
+    const int64x2_t err2 = vaddl_s32(vget_low_s32(err1), vget_high_s32(err1));
+    error = vaddq_s64(error, err2);
+    coeff += 8;
+    dqcoeff += 8;
+    block_size -= 8;
+  } while (block_size != 0);
+
+  return vgetq_lane_s64(error, 0) + vgetq_lane_s64(error, 1);
+}

third_party/aom/av1/encoder/arm/neon/quantize_neon.c

@@ -0,0 +1,118 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include <math.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/quant_common.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/rd.h"
+
+void av1_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan,
+                          const int16_t *iscan) {
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    int i;
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+    // process dc and the first seven ac coeffs
+    {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo =
+          vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi =
+          vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 =
+          vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    // now process the rest of the ac coeffs
+    for (i = 8; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo =
+          vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi =
+          vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 =
+          vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+    }
+    {
+      const int16x4_t v_eobmax_3210 = vmax_s16(
+          vget_low_s16(v_eobmax_76543210), vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}

third_party/aom/av1/encoder/av1_quantize.h

@@ -0,0 +1,184 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_QUANTIZE_H_
+#define AV1_ENCODER_QUANTIZE_H_
+
+#include "./aom_config.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/scan.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct QUANT_PARAM {
+  int log_scale;
+#if CONFIG_NEW_QUANT
+  TX_SIZE tx_size;
+  int dq;
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+  const qm_val_t *qmatrix;
+  const qm_val_t *iqmatrix;
+#endif  // CONFIG_AOM_QM
+} QUANT_PARAM;
+
+typedef void (*AV1_QUANT_FACADE)(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                 const MACROBLOCK_PLANE *p,
+                                 tran_low_t *qcoeff_ptr,
+                                 const MACROBLOCKD_PLANE *pd,
+                                 tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                 const SCAN_ORDER *sc,
+                                 const QUANT_PARAM *qparam);
+
+typedef struct {
+#if CONFIG_NEW_QUANT
+  DECLARE_ALIGNED(
+      16, tran_low_t,
+      y_cuml_bins_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS][NUQ_KNOTS]);
+  DECLARE_ALIGNED(
+      16, tran_low_t,
+      uv_cuml_bins_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS][NUQ_KNOTS]);
+#endif  // CONFIG_NEW_QUANT
+  // 0: dc 1: ac 2-8: ac repeated to SIMD width
+  DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
+
+  // TODO(jingning): in progress of re-working the quantization. will decide
+  // if we want to deprecate the current use of y_quant.
+  DECLARE_ALIGNED(16, int16_t, y_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round_fp[QINDEX_RANGE][8]);
+
+  DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
+} QUANTS;
+
+struct AV1_COMP;
+struct AV1Common;
+
+void av1_frame_init_quantizer(struct AV1_COMP *cpi);
+
+void av1_init_plane_quantizers(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                               int segment_id);
+
+void av1_init_quantizer(struct AV1_COMP *cpi);
+
+void av1_set_quantizer(struct AV1Common *cm, int q);
+
+int av1_quantizer_to_qindex(int quantizer);
+
+int av1_qindex_to_quantizer(int qindex);
+
+void av1_quantize_skip(intptr_t n_coeffs, tran_low_t *qcoeff_ptr,
+                       tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr);
+
+void av1_quantize_fp_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+void av1_quantize_b_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                           const MACROBLOCKD_PLANE *pd, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr, const SCAN_ORDER *sc,
+                           const QUANT_PARAM *qparam);
+
+void av1_quantize_dc_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+#if CONFIG_NEW_QUANT
+void av1_quantize_fp_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam);
+
+void av1_quantize_b_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               const MACROBLOCK_PLANE *p,
+                               tran_low_t *qcoeff_ptr,
+                               const MACROBLOCKD_PLANE *pd,
+                               tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                               const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+void av1_quantize_dc_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam);
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_quantize_fp_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_b_facade(const tran_low_t *coeff_ptr,
+                                  intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                  tran_low_t *qcoeff_ptr,
+                                  const MACROBLOCKD_PLANE *pd,
+                                  tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                  const SCAN_ORDER *sc,
+                                  const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_dc_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam);
+
+#if CONFIG_NEW_QUANT
+void av1_highbd_quantize_fp_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_b_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_dc_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+#endif  // CONFIG_NEW_QUANT
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_QUANTIZE_H_

third_party/aom/av1/encoder/bitstream.h

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_BITSTREAM_H_
+#define AV1_ENCODER_BITSTREAM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+#if CONFIG_REFERENCE_BUFFER
+void write_sequence_header(SequenceHeader *seq_params);
+#endif
+
+void av1_pack_bitstream(AV1_COMP *const cpi, uint8_t *dest, size_t *size);
+
+void av1_encode_token_init(void);
+
+static INLINE int av1_preserve_existing_gf(AV1_COMP *cpi) {
+#if CONFIG_EXT_REFS
+  // Do not swap gf and arf indices for internal overlay frames
+  return !cpi->multi_arf_allowed && cpi->rc.is_src_frame_alt_ref &&
+         !cpi->rc.is_src_frame_ext_arf;
+#else
+  return !cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
+         cpi->rc.is_src_frame_alt_ref;
+#endif  // CONFIG_EXT_REFS
+}
+
+void av1_write_tx_type(const AV1_COMMON *const cm, const MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                       const int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                       int block, int plane,
+#endif
+                       aom_writer *w);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_BITSTREAM_H_

third_party/aom/av1/encoder/block.h

@@ -0,0 +1,241 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_BLOCK_H_
+#define AV1_ENCODER_BLOCK_H_
+
+#include "av1/common/entropymv.h"
+#include "av1/common/entropy.h"
+#if CONFIG_PVQ
+#include "av1/encoder/encint.h"
+#endif
+#if CONFIG_REF_MV
+#include "av1/common/mvref_common.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_PVQ
+// Maximum possible # of tx blocks in luma plane, which is currently 256,
+// since there can be 16x16 of 4x4 tx.
+#define MAX_PVQ_BLOCKS_IN_SB (MAX_SB_SQUARE >> 2 * OD_LOG_BSIZE0)
+#endif
+
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} DIFF;
+
+typedef struct macroblock_plane {
+  DECLARE_ALIGNED(16, int16_t, src_diff[MAX_SB_SQUARE]);
+#if CONFIG_PVQ
+  DECLARE_ALIGNED(16, int16_t, src_int16[MAX_SB_SQUARE]);
+#endif
+  tran_low_t *qcoeff;
+  tran_low_t *coeff;
+  uint16_t *eobs;
+#if CONFIG_LV_MAP
+  uint8_t *txb_entropy_ctx;
+#endif
+  struct buf_2d src;
+
+  // Quantizer setings
+  const int16_t *quant_fp;
+  const int16_t *round_fp;
+  const int16_t *quant;
+  const int16_t *quant_shift;
+  const int16_t *zbin;
+  const int16_t *round;
+#if CONFIG_NEW_QUANT
+  const cuml_bins_type_nuq *cuml_bins_nuq[QUANT_PROFILES];
+#endif  // CONFIG_NEW_QUANT
+} MACROBLOCK_PLANE;
+
+/* The [2] dimension is for whether we skip the EOB node (i.e. if previous
+ * coefficient in this block was zero) or not. */
+typedef unsigned int av1_coeff_cost[PLANE_TYPES][REF_TYPES][COEF_BANDS][2]
+                                   [COEFF_CONTEXTS][ENTROPY_TOKENS];
+
+typedef struct {
+  int_mv ref_mvs[MODE_CTX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  int16_t mode_context[MODE_CTX_REF_FRAMES];
+#if CONFIG_LV_MAP
+  // TODO(angiebird): Reduce the buffer size according to sb_type
+  tran_low_t tcoeff[MAX_MB_PLANE][MAX_SB_SQUARE];
+  uint16_t eobs[MAX_MB_PLANE][MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+  uint8_t txb_skip_ctx[MAX_MB_PLANE]
+                      [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+  int dc_sign_ctx[MAX_MB_PLANE]
+                 [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+#endif
+#if CONFIG_REF_MV
+  uint8_t ref_mv_count[MODE_CTX_REF_FRAMES];
+  CANDIDATE_MV ref_mv_stack[MODE_CTX_REF_FRAMES][MAX_REF_MV_STACK_SIZE];
+#if CONFIG_EXT_INTER
+  int16_t compound_mode_context[MODE_CTX_REF_FRAMES];
+#endif  // CONFIG_EXT_INTER
+#endif
+} MB_MODE_INFO_EXT;
+
+typedef struct {
+  int col_min;
+  int col_max;
+  int row_min;
+  int row_max;
+} MvLimits;
+
+#if CONFIG_PALETTE
+typedef struct {
+  uint8_t best_palette_color_map[MAX_SB_SQUARE];
+  float kmeans_data_buf[2 * MAX_SB_SQUARE];
+} PALETTE_BUFFER;
+#endif  // CONFIG_PALETTE
+
+typedef struct macroblock MACROBLOCK;
+struct macroblock {
+  struct macroblock_plane plane[MAX_MB_PLANE];
+
+  MACROBLOCKD e_mbd;
+  MB_MODE_INFO_EXT *mbmi_ext;
+  int skip_block;
+  int qindex;
+
+  // The equivalent error at the current rdmult of one whole bit (not one
+  // bitcost unit).
+  int errorperbit;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for large blocks.
+  int sadperbit16;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for sub-8x8 blocks.
+  int sadperbit4;
+  int rddiv;
+  int rdmult;
+  int mb_energy;
+  int *m_search_count_ptr;
+  int *ex_search_count_ptr;
+
+#if CONFIG_VAR_TX
+  unsigned int txb_split_count;
+#endif
+
+  // These are set to their default values at the beginning, and then adjusted
+  // further in the encoding process.
+  BLOCK_SIZE min_partition_size;
+  BLOCK_SIZE max_partition_size;
+
+  int mv_best_ref_index[TOTAL_REFS_PER_FRAME];
+  unsigned int max_mv_context[TOTAL_REFS_PER_FRAME];
+  unsigned int source_variance;
+  unsigned int pred_sse[TOTAL_REFS_PER_FRAME];
+  int pred_mv_sad[TOTAL_REFS_PER_FRAME];
+
+#if CONFIG_REF_MV
+  int *nmvjointcost;
+  int nmv_vec_cost[NMV_CONTEXTS][MV_JOINTS];
+  int *nmvcost[NMV_CONTEXTS][2];
+  int *nmvcost_hp[NMV_CONTEXTS][2];
+  int **mv_cost_stack[NMV_CONTEXTS];
+  int *nmvjointsadcost;
+#else
+  int nmvjointcost[MV_JOINTS];
+  int *nmvcost[2];
+  int *nmvcost_hp[2];
+  int nmvjointsadcost[MV_JOINTS];
+#endif
+
+  int **mvcost;
+  int *nmvsadcost[2];
+  int *nmvsadcost_hp[2];
+  int **mvsadcost;
+#if CONFIG_MOTION_VAR
+  int32_t *wsrc_buf;
+  int32_t *mask_buf;
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_PALETTE
+  PALETTE_BUFFER *palette_buffer;
+#endif  // CONFIG_PALETTE
+
+  // These define limits to motion vector components to prevent them
+  // from extending outside the UMV borders
+  MvLimits mv_limits;
+
+#if CONFIG_VAR_TX
+  uint8_t blk_skip[MAX_MB_PLANE][MAX_MIB_SIZE * MAX_MIB_SIZE * 8];
+#if CONFIG_REF_MV
+  uint8_t blk_skip_drl[MAX_MB_PLANE][MAX_MIB_SIZE * MAX_MIB_SIZE * 8];
+#endif
+#endif
+
+  int skip;
+
+#if CONFIG_CB4X4
+  int skip_chroma_rd;
+#endif
+
+  // note that token_costs is the cost when eob node is skipped
+  av1_coeff_cost token_costs[TX_SIZES];
+
+  int optimize;
+
+  // Used to store sub partition's choices.
+  MV pred_mv[TOTAL_REFS_PER_FRAME];
+
+  // Store the best motion vector during motion search
+  int_mv best_mv;
+  // Store the second best motion vector during full-pixel motion search
+  int_mv second_best_mv;
+
+  // use default transform and skip transform type search for intra modes
+  int use_default_intra_tx_type;
+  // use default transform and skip transform type search for inter modes
+  int use_default_inter_tx_type;
+#if CONFIG_PVQ
+  int rate;
+  // 1 if neither AC nor DC is coded. Only used during RDO.
+  int pvq_skip[MAX_MB_PLANE];
+  PVQ_QUEUE *pvq_q;
+
+  // Storage for PVQ tx block encodings in a superblock.
+  // There can be max 16x16 of 4x4 blocks (and YUV) encode by PVQ
+  // 256 is the max # of 4x4 blocks in a SB (64x64), which comes from:
+  // 1) Since PVQ is applied to each trasnform-ed block
+  // 2) 4x4 is the smallest tx size in AV1
+  // 3) AV1 allows using smaller tx size than block (i.e. partition) size
+  // TODO(yushin) : The memory usage could be improved a lot, since this has
+  // storage for 10 bands and 128 coefficients for every 4x4 block,
+  PVQ_INFO pvq[MAX_PVQ_BLOCKS_IN_SB][MAX_MB_PLANE];
+  daala_enc_ctx daala_enc;
+  int pvq_speed;
+  int pvq_coded;  // Indicates whether pvq_info needs be stored to tokenize
+#endif
+#if CONFIG_DAALA_DIST
+  // Keep rate of each 4x4 block in the current macroblock during RDO
+  // This is needed when using the 8x8 Daala distortion metric during RDO,
+  // because it evaluates distortion in a different order than the underlying
+  // 4x4 blocks are coded.
+  int rate_4x4[256];
+#endif
+#if CONFIG_CFL
+  // Whether luma needs to be stored during RDO.
+  int cfl_store_y;
+#endif
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_BLOCK_H_

third_party/aom/av1/encoder/blockiness.c

@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "av1/common/common.h"
+#include "av1/common/filter.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+static int horizontal_filter(const uint8_t *s) {
+  return (s[1] - s[-2]) * 2 + (s[-1] - s[0]) * 6;
+}
+
+static int vertical_filter(const uint8_t *s, int p) {
+  return (s[p] - s[-2 * p]) * 2 + (s[-p] - s[0]) * 6;
+}
+
+static int variance(int sum, int sum_squared, int size) {
+  return sum_squared / size - (sum / size) * (sum / size);
+}
+// Calculate a blockiness level for a vertical block edge.
+// This function returns a new blockiness metric that's defined as
+
+//              p0 p1 p2 p3
+//              q0 q1 q2 q3
+// block edge ->
+//              r0 r1 r2 r3
+//              s0 s1 s2 s3
+
+// blockiness =  p0*-2+q0*6+r0*-6+s0*2 +
+//               p1*-2+q1*6+r1*-6+s1*2 +
+//               p2*-2+q2*6+r2*-6+s2*2 +
+//               p3*-2+q3*6+r3*-6+s3*2 ;
+
+// reconstructed_blockiness = abs(blockiness from reconstructed buffer -
+//                                blockiness from source buffer,0)
+//
+// I make the assumption that flat blocks are much more visible than high
+// contrast blocks. As such, I scale the result of the blockiness calc
+// by dividing the blockiness by the variance of the pixels on either side
+// of the edge as follows:
+// var_0 = (q0^2+q1^2+q2^2+q3^2) - ((q0 + q1 + q2 + q3) / 4 )^2
+// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
+// The returned blockiness is the scaled value
+// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
+static int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r,
+                               int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, s += sp, r += rp) {
+    s_blockiness += horizontal_filter(s);
+    r_blockiness += horizontal_filter(r);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-1];
+    sum_sq_1 += s[-1] * s[-1];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// Calculate a blockiness level for a horizontal block edge
+// same as above.
+static int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r,
+                                 int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, ++s, ++r) {
+    s_blockiness += vertical_filter(s, sp);
+    r_blockiness += vertical_filter(r, rp);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-sp];
+    sum_sq_1 += s[-sp] * s[-sp];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// This function returns the blockiness for the entire frame currently by
+// looking at all borders in steps of 4.
+double av1_get_blockiness(const unsigned char *img1, int img1_pitch,
+                          const unsigned char *img2, int img2_pitch, int width,
+                          int height) {
+  double blockiness = 0;
+  int i, j;
+  aom_clear_system_state();
+  for (i = 0; i < height;
+       i += 4, img1 += img1_pitch * 4, img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4) {
+      if (i > 0 && i < height && j > 0 && j < width) {
+        blockiness +=
+            blockiness_vertical(img1 + j, img1_pitch, img2 + j, img2_pitch, 4);
+        blockiness += blockiness_horizontal(img1 + j, img1_pitch, img2 + j,
+                                            img2_pitch, 4);
+      }
+    }
+  }
+  blockiness /= width * height / 16;
+  return blockiness;
+}

third_party/aom/av1/encoder/context_tree.c

@@ -0,0 +1,331 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/encoder.h"
+
+static const BLOCK_SIZE square[MAX_SB_SIZE_LOG2 - 1] = {
+#if CONFIG_CB4X4
+  BLOCK_4X4,
+#endif
+  BLOCK_8X8,     BLOCK_16X16, BLOCK_32X32, BLOCK_64X64,
+#if CONFIG_EXT_PARTITION
+  BLOCK_128X128,
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static void alloc_mode_context(AV1_COMMON *cm, int num_4x4_blk,
+#if CONFIG_EXT_PARTITION_TYPES
+                               PARTITION_TYPE partition,
+#endif
+                               PICK_MODE_CONTEXT *ctx) {
+  const int num_blk = (num_4x4_blk < 4 ? 4 : num_4x4_blk);
+  const int num_pix = num_blk * tx_size_2d[0];
+  int i;
+#if CONFIG_CB4X4 && CONFIG_VAR_TX
+  ctx->num_4x4_blk = num_blk / 4;
+#else
+  ctx->num_4x4_blk = num_blk;
+#endif
+
+#if CONFIG_EXT_PARTITION_TYPES
+  ctx->partition = partition;
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VAR_TX
+    CHECK_MEM_ERROR(cm, ctx->blk_skip[i], aom_calloc(num_blk, sizeof(uint8_t)));
+#endif
+    CHECK_MEM_ERROR(cm, ctx->coeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->coeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->qcoeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->qcoeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->dqcoeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->dqcoeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->eobs[i],
+                    aom_memalign(32, num_blk * sizeof(*ctx->eobs[i])));
+#if CONFIG_LV_MAP
+    CHECK_MEM_ERROR(
+        cm, ctx->txb_entropy_ctx[i],
+        aom_memalign(32, num_blk * sizeof(*ctx->txb_entropy_ctx[i])));
+#endif
+
+#if CONFIG_PVQ
+    CHECK_MEM_ERROR(cm, ctx->pvq_ref_coeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->pvq_ref_coeff[i])));
+#endif
+  }
+
+#if CONFIG_PALETTE
+  if (cm->allow_screen_content_tools) {
+    for (i = 0; i < 2; ++i) {
+      CHECK_MEM_ERROR(
+          cm, ctx->color_index_map[i],
+          aom_memalign(32, num_pix * sizeof(*ctx->color_index_map[i])));
+    }
+  }
+#endif  // CONFIG_PALETTE
+}
+
+static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VAR_TX
+    aom_free(ctx->blk_skip[i]);
+    ctx->blk_skip[i] = 0;
+#endif
+    aom_free(ctx->coeff[i]);
+    ctx->coeff[i] = 0;
+    aom_free(ctx->qcoeff[i]);
+    ctx->qcoeff[i] = 0;
+    aom_free(ctx->dqcoeff[i]);
+    ctx->dqcoeff[i] = 0;
+#if CONFIG_PVQ
+    aom_free(ctx->pvq_ref_coeff[i]);
+    ctx->pvq_ref_coeff[i] = 0;
+#endif
+    aom_free(ctx->eobs[i]);
+    ctx->eobs[i] = 0;
+#if CONFIG_LV_MAP
+    aom_free(ctx->txb_entropy_ctx[i]);
+    ctx->txb_entropy_ctx[i] = 0;
+#endif
+  }
+
+#if CONFIG_PALETTE
+  for (i = 0; i < 2; ++i) {
+    aom_free(ctx->color_index_map[i]);
+    ctx->color_index_map[i] = 0;
+  }
+#endif  // CONFIG_PALETTE
+}
+
+static void alloc_tree_contexts(AV1_COMMON *cm, PC_TREE *tree,
+                                int num_4x4_blk) {
+#if CONFIG_EXT_PARTITION_TYPES
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_NONE, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->vertical[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->horizontal[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->vertical[1]);
+
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_A,
+                     &tree->horizontala[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_A,
+                     &tree->horizontala[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ_A,
+                     &tree->horizontala[2]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ_B,
+                     &tree->horizontalb[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_B,
+                     &tree->horizontalb[1]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_B,
+                     &tree->horizontalb[2]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_A,
+                     &tree->verticala[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_A,
+                     &tree->verticala[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT_A,
+                     &tree->verticala[2]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT_B,
+                     &tree->verticalb[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_B,
+                     &tree->verticalb[1]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_B,
+                     &tree->verticalb[2]);
+#ifdef CONFIG_SUPERTX
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ,
+                     &tree->horizontal_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT, &tree->vertical_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_SPLIT, &tree->split_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ_A,
+                     &tree->horizontala_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ_B,
+                     &tree->horizontalb_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT_A,
+                     &tree->verticala_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT_B,
+                     &tree->verticalb_supertx);
+#endif  // CONFIG_SUPERTX
+#else
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk / 2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, &tree->vertical[0]);
+#ifdef CONFIG_SUPERTX
+  alloc_mode_context(cm, num_4x4_blk, &tree->horizontal_supertx);
+  alloc_mode_context(cm, num_4x4_blk, &tree->vertical_supertx);
+  alloc_mode_context(cm, num_4x4_blk, &tree->split_supertx);
+#endif
+
+  if (num_4x4_blk > 4) {
+    alloc_mode_context(cm, num_4x4_blk / 2, &tree->horizontal[1]);
+    alloc_mode_context(cm, num_4x4_blk / 2, &tree->vertical[1]);
+  } else {
+    memset(&tree->horizontal[1], 0, sizeof(tree->horizontal[1]));
+    memset(&tree->vertical[1], 0, sizeof(tree->vertical[1]));
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+#if CONFIG_EXT_PARTITION_TYPES
+  int i;
+  for (i = 0; i < 3; i++) {
+    free_mode_context(&tree->horizontala[i]);
+    free_mode_context(&tree->horizontalb[i]);
+    free_mode_context(&tree->verticala[i]);
+    free_mode_context(&tree->verticalb[i]);
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
+#ifdef CONFIG_SUPERTX
+  free_mode_context(&tree->horizontal_supertx);
+  free_mode_context(&tree->vertical_supertx);
+  free_mode_context(&tree->split_supertx);
+#if CONFIG_EXT_PARTITION_TYPES
+  free_mode_context(&tree->horizontala_supertx);
+  free_mode_context(&tree->horizontalb_supertx);
+  free_mode_context(&tree->verticala_supertx);
+  free_mode_context(&tree->verticalb_supertx);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#endif  // CONFIG_SUPERTX
+}
+
+// This function sets up a tree of contexts such that at each square
+// partition level. There are contexts for none, horizontal, vertical, and
+// split.  Along with a block_size value and a selected block_size which
+// represents the state of our search.
+void av1_setup_pc_tree(AV1_COMMON *cm, ThreadData *td) {
+  int i, j;
+// TODO(jingning): The pc_tree allocation is redundant. We can take out all
+// the leaf nodes after cb4x4 mode is enabled.
+#if CONFIG_CB4X4
+#if CONFIG_EXT_PARTITION
+  const int tree_nodes_inc = 1024;
+#else
+  const int tree_nodes_inc = 256;
+#endif  // CONFIG_EXT_PARTITION
+  const int leaf_factor = 4;
+#else
+  const int tree_nodes_inc = 0;
+  const int leaf_factor = 1;
+#endif
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 256 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 64 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int pc_tree_index = 0;
+  PC_TREE *this_pc;
+  PICK_MODE_CONTEXT *this_leaf;
+  int square_index = 1;
+  int nodes;
+
+  aom_free(td->leaf_tree);
+  CHECK_MEM_ERROR(cm, td->leaf_tree,
+                  aom_calloc(leaf_nodes, sizeof(*td->leaf_tree)));
+  aom_free(td->pc_tree);
+  CHECK_MEM_ERROR(cm, td->pc_tree,
+                  aom_calloc(tree_nodes, sizeof(*td->pc_tree)));
+
+  this_pc = &td->pc_tree[0];
+  this_leaf = &td->leaf_tree[0];
+
+  // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
+  // context so we only need to allocate 1 for each 8x8 block.
+  for (i = 0; i < leaf_nodes; ++i) {
+#if CONFIG_EXT_PARTITION_TYPES
+    alloc_mode_context(cm, 4, PARTITION_NONE, &td->leaf_tree[i]);
+#else
+    alloc_mode_context(cm, 16, &td->leaf_tree[i]);
+#endif
+  }
+
+  // Sets up all the leaf nodes in the tree.
+  for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
+    PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+#if CONFIG_CB4X4
+    alloc_tree_contexts(cm, tree, 16);
+#else
+    alloc_tree_contexts(cm, tree, 4);
+#endif
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++) tree->leaf_split[j] = tree->leaf_split[0];
+  }
+
+  // Each node has 4 leaf nodes, fill each block_size level of the tree
+  // from leafs to the root.
+  for (nodes = leaf_nodes >> 2; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+#if CONFIG_CB4X4
+      alloc_tree_contexts(cm, tree, 16 << (2 * square_index));
+#else
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+#endif
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++) tree->split[j] = this_pc++;
+      ++pc_tree_index;
+    }
+    ++square_index;
+  }
+
+  // Set up the root node for the largest superblock size
+  i = MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2;
+  td->pc_root[i] = &td->pc_tree[tree_nodes - 1];
+  td->pc_root[i]->none.best_mode_index = 2;
+  // Set up the root nodes for the rest of the possible superblock sizes
+  while (--i >= 0) {
+    td->pc_root[i] = td->pc_root[i + 1]->split[0];
+    td->pc_root[i]->none.best_mode_index = 2;
+  }
+}
+
+void av1_free_pc_tree(ThreadData *td) {
+#if CONFIG_CB4X4
+#if CONFIG_EXT_PARTITION
+  const int tree_nodes_inc = 1024;
+#else
+  const int tree_nodes_inc = 256;
+#endif  // CONFIG_EXT_PARTITION
+  const int leaf_factor = 4;
+#else
+  const int tree_nodes_inc = 0;
+  const int leaf_factor = 1;
+#endif
+
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 256 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 64 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int i;
+
+  // Set up all 4x4 mode contexts
+  for (i = 0; i < leaf_nodes; ++i) free_mode_context(&td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (i = 0; i < tree_nodes; ++i) free_tree_contexts(&td->pc_tree[i]);
+
+  aom_free(td->pc_tree);
+  td->pc_tree = NULL;
+  aom_free(td->leaf_tree);
+  td->leaf_tree = NULL;
+}

third_party/aom/av1/encoder/context_tree.h

@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_CONTEXT_TREE_H_
+#define AV1_ENCODER_CONTEXT_TREE_H_
+
+#include "av1/common/blockd.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1_COMP;
+struct AV1Common;
+struct ThreadData;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  MB_MODE_INFO_EXT mbmi_ext;
+#if CONFIG_PALETTE
+  uint8_t *color_index_map[2];
+#endif  // CONFIG_PALETTE
+#if CONFIG_VAR_TX
+  uint8_t *blk_skip[MAX_MB_PLANE];
+#endif
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  tran_low_t *coeff[MAX_MB_PLANE];
+  tran_low_t *qcoeff[MAX_MB_PLANE];
+  tran_low_t *dqcoeff[MAX_MB_PLANE];
+#if CONFIG_PVQ
+  tran_low_t *pvq_ref_coeff[MAX_MB_PLANE];
+#endif
+  uint16_t *eobs[MAX_MB_PLANE];
+#if CONFIG_LV_MAP
+  uint8_t *txb_entropy_ctx[MAX_MB_PLANE];
+#endif
+
+  int num_4x4_blk;
+  int skip;
+  int pred_pixel_ready;
+  // For current partition, only if all Y, U, and V transform blocks'
+  // coefficients are quantized to 0, skippable is set to 0.
+  int skippable;
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+
+  // TODO(jingning) Use RD_COST struct here instead. This involves a boarder
+  // scope of refactoring.
+  int rate;
+  int64_t dist;
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[TOTAL_REFS_PER_FRAME];
+  InterpFilter pred_interp_filter;
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPE partition;
+#endif
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+#if CONFIG_EXT_PARTITION_TYPES
+  PICK_MODE_CONTEXT horizontala[3];
+  PICK_MODE_CONTEXT horizontalb[3];
+  PICK_MODE_CONTEXT verticala[3];
+  PICK_MODE_CONTEXT verticalb[3];
+#endif
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+#ifdef CONFIG_SUPERTX
+  PICK_MODE_CONTEXT horizontal_supertx;
+  PICK_MODE_CONTEXT vertical_supertx;
+  PICK_MODE_CONTEXT split_supertx;
+#if CONFIG_EXT_PARTITION_TYPES
+  PICK_MODE_CONTEXT horizontala_supertx;
+  PICK_MODE_CONTEXT horizontalb_supertx;
+  PICK_MODE_CONTEXT verticala_supertx;
+  PICK_MODE_CONTEXT verticalb_supertx;
+#endif
+#endif
+} PC_TREE;
+
+void av1_setup_pc_tree(struct AV1Common *cm, struct ThreadData *td);
+void av1_free_pc_tree(struct ThreadData *td);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_CONTEXT_TREE_H_ */

third_party/aom/av1/encoder/corner_detect.c

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "third_party/fastfeat/fast.h"
+
+#include "av1/encoder/corner_detect.h"
+
+// Fast_9 wrapper
+#define FAST_BARRIER 18
+int fast_corner_detect(unsigned char *buf, int width, int height, int stride,
+                       int *points, int max_points) {
+  int num_points;
+  xy *const frm_corners_xy = fast9_detect_nonmax(buf, width, height, stride,
+                                                 FAST_BARRIER, &num_points);
+  num_points = (num_points <= max_points ? num_points : max_points);
+  if (num_points > 0 && frm_corners_xy) {
+    memcpy(points, frm_corners_xy, sizeof(*frm_corners_xy) * num_points);
+    free(frm_corners_xy);
+    return num_points;
+  }
+  free(frm_corners_xy);
+  return 0;
+}

third_party/aom/av1/encoder/corner_detect.h

@@ -0,0 +1,22 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_CORNER_DETECT_H_
+#define AV1_ENCODER_CORNER_DETECT_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+
+int fast_corner_detect(unsigned char *buf, int width, int height, int stride,
+                       int *points, int max_points);
+
+#endif  // AV1_ENCODER_CORNER_DETECT_H_

third_party/aom/av1/encoder/corner_match.c

@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+
+#include "av1/encoder/corner_match.h"
+
+#define MATCH_SZ 13
+#define MATCH_SZ_BY2 ((MATCH_SZ - 1) / 2)
+#define MATCH_SZ_SQ (MATCH_SZ * MATCH_SZ)
+#define SEARCH_SZ 9
+#define SEARCH_SZ_BY2 ((SEARCH_SZ - 1) / 2)
+
+#define THRESHOLD_NCC 0.75
+
+/* Compute var(im) * MATCH_SZ_SQ over a MATCH_SZ by MATCH_SZ window of im,
+   centered at (x, y).
+*/
+static double compute_variance(unsigned char *im, int stride, int x, int y) {
+  int sum = 0.0;
+  int sumsq = 0.0;
+  int var;
+  int i, j;
+  for (i = 0; i < MATCH_SZ; ++i)
+    for (j = 0; j < MATCH_SZ; ++j) {
+      sum += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
+      sumsq += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)] *
+               im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
+    }
+  var = sumsq * MATCH_SZ_SQ - sum * sum;
+  return (double)var;
+}
+
+/* Compute corr(im1, im2) * MATCH_SZ * stddev(im1), where the
+   correlation/standard deviation are taken over MATCH_SZ by MATCH_SZ windows
+   of each image, centered at (x1, y1) and (x2, y2) respectively.
+*/
+static double compute_cross_correlation(unsigned char *im1, int stride1, int x1,
+                                        int y1, unsigned char *im2, int stride2,
+                                        int x2, int y2) {
+  int v1, v2;
+  int sum1 = 0;
+  int sum2 = 0;
+  int sumsq2 = 0;
+  int cross = 0;
+  int var2, cov;
+  int i, j;
+  for (i = 0; i < MATCH_SZ; ++i)
+    for (j = 0; j < MATCH_SZ; ++j) {
+      v1 = im1[(i + y1 - MATCH_SZ_BY2) * stride1 + (j + x1 - MATCH_SZ_BY2)];
+      v2 = im2[(i + y2 - MATCH_SZ_BY2) * stride2 + (j + x2 - MATCH_SZ_BY2)];
+      sum1 += v1;
+      sum2 += v2;
+      sumsq2 += v2 * v2;
+      cross += v1 * v2;
+    }
+  var2 = sumsq2 * MATCH_SZ_SQ - sum2 * sum2;
+  cov = cross * MATCH_SZ_SQ - sum1 * sum2;
+  return cov / sqrt((double)var2);
+}
+
+static int is_eligible_point(int pointx, int pointy, int width, int height) {
+  return (pointx >= MATCH_SZ_BY2 && pointy >= MATCH_SZ_BY2 &&
+          pointx + MATCH_SZ_BY2 < width && pointy + MATCH_SZ_BY2 < height);
+}
+
+static int is_eligible_distance(int point1x, int point1y, int point2x,
+                                int point2y, int width, int height) {
+  const int thresh = (width < height ? height : width) >> 4;
+  return ((point1x - point2x) * (point1x - point2x) +
+          (point1y - point2y) * (point1y - point2y)) <= thresh * thresh;
+}
+
+static void improve_correspondence(unsigned char *frm, unsigned char *ref,
+                                   int width, int height, int frm_stride,
+                                   int ref_stride,
+                                   Correspondence *correspondences,
+                                   int num_correspondences) {
+  int i;
+  for (i = 0; i < num_correspondences; ++i) {
+    int x, y, best_x = 0, best_y = 0;
+    double best_match_ncc = 0.0;
+    for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y) {
+      for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
+        double match_ncc;
+        if (!is_eligible_point(correspondences[i].rx + x,
+                               correspondences[i].ry + y, width, height))
+          continue;
+        if (!is_eligible_distance(correspondences[i].x, correspondences[i].y,
+                                  correspondences[i].rx + x,
+                                  correspondences[i].ry + y, width, height))
+          continue;
+        match_ncc = compute_cross_correlation(
+            frm, frm_stride, correspondences[i].x, correspondences[i].y, ref,
+            ref_stride, correspondences[i].rx + x, correspondences[i].ry + y);
+        if (match_ncc > best_match_ncc) {
+          best_match_ncc = match_ncc;
+          best_y = y;
+          best_x = x;
+        }
+      }
+    }
+    correspondences[i].rx += best_x;
+    correspondences[i].ry += best_y;
+  }
+  for (i = 0; i < num_correspondences; ++i) {
+    int x, y, best_x = 0, best_y = 0;
+    double best_match_ncc = 0.0;
+    for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y)
+      for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
+        double match_ncc;
+        if (!is_eligible_point(correspondences[i].x + x,
+                               correspondences[i].y + y, width, height))
+          continue;
+        if (!is_eligible_distance(
+                correspondences[i].x + x, correspondences[i].y + y,
+                correspondences[i].rx, correspondences[i].ry, width, height))
+          continue;
+        match_ncc = compute_cross_correlation(
+            ref, ref_stride, correspondences[i].rx, correspondences[i].ry, frm,
+            frm_stride, correspondences[i].x + x, correspondences[i].y + y);
+        if (match_ncc > best_match_ncc) {
+          best_match_ncc = match_ncc;
+          best_y = y;
+          best_x = x;
+        }
+      }
+    correspondences[i].x += best_x;
+    correspondences[i].y += best_y;
+  }
+}
+
+int determine_correspondence(unsigned char *frm, int *frm_corners,
+                             int num_frm_corners, unsigned char *ref,
+                             int *ref_corners, int num_ref_corners, int width,
+                             int height, int frm_stride, int ref_stride,
+                             int *correspondence_pts) {
+  // TODO(sarahparker) Improve this to include 2-way match
+  int i, j;
+  Correspondence *correspondences = (Correspondence *)correspondence_pts;
+  int num_correspondences = 0;
+  for (i = 0; i < num_frm_corners; ++i) {
+    double best_match_ncc = 0.0;
+    double template_norm;
+    int best_match_j = -1;
+    if (!is_eligible_point(frm_corners[2 * i], frm_corners[2 * i + 1], width,
+                           height))
+      continue;
+    for (j = 0; j < num_ref_corners; ++j) {
+      double match_ncc;
+      if (!is_eligible_point(ref_corners[2 * j], ref_corners[2 * j + 1], width,
+                             height))
+        continue;
+      if (!is_eligible_distance(frm_corners[2 * i], frm_corners[2 * i + 1],
+                                ref_corners[2 * j], ref_corners[2 * j + 1],
+                                width, height))
+        continue;
+      match_ncc = compute_cross_correlation(
+          frm, frm_stride, frm_corners[2 * i], frm_corners[2 * i + 1], ref,
+          ref_stride, ref_corners[2 * j], ref_corners[2 * j + 1]);
+      if (match_ncc > best_match_ncc) {
+        best_match_ncc = match_ncc;
+        best_match_j = j;
+      }
+    }
+    // Note: We want to test if the best correlation is >= THRESHOLD_NCC,
+    // but need to account for the normalization in compute_cross_correlation.
+    template_norm = compute_variance(frm, frm_stride, frm_corners[2 * i],
+                                     frm_corners[2 * i + 1]);
+    if (best_match_ncc > THRESHOLD_NCC * sqrt(template_norm)) {
+      correspondences[num_correspondences].x = frm_corners[2 * i];
+      correspondences[num_correspondences].y = frm_corners[2 * i + 1];
+      correspondences[num_correspondences].rx = ref_corners[2 * best_match_j];
+      correspondences[num_correspondences].ry =
+          ref_corners[2 * best_match_j + 1];
+      num_correspondences++;
+    }
+  }
+  improve_correspondence(frm, ref, width, height, frm_stride, ref_stride,
+                         correspondences, num_correspondences);
+  return num_correspondences;
+}

third_party/aom/av1/encoder/corner_match.h

@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_ENCODER_CORNER_MATCH_H_
+#define AV1_ENCODER_CORNER_MATCH_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+
+typedef struct {
+  int x, y;
+  int rx, ry;
+} Correspondence;
+
+int determine_correspondence(unsigned char *frm, int *frm_corners,
+                             int num_frm_corners, unsigned char *ref,
+                             int *ref_corners, int num_ref_corners, int width,
+                             int height, int frm_stride, int ref_stride,
+                             int *correspondence_pts);
+
+#endif  // AV1_ENCODER_CORNER_MATCH_H_

third_party/aom/av1/encoder/cost.c

@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+
+#include "av1/encoder/cost.h"
+#include "av1/common/entropy.h"
+
+/* round(-log2(i/256.) * (1 << AV1_PROB_COST_SHIFT))
+   Begins with a bogus entry for simpler addressing. */
+const uint16_t av1_prob_cost[256] = {
+  4096, 4096, 3584, 3284, 3072, 2907, 2772, 2659, 2560, 2473, 2395, 2325, 2260,
+  2201, 2147, 2096, 2048, 2003, 1961, 1921, 1883, 1847, 1813, 1780, 1748, 1718,
+  1689, 1661, 1635, 1609, 1584, 1559, 1536, 1513, 1491, 1470, 1449, 1429, 1409,
+  1390, 1371, 1353, 1335, 1318, 1301, 1284, 1268, 1252, 1236, 1221, 1206, 1192,
+  1177, 1163, 1149, 1136, 1123, 1110, 1097, 1084, 1072, 1059, 1047, 1036, 1024,
+  1013, 1001, 990,  979,  968,  958,  947,  937,  927,  917,  907,  897,  887,
+  878,  868,  859,  850,  841,  832,  823,  814,  806,  797,  789,  780,  772,
+  764,  756,  748,  740,  732,  724,  717,  709,  702,  694,  687,  680,  673,
+  665,  658,  651,  644,  637,  631,  624,  617,  611,  604,  598,  591,  585,
+  578,  572,  566,  560,  554,  547,  541,  535,  530,  524,  518,  512,  506,
+  501,  495,  489,  484,  478,  473,  467,  462,  456,  451,  446,  441,  435,
+  430,  425,  420,  415,  410,  405,  400,  395,  390,  385,  380,  375,  371,
+  366,  361,  356,  352,  347,  343,  338,  333,  329,  324,  320,  316,  311,
+  307,  302,  298,  294,  289,  285,  281,  277,  273,  268,  264,  260,  256,
+  252,  248,  244,  240,  236,  232,  228,  224,  220,  216,  212,  209,  205,
+  201,  197,  194,  190,  186,  182,  179,  175,  171,  168,  164,  161,  157,
+  153,  150,  146,  143,  139,  136,  132,  129,  125,  122,  119,  115,  112,
+  109,  105,  102,  99,   95,   92,   89,   86,   82,   79,   76,   73,   70,
+  66,   63,   60,   57,   54,   51,   48,   45,   42,   38,   35,   32,   29,
+  26,   23,   20,   18,   15,   12,   9,    6,    3
+};
+
+static void cost(int *costs, aom_tree tree, const aom_prob *probs, int i,
+                 int c) {
+  const aom_prob prob = probs[i / 2];
+  int b;
+
+  assert(prob != 0);
+  for (b = 0; b <= 1; ++b) {
+    const int cc = c + av1_cost_bit(prob, b);
+    const aom_tree_index ii = tree[i + b];
+
+    if (ii <= 0)
+      costs[-ii] = cc;
+    else
+      cost(costs, tree, probs, ii, cc);
+  }
+}
+
+void av1_cost_tokens(int *costs, const aom_prob *probs, aom_tree tree) {
+  cost(costs, tree, probs, 0, 0);
+}
+
+void av1_cost_tokens_skip(int *costs, const aom_prob *probs, aom_tree tree) {
+  assert(tree[0] <= 0 && tree[1] > 0);
+
+  costs[-tree[0]] = av1_cost_bit(probs[0], 0);
+  cost(costs, tree, probs, 2, 0);
+}

third_party/aom/av1/encoder/cost.h

@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_COST_H_
+#define AV1_ENCODER_COST_H_
+
+#include "aom_dsp/prob.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const uint16_t av1_prob_cost[256];
+
+// The factor to scale from cost in bits to cost in av1_prob_cost units.
+#define AV1_PROB_COST_SHIFT 9
+
+#define av1_cost_zero(prob) (av1_prob_cost[prob])
+
+#define av1_cost_one(prob) av1_cost_zero(256 - (prob))
+
+#define av1_cost_bit(prob, bit) av1_cost_zero((bit) ? 256 - (prob) : (prob))
+
+// Cost of coding an n bit literal, using 128 (i.e. 50%) probability
+// for each bit.
+#define av1_cost_literal(n) ((n) * (1 << AV1_PROB_COST_SHIFT))
+
+static INLINE unsigned int cost_branch256(const unsigned int ct[2],
+                                          aom_prob p) {
+  return ct[0] * av1_cost_zero(p) + ct[1] * av1_cost_one(p);
+}
+
+static INLINE int treed_cost(aom_tree tree, const aom_prob *probs, int bits,
+                             int len) {
+  int cost = 0;
+  aom_tree_index i = 0;
+
+  do {
+    const int bit = (bits >> --len) & 1;
+    cost += av1_cost_bit(probs[i >> 1], bit);
+    i = tree[i + bit];
+  } while (len);
+
+  return cost;
+}
+
+void av1_cost_tokens(int *costs, const aom_prob *probs, aom_tree tree);
+void av1_cost_tokens_skip(int *costs, const aom_prob *probs, aom_tree tree);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_COST_H_

third_party/aom/av1/encoder/daala_compat_enc.c

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "encint.h"
+
+void od_encode_checkpoint(const daala_enc_ctx *enc, od_rollback_buffer *rbuf) {
+#if CONFIG_DAALA_EC
+  od_ec_enc_checkpoint(&rbuf->ec, &enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  OD_COPY(&rbuf->adapt, enc->state.adapt, 1);
+}
+
+void od_encode_rollback(daala_enc_ctx *enc, const od_rollback_buffer *rbuf) {
+#if CONFIG_DAALA_EC
+  od_ec_enc_rollback(&enc->w.ec, &rbuf->ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  OD_COPY(enc->state.adapt, &rbuf->adapt, 1);
+}

third_party/aom/av1/encoder/encint.h

@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+/* clang-format off */
+
+#if !defined(_encint_H)
+# define _encint_H (1)
+
+typedef struct daala_enc_ctx od_enc_ctx;
+typedef struct od_params_ctx od_params_ctx;
+typedef struct od_rollback_buffer od_rollback_buffer;
+
+# include "aom_dsp/entenc.h"
+# include "av1/common/odintrin.h"
+# include "av1/common/pvq_state.h"
+
+struct daala_enc_ctx{
+  /* Stores context-adaptive CDFs for PVQ. */
+  od_state state;
+  /* AOM entropy encoder. */
+  aom_writer w;
+  int use_activity_masking;
+  /* Mode of quantization matrice : FLAT (0) or HVS (1) */
+  int qm;
+  /*Normalized PVQ lambda for use where we've already performed
+     quantization.*/
+  double pvq_norm_lambda;
+  double pvq_norm_lambda_dc;
+};
+
+// from daalaenc.h
+/**The encoder context.*/
+typedef struct daala_enc_ctx daala_enc_ctx;
+
+/** Holds important encoder information so we can roll back decisions */
+struct od_rollback_buffer {
+  od_ec_enc ec;
+  od_adapt_ctx adapt;
+};
+
+void od_encode_checkpoint(const daala_enc_ctx *enc, od_rollback_buffer *rbuf);
+void od_encode_rollback(daala_enc_ctx *enc, const od_rollback_buffer *rbuf);
+
+#endif

third_party/aom/av1/encoder/encodeframe.h

@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEFRAME_H_
+#define AV1_ENCODER_ENCODEFRAME_H_
+
+#include "aom/aom_integer.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct macroblock;
+struct yv12_buffer_config;
+struct AV1_COMP;
+struct ThreadData;
+
+// Constants used in SOURCE_VAR_BASED_PARTITION
+#define VAR_HIST_MAX_BG_VAR 1000
+#define VAR_HIST_FACTOR 10
+#define VAR_HIST_BINS (VAR_HIST_MAX_BG_VAR / VAR_HIST_FACTOR + 1)
+#define VAR_HIST_LARGE_CUT_OFF 75
+#define VAR_HIST_SMALL_CUT_OFF 45
+
+void av1_setup_src_planes(struct macroblock *x,
+                          const struct yv12_buffer_config *src, int mi_row,
+                          int mi_col);
+
+void av1_encode_frame(struct AV1_COMP *cpi);
+
+void av1_init_tile_data(struct AV1_COMP *cpi);
+void av1_encode_tile(struct AV1_COMP *cpi, struct ThreadData *td, int tile_row,
+                     int tile_col);
+
+void av1_set_variance_partition_thresholds(struct AV1_COMP *cpi, int q);
+
+void av1_update_tx_type_count(const struct AV1Common *cm, MACROBLOCKD *xd,
+#if CONFIG_TXK_SEL
+                              int block, int plane,
+#endif
+                              BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              FRAME_COUNTS *counts);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEFRAME_H_

third_party/aom/av1/encoder/encodemb.h

@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEMB_H_
+#define AV1_ENCODER_ENCODEMB_H_
+
+#include "./aom_config.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct optimize_ctx {
+  ENTROPY_CONTEXT ta[MAX_MB_PLANE][2 * MAX_MIB_SIZE];
+  ENTROPY_CONTEXT tl[MAX_MB_PLANE][2 * MAX_MIB_SIZE];
+};
+
+struct encode_b_args {
+  AV1_COMMON *cm;
+  MACROBLOCK *x;
+  struct optimize_ctx *ctx;
+  int8_t *skip;
+  ENTROPY_CONTEXT *ta;
+  ENTROPY_CONTEXT *tl;
+  int8_t enable_optimize_b;
+};
+
+typedef enum AV1_XFORM_QUANT {
+  AV1_XFORM_QUANT_FP = 0,
+  AV1_XFORM_QUANT_B = 1,
+  AV1_XFORM_QUANT_DC = 2,
+  AV1_XFORM_QUANT_SKIP_QUANT,
+  AV1_XFORM_QUANT_TYPES,
+} AV1_XFORM_QUANT;
+
+void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row,
+                   int mi_col);
+#if CONFIG_SUPERTX
+void av1_encode_sb_supertx(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize);
+#endif  // CONFIG_SUPERTX
+void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize);
+void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
+                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
+                     TX_SIZE tx_size, int ctx, AV1_XFORM_QUANT xform_quant_idx);
+
+int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
+                   TX_SIZE tx_size, int ctx);
+
+void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
+                      int blk_col, int blk_row, TX_SIZE tx_size);
+
+void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+void av1_set_txb_context(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
+                         ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l);
+
+void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg);
+
+void av1_encode_intra_block_plane(AV1_COMMON *cm, MACROBLOCK *x,
+                                  BLOCK_SIZE bsize, int plane,
+                                  int enable_optimize_b, int mi_row,
+                                  int mi_col);
+
+#if CONFIG_PVQ
+PVQ_SKIP_TYPE av1_pvq_encode_helper(MACROBLOCK *x, tran_low_t *const coeff,
+                                    tran_low_t *ref_coeff,
+                                    tran_low_t *const dqcoeff, uint16_t *eob,
+                                    const int16_t *quant, int plane,
+                                    int tx_size, TX_TYPE tx_type, int *rate,
+                                    int speed, PVQ_INFO *pvq_info);
+
+void av1_store_pvq_enc_info(PVQ_INFO *pvq_info, int *qg, int *theta, int *k,
+                            od_coeff *y, int nb_bands, const int *off,
+                            int *size, int skip_rest, int skip_dir, int bs);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEMB_H_

third_party/aom/av1/encoder/encodemv.c

@@ -0,0 +1,497 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "av1/common/common.h"
+#include "av1/common/entropymode.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/subexp.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+static struct av1_token mv_joint_encodings[MV_JOINTS];
+static struct av1_token mv_class_encodings[MV_CLASSES];
+static struct av1_token mv_fp_encodings[MV_FP_SIZE];
+
+void av1_entropy_mv_init(void) {
+  av1_tokens_from_tree(mv_joint_encodings, av1_mv_joint_tree);
+  av1_tokens_from_tree(mv_class_encodings, av1_mv_class_tree);
+  av1_tokens_from_tree(mv_fp_encodings, av1_mv_fp_tree);
+}
+
+static void encode_mv_component(aom_writer *w, int comp, nmv_component *mvcomp,
+                                int usehp) {
+  int offset;
+  const int sign = comp < 0;
+  const int mag = sign ? -comp : comp;
+  const int mv_class = av1_get_mv_class(mag - 1, &offset);
+  const int d = offset >> 3;         // int mv data
+  const int fr = (offset >> 1) & 3;  // fractional mv data
+  const int hp = offset & 1;         // high precision mv data
+
+  assert(comp != 0);
+
+  // Sign
+  aom_write(w, sign, mvcomp->sign);
+
+// Class
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, mv_class, mvcomp->class_cdf, MV_CLASSES);
+#else
+  av1_write_token(w, av1_mv_class_tree, mvcomp->classes,
+                  &mv_class_encodings[mv_class]);
+#endif
+
+  // Integer bits
+  if (mv_class == MV_CLASS_0) {
+    aom_write(w, d, mvcomp->class0[0]);
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < n; ++i) aom_write(w, (d >> i) & 1, mvcomp->bits[i]);
+  }
+
+// Fractional bits
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(
+      w, fr, mv_class == MV_CLASS_0 ? mvcomp->class0_fp_cdf[d] : mvcomp->fp_cdf,
+      MV_FP_SIZE);
+#else
+  av1_write_token(w, av1_mv_fp_tree,
+                  mv_class == MV_CLASS_0 ? mvcomp->class0_fp[d] : mvcomp->fp,
+                  &mv_fp_encodings[fr]);
+#endif
+
+  // High precision bit
+  if (usehp)
+    aom_write(w, hp, mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
+}
+
+static void build_nmv_component_cost_table(int *mvcost,
+                                           const nmv_component *const mvcomp,
+                                           int usehp) {
+  int i, v;
+  int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
+  int bits_cost[MV_OFFSET_BITS][2];
+  int class0_fp_cost[CLASS0_SIZE][MV_FP_SIZE], fp_cost[MV_FP_SIZE];
+  int class0_hp_cost[2], hp_cost[2];
+
+  sign_cost[0] = av1_cost_zero(mvcomp->sign);
+  sign_cost[1] = av1_cost_one(mvcomp->sign);
+  av1_cost_tokens(class_cost, mvcomp->classes, av1_mv_class_tree);
+  av1_cost_tokens(class0_cost, mvcomp->class0, av1_mv_class0_tree);
+  for (i = 0; i < MV_OFFSET_BITS; ++i) {
+    bits_cost[i][0] = av1_cost_zero(mvcomp->bits[i]);
+    bits_cost[i][1] = av1_cost_one(mvcomp->bits[i]);
+  }
+
+  for (i = 0; i < CLASS0_SIZE; ++i)
+    av1_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], av1_mv_fp_tree);
+  av1_cost_tokens(fp_cost, mvcomp->fp, av1_mv_fp_tree);
+
+  if (usehp) {
+    class0_hp_cost[0] = av1_cost_zero(mvcomp->class0_hp);
+    class0_hp_cost[1] = av1_cost_one(mvcomp->class0_hp);
+    hp_cost[0] = av1_cost_zero(mvcomp->hp);
+    hp_cost[1] = av1_cost_one(mvcomp->hp);
+  }
+  mvcost[0] = 0;
+  for (v = 1; v <= MV_MAX; ++v) {
+    int z, c, o, d, e, f, cost = 0;
+    z = v - 1;
+    c = av1_get_mv_class(z, &o);
+    cost += class_cost[c];
+    d = (o >> 3);     /* int mv data */
+    f = (o >> 1) & 3; /* fractional pel mv data */
+    e = (o & 1);      /* high precision mv data */
+    if (c == MV_CLASS_0) {
+      cost += class0_cost[d];
+    } else {
+      const int b = c + CLASS0_BITS - 1; /* number of bits */
+      for (i = 0; i < b; ++i) cost += bits_cost[i][((d >> i) & 1)];
+    }
+    if (c == MV_CLASS_0) {
+      cost += class0_fp_cost[d][f];
+    } else {
+      cost += fp_cost[f];
+    }
+    if (usehp) {
+      if (c == MV_CLASS_0) {
+        cost += class0_hp_cost[e];
+      } else {
+        cost += hp_cost[e];
+      }
+    }
+    mvcost[v] = cost + sign_cost[0];
+    mvcost[-v] = cost + sign_cost[1];
+  }
+}
+
+static void update_mv(aom_writer *w, const unsigned int ct[2], aom_prob *cur_p,
+                      aom_prob upd_p) {
+  (void)upd_p;
+#if CONFIG_TILE_GROUPS
+  // Just use the default maximum number of tile groups to avoid passing in the
+  // actual
+  // number
+  av1_cond_prob_diff_update(w, cur_p, ct, DEFAULT_MAX_NUM_TG);
+#else
+  av1_cond_prob_diff_update(w, cur_p, ct, 1);
+#endif
+}
+
+#if !CONFIG_EC_ADAPT
+static void write_mv_update(const aom_tree_index *tree,
+                            aom_prob probs[/*n - 1*/],
+                            const unsigned int counts[/*n - 1*/], int n,
+                            aom_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  av1_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    update_mv(w, branch_ct[i], &probs[i], MV_UPDATE_PROB);
+}
+#endif
+
+void av1_write_nmv_probs(AV1_COMMON *cm, int usehp, aom_writer *w,
+                         nmv_context_counts *const nmv_counts) {
+  int i;
+#if CONFIG_REF_MV
+  int nmv_ctx = 0;
+  for (nmv_ctx = 0; nmv_ctx < NMV_CONTEXTS; ++nmv_ctx) {
+    nmv_context *const mvc = &cm->fc->nmvc[nmv_ctx];
+    nmv_context_counts *const counts = &nmv_counts[nmv_ctx];
+#if !CONFIG_EC_ADAPT
+    write_mv_update(av1_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS,
+                    w);
+
+    for (i = 0; i < 2; ++i) {
+      int j;
+      nmv_component *comp = &mvc->comps[i];
+      nmv_component_counts *comp_counts = &counts->comps[i];
+
+      update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+      write_mv_update(av1_mv_class_tree, comp->classes, comp_counts->classes,
+                      MV_CLASSES, w);
+      write_mv_update(av1_mv_class0_tree, comp->class0, comp_counts->class0,
+                      CLASS0_SIZE, w);
+      for (j = 0; j < MV_OFFSET_BITS; ++j)
+        update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+    }
+
+    for (i = 0; i < 2; ++i) {
+      int j;
+      for (j = 0; j < CLASS0_SIZE; ++j)
+        write_mv_update(av1_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                        counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+
+      write_mv_update(av1_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                      MV_FP_SIZE, w);
+    }
+#endif
+
+    if (usehp) {
+      for (i = 0; i < 2; ++i) {
+        update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                  MV_UPDATE_PROB);
+        update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+      }
+    }
+  }
+#else
+  nmv_context *const mvc = &cm->fc->nmvc;
+  nmv_context_counts *const counts = nmv_counts;
+
+#if !CONFIG_EC_ADAPT
+  write_mv_update(av1_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS, w);
+
+  for (i = 0; i < 2; ++i) {
+    int j;
+    nmv_component *comp = &mvc->comps[i];
+    nmv_component_counts *comp_counts = &counts->comps[i];
+
+    update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+    write_mv_update(av1_mv_class_tree, comp->classes, comp_counts->classes,
+                    MV_CLASSES, w);
+    write_mv_update(av1_mv_class0_tree, comp->class0, comp_counts->class0,
+                    CLASS0_SIZE, w);
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    int j;
+    for (j = 0; j < CLASS0_SIZE; ++j) {
+      write_mv_update(av1_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                      counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+    }
+    write_mv_update(av1_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                    MV_FP_SIZE, w);
+  }
+#endif  // !CONFIG_EC_ADAPT
+
+  if (usehp) {
+    for (i = 0; i < 2; ++i) {
+      update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                MV_UPDATE_PROB);
+      update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+    }
+  }
+#endif
+}
+
+void av1_encode_mv(AV1_COMP *cpi, aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx, int usehp) {
+  const MV diff = { mv->row - ref->row, mv->col - ref->col };
+  const MV_JOINT_TYPE j = av1_get_mv_joint(&diff);
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, j, mvctx->joint_cdf, MV_JOINTS);
+#else
+  av1_write_token(w, av1_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+#endif
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);
+
+  // If auto_mv_step_size is enabled then keep track of the largest
+  // motion vector component used.
+  if (cpi->sf.mv.auto_mv_step_size) {
+    unsigned int maxv = AOMMAX(abs(mv->row), abs(mv->col)) >> 3;
+    cpi->max_mv_magnitude = AOMMAX(maxv, cpi->max_mv_magnitude);
+  }
+}
+
+#if CONFIG_INTRABC
+void av1_encode_dv(aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx) {
+  const MV diff = { mv->row - ref->row, mv->col - ref->col };
+  const MV_JOINT_TYPE j = av1_get_mv_joint(&diff);
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, j, mvctx->joint_cdf, MV_JOINTS);
+#else
+  av1_write_token(w, av1_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+#endif
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], 0);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], 0);
+}
+#endif  // CONFIG_INTRABC
+
+void av1_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context *ctx, int usehp) {
+  av1_cost_tokens(mvjoint, ctx->joints, av1_mv_joint_tree);
+  build_nmv_component_cost_table(mvcost[0], &ctx->comps[0], usehp);
+  build_nmv_component_cost_table(mvcost[1], &ctx->comps[1], usehp);
+}
+
+#if CONFIG_EXT_INTER
+static void inc_mvs(const MB_MODE_INFO *mbmi, const MB_MODE_INFO_EXT *mbmi_ext,
+                    const int_mv mvs[2],
+#if CONFIG_REF_MV
+                    const int_mv pred_mvs[2],
+#endif
+                    nmv_context_counts *nmv_counts) {
+  int i;
+  PREDICTION_MODE mode = mbmi->mode;
+#if !CONFIG_REF_MV
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  if (mode == NEWMV || mode == NEW_NEWMV) {
+    for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+      const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0].as_mv;
+      const MV diff = { mvs[i].as_mv.row - ref->row,
+                        mvs[i].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx =
+          av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                      mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+      nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+      (void)pred_mvs;
+#endif
+      av1_inc_mv(&diff, counts, 1);
+    }
+  } else if (mode == NEAREST_NEWMV || mode == NEAR_NEWMV) {
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0].as_mv;
+    const MV diff = { mvs[1].as_mv.row - ref->row,
+                      mvs[1].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  } else if (mode == NEW_NEARESTMV || mode == NEW_NEARMV) {
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0].as_mv;
+    const MV diff = { mvs[0].as_mv.row - ref->row,
+                      mvs[0].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+
+static void inc_mvs_sub8x8(const MODE_INFO *mi, int block, const int_mv mvs[2],
+#if CONFIG_REF_MV
+                           const MB_MODE_INFO_EXT *mbmi_ext,
+#endif
+                           nmv_context_counts *nmv_counts) {
+  int i;
+  PREDICTION_MODE mode = mi->bmi[block].as_mode;
+#if CONFIG_REF_MV
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+#else
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  if (mode == NEWMV || mode == NEW_NEWMV) {
+    for (i = 0; i < 1 + has_second_ref(&mi->mbmi); ++i) {
+      const MV *ref = &mi->bmi[block].ref_mv[i].as_mv;
+      const MV diff = { mvs[i].as_mv.row - ref->row,
+                        mvs[i].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx =
+          av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                      mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+      nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+      av1_inc_mv(&diff, counts, 1);
+    }
+  } else if (mode == NEAREST_NEWMV || mode == NEAR_NEWMV) {
+    const MV *ref = &mi->bmi[block].ref_mv[1].as_mv;
+    const MV diff = { mvs[1].as_mv.row - ref->row,
+                      mvs[1].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  } else if (mode == NEW_NEARESTMV || mode == NEW_NEARMV) {
+    const MV *ref = &mi->bmi[block].ref_mv[0].as_mv;
+    const MV diff = { mvs[0].as_mv.row - ref->row,
+                      mvs[0].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+#else
+static void inc_mvs(const MB_MODE_INFO *mbmi, const MB_MODE_INFO_EXT *mbmi_ext,
+                    const int_mv mvs[2],
+#if CONFIG_REF_MV
+                    const int_mv pred_mvs[2],
+#endif
+                    nmv_context_counts *nmv_counts) {
+  int i;
+#if !CONFIG_REF_MV
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+    const MV *ref = &pred_mvs[i].as_mv;
+#else
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0].as_mv;
+#endif
+    const MV diff = { mvs[i].as_mv.row - ref->row,
+                      mvs[i].as_mv.col - ref->col };
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+#endif  // CONFIG_EXT_INTER
+
+void av1_update_mv_count(ThreadData *td) {
+  const MACROBLOCKD *xd = &td->mb.e_mbd;
+  const MODE_INFO *mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MB_MODE_INFO_EXT *mbmi_ext = td->mb.mbmi_ext;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  if (mbmi->sb_type < BLOCK_8X8 && !unify_bsize) {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[mbmi->sb_type];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[mbmi->sb_type];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int i = idy * 2 + idx;
+
+#if CONFIG_EXT_INTER
+        if (have_newmv_in_inter_mode(mi->bmi[i].as_mode))
+          inc_mvs_sub8x8(mi, i, mi->bmi[i].as_mv,
+#if CONFIG_REF_MV
+                         mbmi_ext, td->counts->mv);
+#else
+                         &td->counts->mv);
+#endif
+#else
+        if (mi->bmi[i].as_mode == NEWMV)
+          inc_mvs(mbmi, mbmi_ext, mi->bmi[i].as_mv,
+#if CONFIG_REF_MV
+                  mi->bmi[i].pred_mv, td->counts->mv);
+#else
+                  &td->counts->mv);
+#endif
+#endif  // CONFIG_EXT_INTER
+      }
+    }
+  } else {
+#if CONFIG_EXT_INTER
+    if (have_newmv_in_inter_mode(mbmi->mode))
+#else
+    if (mbmi->mode == NEWMV)
+#endif  // CONFIG_EXT_INTER
+      inc_mvs(mbmi, mbmi_ext, mbmi->mv,
+#if CONFIG_REF_MV
+              mbmi->pred_mv, td->counts->mv);
+#else
+              &td->counts->mv);
+#endif
+  }
+}

third_party/aom/av1/encoder/encodemv.h

@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEMV_H_
+#define AV1_ENCODER_ENCODEMV_H_
+
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_entropy_mv_init(void);
+
+void av1_write_nmv_probs(AV1_COMMON *cm, int usehp, aom_writer *w,
+                         nmv_context_counts *const counts);
+
+void av1_encode_mv(AV1_COMP *cpi, aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx, int usehp);
+
+void av1_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context *mvctx, int usehp);
+
+void av1_update_mv_count(ThreadData *td);
+
+#if CONFIG_INTRABC
+void av1_encode_dv(aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx);
+#endif  // CONFIG_INTRABC
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEMV_H_

third_party/aom/av1/encoder/encoder.h

@@ -0,0 +1,883 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODER_H_
+#define AV1_ENCODER_ENCODER_H_
+
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "aom/aomcx.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/thread_common.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#if CONFIG_ANS
+#include "aom_dsp/ans.h"
+#include "aom_dsp/buf_ans.h"
+#endif
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/lookahead.h"
+#include "av1/encoder/mbgraph.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/speed_features.h"
+#include "av1/encoder/tokenize.h"
+#include "av1/encoder/variance_tree.h"
+#if CONFIG_XIPHRC
+#include "av1/encoder/ratectrl_xiph.h"
+#endif
+
+#if CONFIG_INTERNAL_STATS
+#include "aom_dsp/ssim.h"
+#endif
+#include "aom_dsp/variance.h"
+#include "aom/internal/aom_codec_internal.h"
+#include "aom_util/aom_thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  int nmvjointcost[MV_JOINTS];
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+
+#if CONFIG_REF_MV
+  int nmv_vec_cost[NMV_CONTEXTS][MV_JOINTS];
+  int nmv_costs[NMV_CONTEXTS][2][MV_VALS];
+  int nmv_costs_hp[NMV_CONTEXTS][2][MV_VALS];
+#endif
+
+  // 0 = Intra, Last, GF, ARF
+  signed char last_ref_lf_deltas[TOTAL_REFS_PER_FRAME];
+  // 0 = ZERO_MV, MV
+  signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
+
+  FRAME_CONTEXT fc;
+} CODING_CONTEXT;
+
+typedef enum {
+  // regular inter frame
+  REGULAR_FRAME = 0,
+  // alternate reference frame
+  ARF_FRAME = 1,
+  // overlay frame
+  OVERLAY_FRAME = 2,
+  // golden frame
+  GLD_FRAME = 3,
+#if CONFIG_EXT_REFS
+  // backward reference frame
+  BRF_FRAME = 4,
+  // extra alternate reference frame
+  EXT_ARF_FRAME = 5
+#endif
+} FRAME_CONTEXT_INDEX;
+
+typedef enum {
+  NORMAL = 0,
+  FOURFIVE = 1,
+  THREEFIVE = 2,
+  ONETWO = 3
+} AOM_SCALING;
+
+typedef enum {
+  // Good Quality Fast Encoding. The encoder balances quality with the amount of
+  // time it takes to encode the output. Speed setting controls how fast.
+  GOOD
+} MODE;
+
+typedef enum {
+  FRAMEFLAGS_KEY = 1 << 0,
+  FRAMEFLAGS_GOLDEN = 1 << 1,
+#if CONFIG_EXT_REFS
+  FRAMEFLAGS_BWDREF = 1 << 2,
+  FRAMEFLAGS_ALTREF = 1 << 3,
+#else
+  FRAMEFLAGS_ALTREF = 1 << 2,
+#endif  // CONFIG_EXT_REFS
+} FRAMETYPE_FLAGS;
+
+typedef enum {
+  NO_AQ = 0,
+  VARIANCE_AQ = 1,
+  COMPLEXITY_AQ = 2,
+  CYCLIC_REFRESH_AQ = 3,
+#if CONFIG_DELTA_Q && !CONFIG_EXT_DELTA_Q
+  DELTA_AQ = 4,
+#endif
+  AQ_MODE_COUNT  // This should always be the last member of the enum
+} AQ_MODE;
+#if CONFIG_EXT_DELTA_Q
+typedef enum {
+  NO_DELTA_Q = 0,
+  DELTA_Q_ONLY = 1,
+  DELTA_Q_LF = 2,
+  DELTAQ_MODE_COUNT  // This should always be the last member of the enum
+} DELTAQ_MODE;
+#endif
+typedef enum {
+  RESIZE_NONE = 0,    // No frame resizing allowed.
+  RESIZE_FIXED = 1,   // All frames are coded at the specified dimension.
+  RESIZE_DYNAMIC = 2  // Coded size of each frame is determined by the codec.
+} RESIZE_TYPE;
+
+typedef struct AV1EncoderConfig {
+  BITSTREAM_PROFILE profile;
+  aom_bit_depth_t bit_depth;     // Codec bit-depth.
+  int width;                     // width of data passed to the compressor
+  int height;                    // height of data passed to the compressor
+  unsigned int input_bit_depth;  // Input bit depth.
+  double init_framerate;         // set to passed in framerate
+  int64_t target_bandwidth;      // bandwidth to be used in bits per second
+
+  int noise_sensitivity;  // pre processing blur: recommendation 0
+  int sharpness;          // sharpening output: recommendation 0:
+  int speed;
+  // maximum allowed bitrate for any intra frame in % of bitrate target.
+  unsigned int rc_max_intra_bitrate_pct;
+  // maximum allowed bitrate for any inter frame in % of bitrate target.
+  unsigned int rc_max_inter_bitrate_pct;
+  // percent of rate boost for golden frame in CBR mode.
+  unsigned int gf_cbr_boost_pct;
+
+  MODE mode;
+  int pass;
+
+  // Key Framing Operations
+  int auto_key;  // autodetect cut scenes and set the keyframes
+  int key_freq;  // maximum distance to key frame.
+
+  int lag_in_frames;  // how many frames lag before we start encoding
+
+  // ----------------------------------------------------------------
+  // DATARATE CONTROL OPTIONS
+
+  // vbr, cbr, constrained quality or constant quality
+  enum aom_rc_mode rc_mode;
+
+  // buffer targeting aggressiveness
+  int under_shoot_pct;
+  int over_shoot_pct;
+
+  // buffering parameters
+  int64_t starting_buffer_level_ms;
+  int64_t optimal_buffer_level_ms;
+  int64_t maximum_buffer_size_ms;
+
+  // Frame drop threshold.
+  int drop_frames_water_mark;
+
+  // controlling quality
+  int fixed_q;
+  int worst_allowed_q;
+  int best_allowed_q;
+  int cq_level;
+  AQ_MODE aq_mode;  // Adaptive Quantization mode
+#if CONFIG_EXT_DELTA_Q
+  DELTAQ_MODE deltaq_mode;
+#endif
+#if CONFIG_AOM_QM
+  int using_qm;
+  int qm_minlevel;
+  int qm_maxlevel;
+#endif
+#if CONFIG_TILE_GROUPS
+  unsigned int num_tile_groups;
+  unsigned int mtu;
+#endif
+
+#if CONFIG_TEMPMV_SIGNALING
+  unsigned int disable_tempmv;
+#endif
+  // Internal frame size scaling.
+  RESIZE_TYPE resize_mode;
+  int scaled_frame_width;
+  int scaled_frame_height;
+
+  // Enable feature to reduce the frame quantization every x frames.
+  int frame_periodic_boost;
+
+  // two pass datarate control
+  int two_pass_vbrbias;  // two pass datarate control tweaks
+  int two_pass_vbrmin_section;
+  int two_pass_vbrmax_section;
+  // END DATARATE CONTROL OPTIONS
+  // ----------------------------------------------------------------
+
+  int enable_auto_arf;
+#if CONFIG_EXT_REFS
+  int enable_auto_brf;  // (b)ackward (r)ef (f)rame
+#endif                  // CONFIG_EXT_REFS
+
+  /* Bitfield defining the error resiliency features to enable.
+   * Can provide decodable frames after losses in previous
+   * frames and decodable partitions after losses in the same frame.
+   */
+  unsigned int error_resilient_mode;
+
+  /* Bitfield defining the parallel decoding mode where the
+   * decoding in successive frames may be conducted in parallel
+   * just by decoding the frame headers.
+   */
+  unsigned int frame_parallel_decoding_mode;
+
+  int arnr_max_frames;
+  int arnr_strength;
+
+  int min_gf_interval;
+  int max_gf_interval;
+
+  int tile_columns;
+  int tile_rows;
+#if CONFIG_DEPENDENT_HORZTILES
+  int dependent_horz_tiles;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  int loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  int max_threads;
+
+  aom_fixed_buf_t two_pass_stats_in;
+  struct aom_codec_pkt_list *output_pkt_list;
+
+#if CONFIG_FP_MB_STATS
+  aom_fixed_buf_t firstpass_mb_stats_in;
+#endif
+
+  aom_tune_metric tuning;
+  aom_tune_content content;
+#if CONFIG_HIGHBITDEPTH
+  int use_highbitdepth;
+#endif
+  aom_color_space_t color_space;
+  int color_range;
+  int render_width;
+  int render_height;
+
+#if CONFIG_EXT_PARTITION
+  aom_superblock_size_t superblock_size;
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  int ans_window_size_log2;
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+#if CONFIG_EXT_TILE
+  unsigned int tile_encoding_mode;
+#endif  // CONFIG_EXT_TILE
+
+  unsigned int motion_vector_unit_test;
+} AV1EncoderConfig;
+
+static INLINE int is_lossless_requested(const AV1EncoderConfig *cfg) {
+  return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
+}
+
+// TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
+typedef struct TileDataEnc {
+  TileInfo tile_info;
+  int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+  int mode_map[BLOCK_SIZES][MAX_MODES];
+  int m_search_count;
+  int ex_search_count;
+#if CONFIG_PVQ
+  PVQ_QUEUE pvq_q;
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  DECLARE_ALIGNED(16, FRAME_CONTEXT, tctx);
+#endif
+} TileDataEnc;
+
+typedef struct RD_COUNTS {
+  av1_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
+  int64_t comp_pred_diff[REFERENCE_MODES];
+#if CONFIG_GLOBAL_MOTION
+  // Stores number of 4x4 blocks using global motion per reference frame.
+  int global_motion_used[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_GLOBAL_MOTION
+} RD_COUNTS;
+
+typedef struct ThreadData {
+  MACROBLOCK mb;
+  RD_COUNTS rd_counts;
+  FRAME_COUNTS *counts;
+
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root[MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2 + 1];
+
+  VAR_TREE *var_tree;
+  VAR_TREE *var_root[MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2 + 1];
+} ThreadData;
+
+struct EncWorkerData;
+
+typedef struct ActiveMap {
+  int enabled;
+  int update;
+  unsigned char *map;
+} ActiveMap;
+
+#define NUM_STAT_TYPES 4  // types of stats: Y, U, V and ALL
+
+typedef struct IMAGE_STAT {
+  double stat[NUM_STAT_TYPES];
+  double worst;
+} ImageStat;
+
+#undef NUM_STAT_TYPES
+
+typedef struct {
+  int ref_count;
+  YV12_BUFFER_CONFIG buf;
+} EncRefCntBuffer;
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+typedef struct SUBFRAME_STATS {
+  av1_coeff_probs_model coef_probs_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+  av1_coeff_count coef_counts_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+  unsigned int eob_counts_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES][REF_TYPES]
+                             [COEF_BANDS][COEFF_CONTEXTS];
+  av1_coeff_probs_model enc_starting_coef_probs[TX_SIZES][PLANE_TYPES];
+} SUBFRAME_STATS;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+typedef struct TileBufferEnc {
+  uint8_t *data;
+  size_t size;
+} TileBufferEnc;
+
+typedef struct AV1_COMP {
+  QUANTS quants;
+  ThreadData td;
+  MB_MODE_INFO_EXT *mbmi_ext_base;
+  DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);   // 8: SIMD width
+  DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);  // 8: SIMD width
+#if CONFIG_NEW_QUANT
+  DECLARE_ALIGNED(16, dequant_val_type_nuq,
+                  y_dequant_val_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS]);
+  DECLARE_ALIGNED(16, dequant_val_type_nuq,
+                  uv_dequant_val_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS]);
+#endif  // CONFIG_NEW_QUANT
+  AV1_COMMON common;
+  AV1EncoderConfig oxcf;
+  struct lookahead_ctx *lookahead;
+  struct lookahead_entry *alt_ref_source;
+
+  YV12_BUFFER_CONFIG *source;
+  YV12_BUFFER_CONFIG *last_source;  // NULL for first frame and alt_ref frames
+  YV12_BUFFER_CONFIG *un_scaled_source;
+  YV12_BUFFER_CONFIG scaled_source;
+  YV12_BUFFER_CONFIG *unscaled_last_source;
+  YV12_BUFFER_CONFIG scaled_last_source;
+
+  // Up-sampled reference buffers
+  // NOTE(zoeliu): It is needed to allocate sufficient space to the up-sampled
+  // reference buffers, which should include the up-sampled version of all the
+  // possibly stored references plus the currently coded frame itself.
+  EncRefCntBuffer upsampled_ref_bufs[REF_FRAMES + 1];
+  int upsampled_ref_idx[REF_FRAMES + 1];
+
+  // For a still frame, this flag is set to 1 to skip partition search.
+  int partition_search_skippable_frame;
+
+  int scaled_ref_idx[TOTAL_REFS_PER_FRAME];
+#if CONFIG_EXT_REFS
+  int lst_fb_idxes[LAST_REF_FRAMES];
+#else
+  int lst_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  int gld_fb_idx;
+#if CONFIG_EXT_REFS
+  int bwd_fb_idx;  // BWD_REF_FRAME
+#endif             // CONFIG_EXT_REFS
+  int alt_fb_idx;
+
+  int last_show_frame_buf_idx;  // last show frame buffer index
+
+  int refresh_last_frame;
+  int refresh_golden_frame;
+#if CONFIG_EXT_REFS
+  int refresh_bwd_ref_frame;
+#endif  // CONFIG_EXT_REFS
+  int refresh_alt_ref_frame;
+
+  int ext_refresh_frame_flags_pending;
+  int ext_refresh_last_frame;
+  int ext_refresh_golden_frame;
+  int ext_refresh_alt_ref_frame;
+
+  int ext_refresh_frame_context_pending;
+  int ext_refresh_frame_context;
+
+  YV12_BUFFER_CONFIG last_frame_uf;
+#if CONFIG_LOOP_RESTORATION
+  YV12_BUFFER_CONFIG last_frame_db;
+  YV12_BUFFER_CONFIG trial_frame_rst;
+  uint8_t *extra_rstbuf;  // Extra buffers used in restoration search
+  RestorationInfo rst_search[MAX_MB_PLANE];  // Used for encoder side search
+#endif                                       // CONFIG_LOOP_RESTORATION
+
+  // Ambient reconstruction err target for force key frames
+  int64_t ambient_err;
+
+  RD_OPT rd;
+
+  CODING_CONTEXT coding_context;
+
+#if CONFIG_REF_MV
+  int nmv_costs[NMV_CONTEXTS][2][MV_VALS];
+  int nmv_costs_hp[NMV_CONTEXTS][2][MV_VALS];
+#endif
+
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+  int nmvsadcosts[2][MV_VALS];
+  int nmvsadcosts_hp[2][MV_VALS];
+
+  int64_t last_time_stamp_seen;
+  int64_t last_end_time_stamp_seen;
+  int64_t first_time_stamp_ever;
+
+  RATE_CONTROL rc;
+#if CONFIG_XIPHRC
+  od_rc_state od_rc;
+#endif
+  double framerate;
+
+  // NOTE(zoeliu): Any inter frame allows maximum of REF_FRAMES inter
+  // references; Plus the currently coded frame itself, it is needed to allocate
+  // sufficient space to the size of the maximum possible number of frames.
+  int interp_filter_selected[REF_FRAMES + 1][SWITCHABLE];
+
+  struct aom_codec_pkt_list *output_pkt_list;
+
+  MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
+  int mbgraph_n_frames;  // number of frames filled in the above
+  int static_mb_pct;     // % forced skip mbs by segmentation
+  int ref_frame_flags;
+
+  SPEED_FEATURES sf;
+
+  unsigned int max_mv_magnitude;
+  int mv_step_param;
+
+  int allow_comp_inter_inter;
+
+  uint8_t *segmentation_map;
+
+  CYCLIC_REFRESH *cyclic_refresh;
+  ActiveMap active_map;
+
+  fractional_mv_step_fp *find_fractional_mv_step;
+  av1_full_search_fn_t full_search_sad;  // It is currently unused.
+  av1_diamond_search_fn_t diamond_search_sad;
+  aom_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
+  uint64_t time_receive_data;
+  uint64_t time_compress_data;
+  uint64_t time_pick_lpf;
+  uint64_t time_encode_sb_row;
+
+#if CONFIG_FP_MB_STATS
+  int use_fp_mb_stats;
+#endif
+
+  TWO_PASS twopass;
+
+  YV12_BUFFER_CONFIG alt_ref_buffer;
+
+#if CONFIG_INTERNAL_STATS
+  unsigned int mode_chosen_counts[MAX_MODES];
+
+  int count;
+  uint64_t total_sq_error;
+  uint64_t total_samples;
+  ImageStat psnr;
+
+  double total_blockiness;
+  double worst_blockiness;
+
+  int bytes;
+  double summed_quality;
+  double summed_weights;
+  unsigned int tot_recode_hits;
+  double worst_ssim;
+
+  ImageStat fastssim;
+  ImageStat psnrhvs;
+
+  int b_calculate_blockiness;
+  int b_calculate_consistency;
+
+  double total_inconsistency;
+  double worst_consistency;
+  Ssimv *ssim_vars;
+  Metrics metrics;
+#endif
+  int b_calculate_psnr;
+
+  int droppable;
+
+  int initial_width;
+  int initial_height;
+  int initial_mbs;  // Number of MBs in the full-size frame; to be used to
+                    // normalize the firstpass stats. This will differ from the
+                    // number of MBs in the current frame when the frame is
+                    // scaled.
+
+  // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
+  DIFF *source_diff_var;
+  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
+  unsigned int source_var_thresh;
+  int frames_till_next_var_check;
+
+  int frame_flags;
+
+  search_site_config ss_cfg;
+
+  int mbmode_cost[BLOCK_SIZE_GROUPS][INTRA_MODES];
+#if CONFIG_REF_MV
+  int newmv_mode_cost[NEWMV_MODE_CONTEXTS][2];
+  int zeromv_mode_cost[ZEROMV_MODE_CONTEXTS][2];
+  int refmv_mode_cost[REFMV_MODE_CONTEXTS][2];
+  int drl_mode_cost0[DRL_MODE_CONTEXTS][2];
+#endif
+
+  unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
+#if CONFIG_EXT_INTER
+  unsigned int inter_compound_mode_cost[INTER_MODE_CONTEXTS]
+                                       [INTER_COMPOUND_MODES];
+  unsigned int interintra_mode_cost[BLOCK_SIZE_GROUPS][INTERINTRA_MODES];
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int motion_mode_cost[BLOCK_SIZES][MOTION_MODES];
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  int motion_mode_cost1[BLOCK_SIZES][2];
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int intra_uv_mode_cost[INTRA_MODES][INTRA_MODES];
+  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
+#if CONFIG_EXT_PARTITION_TYPES
+  int partition_cost[PARTITION_CONTEXTS + CONFIG_UNPOISON_PARTITION_CTX]
+                    [EXT_PARTITION_TYPES];
+#else
+  int partition_cost[PARTITION_CONTEXTS + CONFIG_UNPOISON_PARTITION_CTX]
+                    [PARTITION_TYPES];
+#endif
+#if CONFIG_PALETTE
+  int palette_y_size_cost[PALETTE_BLOCK_SIZES][PALETTE_SIZES];
+  int palette_uv_size_cost[PALETTE_BLOCK_SIZES][PALETTE_SIZES];
+  int palette_y_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
+                          [PALETTE_COLORS];
+  int palette_uv_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
+                           [PALETTE_COLORS];
+#endif  // CONFIG_PALETTE
+  int tx_size_cost[TX_SIZES - 1][TX_SIZE_CONTEXTS][TX_SIZES];
+#if CONFIG_EXT_TX
+  int inter_tx_type_costs[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES];
+  int intra_tx_type_costs[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                         [TX_TYPES];
+#else
+  int intra_tx_type_costs[EXT_TX_SIZES][TX_TYPES][TX_TYPES];
+  int inter_tx_type_costs[EXT_TX_SIZES][TX_TYPES];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  int intra_filter_cost[INTRA_FILTERS + 1][INTRA_FILTERS];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_LOOP_RESTORATION
+  int switchable_restore_cost[RESTORE_SWITCHABLE_TYPES];
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_GLOBAL_MOTION
+  int gmtype_cost[TRANS_TYPES];
+  int gmparams_cost[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_GLOBAL_MOTION
+
+  int multi_arf_allowed;
+  int multi_arf_enabled;
+  int multi_arf_last_grp_enabled;
+
+  TileDataEnc *tile_data;
+  int allocated_tiles;  // Keep track of memory allocated for tiles.
+
+  TOKENEXTRA *tile_tok[MAX_TILE_ROWS][MAX_TILE_COLS];
+  unsigned int tok_count[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  TileBufferEnc tile_buffers[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  int resize_pending;
+  int resize_state;
+  int resize_scale_num;
+  int resize_scale_den;
+  int resize_avg_qp;
+  int resize_buffer_underflow;
+  int resize_count;
+
+  // VAR_BASED_PARTITION thresholds
+  // 0 - threshold_128x128;
+  // 1 - threshold_64x64;
+  // 2 - threshold_32x32;
+  // 3 - threshold_16x16;
+  // 4 - threshold_8x8;
+  int64_t vbp_thresholds[5];
+  int64_t vbp_threshold_minmax;
+  int64_t vbp_threshold_sad;
+  BLOCK_SIZE vbp_bsize_min;
+
+  // VARIANCE_AQ segment map refresh
+  int vaq_refresh;
+
+  // Multi-threading
+  int num_workers;
+  AVxWorker *workers;
+  struct EncWorkerData *tile_thr_data;
+  AV1LfSync lf_row_sync;
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  SUBFRAME_STATS subframe_stats;
+  // TODO(yaowu): minimize the size of count buffers
+  SUBFRAME_STATS wholeframe_stats;
+  av1_coeff_stats branch_ct_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#if CONFIG_ANS
+  struct BufAnsCoder buf_ans;
+#endif
+#if CONFIG_EXT_REFS
+  int refresh_frame_mask;
+  int existing_fb_idx_to_show;
+  int is_arf_filter_off[MAX_EXT_ARFS + 1];
+  int num_extra_arfs;
+  int arf_map[MAX_EXT_ARFS + 1];
+#endif  // CONFIG_EXT_REFS
+#if CONFIG_GLOBAL_MOTION
+  int global_motion_search_done;
+#endif
+#if CONFIG_REFERENCE_BUFFER
+  SequenceHeader seq_params;
+#endif
+#if CONFIG_LV_MAP
+  tran_low_t *tcoeff_buf[MAX_MB_PLANE];
+#endif
+} AV1_COMP;
+
+void av1_initialize_enc(void);
+
+struct AV1_COMP *av1_create_compressor(AV1EncoderConfig *oxcf,
+                                       BufferPool *const pool);
+void av1_remove_compressor(AV1_COMP *cpi);
+
+void av1_change_config(AV1_COMP *cpi, const AV1EncoderConfig *oxcf);
+
+// receive a frames worth of data. caller can assume that a copy of this
+// frame is made and not just a copy of the pointer..
+int av1_receive_raw_frame(AV1_COMP *cpi, aom_enc_frame_flags_t frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time_stamp);
+
+int av1_get_compressed_data(AV1_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest, int64_t *time_stamp,
+                            int64_t *time_end, int flush);
+
+int av1_get_preview_raw_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *dest);
+
+int av1_get_last_show_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *frame);
+
+int av1_use_as_reference(AV1_COMP *cpi, int ref_frame_flags);
+
+void av1_update_reference(AV1_COMP *cpi, int ref_frame_flags);
+
+int av1_copy_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd);
+
+int av1_set_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd);
+
+int av1_update_entropy(AV1_COMP *cpi, int update);
+
+int av1_set_active_map(AV1_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int av1_get_active_map(AV1_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int av1_set_internal_size(AV1_COMP *cpi, AOM_SCALING horiz_mode,
+                          AOM_SCALING vert_mode);
+
+int av1_set_size_literal(AV1_COMP *cpi, unsigned int width,
+                         unsigned int height);
+
+int av1_get_quantizer(struct AV1_COMP *cpi);
+
+void av1_full_to_model_counts(av1_coeff_count_model *model_count,
+                              av1_coeff_count *full_count);
+
+static INLINE int frame_is_kf_gf_arf(const AV1_COMP *cpi) {
+  return frame_is_intra_only(&cpi->common) || cpi->refresh_alt_ref_frame ||
+         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
+}
+
+static INLINE int get_ref_frame_map_idx(const AV1_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+#if CONFIG_EXT_REFS
+  if (ref_frame >= LAST_FRAME && ref_frame <= LAST3_FRAME)
+    return cpi->lst_fb_idxes[ref_frame - 1];
+#else
+  if (ref_frame == LAST_FRAME) return cpi->lst_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  else if (ref_frame == GOLDEN_FRAME)
+    return cpi->gld_fb_idx;
+#if CONFIG_EXT_REFS
+  else if (ref_frame == BWDREF_FRAME)
+    return cpi->bwd_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  else
+    return cpi->alt_fb_idx;
+}
+
+static INLINE int get_ref_frame_buf_idx(const AV1_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
+  return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
+    const AV1_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf
+                                : NULL;
+}
+
+static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(
+    const AV1_COMP *cpi, const MV_REFERENCE_FRAME ref_frame) {
+  // Use up-sampled reference frames.
+  const int buf_idx =
+      cpi->upsampled_ref_idx[get_ref_frame_map_idx(cpi, ref_frame)];
+  return &cpi->upsampled_ref_bufs[buf_idx].buf;
+}
+
+#if CONFIG_EXT_REFS
+static INLINE int enc_is_ref_frame_buf(AV1_COMP *cpi, RefCntBuffer *frame_buf) {
+  MV_REFERENCE_FRAME ref_frame;
+  AV1_COMMON *const cm = &cpi->common;
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+    if (buf_idx == INVALID_IDX) continue;
+    if (frame_buf == &cm->buffer_pool->frame_bufs[buf_idx]) break;
+  }
+  return (ref_frame <= ALTREF_FRAME);
+}
+#endif  // CONFIG_EXT_REFS
+
+static INLINE unsigned int get_token_alloc(int mb_rows, int mb_cols) {
+  // We assume 3 planes all at full resolution. We assume up to 1 token per
+  // pixel, and then allow a head room of 1 EOSB token per 4x4 block per plane,
+  // plus EOSB_TOKEN per plane.
+  return mb_rows * mb_cols * (16 * 16 + 17) * 3;
+}
+
+// Get the allocated token size for a tile. It does the same calculation as in
+// the frame token allocation.
+static INLINE unsigned int allocated_tokens(TileInfo tile) {
+#if CONFIG_CB4X4
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 2) >> 2;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 2) >> 2;
+#else
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
+#endif
+
+  return get_token_alloc(tile_mb_rows, tile_mb_cols);
+}
+
+void av1_alloc_compressor_data(AV1_COMP *cpi);
+
+void av1_scale_references(AV1_COMP *cpi);
+
+void av1_update_reference_frames(AV1_COMP *cpi);
+
+void av1_set_high_precision_mv(AV1_COMP *cpi, int allow_high_precision_mv);
+#if CONFIG_TEMPMV_SIGNALING
+void av1_set_temporal_mv_prediction(AV1_COMP *cpi, int allow_tempmv_prediction);
+#endif
+
+YV12_BUFFER_CONFIG *av1_scale_if_required_fast(AV1_COMMON *cm,
+                                               YV12_BUFFER_CONFIG *unscaled,
+                                               YV12_BUFFER_CONFIG *scaled);
+
+YV12_BUFFER_CONFIG *av1_scale_if_required(AV1_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled);
+
+void av1_apply_encoding_flags(AV1_COMP *cpi, aom_enc_frame_flags_t flags);
+
+static INLINE int is_altref_enabled(const AV1_COMP *const cpi) {
+  return cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.enable_auto_arf;
+}
+
+// TODO(zoeliu): To set up cpi->oxcf.enable_auto_brf
+#if 0 && CONFIG_EXT_REFS
+static INLINE int is_bwdref_enabled(const AV1_COMP *const cpi) {
+  // NOTE(zoeliu): The enabling of bi-predictive frames depends on the use of
+  //               alt_ref, and now will be off when the alt_ref interval is
+  //               not sufficiently large.
+  return is_altref_enabled(cpi) && cpi->oxcf.enable_auto_brf;
+}
+#endif  // CONFIG_EXT_REFS
+
+static INLINE void set_ref_ptrs(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                MV_REFERENCE_FRAME ref0,
+                                MV_REFERENCE_FRAME ref1) {
+  xd->block_refs[0] =
+      &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME : 0];
+  xd->block_refs[1] =
+      &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME : 0];
+}
+
+static INLINE int get_chessboard_index(int frame_index) {
+  return frame_index & 0x1;
+}
+
+static INLINE int *cond_cost_list(const struct AV1_COMP *cpi, int *cost_list) {
+  return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
+}
+
+void av1_new_framerate(AV1_COMP *cpi, double framerate);
+
+#define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
+
+// Update up-sampled reference frame index.
+static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
+                               int new_uidx) {
+  const int ref_index = *uidx;
+
+  if (ref_index >= 0 && ubufs[ref_index].ref_count > 0)
+    ubufs[ref_index].ref_count--;
+
+  *uidx = new_uidx;
+  ubufs[new_uidx].ref_count++;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODER_H_

third_party/aom/av1/encoder/encodetxb.c

@@ -0,0 +1,784 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/scan.h"
+#include "av1/common/blockd.h"
+#include "av1/common/idct.h"
+#include "av1/common/pred_common.h"
+#include "av1/encoder/bitstream.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodetxb.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/subexp.h"
+#include "av1/encoder/tokenize.h"
+
+void av1_alloc_txb_buf(AV1_COMP *cpi) {
+#if 0
+  AV1_COMMON *cm = &cpi->common;
+  int mi_block_size = 1 << MI_SIZE_LOG2;
+  // TODO(angiebird): Make sure cm->subsampling_x/y is set correctly, and then
+  // use precise buffer size according to cm->subsampling_x/y
+  int pixel_stride = mi_block_size * cm->mi_cols;
+  int pixel_height = mi_block_size * cm->mi_rows;
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    CHECK_MEM_ERROR(
+        cm, cpi->tcoeff_buf[i],
+        aom_malloc(sizeof(*cpi->tcoeff_buf[i]) * pixel_stride * pixel_height));
+  }
+#else
+  (void)cpi;
+#endif
+}
+
+void av1_free_txb_buf(AV1_COMP *cpi) {
+#if 0
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    aom_free(cpi->tcoeff_buf[i]);
+  }
+#else
+  (void)cpi;
+#endif
+}
+
+static void write_golomb(aom_writer *w, int level) {
+  int x = level + 1;
+  int i = x;
+  int length = 0;
+
+  while (i) {
+    i >>= 1;
+    ++length;
+  }
+  assert(length > 0);
+
+  for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);
+
+  for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
+}
+
+void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                          aom_writer *w, int block, int plane,
+                          const tran_low_t *tcoeff, uint16_t eob,
+                          TXB_CTX *txb_ctx) {
+  aom_prob *nz_map;
+  aom_prob *eob_flag;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+  int c;
+  int is_nz;
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  const int seg_eob = tx_size_2d[tx_size];
+  uint8_t txb_mask[32 * 32] = { 0 };
+  uint16_t update_eob = 0;
+
+  aom_write(w, eob == 0, cm->fc->txb_skip[tx_size][txb_ctx->txb_skip_ctx]);
+
+  if (eob == 0) return;
+#if CONFIG_TXK_SEL
+  av1_write_tx_type(cm, xd, block, plane, w);
+#endif
+
+  nz_map = cm->fc->nz_map[tx_size][plane_type];
+  eob_flag = cm->fc->eob_flag[tx_size][plane_type];
+
+  for (c = 0; c < eob; ++c) {
+    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeff, scan[c], bwl);
+
+    tran_low_t v = tcoeff[scan[c]];
+    is_nz = (v != 0);
+
+    if (c == seg_eob - 1) break;
+
+    aom_write(w, is_nz, nz_map[coeff_ctx]);
+
+    if (is_nz) {
+      aom_write(w, c == (eob - 1), eob_flag[eob_ctx]);
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  int i;
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    aom_prob *coeff_base = cm->fc->coeff_base[tx_size][plane_type][i];
+
+    update_eob = 0;
+    for (c = eob - 1; c >= 0; --c) {
+      tran_low_t v = tcoeff[scan[c]];
+      tran_low_t level = abs(v);
+      int sign = (v < 0) ? 1 : 0;
+      int ctx;
+
+      if (level <= i) continue;
+
+      ctx = get_base_ctx(tcoeff, scan[c], bwl, i + 1);
+
+      if (level == i + 1) {
+        aom_write(w, 1, coeff_base[ctx]);
+        if (c == 0) {
+          aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
+        } else {
+          aom_write_bit(w, sign);
+        }
+        continue;
+      }
+      aom_write(w, 0, coeff_base[ctx]);
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t v = tcoeff[scan[c]];
+    tran_low_t level = abs(v);
+    int sign = (v < 0) ? 1 : 0;
+    int idx;
+    int ctx;
+
+    if (level <= NUM_BASE_LEVELS) continue;
+
+    if (c == 0) {
+      aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
+    } else {
+      aom_write_bit(w, sign);
+    }
+
+    // level is above 1.
+    ctx = get_level_ctx(tcoeff, scan[c], bwl);
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+        aom_write(w, 1, cm->fc->coeff_lps[tx_size][plane_type][ctx]);
+        break;
+      }
+      aom_write(w, 0, cm->fc->coeff_lps[tx_size][plane_type][ctx]);
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // use 0-th order Golomb code to handle the residual level.
+    write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
+  }
+}
+
+void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
+                         aom_writer *w, int plane) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+#if CONFIG_CB4X4
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
+#endif
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  const int bkw = tx_size_wide_unit[tx_size];
+  const int bkh = tx_size_high_unit[tx_size];
+  const int step = tx_size_wide_unit[tx_size] * tx_size_high_unit[tx_size];
+  int row, col;
+  int block = 0;
+  for (row = 0; row < max_blocks_high; row += bkh) {
+    for (col = 0; col < max_blocks_wide; col += bkw) {
+      tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
+      uint16_t eob = x->mbmi_ext->eobs[plane][block];
+      TXB_CTX txb_ctx = { x->mbmi_ext->txb_skip_ctx[plane][block],
+                          x->mbmi_ext->dc_sign_ctx[plane][block] };
+      av1_write_coeffs_txb(cm, xd, w, block, plane, tcoeff, eob, &txb_ctx);
+      block += step;
+    }
+  }
+}
+
+static INLINE void get_base_ctx_set(const tran_low_t *tcoeffs,
+                                    int c,  // raster order
+                                    const int bwl,
+                                    int ctx_set[NUM_BASE_LEVELS]) {
+  const int row = c >> bwl;
+  const int col = c - (row << bwl);
+  const int stride = 1 << bwl;
+  int mag[NUM_BASE_LEVELS] = { 0 };
+  int idx;
+  tran_low_t abs_coeff;
+  int i;
+
+  for (idx = 0; idx < BASE_CONTEXT_POSITION_NUM; ++idx) {
+    int ref_row = row + base_ref_offset[idx][0];
+    int ref_col = col + base_ref_offset[idx][1];
+    int pos = (ref_row << bwl) + ref_col;
+
+    if (ref_row < 0 || ref_col < 0 || ref_row >= stride || ref_col >= stride)
+      continue;
+
+    abs_coeff = abs(tcoeffs[pos]);
+
+    for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+      ctx_set[i] += abs_coeff > i;
+      if (base_ref_offset[idx][0] >= 0 && base_ref_offset[idx][1] >= 0)
+        mag[i] |= abs_coeff > (i + 1);
+    }
+  }
+
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    ctx_set[i] = (ctx_set[i] + 1) >> 1;
+
+    if (row == 0 && col == 0)
+      ctx_set[i] = (ctx_set[i] << 1) + mag[i];
+    else if (row == 0)
+      ctx_set[i] = 8 + (ctx_set[i] << 1) + mag[i];
+    else if (col == 0)
+      ctx_set[i] = 18 + (ctx_set[i] << 1) + mag[i];
+    else
+      ctx_set[i] = 28 + (ctx_set[i] << 1) + mag[i];
+  }
+  return;
+}
+
+int av1_cost_coeffs_txb(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                        int block, TXB_CTX *txb_ctx) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const struct macroblock_plane *p = &x->plane[plane];
+  const int eob = p->eobs[block];
+  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  int c, cost;
+  const int seg_eob = AOMMIN(eob, tx_size_2d[tx_size] - 1);
+  int txb_skip_ctx = txb_ctx->txb_skip_ctx;
+  aom_prob *nz_map = xd->fc->nz_map[tx_size][plane_type];
+
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  // txb_mask is only initialized for once here. After that, it will be set when
+  // coding zero map and then reset when coding level 1 info.
+  uint8_t txb_mask[32 * 32] = { 0 };
+  aom_prob(*coeff_base)[COEFF_BASE_CONTEXTS] =
+      xd->fc->coeff_base[tx_size][plane_type];
+
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+
+  cost = 0;
+
+  if (eob == 0) {
+    cost = av1_cost_bit(xd->fc->txb_skip[tx_size][txb_skip_ctx], 1);
+    return cost;
+  }
+
+  cost = av1_cost_bit(xd->fc->txb_skip[tx_size][txb_skip_ctx], 0);
+
+#if CONFIG_TXK_SEL
+  cost += av1_tx_type_cost(cpi, xd, mbmi->sb_type, plane, tx_size, tx_type);
+#endif
+
+  for (c = 0; c < eob; ++c) {
+    tran_low_t v = qcoeff[scan[c]];
+    int is_nz = (v != 0);
+    int level = abs(v);
+
+    if (c < seg_eob) {
+      int coeff_ctx = get_nz_map_ctx(qcoeff, txb_mask, scan[c], bwl);
+      cost += av1_cost_bit(nz_map[coeff_ctx], is_nz);
+    }
+
+    if (is_nz) {
+      int ctx_ls[NUM_BASE_LEVELS] = { 0 };
+      int sign = (v < 0) ? 1 : 0;
+
+      // sign bit cost
+      if (c == 0) {
+        int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+
+        cost += av1_cost_bit(xd->fc->dc_sign[plane_type][dc_sign_ctx], sign);
+      } else {
+        cost += av1_cost_bit(128, sign);
+      }
+
+      get_base_ctx_set(qcoeff, scan[c], bwl, ctx_ls);
+
+      int i;
+      for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+        if (level <= i) continue;
+
+        if (level == i + 1) {
+          cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 1);
+          continue;
+        }
+        cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 0);
+      }
+
+      if (level > NUM_BASE_LEVELS) {
+        int idx;
+        int ctx;
+
+        ctx = get_level_ctx(qcoeff, scan[c], bwl);
+
+        for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+          if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+            cost +=
+                av1_cost_bit(xd->fc->coeff_lps[tx_size][plane_type][ctx], 1);
+            break;
+          }
+          cost += av1_cost_bit(xd->fc->coeff_lps[tx_size][plane_type][ctx], 0);
+        }
+
+        if (idx >= COEFF_BASE_RANGE) {
+          // residual cost
+          int r = level - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
+          int ri = r;
+          int length = 0;
+
+          while (ri) {
+            ri >>= 1;
+            ++length;
+          }
+
+          for (ri = 0; ri < length - 1; ++ri) cost += av1_cost_bit(128, 0);
+
+          for (ri = length - 1; ri >= 0; --ri)
+            cost += av1_cost_bit(128, (r >> ri) & 0x01);
+        }
+      }
+
+      if (c < seg_eob) {
+        int eob_ctx = get_eob_ctx(qcoeff, scan[c], bwl);
+        cost += av1_cost_bit(xd->fc->eob_flag[tx_size][plane_type][eob_ctx],
+                             c == (eob - 1));
+      }
+    }
+
+    txb_mask[scan[c]] = 1;
+  }
+
+  return cost;
+}
+
+typedef struct TxbParams {
+  const AV1_COMP *cpi;
+  ThreadData *td;
+  int rate;
+} TxbParams;
+
+int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
+                                const SCAN_ORDER *scan_order, int eob) {
+  const int16_t *scan = scan_order->scan;
+  int cul_level = 0;
+  int c;
+  for (c = 0; c < eob; ++c) {
+    cul_level += abs(qcoeff[scan[c]]);
+  }
+
+  cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
+  set_dc_sign(&cul_level, qcoeff[0]);
+
+  return cul_level;
+}
+
+static void update_txb_context(int plane, int block, int blk_row, int blk_col,
+                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                               void *arg) {
+  TxbParams *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  const uint16_t eob = p->eobs[block];
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  (void)plane_bsize;
+
+  int cul_level = av1_get_txb_entropy_context(qcoeff, scan_order, eob);
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, blk_col, blk_row);
+}
+
+static void update_and_record_txb_context(int plane, int block, int blk_row,
+                                          int blk_col, BLOCK_SIZE plane_bsize,
+                                          TX_SIZE tx_size, void *arg) {
+  TxbParams *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int eob = p->eobs[block], update_eob = 0;
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
+  const int segment_id = mbmi->segment_id;
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
+  int c, i;
+  TXB_CTX txb_ctx;
+  get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + blk_col,
+              pd->left_context + blk_row, &txb_ctx);
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  int cul_level = 0;
+  unsigned int(*nz_map_count)[SIG_COEF_CONTEXTS][2];
+  uint8_t txb_mask[32 * 32] = { 0 };
+
+  nz_map_count = &td->counts->nz_map[tx_size][plane_type];
+
+  memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
+
+  ++td->counts->txb_skip[tx_size][txb_ctx.txb_skip_ctx][eob == 0];
+  x->mbmi_ext->txb_skip_ctx[plane][block] = txb_ctx.txb_skip_ctx;
+
+  x->mbmi_ext->eobs[plane][block] = eob;
+
+  if (eob == 0) {
+    av1_set_contexts(xd, pd, plane, tx_size, 0, blk_col, blk_row);
+    return;
+  }
+
+#if CONFIG_TXK_SEL
+  av1_update_tx_type_count(cm, xd, block, plane, mbmi->sb_type, tx_size,
+                           td->counts);
+#endif
+
+  for (c = 0; c < eob; ++c) {
+    tran_low_t v = qcoeff[scan[c]];
+    int is_nz = (v != 0);
+    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeff, scan[c], bwl);
+
+    if (c == seg_eob - 1) break;
+
+    ++(*nz_map_count)[coeff_ctx][is_nz];
+
+    if (is_nz) {
+      ++td->counts->eob_flag[tx_size][plane_type][eob_ctx][c == (eob - 1)];
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  // Reverse process order to handle coefficient level and sign.
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    update_eob = 0;
+    for (c = eob - 1; c >= 0; --c) {
+      tran_low_t v = qcoeff[scan[c]];
+      tran_low_t level = abs(v);
+      int ctx;
+
+      if (level <= i) continue;
+
+      ctx = get_base_ctx(tcoeff, scan[c], bwl, i + 1);
+
+      if (level == i + 1) {
+        ++td->counts->coeff_base[tx_size][plane_type][i][ctx][1];
+        if (c == 0) {
+          int dc_sign_ctx = txb_ctx.dc_sign_ctx;
+
+          ++td->counts->dc_sign[plane_type][dc_sign_ctx][v < 0];
+          x->mbmi_ext->dc_sign_ctx[plane][block] = dc_sign_ctx;
+        }
+        cul_level += level;
+        continue;
+      }
+      ++td->counts->coeff_base[tx_size][plane_type][i][ctx][0];
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t v = qcoeff[scan[c]];
+    tran_low_t level = abs(v);
+    int idx;
+    int ctx;
+
+    if (level <= NUM_BASE_LEVELS) continue;
+
+    cul_level += level;
+    if (c == 0) {
+      int dc_sign_ctx = txb_ctx.dc_sign_ctx;
+
+      ++td->counts->dc_sign[plane_type][dc_sign_ctx][v < 0];
+      x->mbmi_ext->dc_sign_ctx[plane][block] = dc_sign_ctx;
+    }
+
+    // level is above 1.
+    ctx = get_level_ctx(tcoeff, scan[c], bwl);
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+        ++td->counts->coeff_lps[tx_size][plane_type][ctx][1];
+        break;
+      }
+      ++td->counts->coeff_lps[tx_size][plane_type][ctx][0];
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // use 0-th order Golomb code to handle the residual level.
+  }
+
+  cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
+
+  // DC value
+  set_dc_sign(&cul_level, tcoeff[0]);
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, blk_col, blk_row);
+
+#if CONFIG_ADAPT_SCAN
+  // Since dqcoeff is not available here, we pass qcoeff into
+  // av1_update_scan_count_facade(). The update behavior should be the same
+  // because av1_update_scan_count_facade() only cares if coefficients are zero
+  // or not.
+  av1_update_scan_count_facade((AV1_COMMON *)cm, td->counts, tx_size, tx_type,
+                               qcoeff, eob);
+#endif
+}
+
+void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
+                            RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                            int mi_row, int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct TxbParams arg = { cpi, td, 0 };
+  (void)rate;
+  (void)mi_row;
+  (void)mi_col;
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+    td->counts->skip[ctx][0] += skip_inc;
+    av1_foreach_transformed_block(xd, bsize, mi_row, mi_col,
+                                  update_and_record_txb_context, &arg);
+  } else if (dry_run == DRY_RUN_NORMAL) {
+    av1_foreach_transformed_block(xd, bsize, mi_row, mi_col, update_txb_context,
+                                  &arg);
+  } else {
+    printf("DRY_RUN_COSTCOEFFS is not supported yet\n");
+    assert(0);
+  }
+}
+
+static void find_new_prob(unsigned int *branch_cnt, aom_prob *oldp,
+                          int *savings, int *update, aom_writer *const bc) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  int u = 0;
+  aom_prob newp = get_binary_prob(branch_cnt[0], branch_cnt[1]);
+  int s = av1_prob_diff_update_savings_search(branch_cnt, *oldp, &newp, upd, 1);
+
+  if (s > 0 && newp != *oldp) u = 1;
+
+  if (u)
+    *savings += s - (int)(av1_cost_zero(upd));  // TODO(jingning): 1?
+  else
+    *savings -= (int)(av1_cost_zero(upd));
+
+  if (update) {
+    ++update[u];
+    return;
+  }
+
+  aom_write(bc, u, upd);
+  if (u) {
+    /* send/use new probability */
+    av1_write_prob_diff_update(bc, newp, *oldp);
+    *oldp = newp;
+  }
+}
+
+static void write_txb_probs(aom_writer *const bc, AV1_COMP *cpi,
+                            TX_SIZE tx_size) {
+  FRAME_CONTEXT *fc = cpi->common.fc;
+  FRAME_COUNTS *counts = cpi->td.counts;
+  int savings = 0;
+  int update[2] = { 0, 0 };
+  int plane, ctx, level;
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx) {
+    find_new_prob(counts->txb_skip[tx_size][ctx], &fc->txb_skip[tx_size][ctx],
+                  &savings, update, bc);
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->nz_map[tx_size][plane][ctx],
+                    &fc->nz_map[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->eob_flag[tx_size][plane][ctx],
+                    &fc->eob_flag[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane) {
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx) {
+        find_new_prob(counts->coeff_base[tx_size][plane][level][ctx],
+                      &fc->coeff_base[tx_size][plane][level][ctx], &savings,
+                      update, bc);
+      }
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx) {
+      find_new_prob(counts->coeff_lps[tx_size][plane][ctx],
+                    &fc->coeff_lps[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  // Decide if to update the model for this tx_size
+  if (update[1] == 0 || savings < 0) {
+    aom_write_bit(bc, 0);
+    return;
+  }
+  aom_write_bit(bc, 1);
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx) {
+    find_new_prob(counts->txb_skip[tx_size][ctx], &fc->txb_skip[tx_size][ctx],
+                  &savings, NULL, bc);
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->nz_map[tx_size][plane][ctx],
+                    &fc->nz_map[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->eob_flag[tx_size][plane][ctx],
+                    &fc->eob_flag[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane) {
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx) {
+        find_new_prob(counts->coeff_base[tx_size][plane][level][ctx],
+                      &fc->coeff_base[tx_size][plane][level][ctx], &savings,
+                      NULL, bc);
+      }
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx) {
+      find_new_prob(counts->coeff_lps[tx_size][plane][ctx],
+                    &fc->coeff_lps[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+}
+
+void av1_write_txb_probs(AV1_COMP *cpi, aom_writer *w) {
+  const TX_MODE tx_mode = cpi->common.tx_mode;
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  int ctx, plane;
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
+      av1_cond_prob_diff_update(w, &cpi->common.fc->dc_sign[plane][ctx],
+                                cpi->td.counts->dc_sign[plane][ctx], 1);
+
+  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+    write_txb_probs(w, cpi, tx_size);
+}
+
+#if CONFIG_TXK_SEL
+int64_t av1_search_txk_type(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                            int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                            const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                            int use_fast_coef_costing, RD_STATS *rd_stats) {
+  const AV1_COMMON *cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  TX_TYPE txk_start = DCT_DCT;
+  TX_TYPE txk_end = TX_TYPES - 1;
+  TX_TYPE best_tx_type = txk_start;
+  int64_t best_rd = INT64_MAX;
+  const int coeff_ctx = combine_entropy_contexts(*a, *l);
+  TX_TYPE tx_type;
+  for (tx_type = txk_start; tx_type <= txk_end; ++tx_type) {
+    if (plane == 0) mbmi->txk_type[block] = tx_type;
+    TX_TYPE ref_tx_type =
+        get_tx_type(get_plane_type(plane), xd, block, tx_size);
+    if (tx_type != ref_tx_type) {
+      // use get_tx_type() to check if the tx_type is valid for the current mode
+      // if it's not, we skip it here.
+      continue;
+    }
+    RD_STATS this_rd_stats;
+    av1_invalid_rd_stats(&this_rd_stats);
+    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                    coeff_ctx, AV1_XFORM_QUANT_FP);
+    if (x->plane[plane].eobs[block] && !xd->lossless[mbmi->segment_id])
+      av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+    av1_dist_block(cpi, x, plane, plane_bsize, block, blk_row, blk_col, tx_size,
+                   &this_rd_stats.dist, &this_rd_stats.sse,
+                   OUTPUT_HAS_PREDICTED_PIXELS);
+    const SCAN_ORDER *scan_order =
+        get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+    this_rd_stats.rate = av1_cost_coeffs(
+        cpi, x, plane, block, tx_size, scan_order, a, l, use_fast_coef_costing);
+    int rd =
+        RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate, this_rd_stats.dist);
+    if (rd < best_rd) {
+      best_rd = rd;
+      *rd_stats = this_rd_stats;
+      best_tx_type = tx_type;
+    }
+  }
+  if (plane == 0) mbmi->txk_type[block] = best_tx_type;
+  // TODO(angiebird): Instead of re-call av1_xform_quant and av1_optimize_b,
+  // copy the best result in the above tx_type search for loop
+  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                  coeff_ctx, AV1_XFORM_QUANT_FP);
+  if (x->plane[plane].eobs[block] && !xd->lossless[mbmi->segment_id])
+    av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+  if (!is_inter_block(mbmi)) {
+    // intra mode needs decoded result such that the next transform block
+    // can use it for prediction.
+    av1_inverse_transform_block_facade(xd, plane, block, blk_row, blk_col,
+                                       x->plane[plane].eobs[block]);
+  }
+  return best_rd;
+}
+#endif  // CONFIG_TXK_SEL

third_party/aom/av1/encoder/encodetxb.h

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef ENCODETXB_H_
+#define ENCODETXB_H_
+
+#include "./aom_config.h"
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/txb_common.h"
+#include "av1/encoder/block.h"
+#include "av1/encoder/encoder.h"
+#include "aom_dsp/bitwriter.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
+void av1_alloc_txb_buf(AV1_COMP *cpi);
+void av1_free_txb_buf(AV1_COMP *cpi);
+int av1_cost_coeffs_txb(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                        int block, TXB_CTX *txb_ctx);
+void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                          aom_writer *w, int block, int plane,
+                          const tran_low_t *tcoeff, uint16_t eob,
+                          TXB_CTX *txb_ctx);
+void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
+                         aom_writer *w, int plane);
+int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
+                                const SCAN_ORDER *scan_order, int eob);
+void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
+                            RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                            const int mi_row, const int mi_col);
+void av1_write_txb_probs(AV1_COMP *cpi, aom_writer *w);
+
+#if CONFIG_TXK_SEL
+int64_t av1_search_txk_type(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                            int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                            const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                            int use_fast_coef_costing, RD_STATS *rd_stats);
+#endif
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // COEFFS_CODING_H_

third_party/aom/av1/encoder/ethread.c

@@ -0,0 +1,176 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/ethread.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+static void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
+  int i, j, k, l, m, n;
+
+  for (i = 0; i < REFERENCE_MODES; i++)
+    td->rd_counts.comp_pred_diff[i] += td_t->rd_counts.comp_pred_diff[i];
+
+#if CONFIG_GLOBAL_MOTION
+  for (i = 0; i < TOTAL_REFS_PER_FRAME; i++)
+    td->rd_counts.global_motion_used[i] +=
+        td_t->rd_counts.global_motion_used[i];
+#endif  // CONFIG_GLOBAL_MOTION
+
+  for (i = 0; i < TX_SIZES; i++)
+    for (j = 0; j < PLANE_TYPES; j++)
+      for (k = 0; k < REF_TYPES; k++)
+        for (l = 0; l < COEF_BANDS; l++)
+          for (m = 0; m < COEFF_CONTEXTS; m++)
+            for (n = 0; n < ENTROPY_TOKENS; n++)
+              td->rd_counts.coef_counts[i][j][k][l][m][n] +=
+                  td_t->rd_counts.coef_counts[i][j][k][l][m][n];
+}
+
+static int enc_worker_hook(EncWorkerData *const thread_data, void *unused) {
+  AV1_COMP *const cpi = thread_data->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  int t;
+
+  (void)unused;
+
+  for (t = thread_data->start; t < tile_rows * tile_cols;
+       t += cpi->num_workers) {
+    int tile_row = t / tile_cols;
+    int tile_col = t % tile_cols;
+
+    av1_encode_tile(cpi, thread_data->td, tile_row, tile_col);
+  }
+
+  return 0;
+}
+
+void av1_encode_tiles_mt(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int tile_cols = cm->tile_cols;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int num_workers = AOMMIN(cpi->oxcf.max_threads, tile_cols);
+  int i;
+
+  av1_init_tile_data(cpi);
+
+  // Only run once to create threads and allocate thread data.
+  if (cpi->num_workers == 0) {
+    CHECK_MEM_ERROR(cm, cpi->workers,
+                    aom_malloc(num_workers * sizeof(*cpi->workers)));
+
+    CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+                    aom_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
+
+    for (i = 0; i < num_workers; i++) {
+      AVxWorker *const worker = &cpi->workers[i];
+      EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
+
+      ++cpi->num_workers;
+      winterface->init(worker);
+
+      thread_data->cpi = cpi;
+
+      if (i < num_workers - 1) {
+        // Allocate thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td,
+                        aom_memalign(32, sizeof(*thread_data->td)));
+        av1_zero(*thread_data->td);
+
+        // Set up pc_tree.
+        thread_data->td->leaf_tree = NULL;
+        thread_data->td->pc_tree = NULL;
+        av1_setup_pc_tree(cm, thread_data->td);
+
+        // Set up variance tree if needed.
+        if (cpi->sf.partition_search_type == VAR_BASED_PARTITION)
+          av1_setup_var_tree(cm, thread_data->td);
+
+        // Allocate frame counters in thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td->counts,
+                        aom_calloc(1, sizeof(*thread_data->td->counts)));
+
+        // Create threads
+        if (!winterface->reset(worker))
+          aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                             "Tile encoder thread creation failed");
+      } else {
+        // Main thread acts as a worker and uses the thread data in cpi.
+        thread_data->td = &cpi->td;
+      }
+
+      winterface->sync(worker);
+    }
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *thread_data;
+
+    worker->hook = (AVxWorkerHook)enc_worker_hook;
+    worker->data1 = &cpi->tile_thr_data[i];
+    worker->data2 = NULL;
+    thread_data = (EncWorkerData *)worker->data1;
+
+    // Before encoding a frame, copy the thread data from cpi.
+    if (thread_data->td != &cpi->td) {
+      thread_data->td->mb = cpi->td.mb;
+      thread_data->td->rd_counts = cpi->td.rd_counts;
+    }
+    if (thread_data->td->counts != &cpi->common.counts) {
+      memcpy(thread_data->td->counts, &cpi->common.counts,
+             sizeof(cpi->common.counts));
+    }
+
+#if CONFIG_PALETTE
+    // Allocate buffers used by palette coding mode.
+    if (cpi->common.allow_screen_content_tools && i < num_workers - 1) {
+      MACROBLOCK *x = &thread_data->td->mb;
+      CHECK_MEM_ERROR(cm, x->palette_buffer,
+                      aom_memalign(16, sizeof(*x->palette_buffer)));
+    }
+#endif  // CONFIG_PALETTE
+  }
+
+  // Encode a frame
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+
+    // Set the starting tile for each thread.
+    thread_data->start = i;
+
+    if (i == cpi->num_workers - 1)
+      winterface->execute(worker);
+    else
+      winterface->launch(worker);
+  }
+
+  // Encoding ends.
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    winterface->sync(worker);
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+
+    // Accumulate counters.
+    if (i < cpi->num_workers - 1) {
+      av1_accumulate_frame_counts(&cm->counts, thread_data->td->counts);
+      accumulate_rd_opt(&cpi->td, thread_data->td);
+    }
+  }
+}

third_party/aom/av1/encoder/ethread.h

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ETHREAD_H_
+#define AV1_ENCODER_ETHREAD_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1_COMP;
+struct ThreadData;
+
+typedef struct EncWorkerData {
+  struct AV1_COMP *cpi;
+  struct ThreadData *td;
+  int start;
+} EncWorkerData;
+
+void av1_encode_tiles_mt(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ETHREAD_H_

third_party/aom/av1/encoder/extend.c

@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/common.h"
+#include "av1/encoder/extend.h"
+
+static void copy_and_extend_plane(const uint8_t *src, int src_pitch,
+                                  uint8_t *dst, int dst_pitch, int w, int h,
+                                  int extend_top, int extend_left,
+                                  int extend_bottom, int extend_right) {
+  int i, linesize;
+
+  // copy the left and right most columns out
+  const uint8_t *src_ptr1 = src;
+  const uint8_t *src_ptr2 = src + w - 1;
+  uint8_t *dst_ptr1 = dst - extend_left;
+  uint8_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h)-extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += dst_pitch;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_copy_and_extend_plane(const uint8_t *src8, int src_pitch,
+                                         uint8_t *dst8, int dst_pitch, int w,
+                                         int h, int extend_top, int extend_left,
+                                         int extend_bottom, int extend_right) {
+  int i, linesize;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+
+  // copy the left and right most columns out
+  const uint16_t *src_ptr1 = src;
+  const uint16_t *src_ptr2 = src + w - 1;
+  uint16_t *dst_ptr1 = dst - extend_left;
+  uint16_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    aom_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w * sizeof(src_ptr1[0]));
+    aom_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h)-extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(src_ptr1[0]));
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(src_ptr2[0]));
+    dst_ptr2 += dst_pitch;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst) {
+  // Extend src frame in buffer
+  // Altref filtering assumes 16 pixel extension
+  const int et_y = 16;
+  const int el_y = 16;
+  // Motion estimation may use src block variance with the block size up
+  // to 64x64, so the right and bottom need to be extended to 64 multiple
+  // or up to 16, whichever is greater.
+  const int er_y =
+      AOMMAX(src->y_width + 16, ALIGN_POWER_OF_TWO(src->y_width, 6)) -
+      src->y_crop_width;
+  const int eb_y =
+      AOMMAX(src->y_height + 16, ALIGN_POWER_OF_TWO(src->y_height, 6)) -
+      src->y_crop_height;
+  const int uv_width_subsampling = (src->uv_width != src->y_width);
+  const int uv_height_subsampling = (src->uv_height != src->y_height);
+  const int et_uv = et_y >> uv_height_subsampling;
+  const int el_uv = el_y >> uv_width_subsampling;
+  const int eb_uv = eb_y >> uv_height_subsampling;
+  const int er_uv = er_y >> uv_width_subsampling;
+
+#if CONFIG_HIGHBITDEPTH
+  if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_copy_and_extend_plane(src->y_buffer, src->y_stride, dst->y_buffer,
+                                 dst->y_stride, src->y_crop_width,
+                                 src->y_crop_height, et_y, el_y, eb_y, er_y);
+
+    highbd_copy_and_extend_plane(
+        src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
+        src->uv_crop_width, src->uv_crop_height, et_uv, el_uv, eb_uv, er_uv);
+
+    highbd_copy_and_extend_plane(
+        src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
+        src->uv_crop_width, src->uv_crop_height, et_uv, el_uv, eb_uv, er_uv);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  copy_and_extend_plane(src->y_buffer, src->y_stride, dst->y_buffer,
+                        dst->y_stride, src->y_crop_width, src->y_crop_height,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer, src->uv_stride, dst->u_buffer,
+                        dst->uv_stride, src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer, src->uv_stride, dst->v_buffer,
+                        dst->uv_stride, src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+}
+
+void av1_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst, int srcy,
+                                         int srcx, int srch, int srcw) {
+  // If the side is not touching the bounder then don't extend.
+  const int et_y = srcy ? 0 : dst->border;
+  const int el_y = srcx ? 0 : dst->border;
+  const int eb_y = srcy + srch != src->y_height
+                       ? 0
+                       : dst->border + dst->y_height - src->y_height;
+  const int er_y = srcx + srcw != src->y_width
+                       ? 0
+                       : dst->border + dst->y_width - src->y_width;
+  const int src_y_offset = srcy * src->y_stride + srcx;
+  const int dst_y_offset = srcy * dst->y_stride + srcx;
+
+  const int et_uv = ROUND_POWER_OF_TWO(et_y, 1);
+  const int el_uv = ROUND_POWER_OF_TWO(el_y, 1);
+  const int eb_uv = ROUND_POWER_OF_TWO(eb_y, 1);
+  const int er_uv = ROUND_POWER_OF_TWO(er_y, 1);
+  const int src_uv_offset = ((srcy * src->uv_stride) >> 1) + (srcx >> 1);
+  const int dst_uv_offset = ((srcy * dst->uv_stride) >> 1) + (srcx >> 1);
+  const int srch_uv = ROUND_POWER_OF_TWO(srch, 1);
+  const int srcw_uv = ROUND_POWER_OF_TWO(srcw, 1);
+
+  copy_and_extend_plane(src->y_buffer + src_y_offset, src->y_stride,
+                        dst->y_buffer + dst_y_offset, dst->y_stride, srcw, srch,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer + src_uv_offset, src->uv_stride,
+                        dst->u_buffer + dst_uv_offset, dst->uv_stride, srcw_uv,
+                        srch_uv, et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer + src_uv_offset, src->uv_stride,
+                        dst->v_buffer + dst_uv_offset, dst->uv_stride, srcw_uv,
+                        srch_uv, et_uv, el_uv, eb_uv, er_uv);
+}

third_party/aom/av1/encoder/extend.h

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_EXTEND_H_
+#define AV1_ENCODER_EXTEND_H_
+
+#include "aom_scale/yv12config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst);
+
+void av1_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst, int srcy,
+                                         int srcx, int srch, int srcw);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_EXTEND_H_

third_party/aom/av1/encoder/firstpass.h

@@ -0,0 +1,202 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_FIRSTPASS_H_
+#define AV1_ENCODER_FIRSTPASS_H_
+
+#include "av1/encoder/lookahead.h"
+#include "av1/encoder/ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_FP_MB_STATS
+
+#define FPMB_DCINTRA_MASK 0x01
+
+#define FPMB_MOTION_ZERO_MASK 0x02
+#define FPMB_MOTION_LEFT_MASK 0x04
+#define FPMB_MOTION_RIGHT_MASK 0x08
+#define FPMB_MOTION_UP_MASK 0x10
+#define FPMB_MOTION_DOWN_MASK 0x20
+
+#define FPMB_ERROR_SMALL_MASK 0x40
+#define FPMB_ERROR_LARGE_MASK 0x80
+#define FPMB_ERROR_SMALL_TH 2000
+#define FPMB_ERROR_LARGE_TH 48000
+
+typedef struct {
+  uint8_t *mb_stats_start;
+  uint8_t *mb_stats_end;
+} FIRSTPASS_MB_STATS;
+#endif
+
+#if CONFIG_EXT_REFS
+// Length of the bi-predictive frame group (BFG)
+// NOTE: Currently each BFG contains one backward ref (BWF) frame plus a certain
+//       number of bi-predictive frames.
+#define BFG_INTERVAL 2
+// The maximum number of extra ALT_REF's
+// NOTE: This number cannot be greater than 2 or the reference frame buffer will
+//       overflow.
+#define MAX_EXT_ARFS 2
+#define MIN_EXT_ARF_INTERVAL 4
+#endif  // CONFIG_EXT_REFS
+
+#define VLOW_MOTION_THRESHOLD 950
+
+typedef struct {
+  double frame;
+  double weight;
+  double intra_error;
+  double coded_error;
+  double sr_coded_error;
+  double pcnt_inter;
+  double pcnt_motion;
+  double pcnt_second_ref;
+  double pcnt_neutral;
+  double intra_skip_pct;
+  double inactive_zone_rows;  // Image mask rows top and bottom.
+  double inactive_zone_cols;  // Image mask columns at left and right edges.
+  double MVr;
+  double mvr_abs;
+  double MVc;
+  double mvc_abs;
+  double MVrv;
+  double MVcv;
+  double mv_in_out_count;
+  double new_mv_count;
+  double duration;
+  double count;
+} FIRSTPASS_STATS;
+
+typedef enum {
+  KF_UPDATE = 0,
+  LF_UPDATE = 1,
+  GF_UPDATE = 2,
+  ARF_UPDATE = 3,
+  OVERLAY_UPDATE = 4,
+#if CONFIG_EXT_REFS
+  BRF_UPDATE = 5,            // Backward Reference Frame
+  LAST_BIPRED_UPDATE = 6,    // Last Bi-predictive Frame
+  BIPRED_UPDATE = 7,         // Bi-predictive Frame, but not the last one
+  INTNL_OVERLAY_UPDATE = 8,  // Internal Overlay Frame
+  FRAME_UPDATE_TYPES = 9
+#else
+  FRAME_UPDATE_TYPES = 5
+#endif  // CONFIG_EXT_REFS
+} FRAME_UPDATE_TYPE;
+
+#define FC_ANIMATION_THRESH 0.15
+typedef enum {
+  FC_NORMAL = 0,
+  FC_GRAPHICS_ANIMATION = 1,
+  FRAME_CONTENT_TYPES = 2
+} FRAME_CONTENT_TYPE;
+
+typedef struct {
+  unsigned char index;
+  RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1];
+  FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1];
+#if CONFIG_EXT_REFS
+  unsigned char brf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char bidir_pred_enabled[(MAX_LAG_BUFFERS * 2) + 1];
+#endif  // CONFIG_EXT_REFS
+  int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1];
+} GF_GROUP;
+
+typedef struct {
+  unsigned int section_intra_rating;
+  FIRSTPASS_STATS total_stats;
+  FIRSTPASS_STATS this_frame_stats;
+  const FIRSTPASS_STATS *stats_in;
+  const FIRSTPASS_STATS *stats_in_start;
+  const FIRSTPASS_STATS *stats_in_end;
+  FIRSTPASS_STATS total_left_stats;
+  int first_pass_done;
+  int64_t bits_left;
+  double modified_error_min;
+  double modified_error_max;
+  double modified_error_left;
+  double mb_av_energy;
+
+#if CONFIG_FP_MB_STATS
+  uint8_t *frame_mb_stats_buf;
+  uint8_t *this_frame_mb_stats;
+  FIRSTPASS_MB_STATS firstpass_mb_stats;
+#endif
+  // An indication of the content type of the current frame
+  FRAME_CONTENT_TYPE fr_content_type;
+
+  // Projected total bits available for a key frame group of frames
+  int64_t kf_group_bits;
+
+  // Error score of frames still to be coded in kf group
+  int64_t kf_group_error_left;
+
+  // The fraction for a kf groups total bits allocated to the inter frames
+  double kfgroup_inter_fraction;
+
+  int sr_update_lag;
+
+  int kf_zeromotion_pct;
+  int last_kfgroup_zeromotion_pct;
+  int gf_zeromotion_pct;
+  int active_worst_quality;
+  int baseline_active_worst_quality;
+  int extend_minq;
+  int extend_maxq;
+  int extend_minq_fast;
+
+  GF_GROUP gf_group;
+} TWO_PASS;
+
+struct AV1_COMP;
+
+void av1_init_first_pass(struct AV1_COMP *cpi);
+void av1_rc_get_first_pass_params(struct AV1_COMP *cpi);
+void av1_first_pass(struct AV1_COMP *cpi, const struct lookahead_entry *source);
+void av1_end_first_pass(struct AV1_COMP *cpi);
+
+void av1_init_second_pass(struct AV1_COMP *cpi);
+void av1_rc_get_second_pass_params(struct AV1_COMP *cpi);
+void av1_twopass_postencode_update(struct AV1_COMP *cpi);
+
+// Post encode update of the rate control parameters for 2-pass
+void av1_twopass_postencode_update(struct AV1_COMP *cpi);
+
+void av1_init_subsampling(struct AV1_COMP *cpi);
+
+void av1_calculate_coded_size(struct AV1_COMP *cpi, int *scaled_frame_width,
+                              int *scaled_frame_height);
+
+#if CONFIG_EXT_REFS
+static INLINE int get_number_of_extra_arfs(int interval, int arf_pending) {
+  if (arf_pending && MAX_EXT_ARFS > 0)
+    return interval >= MIN_EXT_ARF_INTERVAL * (MAX_EXT_ARFS + 1)
+               ? MAX_EXT_ARFS
+               : interval >= MIN_EXT_ARF_INTERVAL * MAX_EXT_ARFS
+                     ? MAX_EXT_ARFS - 1
+                     : 0;
+  else
+    return 0;
+}
+#endif  // CONFIG_EXT_REFS
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_FIRSTPASS_H_

third_party/aom/av1/encoder/generic_encoder.c

@@ -0,0 +1,157 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitwriter.h"
+#include "av1/common/generic_code.h"
+#include "av1/common/odintrin.h"
+#include "pvq_encoder.h"
+
+/** Encodes a value from 0 to N-1 (with N up to 16) based on a cdf and adapts
+ * the cdf accordingly.
+ *
+ * @param [in,out] w     multi-symbol entropy encoder
+ * @param [in]     val   variable being encoded
+ * @param [in,out] cdf   CDF of the variable (Q15)
+ * @param [in]     n     number of values possible
+ * @param [in,out] count number of symbols encoded with that cdf so far
+ * @param [in]     rate  adaptation rate shift (smaller is faster)
+ */
+void aom_encode_cdf_adapt_q15(aom_writer *w, int val, uint16_t *cdf, int n,
+ int *count, int rate) {
+  int i;
+  if (*count == 0) {
+    /* On the first call, we normalize the cdf to (32768 - n). This should
+       eventually be moved to the state init, but for now it makes it much
+       easier to experiment and convert symbols to the Q15 adaptation.*/
+    int ft;
+    ft = cdf[n - 1];
+    for (i = 0; i < n; i++) {
+      cdf[i] = AOM_ICDF(cdf[i]*32768/ft);
+    }
+  }
+  aom_write_cdf(w, val, cdf, n);
+  aom_cdf_adapt_q15(val, cdf, n, count, rate);
+}
+
+/** Encodes a random variable using a "generic" model, assuming that the
+ * distribution is one-sided (zero and up), has a single mode, and decays
+ * exponentially past the model.
+ *
+ * @param [in,out] w     multi-symbol entropy encoder
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @param [in]     integration integration period of ExQ16 (leaky average over
+ * 1<<integration samples)
+ */
+void generic_encode(aom_writer *w, generic_encoder *model, int x,
+ int *ex_q16, int integration) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  lg_q1 = log_ex(*ex_q16);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "%d %d", *ex_q16, lg_q1));
+  /* If expectation is too large, shift x to ensure that
+     all we have past xs=15 is the exponentially decaying tail
+     of the distribution */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16 */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = (x + (1 << shift >> 1)) >> shift;
+  aom_write_symbol_pvq(w, OD_MINI(15, xs), cdf, 16);
+  if (xs >= 15) {
+    int e;
+    unsigned decay;
+    /* Estimate decay based on the assumption that the distribution is close
+       to Laplacian for large values. We should probably have an adaptive
+       estimate instead. Note: The 2* is a kludge that's not fully understood
+       yet. */
+    OD_ASSERT(*ex_q16 < INT_MAX >> 1);
+    e = ((2**ex_q16 >> 8) + (1 << shift >> 1)) >> shift;
+    decay = OD_MAXI(2, OD_MINI(254, 256*e/(e + 256)));
+    /* Encode the tail of the distribution assuming exponential decay. */
+    aom_laplace_encode_special(w, xs - 15, decay);
+  }
+  if (shift != 0) {
+    int special;
+    /* Because of the rounding, there's only half the number of possibilities
+       for xs=0. */
+    special = xs == 0;
+    if (shift - special > 0) {
+      aom_write_literal(w, x - (xs << shift) + (!special << (shift - 1)),
+       shift - special);
+    }
+  }
+  generic_model_update(ex_q16, x, integration);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "enc: %d %d %d %d %d %x", *ex_q16, x, shift, id, xs, enc->rng));
+}
+
+/** Estimates the cost of encoding a value with generic_encode().
+ *
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @return number of bits (approximation)
+ */
+double generic_encode_cost(generic_encoder *model, int x, int *ex_q16) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  int extra;
+  lg_q1 = log_ex(*ex_q16);
+  /* If expectation is too large, shift x to ensure that
+       all we have past xs=15 is the exponentially decaying tail
+       of the distribution */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16 */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = (x + (1 << shift >> 1)) >> shift;
+  extra = 0;
+  if (shift) extra = shift - (xs == 0);
+  xs = OD_MINI(15, xs);
+  /* Shortcut: assume it's going to cost 2 bits for the Laplace coder. */
+  if (xs == 15) extra += 2;
+  return
+      extra - OD_LOG2((double)(cdf[xs] - (xs == 0 ? 0 : cdf[xs - 1]))/cdf[15]);
+}
+
+/*Estimates the cost of encoding a value with a given CDF.*/
+double od_encode_cdf_cost(int val, uint16_t *cdf, int n) {
+  int total_prob;
+  int prev_prob;
+  double val_prob;
+  OD_ASSERT(n > 0);
+  total_prob = cdf[n - 1];
+  if (val == 0) {
+    prev_prob = 0;
+  }
+  else {
+    prev_prob = cdf[val - 1];
+  }
+  val_prob = (cdf[val] - prev_prob) / (double)total_prob;
+  return -OD_LOG2(val_prob);
+}

third_party/aom/av1/encoder/global_motion.c

@@ -0,0 +1,319 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "av1/encoder/global_motion.h"
+
+#include "av1/common/warped_motion.h"
+
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/corner_detect.h"
+#include "av1/encoder/corner_match.h"
+#include "av1/encoder/ransac.h"
+
+#define MAX_CORNERS 4096
+#define MIN_INLIER_PROB 0.1
+
+#define MIN_TRANS_THRESH (1 * GM_TRANS_DECODE_FACTOR)
+
+// Border over which to compute the global motion
+#define ERRORADV_BORDER 0
+
+#define ERRORADV_MAX_THRESH 0.995
+#define ERRORADV_COST_PRODUCT_THRESH 26000
+
+int is_enough_erroradvantage(double best_erroradvantage, int params_cost) {
+  return best_erroradvantage < ERRORADV_MAX_THRESH &&
+         best_erroradvantage * params_cost < ERRORADV_COST_PRODUCT_THRESH;
+}
+
+static void convert_to_params(const double *params, int32_t *model) {
+  int i;
+  int alpha_present = 0;
+  model[0] = (int32_t)floor(params[0] * (1 << GM_TRANS_PREC_BITS) + 0.5);
+  model[1] = (int32_t)floor(params[1] * (1 << GM_TRANS_PREC_BITS) + 0.5);
+  model[0] = (int32_t)clamp(model[0], GM_TRANS_MIN, GM_TRANS_MAX) *
+             GM_TRANS_DECODE_FACTOR;
+  model[1] = (int32_t)clamp(model[1], GM_TRANS_MIN, GM_TRANS_MAX) *
+             GM_TRANS_DECODE_FACTOR;
+
+  for (i = 2; i < 6; ++i) {
+    const int diag_value = ((i == 2 || i == 5) ? (1 << GM_ALPHA_PREC_BITS) : 0);
+    model[i] = (int32_t)floor(params[i] * (1 << GM_ALPHA_PREC_BITS) + 0.5);
+    model[i] =
+        (int32_t)clamp(model[i] - diag_value, GM_ALPHA_MIN, GM_ALPHA_MAX);
+    alpha_present |= (model[i] != 0);
+    model[i] = (model[i] + diag_value) * GM_ALPHA_DECODE_FACTOR;
+  }
+  for (; i < 8; ++i) {
+    model[i] = (int32_t)floor(params[i] * (1 << GM_ROW3HOMO_PREC_BITS) + 0.5);
+    model[i] = (int32_t)clamp(model[i], GM_ROW3HOMO_MIN, GM_ROW3HOMO_MAX) *
+               GM_ROW3HOMO_DECODE_FACTOR;
+    alpha_present |= (model[i] != 0);
+  }
+
+  if (!alpha_present) {
+    if (abs(model[0]) < MIN_TRANS_THRESH && abs(model[1]) < MIN_TRANS_THRESH) {
+      model[0] = 0;
+      model[1] = 0;
+    }
+  }
+}
+
+void convert_model_to_params(const double *params, WarpedMotionParams *model) {
+  convert_to_params(params, model->wmmat);
+  model->wmtype = get_gmtype(model);
+}
+
+// Adds some offset to a global motion parameter and handles
+// all of the necessary precision shifts, clamping, and
+// zero-centering.
+static int32_t add_param_offset(int param_index, int32_t param_value,
+                                int32_t offset) {
+  const int scale_vals[3] = { GM_TRANS_PREC_DIFF, GM_ALPHA_PREC_DIFF,
+                              GM_ROW3HOMO_PREC_DIFF };
+  const int clamp_vals[3] = { GM_TRANS_MAX, GM_ALPHA_MAX, GM_ROW3HOMO_MAX };
+  // type of param: 0 - translation, 1 - affine, 2 - homography
+  const int param_type = (param_index < 2 ? 0 : (param_index < 6 ? 1 : 2));
+  const int is_one_centered = (param_index == 2 || param_index == 5);
+
+  // Make parameter zero-centered and offset the shift that was done to make
+  // it compatible with the warped model
+  param_value = (param_value - (is_one_centered << WARPEDMODEL_PREC_BITS)) >>
+                scale_vals[param_type];
+  // Add desired offset to the rescaled/zero-centered parameter
+  param_value += offset;
+  // Clamp the parameter so it does not overflow the number of bits allotted
+  // to it in the bitstream
+  param_value = (int32_t)clamp(param_value, -clamp_vals[param_type],
+                               clamp_vals[param_type]);
+  // Rescale the parameter to WARPEDMODEL_PRECISION_BITS so it is compatible
+  // with the warped motion library
+  param_value *= (1 << scale_vals[param_type]);
+
+  // Undo the zero-centering step if necessary
+  return param_value + (is_one_centered << WARPEDMODEL_PREC_BITS);
+}
+
+static void force_wmtype(WarpedMotionParams *wm, TransformationType wmtype) {
+  switch (wmtype) {
+    case IDENTITY: wm->wmmat[0] = 0; wm->wmmat[1] = 0;
+    case TRANSLATION:
+      wm->wmmat[2] = 1 << WARPEDMODEL_PREC_BITS;
+      wm->wmmat[3] = 0;
+    case ROTZOOM: wm->wmmat[4] = -wm->wmmat[3]; wm->wmmat[5] = wm->wmmat[2];
+    case AFFINE: wm->wmmat[6] = wm->wmmat[7] = 0; break;
+    case HORTRAPEZOID: wm->wmmat[6] = wm->wmmat[4] = 0; break;
+    case VERTRAPEZOID: wm->wmmat[7] = wm->wmmat[3] = 0; break;
+    case HOMOGRAPHY: break;
+    default: assert(0);
+  }
+  wm->wmtype = wmtype;
+}
+
+double refine_integerized_param(WarpedMotionParams *wm,
+                                TransformationType wmtype,
+#if CONFIG_HIGHBITDEPTH
+                                int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                uint8_t *ref, int r_width, int r_height,
+                                int r_stride, uint8_t *dst, int d_width,
+                                int d_height, int d_stride, int n_refinements) {
+  static const int max_trans_model_params[TRANS_TYPES] = {
+    0, 2, 4, 6, 8, 8, 8
+  };
+  const int border = ERRORADV_BORDER;
+  int i = 0, p;
+  int n_params = max_trans_model_params[wmtype];
+  int32_t *param_mat = wm->wmmat;
+  double step_error;
+  int32_t step;
+  int32_t *param;
+  int32_t curr_param;
+  int32_t best_param;
+  double best_error;
+
+  force_wmtype(wm, wmtype);
+  best_error = av1_warp_erroradv(wm,
+#if CONFIG_HIGHBITDEPTH
+                                 use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                 ref, r_width, r_height, r_stride,
+                                 dst + border * d_stride + border, border,
+                                 border, d_width - 2 * border,
+                                 d_height - 2 * border, d_stride, 0, 0, 16, 16);
+  step = 1 << (n_refinements + 1);
+  for (i = 0; i < n_refinements; i++, step >>= 1) {
+    for (p = 0; p < n_params; ++p) {
+      int step_dir = 0;
+      // Skip searches for parameters that are forced to be 0
+      if (wmtype == HORTRAPEZOID && (p == 4 || p == 6)) continue;
+      if (wmtype == VERTRAPEZOID && (p == 3 || p == 7)) continue;
+      param = param_mat + p;
+      curr_param = *param;
+      best_param = curr_param;
+      // look to the left
+      *param = add_param_offset(p, curr_param, -step);
+      step_error = av1_warp_erroradv(
+          wm,
+#if CONFIG_HIGHBITDEPTH
+          use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+          ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+          border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
+          0, 0, 16, 16);
+      if (step_error < best_error) {
+        best_error = step_error;
+        best_param = *param;
+        step_dir = -1;
+      }
+
+      // look to the right
+      *param = add_param_offset(p, curr_param, step);
+      step_error = av1_warp_erroradv(
+          wm,
+#if CONFIG_HIGHBITDEPTH
+          use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+          ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+          border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
+          0, 0, 16, 16);
+      if (step_error < best_error) {
+        best_error = step_error;
+        best_param = *param;
+        step_dir = 1;
+      }
+      *param = best_param;
+
+      // look to the direction chosen above repeatedly until error increases
+      // for the biggest step size
+      while (step_dir) {
+        *param = add_param_offset(p, best_param, step * step_dir);
+        step_error = av1_warp_erroradv(
+            wm,
+#if CONFIG_HIGHBITDEPTH
+            use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+            ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+            border, border, d_width - 2 * border, d_height - 2 * border,
+            d_stride, 0, 0, 16, 16);
+        if (step_error < best_error) {
+          best_error = step_error;
+          best_param = *param;
+        } else {
+          *param = best_param;
+          step_dir = 0;
+        }
+      }
+    }
+  }
+  force_wmtype(wm, wmtype);
+  wm->wmtype = get_gmtype(wm);
+  return best_error;
+}
+
+static INLINE RansacFunc get_ransac_type(TransformationType type) {
+  switch (type) {
+    case HOMOGRAPHY: return ransac_homography;
+    case HORTRAPEZOID: return ransac_hortrapezoid;
+    case VERTRAPEZOID: return ransac_vertrapezoid;
+    case AFFINE: return ransac_affine;
+    case ROTZOOM: return ransac_rotzoom;
+    case TRANSLATION: return ransac_translation;
+    default: assert(0); return NULL;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static unsigned char *downconvert_frame(YV12_BUFFER_CONFIG *frm,
+                                        int bit_depth) {
+  int i, j;
+  uint16_t *orig_buf = CONVERT_TO_SHORTPTR(frm->y_buffer);
+  uint8_t *buf = malloc(frm->y_height * frm->y_stride * sizeof(*buf));
+
+  for (i = 0; i < frm->y_height; ++i)
+    for (j = 0; j < frm->y_width; ++j)
+      buf[i * frm->y_stride + j] =
+          orig_buf[i * frm->y_stride + j] >> (bit_depth - 8);
+
+  return buf;
+}
+#endif
+
+int compute_global_motion_feature_based(
+    TransformationType type, YV12_BUFFER_CONFIG *frm, YV12_BUFFER_CONFIG *ref,
+#if CONFIG_HIGHBITDEPTH
+    int bit_depth,
+#endif
+    int *num_inliers_by_motion, double *params_by_motion, int num_motions) {
+  int i;
+  int num_frm_corners, num_ref_corners;
+  int num_correspondences;
+  int *correspondences;
+  int frm_corners[2 * MAX_CORNERS], ref_corners[2 * MAX_CORNERS];
+  unsigned char *frm_buffer = frm->y_buffer;
+  unsigned char *ref_buffer = ref->y_buffer;
+  RansacFunc ransac = get_ransac_type(type);
+
+#if CONFIG_HIGHBITDEPTH
+  if (frm->flags & YV12_FLAG_HIGHBITDEPTH) {
+    // The frame buffer is 16-bit, so we need to convert to 8 bits for the
+    // following code. We cache the result until the frame is released.
+    if (frm->y_buffer_8bit)
+      frm_buffer = frm->y_buffer_8bit;
+    else
+      frm_buffer = frm->y_buffer_8bit = downconvert_frame(frm, bit_depth);
+  }
+  if (ref->flags & YV12_FLAG_HIGHBITDEPTH) {
+    if (ref->y_buffer_8bit)
+      ref_buffer = ref->y_buffer_8bit;
+    else
+      ref_buffer = ref->y_buffer_8bit = downconvert_frame(ref, bit_depth);
+  }
+#endif
+
+  // compute interest points in images using FAST features
+  num_frm_corners = fast_corner_detect(frm_buffer, frm->y_width, frm->y_height,
+                                       frm->y_stride, frm_corners, MAX_CORNERS);
+  num_ref_corners = fast_corner_detect(ref_buffer, ref->y_width, ref->y_height,
+                                       ref->y_stride, ref_corners, MAX_CORNERS);
+
+  // find correspondences between the two images
+  correspondences =
+      (int *)malloc(num_frm_corners * 4 * sizeof(*correspondences));
+  num_correspondences = determine_correspondence(
+      frm_buffer, (int *)frm_corners, num_frm_corners, ref_buffer,
+      (int *)ref_corners, num_ref_corners, frm->y_width, frm->y_height,
+      frm->y_stride, ref->y_stride, correspondences);
+
+  ransac(correspondences, num_correspondences, num_inliers_by_motion,
+         params_by_motion, num_motions);
+
+  free(correspondences);
+
+  // Set num_inliers = 0 for motions with too few inliers so they are ignored.
+  for (i = 0; i < num_motions; ++i) {
+    if (num_inliers_by_motion[i] < MIN_INLIER_PROB * num_correspondences) {
+      num_inliers_by_motion[i] = 0;
+    }
+  }
+
+  // Return true if any one of the motions has inliers.
+  for (i = 0; i < num_motions; ++i) {
+    if (num_inliers_by_motion[i] > 0) return 1;
+  }
+  return 0;
+}

third_party/aom/av1/encoder/global_motion.h

@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_GLOBAL_MOTION_H_
+#define AV1_ENCODER_GLOBAL_MOTION_H_
+
+#include "aom/aom_integer.h"
+#include "aom_scale/yv12config.h"
+#include "av1/common/mv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RANSAC_NUM_MOTIONS 1
+
+void convert_model_to_params(const double *params, WarpedMotionParams *model);
+
+int is_enough_erroradvantage(double erroradv, int params_cost);
+
+double refine_integerized_param(WarpedMotionParams *wm,
+                                TransformationType wmtype,
+#if CONFIG_HIGHBITDEPTH
+                                int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                uint8_t *ref, int r_width, int r_height,
+                                int r_stride, uint8_t *dst, int d_width,
+                                int d_height, int d_stride, int n_refinements);
+
+/*
+  Computes "num_motions" candidate global motion parameters between two frames.
+  The array "params_by_motion" should be length 8 * "num_motions". The ordering
+  of each set of parameters is best described  by the homography:
+
+        [x'     (m2 m3 m0   [x
+    z .  y'  =   m4 m5 m1 *  y
+         1]      m6 m7 1)    1]
+
+  where m{i} represents the ith value in any given set of parameters.
+
+  "num_inliers" should be length "num_motions", and will be populated with the
+  number of inlier feature points for each motion. Params for which the
+  num_inliers entry is 0 should be ignored by the caller.
+*/
+int compute_global_motion_feature_based(
+    TransformationType type, YV12_BUFFER_CONFIG *frm, YV12_BUFFER_CONFIG *ref,
+#if CONFIG_HIGHBITDEPTH
+    int bit_depth,
+#endif
+    int *num_inliers_by_motion, double *params_by_motion, int num_motions);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_ENCODER_GLOBAL_MOTION_H_

third_party/aom/av1/encoder/hybrid_fwd_txfm.c

@@ -0,0 +1,499 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "av1/common/idct.h"
+#include "av1/encoder/hybrid_fwd_txfm.h"
+
+#if CONFIG_CB4X4
+static void fwd_txfm_2x2(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type, int lossless) {
+  tran_high_t a1 = src_diff[0];
+  tran_high_t b1 = src_diff[1];
+  tran_high_t c1 = src_diff[diff_stride];
+  tran_high_t d1 = src_diff[1 + diff_stride];
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  a1 = a2 + b2;
+  b1 = a2 - b2;
+  c1 = c2 + d2;
+  d1 = c2 - d2;
+
+  coeff[0] = (tran_low_t)(4 * a1);
+  coeff[1] = (tran_low_t)(4 * b1);
+  coeff[2] = (tran_low_t)(4 * c1);
+  coeff[3] = (tran_low_t)(4 * d1);
+
+  (void)tx_type;
+  (void)lossless;
+}
+#endif
+
+static void fwd_txfm_4x4(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type, int lossless) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_fwht4x4(src_diff, coeff, diff_stride);
+    return;
+  }
+
+  av1_fht4x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_4x8(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht4x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x4(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x16(const int16_t *src_diff, tran_low_t *coeff,
+                          int diff_stride, TX_TYPE tx_type,
+                          FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x8(const int16_t *src_diff, tran_low_t *coeff,
+                          int diff_stride, TX_TYPE tx_type,
+                          FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x32(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_32x16(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht32x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x8(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x16(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_32x32(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht32x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+#if CONFIG_TX64X64
+static void fwd_txfm_64x64(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+#if CONFIG_EXT_TX
+  if (tx_type == IDTX)
+    av1_fwd_idtx_c(src_diff, coeff, diff_stride, 64, tx_type);
+  else
+#endif
+    av1_fht64x64(src_diff, coeff, diff_stride, tx_type);
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_CB4X4
+static void highbd_fwd_txfm_2x2(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type, int lossless,
+                                const int bd) {
+  tran_high_t a1 = src_diff[0];
+  tran_high_t b1 = src_diff[1];
+  tran_high_t c1 = src_diff[diff_stride];
+  tran_high_t d1 = src_diff[1 + diff_stride];
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  a1 = a2 + b2;
+  b1 = a2 - b2;
+  c1 = c2 + d2;
+  d1 = c2 - d2;
+
+  coeff[0] = (tran_low_t)(4 * a1);
+  coeff[1] = (tran_low_t)(4 * b1);
+  coeff[2] = (tran_low_t)(4 * c1);
+  coeff[3] = (tran_low_t)(4 * d1);
+
+  (void)tx_type;
+  (void)lossless;
+  (void)bd;
+}
+#endif
+
+static void highbd_fwd_txfm_4x4(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type, int lossless,
+                                const int bd) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_highbd_fwht4x4(src_diff, coeff, diff_stride);
+    return;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_4x4(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_4x4(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht4x4_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 4, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_4x8(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht4x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x4(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht8x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x16(const int16_t *src_diff, tran_low_t *coeff,
+                                 int diff_stride, TX_TYPE tx_type,
+                                 FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht8x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_16x8(const int16_t *src_diff, tran_low_t *coeff,
+                                 int diff_stride, TX_TYPE tx_type,
+                                 FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht16x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_16x32(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht16x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_32x16(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht32x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x8(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_8x8(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_8x8(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST exists only in C
+      av1_highbd_fht8x8_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 8, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_16x16(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_16x16(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_16x16(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST exists only in C
+      av1_highbd_fht16x16_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 16, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_32x32(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_fwd_txfm2d_32x32(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht32x32_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 32, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_TX64X64
+static void highbd_fwd_txfm_64x64(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_highbd_fht64x64(src_diff, coeff, diff_stride, tx_type);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht64x64(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 64, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff, int diff_stride,
+                  FWD_TXFM_PARAM *fwd_txfm_param) {
+  const int fwd_txfm_opt = FWD_TXFM_OPT_NORMAL;
+  const TX_TYPE tx_type = fwd_txfm_param->tx_type;
+  const TX_SIZE tx_size = fwd_txfm_param->tx_size;
+  const int lossless = fwd_txfm_param->lossless;
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64:
+      fwd_txfm_64x64(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32:
+      fwd_txfm_32x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X16:
+      fwd_txfm_16x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X8:
+      fwd_txfm_8x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_4X8:
+      fwd_txfm_4x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X4:
+      fwd_txfm_8x4(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X16:
+      fwd_txfm_8x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X8:
+      fwd_txfm_16x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X32:
+      fwd_txfm_16x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_32X16:
+      fwd_txfm_32x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_4X4:
+      fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, lossless);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      fwd_txfm_2x2(src_diff, coeff, diff_stride, tx_type, lossless);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, FWD_TXFM_PARAM *fwd_txfm_param) {
+  const int fwd_txfm_opt = FWD_TXFM_OPT_NORMAL;
+  const TX_TYPE tx_type = fwd_txfm_param->tx_type;
+  const TX_SIZE tx_size = fwd_txfm_param->tx_size;
+  const int lossless = fwd_txfm_param->lossless;
+  const int bd = fwd_txfm_param->bd;
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64:
+      highbd_fwd_txfm_64x64(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32:
+      highbd_fwd_txfm_32x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_16X16:
+      highbd_fwd_txfm_16x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_8X8:
+      highbd_fwd_txfm_8x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_4X8:
+      highbd_fwd_txfm_4x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_8X4:
+      highbd_fwd_txfm_8x4(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_8X16:
+      highbd_fwd_txfm_8x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                           bd);
+      break;
+    case TX_16X8:
+      highbd_fwd_txfm_16x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                           bd);
+      break;
+    case TX_16X32:
+      highbd_fwd_txfm_16x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_32X16:
+      highbd_fwd_txfm_32x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_4X4:
+      highbd_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, lossless, bd);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      highbd_fwd_txfm_2x2(src_diff, coeff, diff_stride, tx_type, lossless, bd);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/av1/encoder/hybrid_fwd_txfm.h

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_HYBRID_FWD_TXFM_H_
+#define AV1_ENCODER_HYBRID_FWD_TXFM_H_
+
+#include "./aom_config.h"
+
+typedef enum FWD_TXFM_OPT { FWD_TXFM_OPT_NORMAL } FWD_TXFM_OPT;
+
+typedef struct FWD_TXFM_PARAM {
+  TX_TYPE tx_type;
+  TX_SIZE tx_size;
+  int lossless;
+#if CONFIG_HIGHBITDEPTH
+  int bd;
+#endif  // CONFIG_HIGHBITDEPTH
+} FWD_TXFM_PARAM;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff, int diff_stride,
+                  FWD_TXFM_PARAM *fwd_txfm_param);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, FWD_TXFM_PARAM *fwd_txfm_param);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_HYBRID_FWD_TXFM_H_

third_party/aom/av1/encoder/laplace_encoder.c

@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitwriter.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/pvq.h"
+#include "pvq_encoder.h"
+
+static void aom_encode_pvq_split(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ int count, int sum, int ctx) {
+  int shift;
+  int rest;
+  int fctx;
+  if (sum == 0) return;
+  shift = OD_MAXI(0, OD_ILOG(sum) - 3);
+  if (shift) {
+    rest = count & ((1 << shift) - 1);
+    count >>= shift;
+    sum >>= shift;
+  }
+  fctx = 7*ctx + sum - 1;
+  aom_write_symbol_pvq(w, count, adapt->pvq_split_cdf[fctx], sum + 1);
+  if (shift) aom_write_literal(w, rest, shift);
+}
+
+void aom_encode_band_pvq_splits(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const int *y, int n, int k, int level) {
+  int mid;
+  int i;
+  int count_right;
+  if (n <= 1 || k == 0) return;
+  if (k == 1 && n <= 16) {
+    int cdf_id;
+    int pos;
+    cdf_id = od_pvq_k1_ctx(n, level == 0);
+    for (pos = 0; !y[pos]; pos++);
+    OD_ASSERT(pos < n);
+    aom_write_symbol_pvq(w, pos, adapt->pvq_k1_cdf[cdf_id], n);
+  }
+  else {
+    mid = n >> 1;
+    count_right = k;
+    for (i = 0; i < mid; i++) count_right -= abs(y[i]);
+    aom_encode_pvq_split(w, adapt, count_right, k, od_pvq_size_ctx(n));
+    aom_encode_band_pvq_splits(w, adapt, y, mid, k - count_right, level + 1);
+    aom_encode_band_pvq_splits(w, adapt, y + mid, n - mid, count_right,
+     level + 1);
+  }
+}
+
+/** Encodes the tail of a Laplace-distributed variable, i.e. it doesn't
+ * do anything special for the zero case.
+ *
+ * @param [in,out] enc     range encoder
+ * @param [in]     x       variable to encode (has to be positive)
+ * @param [in]     decay   decay factor of the distribution in Q8 format,
+ * i.e. pdf ~= decay^x
+ */
+void aom_laplace_encode_special(aom_writer *w, int x, unsigned decay) {
+  int shift;
+  int xs;
+  int sym;
+  const uint16_t *cdf;
+  shift = 0;
+  /* We don't want a large decay value because that would require too many
+     symbols. */
+  while (decay > 235) {
+    decay = (decay*decay + 128) >> 8;
+    shift++;
+  }
+  decay = OD_MINI(decay, 254);
+  decay = OD_MAXI(decay, 2);
+  xs = x >> shift;
+  cdf = EXP_CDF_TABLE[(decay + 1) >> 1];
+  OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "decay = %d", decay));
+  do {
+    sym = OD_MINI(xs, 15);
+    {
+      int i;
+      OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "%d %d %d %d %d\n", x, xs, shift,
+       sym, max));
+      for (i = 0; i < 16; i++) {
+        OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "%d ", cdf[i]));
+      }
+      OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "\n"));
+    }
+    aom_write_cdf(w, sym, cdf, 16);
+    xs -= 15;
+  } while (sym >= 15);
+  if (shift) aom_write_literal(w, x & ((1 << shift) - 1), shift);
+}

third_party/aom/av1/encoder/lookahead.c

@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+
+#include "av1/common/common.h"
+
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/lookahead.h"
+
+/* Return the buffer at the given absolute index and increment the index */
+static struct lookahead_entry *pop(struct lookahead_ctx *ctx, int *idx) {
+  int index = *idx;
+  struct lookahead_entry *buf = ctx->buf + index;
+
+  assert(index < ctx->max_sz);
+  if (++index >= ctx->max_sz) index -= ctx->max_sz;
+  *idx = index;
+  return buf;
+}
+
+void av1_lookahead_destroy(struct lookahead_ctx *ctx) {
+  if (ctx) {
+    if (ctx->buf) {
+      int i;
+
+      for (i = 0; i < ctx->max_sz; i++) aom_free_frame_buffer(&ctx->buf[i].img);
+      free(ctx->buf);
+    }
+    free(ctx);
+  }
+}
+
+struct lookahead_ctx *av1_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth) {
+  struct lookahead_ctx *ctx = NULL;
+
+  // Clamp the lookahead queue depth
+  depth = clamp(depth, 1, MAX_LAG_BUFFERS);
+
+  // Allocate memory to keep previous source frames available.
+  depth += MAX_PRE_FRAMES;
+
+  // Allocate the lookahead structures
+  ctx = calloc(1, sizeof(*ctx));
+  if (ctx) {
+    const int legacy_byte_alignment = 0;
+    unsigned int i;
+    ctx->max_sz = depth;
+    ctx->buf = calloc(depth, sizeof(*ctx->buf));
+    if (!ctx->buf) goto bail;
+    for (i = 0; i < depth; i++)
+      if (aom_alloc_frame_buffer(&ctx->buf[i].img, width, height, subsampling_x,
+                                 subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 AOM_BORDER_IN_PIXELS, legacy_byte_alignment))
+        goto bail;
+  }
+  return ctx;
+bail:
+  av1_lookahead_destroy(ctx);
+  return NULL;
+}
+
+#define USE_PARTIAL_COPY 0
+
+int av1_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       aom_enc_frame_flags_t flags) {
+  struct lookahead_entry *buf;
+#if USE_PARTIAL_COPY
+  int row, col, active_end;
+  int mb_rows = (src->y_height + 15) >> 4;
+  int mb_cols = (src->y_width + 15) >> 4;
+#endif
+  int width = src->y_crop_width;
+  int height = src->y_crop_height;
+  int uv_width = src->uv_crop_width;
+  int uv_height = src->uv_crop_height;
+  int subsampling_x = src->subsampling_x;
+  int subsampling_y = src->subsampling_y;
+  int larger_dimensions, new_dimensions;
+
+  if (ctx->sz + 1 + MAX_PRE_FRAMES > ctx->max_sz) return 1;
+  ctx->sz++;
+  buf = pop(ctx, &ctx->write_idx);
+
+  new_dimensions = width != buf->img.y_crop_width ||
+                   height != buf->img.y_crop_height ||
+                   uv_width != buf->img.uv_crop_width ||
+                   uv_height != buf->img.uv_crop_height;
+  larger_dimensions = width > buf->img.y_width || height > buf->img.y_height ||
+                      uv_width > buf->img.uv_width ||
+                      uv_height > buf->img.uv_height;
+  assert(!larger_dimensions || new_dimensions);
+
+#if USE_PARTIAL_COPY
+  // TODO(jkoleszar): This is disabled for now, as
+  // av1_copy_and_extend_frame_with_rect is not subsampling/alpha aware.
+
+  // Only do this partial copy if the following conditions are all met:
+  // 1. Lookahead queue has has size of 1.
+  // 2. Active map is provided.
+  // 3. This is not a key frame, golden nor altref frame.
+  if (!new_dimensions && ctx->max_sz == 1 && active_map && !flags) {
+    for (row = 0; row < mb_rows; ++row) {
+      col = 0;
+
+      while (1) {
+        // Find the first active macroblock in this row.
+        for (; col < mb_cols; ++col) {
+          if (active_map[col]) break;
+        }
+
+        // No more active macroblock in this row.
+        if (col == mb_cols) break;
+
+        // Find the end of active region in this row.
+        active_end = col;
+
+        for (; active_end < mb_cols; ++active_end) {
+          if (!active_map[active_end]) break;
+        }
+
+        // Only copy this active region.
+        av1_copy_and_extend_frame_with_rect(src, &buf->img, row << 4, col << 4,
+                                            16, (active_end - col) << 4);
+
+        // Start again from the end of this active region.
+        col = active_end;
+      }
+
+      active_map += mb_cols;
+    }
+  } else {
+#endif
+    if (larger_dimensions) {
+      YV12_BUFFER_CONFIG new_img;
+      memset(&new_img, 0, sizeof(new_img));
+      if (aom_alloc_frame_buffer(&new_img, width, height, subsampling_x,
+                                 subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 AOM_BORDER_IN_PIXELS, 0))
+        return 1;
+      aom_free_frame_buffer(&buf->img);
+      buf->img = new_img;
+    } else if (new_dimensions) {
+      buf->img.y_crop_width = src->y_crop_width;
+      buf->img.y_crop_height = src->y_crop_height;
+      buf->img.uv_crop_width = src->uv_crop_width;
+      buf->img.uv_crop_height = src->uv_crop_height;
+      buf->img.subsampling_x = src->subsampling_x;
+      buf->img.subsampling_y = src->subsampling_y;
+    }
+    // Partial copy not implemented yet
+    av1_copy_and_extend_frame(src, &buf->img);
+#if USE_PARTIAL_COPY
+  }
+#endif
+
+  buf->ts_start = ts_start;
+  buf->ts_end = ts_end;
+  buf->flags = flags;
+  return 0;
+}
+
+struct lookahead_entry *av1_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain) {
+  struct lookahead_entry *buf = NULL;
+
+  if (ctx && ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = pop(ctx, &ctx->read_idx);
+    ctx->sz--;
+  }
+  return buf;
+}
+
+struct lookahead_entry *av1_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index) {
+  struct lookahead_entry *buf = NULL;
+
+  if (index >= 0) {
+    // Forward peek
+    if (index < ctx->sz) {
+      index += ctx->read_idx;
+      if (index >= ctx->max_sz) index -= ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  } else if (index < 0) {
+    // Backward peek
+    if (-index <= MAX_PRE_FRAMES) {
+      index += (int)(ctx->read_idx);
+      if (index < 0) index += (int)(ctx->max_sz);
+      buf = ctx->buf + index;
+    }
+  }
+
+  return buf;
+}
+
+unsigned int av1_lookahead_depth(struct lookahead_ctx *ctx) { return ctx->sz; }

third_party/aom/av1/encoder/lookahead.h

@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_LOOKAHEAD_H_
+#define AV1_ENCODER_LOOKAHEAD_H_
+
+#include "aom_scale/yv12config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LAG_BUFFERS 25
+
+struct lookahead_entry {
+  YV12_BUFFER_CONFIG img;
+  int64_t ts_start;
+  int64_t ts_end;
+  aom_enc_frame_flags_t flags;
+};
+
+// The max of past frames we want to keep in the queue.
+#define MAX_PRE_FRAMES 1
+
+struct lookahead_ctx {
+  int max_sz;                  /* Absolute size of the queue */
+  int sz;                      /* Number of buffers currently in the queue */
+  int read_idx;                /* Read index */
+  int write_idx;               /* Write index */
+  struct lookahead_entry *buf; /* Buffer list */
+};
+
+/**\brief Initializes the lookahead stage
+ *
+ * The lookahead stage is a queue of frame buffers on which some analysis
+ * may be done when buffers are enqueued.
+ */
+struct lookahead_ctx *av1_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth);
+
+/**\brief Destroys the lookahead stage
+ */
+void av1_lookahead_destroy(struct lookahead_ctx *ctx);
+
+/**\brief Enqueue a source buffer
+ *
+ * This function will copy the source image into a new framebuffer with
+ * the expected stride/border.
+ *
+ * If active_map is non-NULL and there is only one frame in the queue, then copy
+ * only active macroblocks.
+ *
+ * \param[in] ctx         Pointer to the lookahead context
+ * \param[in] src         Pointer to the image to enqueue
+ * \param[in] ts_start    Timestamp for the start of this frame
+ * \param[in] ts_end      Timestamp for the end of this frame
+ * \param[in] flags       Flags set on this frame
+ * \param[in] active_map  Map that specifies which macroblock is active
+ */
+int av1_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       aom_enc_frame_flags_t flags);
+
+/**\brief Get the next source buffer to encode
+ *
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] drain     Flag indicating the buffer should be drained
+ *                      (return a buffer regardless of the current queue depth)
+ *
+ * \retval NULL, if drain set and queue is empty
+ * \retval NULL, if drain not set and queue not of the configured depth
+ */
+struct lookahead_entry *av1_lookahead_pop(struct lookahead_ctx *ctx, int drain);
+
+/**\brief Get a future source buffer to encode
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] index     Index of the frame to be returned, 0 == next frame
+ *
+ * \retval NULL, if no buffer exists at the specified index
+ */
+struct lookahead_entry *av1_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index);
+
+/**\brief Get the number of frames currently in the lookahead queue
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ */
+unsigned int av1_lookahead_depth(struct lookahead_ctx *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_LOOKAHEAD_H_

third_party/aom/av1/encoder/mbgraph.c

@@ -0,0 +1,398 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/system_state.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/common/blockd.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+
+static unsigned int do_16x16_motion_iteration(AV1_COMP *cpi, const MV *ref_mv,
+                                              int mb_row, int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const aom_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+
+  const MvLimits tmp_mv_limits = x->mv_limits;
+  MV ref_full;
+  int cost_list[5];
+
+  // Further step/diamond searches as necessary
+  int step_param = mv_sf->reduce_first_step_size;
+  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  av1_set_mv_search_range(&x->mv_limits, ref_mv);
+
+  ref_full.col = ref_mv->col >> 3;
+  ref_full.row = ref_mv->row >> 3;
+
+  /*cpi->sf.search_method == HEX*/
+  av1_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
+                 cond_cost_list(cpi, cost_list), &v_fn_ptr, 0, ref_mv);
+
+  // Try sub-pixel MC
+  // if (bestsme > error_thresh && bestsme < INT_MAX)
+  {
+    int distortion;
+    unsigned int sse;
+    cpi->find_fractional_mv_step(x, ref_mv, cpi->common.allow_high_precision_mv,
+                                 x->errorperbit, &v_fn_ptr, 0,
+                                 mv_sf->subpel_iters_per_step,
+                                 cond_cost_list(cpi, cost_list), NULL, NULL,
+                                 &distortion, &sse, NULL, 0, 0, 0);
+  }
+
+#if CONFIG_EXT_INTER
+  if (has_second_ref(&xd->mi[0]->mbmi))
+    xd->mi[0]->mbmi.mode = NEW_NEWMV;
+  else
+#endif  // CONFIG_EXT_INTER
+    xd->mi[0]->mbmi.mode = NEWMV;
+
+  xd->mi[0]->mbmi.mv[0] = x->best_mv;
+#if CONFIG_EXT_INTER
+  xd->mi[0]->mbmi.ref_frame[1] = NONE_FRAME;
+#endif  // CONFIG_EXT_INTER
+
+  av1_build_inter_predictors_sby(xd, mb_row, mb_col, NULL, BLOCK_16X16);
+
+  /* restore UMV window */
+  x->mv_limits = tmp_mv_limits;
+
+  return aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                      xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+}
+
+static int do_16x16_motion_search(AV1_COMP *cpi, const MV *ref_mv, int mb_row,
+                                  int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err, tmp_err;
+  MV best_mv;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+  best_mv.col = best_mv.row = 0;
+
+  // Test last reference frame using the previous best mv as the
+  // starting point (best reference) for the search
+  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, mb_row, mb_col);
+  if (tmp_err < err) {
+    err = tmp_err;
+    best_mv = x->best_mv.as_mv;
+  }
+
+  // If the current best reference mv is not centered on 0,0 then do a 0,0
+  // based search as well.
+  if (ref_mv->row != 0 || ref_mv->col != 0) {
+    MV zero_ref_mv = { 0, 0 };
+
+    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, mb_row, mb_col);
+    if (tmp_err < err) {
+      err = tmp_err;
+      best_mv = x->best_mv.as_mv;
+    }
+  }
+
+  x->best_mv.as_mv = best_mv;
+  return err;
+}
+
+static int do_16x16_zerozero_search(AV1_COMP *cpi, int_mv *dst_mv) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+
+  dst_mv->as_int = 0;
+
+  return err;
+}
+static int find_best_16x16_intra(AV1_COMP *cpi, PREDICTION_MODE *pbest_mode) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  PREDICTION_MODE best_mode = -1, mode;
+  unsigned int best_err = INT_MAX;
+
+  // calculate SATD for each intra prediction mode;
+  // we're intentionally not doing 4x4, we just want a rough estimate
+  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+    unsigned int err;
+
+    xd->mi[0]->mbmi.mode = mode;
+    av1_predict_intra_block(xd, 16, 16, BLOCK_16X16, mode, x->plane[0].src.buf,
+                            x->plane[0].src.stride, xd->plane[0].dst.buf,
+                            xd->plane[0].dst.stride, 0, 0, 0);
+    err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+
+    // find best
+    if (err < best_err) {
+      best_err = err;
+      best_mode = mode;
+    }
+  }
+
+  if (pbest_mode) *pbest_mode = best_mode;
+
+  return best_err;
+}
+
+static void update_mbgraph_mb_stats(AV1_COMP *cpi, MBGRAPH_MB_STATS *stats,
+                                    YV12_BUFFER_CONFIG *buf, int mb_y_offset,
+                                    YV12_BUFFER_CONFIG *golden_ref,
+                                    const MV *prev_golden_ref_mv,
+                                    YV12_BUFFER_CONFIG *alt_ref, int mb_row,
+                                    int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int intra_error;
+  AV1_COMMON *cm = &cpi->common;
+
+  // FIXME in practice we're completely ignoring chroma here
+  x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
+  x->plane[0].src.stride = buf->y_stride;
+
+  xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
+  xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
+
+  // do intra 16x16 prediction
+  intra_error = find_best_16x16_intra(cpi, &stats->ref[INTRA_FRAME].m.mode);
+  if (intra_error <= 0) intra_error = 1;
+  stats->ref[INTRA_FRAME].err = intra_error;
+
+  // Golden frame MV search, if it exists and is different than last frame
+  if (golden_ref) {
+    int g_motion_error;
+    xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = golden_ref->y_stride;
+    g_motion_error =
+        do_16x16_motion_search(cpi, prev_golden_ref_mv, mb_row, mb_col);
+    stats->ref[GOLDEN_FRAME].m.mv = x->best_mv;
+    stats->ref[GOLDEN_FRAME].err = g_motion_error;
+  } else {
+    stats->ref[GOLDEN_FRAME].err = INT_MAX;
+    stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
+  }
+
+  // Do an Alt-ref frame MV search, if it exists and is different than
+  // last/golden frame.
+  if (alt_ref) {
+    int a_motion_error;
+    xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = alt_ref->y_stride;
+    a_motion_error =
+        do_16x16_zerozero_search(cpi, &stats->ref[ALTREF_FRAME].m.mv);
+
+    stats->ref[ALTREF_FRAME].err = a_motion_error;
+  } else {
+    stats->ref[ALTREF_FRAME].err = INT_MAX;
+    stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
+  }
+}
+
+static void update_mbgraph_frame_stats(AV1_COMP *cpi,
+                                       MBGRAPH_FRAME_STATS *stats,
+                                       YV12_BUFFER_CONFIG *buf,
+                                       YV12_BUFFER_CONFIG *golden_ref,
+                                       YV12_BUFFER_CONFIG *alt_ref) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  AV1_COMMON *const cm = &cpi->common;
+
+  int mb_col, mb_row, offset = 0;
+  int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
+  MV gld_top_mv = { 0, 0 };
+  MODE_INFO mi_local;
+
+  av1_zero(mi_local);
+  // Set up limit values for motion vectors to prevent them extending outside
+  // the UMV borders.
+  x->mv_limits.row_min = -BORDER_MV_PIXELS_B16;
+  x->mv_limits.row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
+  xd->up_available = 0;
+  xd->plane[0].dst.stride = buf->y_stride;
+  xd->plane[0].pre[0].stride = buf->y_stride;
+  xd->plane[1].dst.stride = buf->uv_stride;
+  xd->mi[0] = &mi_local;
+  mi_local.mbmi.sb_type = BLOCK_16X16;
+  mi_local.mbmi.ref_frame[0] = LAST_FRAME;
+  mi_local.mbmi.ref_frame[1] = NONE_FRAME;
+
+  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
+    MV gld_left_mv = gld_top_mv;
+    int mb_y_in_offset = mb_y_offset;
+    int arf_y_in_offset = arf_y_offset;
+    int gld_y_in_offset = gld_y_offset;
+
+    // Set up limit values for motion vectors to prevent them extending outside
+    // the UMV borders.
+    x->mv_limits.col_min = -BORDER_MV_PIXELS_B16;
+    x->mv_limits.col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
+    xd->left_available = 0;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+      MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
+
+      update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset, golden_ref,
+                              &gld_left_mv, alt_ref, mb_row, mb_col);
+      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
+      if (mb_col == 0) {
+        gld_top_mv = gld_left_mv;
+      }
+      xd->left_available = 1;
+      mb_y_in_offset += 16;
+      gld_y_in_offset += 16;
+      arf_y_in_offset += 16;
+      x->mv_limits.col_min -= 16;
+      x->mv_limits.col_max -= 16;
+    }
+    xd->up_available = 1;
+    mb_y_offset += buf->y_stride * 16;
+    gld_y_offset += golden_ref->y_stride * 16;
+    if (alt_ref) arf_y_offset += alt_ref->y_stride * 16;
+    x->mv_limits.row_min -= 16;
+    x->mv_limits.row_max -= 16;
+    offset += cm->mb_cols;
+  }
+}
+
+// void separate_arf_mbs_byzz
+static void separate_arf_mbs(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int mb_col, mb_row, offset, i;
+  int mi_row, mi_col;
+  int ncnt[4] = { 0 };
+  int n_frames = cpi->mbgraph_n_frames;
+
+  int *arf_not_zz;
+
+  CHECK_MEM_ERROR(
+      cm, arf_not_zz,
+      aom_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz), 1));
+
+  // We are not interested in results beyond the alt ref itself.
+  if (n_frames > cpi->rc.frames_till_gf_update_due)
+    n_frames = cpi->rc.frames_till_gf_update_due;
+
+  // defer cost to reference frames
+  for (i = n_frames - 1; i >= 0; i--) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+
+    for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
+         offset += cm->mb_cols, mb_row++) {
+      for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+        MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
+
+        int altref_err = mb_stats->ref[ALTREF_FRAME].err;
+        int intra_err = mb_stats->ref[INTRA_FRAME].err;
+        int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
+
+        // Test for altref vs intra and gf and that its mv was 0,0.
+        if (altref_err > 1000 || altref_err > intra_err ||
+            altref_err > golden_err) {
+          arf_not_zz[offset + mb_col]++;
+        }
+      }
+    }
+  }
+
+  // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
+  // of bound access in segmentation_map
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      // If any of the blocks in the sequence failed then the MB
+      // goes in segment 0
+      if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
+        ncnt[0]++;
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
+      } else {
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
+        ncnt[1]++;
+      }
+    }
+  }
+
+  // Only bother with segmentation if over 10% of the MBs in static segment
+  // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
+  if (1) {
+    // Note % of blocks that are marked as static
+    if (cm->MBs)
+      cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
+
+    // This error case should not be reachable as this function should
+    // never be called with the common data structure uninitialized.
+    else
+      cpi->static_mb_pct = 0;
+
+    av1_enable_segmentation(&cm->seg);
+  } else {
+    cpi->static_mb_pct = 0;
+    av1_disable_segmentation(&cm->seg);
+  }
+
+  // Free localy allocated storage
+  aom_free(arf_not_zz);
+}
+
+void av1_update_mbgraph_stats(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int i, n_frames = av1_lookahead_depth(cpi->lookahead);
+  YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+
+  assert(golden_ref != NULL);
+
+  // we need to look ahead beyond where the ARF transitions into
+  // being a GF - so exit if we don't look ahead beyond that
+  if (n_frames <= cpi->rc.frames_till_gf_update_due) return;
+
+  if (n_frames > MAX_LAG_BUFFERS) n_frames = MAX_LAG_BUFFERS;
+
+  cpi->mbgraph_n_frames = n_frames;
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    memset(frame_stats->mb_stats, 0,
+           cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
+  }
+
+  // do motion search to find contribution of each reference to data
+  // later on in this GF group
+  // FIXME really, the GF/last MC search should be done forward, and
+  // the ARF MC search backwards, to get optimal results for MV caching
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    struct lookahead_entry *q_cur = av1_lookahead_peek(cpi->lookahead, i);
+
+    assert(q_cur != NULL);
+
+    update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img, golden_ref,
+                               cpi->source);
+  }
+
+  aom_clear_system_state();
+
+  separate_arf_mbs(cpi);
+}

third_party/aom/av1/encoder/mbgraph.h

@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MBGRAPH_H_
+#define AV1_ENCODER_MBGRAPH_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  struct {
+    int err;
+    union {
+      int_mv mv;
+      PREDICTION_MODE mode;
+    } m;
+  } ref[TOTAL_REFS_PER_FRAME];
+} MBGRAPH_MB_STATS;
+
+typedef struct { MBGRAPH_MB_STATS *mb_stats; } MBGRAPH_FRAME_STATS;
+
+struct AV1_COMP;
+
+void av1_update_mbgraph_stats(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_MBGRAPH_H_

third_party/aom/av1/encoder/mcomp.h

@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MCOMP_H_
+#define AV1_ENCODER_MCOMP_H_
+
+#include "av1/encoder/block.h"
+#include "aom_dsp/variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The maximum number of steps in a step search given the largest
+// allowed initial step
+#define MAX_MVSEARCH_STEPS 11
+// Max full pel mv specified in the unit of full pixel
+// Enable the use of motion vector in range [-1023, 1023].
+#define MAX_FULL_PEL_VAL ((1 << (MAX_MVSEARCH_STEPS - 1)) - 1)
+// Maximum size of the first step in full pel units
+#define MAX_FIRST_STEP (1 << (MAX_MVSEARCH_STEPS - 1))
+// Allowed motion vector pixel distance outside image border
+// for Block_16x16
+#define BORDER_MV_PIXELS_B16 (16 + AOM_INTERP_EXTEND)
+
+// motion search site
+typedef struct search_site {
+  MV mv;
+  int offset;
+} search_site;
+
+typedef struct search_site_config {
+  search_site ss[8 * MAX_MVSEARCH_STEPS + 1];
+  int ss_count;
+  int searches_per_step;
+} search_site_config;
+
+void av1_init_dsmotion_compensation(search_site_config *cfg, int stride);
+void av1_init3smotion_compensation(search_site_config *cfg, int stride);
+
+void av1_set_mv_search_range(MvLimits *mv_limits, const MV *mv);
+
+int av1_mv_bit_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                    int *mvcost[2], int weight);
+
+// Utility to compute variance + MV rate cost for a given MV
+int av1_get_mvpred_var(const MACROBLOCK *x, const MV *best_mv,
+                       const MV *center_mv, const aom_variance_fn_ptr_t *vfp,
+                       int use_mvcost);
+int av1_get_mvpred_av_var(const MACROBLOCK *x, const MV *best_mv,
+                          const MV *center_mv, const uint8_t *second_pred,
+                          const aom_variance_fn_ptr_t *vfp, int use_mvcost);
+
+struct AV1_COMP;
+struct SPEED_FEATURES;
+
+int av1_init_search_range(int size);
+
+int av1_refining_search_sad(struct macroblock *x, struct mv *ref_mv,
+                            int sad_per_bit, int distance,
+                            const aom_variance_fn_ptr_t *fn_ptr,
+                            const struct mv *center_mv);
+
+// Runs sequence of diamond searches in smaller steps for RD.
+int av1_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                           MV *mvp_full, int step_param, int sadpb,
+                           int further_steps, int do_refine, int *cost_list,
+                           const aom_variance_fn_ptr_t *fn_ptr,
+                           const MV *ref_mv, MV *dst_mv);
+
+// Perform integral projection based motion estimation.
+unsigned int av1_int_pro_motion_estimation(const struct AV1_COMP *cpi,
+                                           MACROBLOCK *x, BLOCK_SIZE bsize,
+                                           int mi_row, int mi_col);
+
+int av1_hex_search(MACROBLOCK *x, MV *start_mv, int search_param,
+                   int sad_per_bit, int do_init_search, int *cost_list,
+                   const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                   const MV *center_mv);
+
+typedef int(fractional_mv_step_fp)(
+    MACROBLOCK *x, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp,
+    int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
+    int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
+    int *distortion, unsigned int *sse1, const uint8_t *second_pred, int w,
+    int h, int use_upsampled_ref);
+
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned_more;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned_evenmore;
+extern fractional_mv_step_fp av1_return_max_sub_pixel_mv;
+extern fractional_mv_step_fp av1_return_min_sub_pixel_mv;
+
+typedef int (*av1_full_search_fn_t)(const MACROBLOCK *x, const MV *ref_mv,
+                                    int sad_per_bit, int distance,
+                                    const aom_variance_fn_ptr_t *fn_ptr,
+                                    const MV *center_mv, MV *best_mv);
+
+typedef int (*av1_diamond_search_fn_t)(
+    MACROBLOCK *x, const search_site_config *cfg, MV *ref_mv, MV *best_mv,
+    int search_param, int sad_per_bit, int *num00,
+    const aom_variance_fn_ptr_t *fn_ptr, const MV *center_mv);
+
+int av1_refining_search_8p_c(MACROBLOCK *x, int error_per_bit, int search_range,
+                             const aom_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, const uint8_t *second_pred);
+
+struct AV1_COMP;
+
+int av1_full_pixel_search(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full, int step_param,
+                          int error_per_bit, int *cost_list, const MV *ref_mv,
+                          int var_max, int rd);
+
+#if CONFIG_EXT_INTER
+int av1_find_best_masked_sub_pixel_tree(
+    const MACROBLOCK *x, const uint8_t *mask, int mask_stride, MV *bestmv,
+    const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second);
+int av1_find_best_masked_sub_pixel_tree_up(
+    const struct AV1_COMP *cpi, MACROBLOCK *x, const uint8_t *mask,
+    int mask_stride, int mi_row, int mi_col, MV *bestmv, const MV *ref_mv,
+    int allow_hp, int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    int forced_stop, int iters_per_step, int *mvjcost, int *mvcost[2],
+    int *distortion, unsigned int *sse1, int is_second, int use_upsampled_ref);
+int av1_masked_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                                  const uint8_t *mask, int mask_stride,
+                                  MV *mvp_full, int step_param, int sadpb,
+                                  int further_steps, int do_refine,
+                                  const aom_variance_fn_ptr_t *fn_ptr,
+                                  const MV *ref_mv, MV *dst_mv, int is_second);
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR
+int av1_obmc_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                                MV *mvp_full, int step_param, int sadpb,
+                                int further_steps, int do_refine,
+                                const aom_variance_fn_ptr_t *fn_ptr,
+                                const MV *ref_mv, MV *dst_mv, int is_second);
+int av1_find_best_obmc_sub_pixel_tree_up(
+    const struct AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col,
+    MV *bestmv, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second, int use_upsampled_ref);
+#endif  // CONFIG_MOTION_VAR
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_MCOMP_H_

third_party/aom/av1/encoder/mips/msa/error_msa.c

@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+#define BLOCK_ERROR_BLOCKSIZE_MSA(BSize)                                     \
+  static int64_t block_error_##BSize##size_msa(                              \
+      const int16_t *coeff_ptr, const int16_t *dq_coeff_ptr, int64_t *ssz) { \
+    int64_t err = 0;                                                         \
+    uint32_t loop_cnt;                                                       \
+    v8i16 coeff, dq_coeff, coeff_r_h, coeff_l_h;                             \
+    v4i32 diff_r, diff_l, coeff_r_w, coeff_l_w;                              \
+    v2i64 sq_coeff_r, sq_coeff_l;                                            \
+    v2i64 err0, err_dup0, err1, err_dup1;                                    \
+                                                                             \
+    coeff = LD_SH(coeff_ptr);                                                \
+    dq_coeff = LD_SH(dq_coeff_ptr);                                          \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+    DOTP_SW2_SD(coeff_r_w, coeff_l_w, coeff_r_w, coeff_l_w, sq_coeff_r,      \
+                sq_coeff_l);                                                 \
+    DOTP_SW2_SD(diff_r, diff_l, diff_r, diff_l, err0, err1);                 \
+                                                                             \
+    coeff = LD_SH(coeff_ptr + 8);                                            \
+    dq_coeff = LD_SH(dq_coeff_ptr + 8);                                      \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+    DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);              \
+    DPADD_SD2_SD(diff_r, diff_l, err0, err1);                                \
+                                                                             \
+    coeff_ptr += 16;                                                         \
+    dq_coeff_ptr += 16;                                                      \
+                                                                             \
+    for (loop_cnt = ((BSize >> 4) - 1); loop_cnt--;) {                       \
+      coeff = LD_SH(coeff_ptr);                                              \
+      dq_coeff = LD_SH(dq_coeff_ptr);                                        \
+      UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+      ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+      HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+      DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+      DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                             \
+      coeff = LD_SH(coeff_ptr + 8);                                          \
+      dq_coeff = LD_SH(dq_coeff_ptr + 8);                                    \
+      UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+      ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+      HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+      DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+      DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                             \
+      coeff_ptr += 16;                                                       \
+      dq_coeff_ptr += 16;                                                    \
+    }                                                                        \
+                                                                             \
+    err_dup0 = __msa_splati_d(sq_coeff_r, 1);                                \
+    err_dup1 = __msa_splati_d(sq_coeff_l, 1);                                \
+    sq_coeff_r += err_dup0;                                                  \
+    sq_coeff_l += err_dup1;                                                  \
+    *ssz = __msa_copy_s_d(sq_coeff_r, 0);                                    \
+    *ssz += __msa_copy_s_d(sq_coeff_l, 0);                                   \
+                                                                             \
+    err_dup0 = __msa_splati_d(err0, 1);                                      \
+    err_dup1 = __msa_splati_d(err1, 1);                                      \
+    err0 += err_dup0;                                                        \
+    err1 += err_dup1;                                                        \
+    err = __msa_copy_s_d(err0, 0);                                           \
+    err += __msa_copy_s_d(err1, 0);                                          \
+                                                                             \
+    return err;                                                              \
+  }
+
+/* clang-format off */
+BLOCK_ERROR_BLOCKSIZE_MSA(16)
+BLOCK_ERROR_BLOCKSIZE_MSA(64)
+BLOCK_ERROR_BLOCKSIZE_MSA(256)
+BLOCK_ERROR_BLOCKSIZE_MSA(1024)
+/* clang-format on */
+
+int64_t av1_block_error_msa(const tran_low_t *coeff_ptr,
+                            const tran_low_t *dq_coeff_ptr, intptr_t blk_size,
+                            int64_t *ssz) {
+  int64_t err;
+  const int16_t *coeff = (const int16_t *)coeff_ptr;
+  const int16_t *dq_coeff = (const int16_t *)dq_coeff_ptr;
+
+  switch (blk_size) {
+    case 16: err = block_error_16size_msa(coeff, dq_coeff, ssz); break;
+    case 64: err = block_error_64size_msa(coeff, dq_coeff, ssz); break;
+    case 256: err = block_error_256size_msa(coeff, dq_coeff, ssz); break;
+    case 1024: err = block_error_1024size_msa(coeff, dq_coeff, ssz); break;
+    default:
+      err = av1_block_error_c(coeff_ptr, dq_coeff_ptr, blk_size, ssz);
+      break;
+  }
+
+  return err;
+}

third_party/aom/av1/encoder/mips/msa/fdct16x16_msa.c

@@ -0,0 +1,436 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+
+static void fadst16_cols_step1_msa(const int16_t *input, int32_t stride,
+                                   const int32_t *const0, int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r15 = LD_SH(input + 15 * stride);
+  r7 = LD_SH(input + 7 * stride);
+  r8 = LD_SH(input + 8 * stride);
+  SLLI_4V(r0, r15, r7, r8, 2);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 8, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * stride);
+  r4 = LD_SH(input + 4 * stride);
+  r11 = LD_SH(input + 11 * stride);
+  r12 = LD_SH(input + 12 * stride);
+  SLLI_4V(r3, r4, r11, r12, 2);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp2, int_buf, 8);
+  ST_SH2(tp1, tp3, int_buf + 4 * 8, 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+
+  ST_SH2(h0, h1, int_buf + 8 * 8, 8);
+  ST_SH2(h3, h2, int_buf + 12 * 8, 8);
+
+  r9 = LD_SH(input + 9 * stride);
+  r6 = LD_SH(input + 6 * stride);
+  r1 = LD_SH(input + stride);
+  r14 = LD_SH(input + 14 * stride);
+  SLLI_4V(r9, r6, r1, r14, 2);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r13 = LD_SH(input + 13 * stride);
+  r2 = LD_SH(input + 2 * stride);
+  r5 = LD_SH(input + 5 * stride);
+  r10 = LD_SH(input + 10 * stride);
+  SLLI_4V(r13, r2, r5, r10, 2);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_step2_msa_helper(int16_t *int_buf, const int32_t *const0,
+                                     int16_t *out, int16_t *out_ptr) {
+  v8i16 tp0, tp1, tp2, tp3, g5, g7, g13, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h10, h11;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v4i32 k0, k1, k2, k3;
+
+  LD_SH2(int_buf + 3 * 8, 4 * 8, g13, g15);
+  LD_SH2(int_buf + 11 * 8, 4 * 8, g5, g7);
+  LD_SW2(const0 + 4 * 19, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 21);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+
+  tp0 = LD_SH(int_buf + 4 * 8);
+  tp1 = LD_SH(int_buf + 5 * 8);
+  tp3 = LD_SH(int_buf + 10 * 8);
+  tp2 = LD_SH(int_buf + 14 * 8);
+  LD_SW2(const0 + 4 * 22, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 24);
+  MADD_BF(tp0, tp1, tp2, tp3, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  ST_SH(out4, (out + 3 * 16));
+  ST_SH(out5, (out_ptr + 4 * 16));
+
+  h1 = LD_SH(int_buf + 9 * 8);
+  h3 = LD_SH(int_buf + 12 * 8);
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  ST_SH(out12, (out + 2 * 16));
+  ST_SH(out13, (out_ptr + 5 * 16));
+
+  tp0 = LD_SH(int_buf);
+  tp1 = LD_SH(int_buf + 8);
+  tp2 = LD_SH(int_buf + 2 * 8);
+  tp3 = LD_SH(int_buf + 6 * 8);
+
+  BUTTERFLY_4(tp0, tp1, tp3, tp2, out0, out1, h11, h10);
+  out1 = -out1;
+  ST_SH(out0, (out));
+  ST_SH(out1, (out_ptr + 7 * 16));
+
+  h0 = LD_SH(int_buf + 8 * 8);
+  h2 = LD_SH(int_buf + 13 * 8);
+
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+  ST_SH(out8, (out + 16));
+  ST_SH(out9, (out_ptr + 6 * 16));
+
+  /* stage 4 */
+  LD_SW2(const0 + 4 * 25, 4, k0, k1);
+  LD_SW2(const0 + 4 * 27, 4, k2, k3);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  ST_SH(out2, (out + 7 * 16));
+  ST_SH(out3, (out_ptr));
+
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  ST_SH(out6, (out + 4 * 16));
+  ST_SH(out7, (out_ptr + 3 * 16));
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  ST_SH(out10, (out + 6 * 16));
+  ST_SH(out11, (out_ptr + 16));
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  ST_SH(out14, (out + 5 * 16));
+  ST_SH(out15, (out_ptr + 2 * 16));
+}
+
+static void fadst16_cols_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  fadst16_step2_msa_helper(int_buf, const0, out, out + 128);
+}
+
+static void fadst16_transpose_postproc_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+  out += 64;
+
+  /* load input data */
+  input += 128;
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+}
+
+static void fadst16_rows_step1_msa(int16_t *input, const int32_t *const0,
+                                   int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r7 = LD_SH(input + 7 * 8);
+  r8 = LD_SH(input + 8 * 8);
+  r15 = LD_SH(input + 15 * 8);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 4 * 2, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * 8);
+  r4 = LD_SH(input + 4 * 8);
+  r11 = LD_SH(input + 11 * 8);
+  r12 = LD_SH(input + 12 * 8);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp1, int_buf, 4 * 8);
+  ST_SH2(tp2, tp3, int_buf + 8, 4 * 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+  ST_SH2(h0, h3, int_buf + 8 * 8, 4 * 8);
+  ST_SH2(h1, h2, int_buf + 9 * 8, 4 * 8);
+
+  r1 = LD_SH(input + 8);
+  r6 = LD_SH(input + 6 * 8);
+  r9 = LD_SH(input + 9 * 8);
+  r14 = LD_SH(input + 14 * 8);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r2 = LD_SH(input + 2 * 8);
+  r5 = LD_SH(input + 5 * 8);
+  r10 = LD_SH(input + 10 * 8);
+  r13 = LD_SH(input + 13 * 8);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_rows_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  fadst16_step2_msa_helper(int_buf, const0, out, out + 8);
+}
+
+static void fadst16_transpose_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14,
+          l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+  out += 16 * 8;
+
+  /* load input data */
+  input += 128;
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14,
+          l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+}
+
+static void postproc_fdct16x8_1d_row(int16_t *intermediate, int16_t *output) {
+  int16_t *temp = intermediate;
+  int16_t *out = output;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11;
+  v8i16 in12, in13, in14, in15;
+
+  LD_SH8(temp, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  temp = intermediate + 8;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                     in10, in11, in12, in13, in14, in15);
+  FDCT_POSTPROC_2V_NEG_H(in0, in1);
+  FDCT_POSTPROC_2V_NEG_H(in2, in3);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  FDCT_POSTPROC_2V_NEG_H(in6, in7);
+  FDCT_POSTPROC_2V_NEG_H(in8, in9);
+  FDCT_POSTPROC_2V_NEG_H(in10, in11);
+  FDCT_POSTPROC_2V_NEG_H(in12, in13);
+  FDCT_POSTPROC_2V_NEG_H(in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6,
+               tmp7, in8, in9, in10, in11, in12, in13, in14, in15);
+  temp = intermediate;
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, temp, 16);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp0, tmp1,
+                tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  temp = intermediate;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT8x16_ODD(in8, in9, in10, in11, in12, in13, in14, in15, in0, in1, in2, in3,
+               in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, tmp0, in0,
+                     tmp1, in1, tmp2, in2, tmp3, in3);
+  ST_SH8(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, out, 16);
+  TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, tmp4, in4,
+                     tmp5, in5, tmp6, in6, tmp7, in7);
+  out = output + 8;
+  ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, out, 16);
+}
+
+void av1_fht16x16_msa(const int16_t *input, int16_t *output, int32_t stride,
+                      int32_t tx_type) {
+  DECLARE_ALIGNED(32, int16_t, tmp[256]);
+  DECLARE_ALIGNED(32, int16_t, trans_buf[256]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[128]);
+  int32_t i;
+  int16_t *ptmpbuf = &tmp_buf[0];
+  int16_t *trans = &trans_buf[0];
+  const int32_t const_arr[29 * 4] = {
+    52707308,    52707308,    52707308,    52707308,    -1072430300,
+    -1072430300, -1072430300, -1072430300, 795618043,   795618043,
+    795618043,   795618043,   -721080468,  -721080468,  -721080468,
+    -721080468,  459094491,   459094491,   459094491,   459094491,
+    -970646691,  -970646691,  -970646691,  -970646691,  1010963856,
+    1010963856,  1010963856,  1010963856,  -361743294,  -361743294,
+    -361743294,  -361743294,  209469125,   209469125,   209469125,
+    209469125,   -1053094788, -1053094788, -1053094788, -1053094788,
+    1053160324,  1053160324,  1053160324,  1053160324,  639644520,
+    639644520,   639644520,   639644520,   -862444000,  -862444000,
+    -862444000,  -862444000,  1062144356,  1062144356,  1062144356,
+    1062144356,  -157532337,  -157532337,  -157532337,  -157532337,
+    260914709,   260914709,   260914709,   260914709,   -1041559667,
+    -1041559667, -1041559667, -1041559667, 920985831,   920985831,
+    920985831,   920985831,   -551995675,  -551995675,  -551995675,
+    -551995675,  596522295,   596522295,   596522295,   596522295,
+    892853362,   892853362,   892853362,   892853362,   -892787826,
+    -892787826,  -892787826,  -892787826,  410925857,   410925857,
+    410925857,   410925857,   -992012162,  -992012162,  -992012162,
+    -992012162,  992077698,   992077698,   992077698,   992077698,
+    759246145,   759246145,   759246145,   759246145,   -759180609,
+    -759180609,  -759180609,  -759180609,  -759222975,  -759222975,
+    -759222975,  -759222975,  759288511,   759288511,   759288511,
+    759288511
+  };
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case ADST_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        postproc_fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case DCT_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    case ADST_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    default: assert(0); break;
+  }
+}

third_party/aom/av1/encoder/mips/msa/fdct4x4_msa.c

@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+
+void av1_fwht4x4_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4;
+
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+
+  in0 += in1;
+  in3 -= in2;
+  in4 = (in0 - in3) >> 1;
+  SUB2(in4, in1, in4, in2, in1, in2);
+  in0 -= in2;
+  in3 += in1;
+
+  TRANSPOSE4x4_SH_SH(in0, in2, in3, in1, in0, in2, in3, in1);
+
+  in0 += in2;
+  in1 -= in3;
+  in4 = (in0 - in1) >> 1;
+  SUB2(in4, in2, in4, in3, in2, in3);
+  in0 -= in3;
+  in1 += in2;
+
+  SLLI_4V(in0, in1, in2, in3, 2);
+
+  TRANSPOSE4x4_SH_SH(in0, in3, in1, in2, in0, in3, in1, in2);
+
+  ST4x2_UB(in0, output, 4);
+  ST4x2_UB(in3, output + 4, 4);
+  ST4x2_UB(in1, output + 8, 4);
+  ST4x2_UB(in2, output + 12, 4);
+}
+
+void av1_fht4x4_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3;
+
+  LD_SH4(input, stride, in0, in1, in2, in3);
+
+  /* fdct4 pre-process */
+  {
+    v8i16 temp, mask;
+    v16i8 zero = { 0 };
+    v16i8 one = __msa_ldi_b(1);
+
+    mask = (v8i16)__msa_sldi_b(zero, one, 15);
+    SLLI_4V(in0, in1, in2, in3, 4);
+    temp = __msa_ceqi_h(in0, 0);
+    temp = (v8i16)__msa_xori_b((v16u8)temp, 255);
+    temp = mask & temp;
+    in0 += temp;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    default: assert(0); break;
+  }
+
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  SRA_4V(in0, in1, in2, in3, 2);
+  PCKEV_D2_SH(in1, in0, in3, in2, in0, in2);
+  ST_SH2(in0, in2, output, 8);
+}

third_party/aom/av1/encoder/mips/msa/fdct8x8_msa.c

@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+
+void av1_fht8x8_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case ADST_DCT:
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case DCT_ADST:
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case ADST_ADST:
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    default: assert(0); break;
+  }
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}

third_party/aom/av1/encoder/mips/msa/fdct_msa.h

@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_
+#define AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_
+
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+#include "aom_dsp/mips/txfm_macros_msa.h"
+#include "aom_ports/mem.h"
+
+#define AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2,  \
+                  out3, out4, out5, out6, out7)                              \
+  {                                                                          \
+    v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                       \
+    v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                        \
+    v8i16 coeff0_m = { cospi_2_64,  cospi_6_64,  cospi_10_64, cospi_14_64,   \
+                       cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 }; \
+    v8i16 coeff1_m = { cospi_8_64,  -cospi_8_64,  cospi_16_64, -cospi_16_64, \
+                       cospi_24_64, -cospi_24_64, 0,           0 };          \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in7, in0, in4, in3);             \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+                                                                             \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in5, in2, in6, in1);             \
+    BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                   \
+    out7 = -s0_m;                                                            \
+    out0 = s1_m;                                                             \
+                                                                             \
+    SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5, cnst0_m, cnst1_m, cnst2_m, cnst3_m);  \
+                                                                             \
+    ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);       \
+    cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+    cnst1_m = cnst0_m;                                                       \
+                                                                             \
+    ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst2_m,  \
+                          cnst3_m, cnst1_m, out1, out6, s0_m, s1_m);         \
+                                                                             \
+    SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                          \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                                 \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                   \
+    out4 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                   \
+    out2 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                   \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                   \
+                                                                             \
+    out1 = -out1;                                                            \
+    out3 = -out3;                                                            \
+    out5 = -out5;                                                            \
+  }
+
+#define AOM_FADST4(in0, in1, in2, in3, out0, out1, out2, out3)              \
+  {                                                                         \
+    v4i32 s0_m, s1_m, s2_m, s3_m, constant_m;                               \
+    v4i32 in0_r_m, in1_r_m, in2_r_m, in3_r_m;                               \
+                                                                            \
+    UNPCK_R_SH_SW(in0, in0_r_m);                                            \
+    UNPCK_R_SH_SW(in1, in1_r_m);                                            \
+    UNPCK_R_SH_SW(in2, in2_r_m);                                            \
+    UNPCK_R_SH_SW(in3, in3_r_m);                                            \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_4_9);                                   \
+    MUL2(in0_r_m, constant_m, in3_r_m, constant_m, s1_m, s0_m);             \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_1_9);                                   \
+    s0_m += in0_r_m * constant_m;                                           \
+    s1_m -= in1_r_m * constant_m;                                           \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_2_9);                                   \
+    s0_m += in1_r_m * constant_m;                                           \
+    s1_m += in3_r_m * constant_m;                                           \
+                                                                            \
+    s2_m = in0_r_m + in1_r_m - in3_r_m;                                     \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_3_9);                                   \
+    MUL2(in2_r_m, constant_m, s2_m, constant_m, s3_m, in1_r_m);             \
+                                                                            \
+    in0_r_m = s0_m + s3_m;                                                  \
+    s2_m = s1_m - s3_m;                                                     \
+    s3_m = s1_m - s0_m + s3_m;                                              \
+                                                                            \
+    SRARI_W4_SW(in0_r_m, in1_r_m, s2_m, s3_m, DCT_CONST_BITS);              \
+    PCKEV_H4_SH(in0_r_m, in0_r_m, in1_r_m, in1_r_m, s2_m, s2_m, s3_m, s3_m, \
+                out0, out1, out2, out3);                                    \
+  }
+#endif  // AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_

third_party/aom/av1/encoder/mips/msa/temporal_filter_msa.c

@@ -0,0 +1,284 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+static void temporal_filter_apply_8size_msa(uint8_t *frm1_ptr, uint32_t stride,
+                                            uint8_t *frm2_ptr, int32_t filt_sth,
+                                            int32_t filt_wgt, uint32_t *acc,
+                                            uint16_t *cnt) {
+  uint32_t row;
+  uint64_t f0, f1, f2, f3;
+  v16i8 frm2, frm1 = { 0 };
+  v16i8 frm4, frm3 = { 0 };
+  v16u8 frm_r, frm_l;
+  v8i16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 2; row--;) {
+    LD4(frm1_ptr, stride, f0, f1, f2, f3);
+    frm1_ptr += (4 * stride);
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 32;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    INSERT_D2_SB(f0, f1, frm1);
+    INSERT_D2_SB(f2, f3, frm3);
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    UNPCK_UB_SH(frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+    UNPCK_UB_SH(frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+  }
+}
+
+static void temporal_filter_apply_16size_msa(uint8_t *frm1_ptr, uint32_t stride,
+                                             uint8_t *frm2_ptr,
+                                             int32_t filt_sth, int32_t filt_wgt,
+                                             uint32_t *acc, uint16_t *cnt) {
+  uint32_t row;
+  v16i8 frm1, frm2, frm3, frm4;
+  v16u8 frm_r, frm_l;
+  v16i8 zero = { 0 };
+  v8u16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 8; row--;) {
+    LD_SB2(frm1_ptr, stride, frm1, frm3);
+    frm1_ptr += stride;
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 16;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    frm1_ptr += stride;
+    frm2_ptr += 16;
+  }
+}
+
+void av1_temporal_filter_apply_msa(uint8_t *frame1_ptr, uint32_t stride,
+                                   uint8_t *frame2_ptr, uint32_t blk_w,
+                                   uint32_t blk_h, int32_t strength,
+                                   int32_t filt_wgt, uint32_t *accu,
+                                   uint16_t *cnt) {
+  if (8 == (blk_w * blk_h)) {
+    temporal_filter_apply_8size_msa(frame1_ptr, stride, frame2_ptr, strength,
+                                    filt_wgt, accu, cnt);
+  } else if (16 == (blk_w * blk_h)) {
+    temporal_filter_apply_16size_msa(frame1_ptr, stride, frame2_ptr, strength,
+                                     filt_wgt, accu, cnt);
+  } else {
+    av1_temporal_filter_apply_c(frame1_ptr, stride, frame2_ptr, blk_w, blk_h,
+                                strength, filt_wgt, accu, cnt);
+  }
+}

third_party/aom/av1/encoder/palette.c

@@ -0,0 +1,277 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/palette.h"
+
+static float calc_dist(const float *p1, const float *p2, int dim) {
+  float dist = 0;
+  int i;
+  for (i = 0; i < dim; ++i) {
+    const float diff = p1[i] - p2[i];
+    dist += diff * diff;
+  }
+  return dist;
+}
+
+void av1_calc_indices(const float *data, const float *centroids,
+                      uint8_t *indices, int n, int k, int dim) {
+  int i, j;
+  for (i = 0; i < n; ++i) {
+    float min_dist = calc_dist(data + i * dim, centroids, dim);
+    indices[i] = 0;
+    for (j = 1; j < k; ++j) {
+      const float this_dist =
+          calc_dist(data + i * dim, centroids + j * dim, dim);
+      if (this_dist < min_dist) {
+        min_dist = this_dist;
+        indices[i] = j;
+      }
+    }
+  }
+}
+
+// Generate a random number in the range [0, 32768).
+static unsigned int lcg_rand16(unsigned int *state) {
+  *state = (unsigned int)(*state * 1103515245ULL + 12345);
+  return *state / 65536 % 32768;
+}
+
+static void calc_centroids(const float *data, float *centroids,
+                           const uint8_t *indices, int n, int k, int dim) {
+  int i, j, index;
+  int count[PALETTE_MAX_SIZE];
+  unsigned int rand_state = (unsigned int)data[0];
+
+  assert(n <= 32768);
+
+  memset(count, 0, sizeof(count[0]) * k);
+  memset(centroids, 0, sizeof(centroids[0]) * k * dim);
+
+  for (i = 0; i < n; ++i) {
+    index = indices[i];
+    assert(index < k);
+    ++count[index];
+    for (j = 0; j < dim; ++j) {
+      centroids[index * dim + j] += data[i * dim + j];
+    }
+  }
+
+  for (i = 0; i < k; ++i) {
+    if (count[i] == 0) {
+      memcpy(centroids + i * dim, data + (lcg_rand16(&rand_state) % n) * dim,
+             sizeof(centroids[0]) * dim);
+    } else {
+      const float norm = 1.0f / count[i];
+      for (j = 0; j < dim; ++j) centroids[i * dim + j] *= norm;
+    }
+  }
+
+  // Round to nearest integers.
+  for (i = 0; i < k * dim; ++i) {
+    centroids[i] = roundf(centroids[i]);
+  }
+}
+
+static float calc_total_dist(const float *data, const float *centroids,
+                             const uint8_t *indices, int n, int k, int dim) {
+  float dist = 0;
+  int i;
+  (void)k;
+
+  for (i = 0; i < n; ++i)
+    dist += calc_dist(data + i * dim, centroids + indices[i] * dim, dim);
+
+  return dist;
+}
+
+void av1_k_means(const float *data, float *centroids, uint8_t *indices, int n,
+                 int k, int dim, int max_itr) {
+  int i;
+  float this_dist;
+  float pre_centroids[2 * PALETTE_MAX_SIZE];
+  uint8_t pre_indices[MAX_SB_SQUARE];
+
+  av1_calc_indices(data, centroids, indices, n, k, dim);
+  this_dist = calc_total_dist(data, centroids, indices, n, k, dim);
+
+  for (i = 0; i < max_itr; ++i) {
+    const float pre_dist = this_dist;
+    memcpy(pre_centroids, centroids, sizeof(pre_centroids[0]) * k * dim);
+    memcpy(pre_indices, indices, sizeof(pre_indices[0]) * n);
+
+    calc_centroids(data, centroids, indices, n, k, dim);
+    av1_calc_indices(data, centroids, indices, n, k, dim);
+    this_dist = calc_total_dist(data, centroids, indices, n, k, dim);
+
+    if (this_dist > pre_dist) {
+      memcpy(centroids, pre_centroids, sizeof(pre_centroids[0]) * k * dim);
+      memcpy(indices, pre_indices, sizeof(pre_indices[0]) * n);
+      break;
+    }
+    if (!memcmp(centroids, pre_centroids, sizeof(pre_centroids[0]) * k * dim))
+      break;
+  }
+}
+
+static int float_comparer(const void *a, const void *b) {
+  const float fa = *(const float *)a;
+  const float fb = *(const float *)b;
+  return (fa > fb) - (fa < fb);
+}
+
+int av1_remove_duplicates(float *centroids, int num_centroids) {
+  int num_unique;  // number of unique centroids
+  int i;
+  qsort(centroids, num_centroids, sizeof(*centroids), float_comparer);
+  // Remove duplicates.
+  num_unique = 1;
+  for (i = 1; i < num_centroids; ++i) {
+    if (centroids[i] != centroids[i - 1]) {  // found a new unique centroid
+      centroids[num_unique++] = centroids[i];
+    }
+  }
+  return num_unique;
+}
+
+int av1_count_colors(const uint8_t *src, int stride, int rows, int cols) {
+  int n = 0, r, c, i, val_count[256];
+  uint8_t val;
+  memset(val_count, 0, sizeof(val_count));
+
+  for (r = 0; r < rows; ++r) {
+    for (c = 0; c < cols; ++c) {
+      val = src[r * stride + c];
+      ++val_count[val];
+    }
+  }
+
+  for (i = 0; i < 256; ++i) {
+    if (val_count[i]) {
+      ++n;
+    }
+  }
+
+  return n;
+}
+
+#if CONFIG_PALETTE_DELTA_ENCODING
+int av1_get_palette_delta_bits_y(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits) {
+  const int n = pmi->palette_size[0];
+  int max_d = 0, i;
+  *min_bits = bit_depth - 3;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[i] - pmi->palette_colors[i - 1];
+    assert(delta > 0);
+    if (delta > max_d) max_d = delta;
+  }
+  return AOMMAX(av1_ceil_log2(max_d), *min_bits);
+}
+
+int av1_get_palette_delta_bits_u(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits) {
+  const int n = pmi->palette_size[1];
+  int max_d = 0, i;
+  *min_bits = bit_depth - 3;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[PALETTE_MAX_SIZE + i] -
+                      pmi->palette_colors[PALETTE_MAX_SIZE + i - 1];
+    assert(delta >= 0);
+    if (delta > max_d) max_d = delta;
+  }
+  return AOMMAX(av1_ceil_log2(max_d + 1), *min_bits);
+}
+
+int av1_get_palette_delta_bits_v(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *zero_count,
+                                 int *min_bits) {
+  const int n = pmi->palette_size[1];
+  const int max_val = 1 << bit_depth;
+  int max_d = 0, i;
+  *min_bits = bit_depth - 4;
+  *zero_count = 0;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] -
+                      pmi->palette_colors[2 * PALETTE_MAX_SIZE + i - 1];
+    const int v = abs(delta);
+    const int d = AOMMIN(v, max_val - v);
+    if (d > max_d) max_d = d;
+    if (d == 0) ++(*zero_count);
+  }
+  return AOMMAX(av1_ceil_log2(max_d + 1), *min_bits);
+}
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+
+int av1_palette_color_cost_y(const PALETTE_MODE_INFO *const pmi,
+                             int bit_depth) {
+  const int n = pmi->palette_size[0];
+#if CONFIG_PALETTE_DELTA_ENCODING
+  int min_bits = 0;
+  const int bits = av1_get_palette_delta_bits_y(pmi, bit_depth, &min_bits);
+  return av1_cost_bit(128, 0) * (2 + bit_depth + bits * (n - 1));
+#else
+  return bit_depth * n * av1_cost_bit(128, 0);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+}
+
+int av1_palette_color_cost_uv(const PALETTE_MODE_INFO *const pmi,
+                              int bit_depth) {
+  const int n = pmi->palette_size[1];
+#if CONFIG_PALETTE_DELTA_ENCODING
+  int cost = 0;
+  // U channel palette color cost.
+  int min_bits_u = 0;
+  const int bits_u = av1_get_palette_delta_bits_u(pmi, bit_depth, &min_bits_u);
+  cost += av1_cost_bit(128, 0) * (2 + bit_depth + bits_u * (n - 1));
+  // V channel palette color cost.
+  int zero_count = 0, min_bits_v = 0;
+  const int bits_v =
+      av1_get_palette_delta_bits_v(pmi, bit_depth, &zero_count, &min_bits_v);
+  const int bits_using_delta =
+      2 + bit_depth + (bits_v + 1) * (n - 1) - zero_count;
+  const int bits_using_raw = bit_depth * n;
+  cost += av1_cost_bit(128, 0) * (1 + AOMMIN(bits_using_delta, bits_using_raw));
+  return cost;
+#else
+  return 2 * bit_depth * n * av1_cost_bit(128, 0);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+}
+
+#if CONFIG_HIGHBITDEPTH
+int av1_count_colors_highbd(const uint8_t *src8, int stride, int rows, int cols,
+                            int bit_depth) {
+  int n = 0, r, c, i;
+  uint16_t val;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  int val_count[1 << 12];
+
+  assert(bit_depth <= 12);
+  memset(val_count, 0, (1 << 12) * sizeof(val_count[0]));
+  for (r = 0; r < rows; ++r) {
+    for (c = 0; c < cols; ++c) {
+      val = src[r * stride + c];
+      ++val_count[val];
+    }
+  }
+
+  for (i = 0; i < (1 << bit_depth); ++i) {
+    if (val_count[i]) {
+      ++n;
+    }
+  }
+
+  return n;
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/av1/encoder/palette.h

@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_PALETTE_H_
+#define AV1_ENCODER_PALETTE_H_
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Given 'n' 'data' points and 'k' 'centroids' each of dimension 'dim',
+// calculate the centroid 'indices' for the data points.
+void av1_calc_indices(const float *data, const float *centroids,
+                      uint8_t *indices, int n, int k, int dim);
+
+// Given 'n' 'data' points and an initial guess of 'k' 'centroids' each of
+// dimension 'dim', runs up to 'max_itr' iterations of k-means algorithm to get
+// updated 'centroids' and the centroid 'indices' for elements in 'data'.
+// Note: the output centroids are rounded off to nearest integers.
+void av1_k_means(const float *data, float *centroids, uint8_t *indices, int n,
+                 int k, int dim, int max_itr);
+
+// Given a list of centroids, returns the unique number of centroids 'k', and
+// puts these unique centroids in first 'k' indices of 'centroids' array.
+// Ideally, the centroids should be rounded to integers before calling this
+// method.
+int av1_remove_duplicates(float *centroids, int num_centroids);
+
+// Returns the number of colors in 'src'.
+int av1_count_colors(const uint8_t *src, int stride, int rows, int cols);
+#if CONFIG_HIGHBITDEPTH
+// Same as av1_count_colors(), but for high-bitdepth mode.
+int av1_count_colors_highbd(const uint8_t *src8, int stride, int rows, int cols,
+                            int bit_depth);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_PALETTE_DELTA_ENCODING
+// Return the number of bits used to transmit each luma palette color delta.
+int av1_get_palette_delta_bits_y(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits);
+
+// Return the number of bits used to transmit each U palette color delta.
+int av1_get_palette_delta_bits_u(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits);
+
+// Return the number of bits used to transmit each v palette color delta;
+// assign zero_count with the number of deltas being 0.
+int av1_get_palette_delta_bits_v(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *zero_count, int *min_bits);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+
+// Return the rate cost for transmitting luma palette color values.
+int av1_palette_color_cost_y(const PALETTE_MODE_INFO *const pmi, int bit_depth);
+
+// Return the rate cost for transmitting chroma palette color values.
+int av1_palette_color_cost_uv(const PALETTE_MODE_INFO *const pmi,
+                              int bit_depth);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_PALETTE_H_ */

third_party/aom/av1/encoder/pickcdef.c

@@ -0,0 +1,490 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <string.h>
+
+#include "./aom_scale_rtcd.h"
+#include "aom/aom_integer.h"
+#include "av1/common/cdef.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/reconinter.h"
+#include "av1/encoder/encoder.h"
+
+#define TOTAL_STRENGTHS (DERING_STRENGTHS * CLPF_STRENGTHS)
+
+/* Search for the best strength to add as an option, knowing we
+   already selected nb_strengths options. */
+static uint64_t search_one(int *lev, int nb_strengths,
+                           uint64_t mse[][TOTAL_STRENGTHS], int sb_count) {
+  uint64_t tot_mse[TOTAL_STRENGTHS];
+  int i, j;
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  int best_id = 0;
+  memset(tot_mse, 0, sizeof(tot_mse));
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    /* Find best mse among already selected options. */
+    for (gi = 0; gi < nb_strengths; gi++) {
+      if (mse[i][lev[gi]] < best_mse) {
+        best_mse = mse[i][lev[gi]];
+      }
+    }
+    /* Find best mse when adding each possible new option. */
+    for (j = 0; j < TOTAL_STRENGTHS; j++) {
+      uint64_t best = best_mse;
+      if (mse[i][j] < best) best = mse[i][j];
+      tot_mse[j] += best;
+    }
+  }
+  for (j = 0; j < TOTAL_STRENGTHS; j++) {
+    if (tot_mse[j] < best_tot_mse) {
+      best_tot_mse = tot_mse[j];
+      best_id = j;
+    }
+  }
+  lev[nb_strengths] = best_id;
+  return best_tot_mse;
+}
+
+/* Search for the best luma+chroma strength to add as an option, knowing we
+   already selected nb_strengths options. */
+static uint64_t search_one_dual(int *lev0, int *lev1, int nb_strengths,
+                                uint64_t (**mse)[TOTAL_STRENGTHS],
+                                int sb_count) {
+  uint64_t tot_mse[TOTAL_STRENGTHS][TOTAL_STRENGTHS];
+  int i, j;
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  int best_id0 = 0;
+  int best_id1 = 0;
+  memset(tot_mse, 0, sizeof(tot_mse));
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    /* Find best mse among already selected options. */
+    for (gi = 0; gi < nb_strengths; gi++) {
+      uint64_t curr = mse[0][i][lev0[gi]];
+      curr += mse[1][i][lev1[gi]];
+      if (curr < best_mse) {
+        best_mse = curr;
+      }
+    }
+    /* Find best mse when adding each possible new option. */
+    for (j = 0; j < TOTAL_STRENGTHS; j++) {
+      int k;
+      for (k = 0; k < TOTAL_STRENGTHS; k++) {
+        uint64_t best = best_mse;
+        uint64_t curr = mse[0][i][j];
+        curr += mse[1][i][k];
+        if (curr < best) best = curr;
+        tot_mse[j][k] += best;
+      }
+    }
+  }
+  for (j = 0; j < TOTAL_STRENGTHS; j++) {
+    int k;
+    for (k = 0; k < TOTAL_STRENGTHS; k++) {
+      if (tot_mse[j][k] < best_tot_mse) {
+        best_tot_mse = tot_mse[j][k];
+        best_id0 = j;
+        best_id1 = k;
+      }
+    }
+  }
+  lev0[nb_strengths] = best_id0;
+  lev1[nb_strengths] = best_id1;
+  return best_tot_mse;
+}
+
+/* Search for the set of strengths that minimizes mse. */
+static uint64_t joint_strength_search(int *best_lev, int nb_strengths,
+                                      uint64_t mse[][TOTAL_STRENGTHS],
+                                      int sb_count) {
+  uint64_t best_tot_mse;
+  int i;
+  best_tot_mse = (uint64_t)1 << 63;
+  /* Greedy search: add one strength options at a time. */
+  for (i = 0; i < nb_strengths; i++) {
+    best_tot_mse = search_one(best_lev, i, mse, sb_count);
+  }
+  /* Trying to refine the greedy search by reconsidering each
+     already-selected option. */
+  for (i = 0; i < 4 * nb_strengths; i++) {
+    int j;
+    for (j = 0; j < nb_strengths - 1; j++) best_lev[j] = best_lev[j + 1];
+    best_tot_mse = search_one(best_lev, nb_strengths - 1, mse, sb_count);
+  }
+  return best_tot_mse;
+}
+
+/* Search for the set of luma+chroma strengths that minimizes mse. */
+static uint64_t joint_strength_search_dual(int *best_lev0, int *best_lev1,
+                                           int nb_strengths,
+                                           uint64_t (**mse)[TOTAL_STRENGTHS],
+                                           int sb_count) {
+  uint64_t best_tot_mse;
+  int i;
+  best_tot_mse = (uint64_t)1 << 63;
+  /* Greedy search: add one strength options at a time. */
+  for (i = 0; i < nb_strengths; i++) {
+    best_tot_mse = search_one_dual(best_lev0, best_lev1, i, mse, sb_count);
+  }
+  /* Trying to refine the greedy search by reconsidering each
+     already-selected option. */
+  for (i = 0; i < 4 * nb_strengths; i++) {
+    int j;
+    for (j = 0; j < nb_strengths - 1; j++) {
+      best_lev0[j] = best_lev0[j + 1];
+      best_lev1[j] = best_lev1[j + 1];
+    }
+    best_tot_mse =
+        search_one_dual(best_lev0, best_lev1, nb_strengths - 1, mse, sb_count);
+  }
+  return best_tot_mse;
+}
+
+/* FIXME: SSE-optimize this. */
+static void copy_sb16_16(uint16_t *dst, int dstride, const uint16_t *src,
+                         int src_voffset, int src_hoffset, int sstride,
+                         int vsize, int hsize) {
+  int r, c;
+  const uint16_t *base = &src[src_voffset * sstride + src_hoffset];
+  for (r = 0; r < vsize; r++) {
+    for (c = 0; c < hsize; c++) {
+      dst[r * dstride + c] = base[r * sstride + c];
+    }
+  }
+}
+
+static INLINE uint64_t dist_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                      int sstride, int coeff_shift) {
+  uint64_t svar = 0;
+  uint64_t dvar = 0;
+  uint64_t sum_s = 0;
+  uint64_t sum_d = 0;
+  uint64_t sum_s2 = 0;
+  uint64_t sum_d2 = 0;
+  uint64_t sum_sd = 0;
+  int i, j;
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++) {
+      sum_s += src[i * sstride + j];
+      sum_d += dst[i * dstride + j];
+      sum_s2 += src[i * sstride + j] * src[i * sstride + j];
+      sum_d2 += dst[i * dstride + j] * dst[i * dstride + j];
+      sum_sd += src[i * sstride + j] * dst[i * dstride + j];
+    }
+  }
+  /* Compute the variance -- the calculation cannot go negative. */
+  svar = sum_s2 - ((sum_s * sum_s + 32) >> 6);
+  dvar = sum_d2 - ((sum_d * sum_d + 32) >> 6);
+  return (uint64_t)floor(
+      .5 +
+      (sum_d2 + sum_s2 - 2 * sum_sd) * .5 *
+          (svar + dvar + (400 << 2 * coeff_shift)) /
+          (sqrt((20000 << 4 * coeff_shift) + svar * (double)dvar)));
+}
+
+static INLINE uint64_t mse_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                     int sstride) {
+  uint64_t sum = 0;
+  int i, j;
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++) {
+      int e = dst[i * dstride + j] - src[i * sstride + j];
+      sum += e * e;
+    }
+  }
+  return sum;
+}
+
+static INLINE uint64_t mse_4x4_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                     int sstride) {
+  uint64_t sum = 0;
+  int i, j;
+  for (i = 0; i < 4; i++) {
+    for (j = 0; j < 4; j++) {
+      int e = dst[i * dstride + j] - src[i * sstride + j];
+      sum += e * e;
+    }
+  }
+  return sum;
+}
+
+/* Compute MSE only on the blocks we filtered. */
+uint64_t compute_dering_dist(uint16_t *dst, int dstride, uint16_t *src,
+                             dering_list *dlist, int dering_count,
+                             BLOCK_SIZE bsize, int coeff_shift, int pli) {
+  uint64_t sum = 0;
+  int bi, bx, by;
+  if (bsize == BLOCK_8X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      if (pli == 0) {
+        sum += dist_8x8_16bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                              &src[bi << (3 + 3)], 8, coeff_shift);
+      } else {
+        sum += mse_8x8_16bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                             &src[bi << (3 + 3)], 8);
+      }
+    }
+  } else if (bsize == BLOCK_4X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 3) * dstride + (bx << 2)], dstride,
+                           &src[bi << (3 + 2)], 4);
+      sum += mse_4x4_16bit(&dst[((by << 3) + 4) * dstride + (bx << 2)], dstride,
+                           &src[(bi << (3 + 2)) + 4 * 4], 4);
+    }
+  } else if (bsize == BLOCK_8X4) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 3)], dstride,
+                           &src[bi << (2 + 3)], 8);
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 3) + 4], dstride,
+                           &src[(bi << (2 + 3)) + 4], 8);
+    }
+  } else {
+    assert(bsize == BLOCK_4X4);
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 2)], dstride,
+                           &src[bi << (2 + 2)], 4);
+    }
+  }
+  return sum >> 2 * coeff_shift;
+}
+
+void av1_cdef_search(YV12_BUFFER_CONFIG *frame, const YV12_BUFFER_CONFIG *ref,
+                     AV1_COMMON *cm, MACROBLOCKD *xd) {
+  int r, c;
+  int sbr, sbc;
+  uint16_t *src[3];
+  uint16_t *ref_coeff[3];
+  dering_list dlist[MAX_MIB_SIZE * MAX_MIB_SIZE];
+  int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int stride[3];
+  int bsize[3];
+  int mi_wide_l2[3];
+  int mi_high_l2[3];
+  int xdec[3];
+  int ydec[3];
+  int pli;
+  int dering_count;
+  int coeff_shift = AOMMAX(cm->bit_depth - 8, 0);
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  uint64_t tot_mse;
+  int sb_count;
+  int nvsb = (cm->mi_rows + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  int nhsb = (cm->mi_cols + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  int *sb_index = aom_malloc(nvsb * nhsb * sizeof(*sb_index));
+  int *selected_strength = aom_malloc(nvsb * nhsb * sizeof(*sb_index));
+  uint64_t(*mse[2])[TOTAL_STRENGTHS];
+  int clpf_damping = 3 + (cm->base_qindex >> 6);
+  int dering_damping = 6;
+  int i;
+  int nb_strengths;
+  int nb_strength_bits;
+  int quantizer;
+  double lambda;
+  int nplanes = 3;
+  DECLARE_ALIGNED(32, uint16_t, inbuf[OD_DERING_INBUF_SIZE]);
+  uint16_t *in;
+  DECLARE_ALIGNED(32, uint16_t, tmp_dst[MAX_SB_SQUARE]);
+  int chroma_dering =
+      xd->plane[1].subsampling_x == xd->plane[1].subsampling_y &&
+      xd->plane[2].subsampling_x == xd->plane[2].subsampling_y;
+  quantizer =
+      av1_ac_quant(cm->base_qindex, 0, cm->bit_depth) >> (cm->bit_depth - 8);
+  lambda = .12 * quantizer * quantizer / 256.;
+
+  av1_setup_dst_planes(xd->plane, cm->sb_size, frame, 0, 0);
+  mse[0] = aom_malloc(sizeof(**mse) * nvsb * nhsb);
+  mse[1] = aom_malloc(sizeof(**mse) * nvsb * nhsb);
+  for (pli = 0; pli < nplanes; pli++) {
+    uint8_t *ref_buffer;
+    int ref_stride;
+    switch (pli) {
+      case 0:
+        ref_buffer = ref->y_buffer;
+        ref_stride = ref->y_stride;
+        break;
+      case 1:
+        ref_buffer = ref->u_buffer;
+        ref_stride = ref->uv_stride;
+        break;
+      case 2:
+        ref_buffer = ref->v_buffer;
+        ref_stride = ref->uv_stride;
+        break;
+    }
+    src[pli] = aom_memalign(
+        32, sizeof(*src) * cm->mi_rows * cm->mi_cols * MI_SIZE * MI_SIZE);
+    ref_coeff[pli] = aom_memalign(
+        32, sizeof(*ref_coeff) * cm->mi_rows * cm->mi_cols * MI_SIZE * MI_SIZE);
+    xdec[pli] = xd->plane[pli].subsampling_x;
+    ydec[pli] = xd->plane[pli].subsampling_y;
+    bsize[pli] = ydec[pli] ? (xdec[pli] ? BLOCK_4X4 : BLOCK_8X4)
+                           : (xdec[pli] ? BLOCK_4X8 : BLOCK_8X8);
+    stride[pli] = cm->mi_cols << MI_SIZE_LOG2;
+    mi_wide_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_x;
+    mi_high_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_y;
+
+    const int frame_height =
+        (cm->mi_rows * MI_SIZE) >> xd->plane[pli].subsampling_y;
+    const int frame_width =
+        (cm->mi_cols * MI_SIZE) >> xd->plane[pli].subsampling_x;
+
+    for (r = 0; r < frame_height; ++r) {
+      for (c = 0; c < frame_width; ++c) {
+#if CONFIG_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          src[pli][r * stride[pli] + c] = CONVERT_TO_SHORTPTR(
+              xd->plane[pli].dst.buf)[r * xd->plane[pli].dst.stride + c];
+          ref_coeff[pli][r * stride[pli] + c] =
+              CONVERT_TO_SHORTPTR(ref_buffer)[r * ref_stride + c];
+        } else {
+#endif
+          src[pli][r * stride[pli] + c] =
+              xd->plane[pli].dst.buf[r * xd->plane[pli].dst.stride + c];
+          ref_coeff[pli][r * stride[pli] + c] = ref_buffer[r * ref_stride + c];
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif
+      }
+    }
+  }
+  in = inbuf + OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER;
+  sb_count = 0;
+  for (sbr = 0; sbr < nvsb; ++sbr) {
+    for (sbc = 0; sbc < nhsb; ++sbc) {
+      int nvb, nhb;
+      int gi;
+      int dirinit = 0;
+      nhb = AOMMIN(MAX_MIB_SIZE, cm->mi_cols - MAX_MIB_SIZE * sbc);
+      nvb = AOMMIN(MAX_MIB_SIZE, cm->mi_rows - MAX_MIB_SIZE * sbr);
+      cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                          MAX_MIB_SIZE * sbc]
+          ->mbmi.cdef_strength = -1;
+      if (sb_all_skip(cm, sbr * MAX_MIB_SIZE, sbc * MAX_MIB_SIZE)) continue;
+      dering_count = sb_compute_dering_list(cm, sbr * MAX_MIB_SIZE,
+                                            sbc * MAX_MIB_SIZE, dlist, 1);
+      for (pli = 0; pli < nplanes; pli++) {
+        for (i = 0; i < OD_DERING_INBUF_SIZE; i++)
+          inbuf[i] = OD_DERING_VERY_LARGE;
+        for (gi = 0; gi < TOTAL_STRENGTHS; gi++) {
+          int threshold;
+          uint64_t curr_mse;
+          int clpf_strength;
+          threshold = gi / CLPF_STRENGTHS;
+          if (pli > 0 && !chroma_dering) threshold = 0;
+          /* We avoid filtering the pixels for which some of the pixels to
+             average
+             are outside the frame. We could change the filter instead, but it
+             would add special cases for any future vectorization. */
+          int yoff = OD_FILT_VBORDER * (sbr != 0);
+          int xoff = OD_FILT_HBORDER * (sbc != 0);
+          int ysize = (nvb << mi_high_l2[pli]) +
+                      OD_FILT_VBORDER * (sbr != nvsb - 1) + yoff;
+          int xsize = (nhb << mi_wide_l2[pli]) +
+                      OD_FILT_HBORDER * (sbc != nhsb - 1) + xoff;
+          clpf_strength = gi % CLPF_STRENGTHS;
+          if (clpf_strength == 0)
+            copy_sb16_16(&in[(-yoff * OD_FILT_BSTRIDE - xoff)], OD_FILT_BSTRIDE,
+                         src[pli],
+                         (sbr * MAX_MIB_SIZE << mi_high_l2[pli]) - yoff,
+                         (sbc * MAX_MIB_SIZE << mi_wide_l2[pli]) - xoff,
+                         stride[pli], ysize, xsize);
+          od_dering(clpf_strength ? NULL : (uint8_t *)in, OD_FILT_BSTRIDE,
+                    tmp_dst, in, xdec[pli], ydec[pli], dir, &dirinit, var, pli,
+                    dlist, dering_count, threshold,
+                    clpf_strength + (clpf_strength == 3), clpf_damping,
+                    dering_damping, coeff_shift, clpf_strength != 0, 1);
+          curr_mse = compute_dering_dist(
+              ref_coeff[pli] +
+                  (sbr * MAX_MIB_SIZE << mi_high_l2[pli]) * stride[pli] +
+                  (sbc * MAX_MIB_SIZE << mi_wide_l2[pli]),
+              stride[pli], tmp_dst, dlist, dering_count, bsize[pli],
+              coeff_shift, pli);
+          if (pli < 2)
+            mse[pli][sb_count][gi] = curr_mse;
+          else
+            mse[1][sb_count][gi] += curr_mse;
+          sb_index[sb_count] =
+              MAX_MIB_SIZE * sbr * cm->mi_stride + MAX_MIB_SIZE * sbc;
+        }
+      }
+      sb_count++;
+    }
+  }
+  nb_strength_bits = 0;
+  /* Search for different number of signalling bits. */
+  for (i = 0; i <= 3; i++) {
+    int j;
+    int best_lev0[CDEF_MAX_STRENGTHS];
+    int best_lev1[CDEF_MAX_STRENGTHS] = { 0 };
+    nb_strengths = 1 << i;
+    if (nplanes >= 3)
+      tot_mse = joint_strength_search_dual(best_lev0, best_lev1, nb_strengths,
+                                           mse, sb_count);
+    else
+      tot_mse =
+          joint_strength_search(best_lev0, nb_strengths, mse[0], sb_count);
+    /* Count superblock signalling cost. */
+    tot_mse += (uint64_t)(sb_count * lambda * i);
+    /* Count header signalling cost. */
+    tot_mse += (uint64_t)(nb_strengths * lambda * CDEF_STRENGTH_BITS);
+    if (tot_mse < best_tot_mse) {
+      best_tot_mse = tot_mse;
+      nb_strength_bits = i;
+      for (j = 0; j < 1 << nb_strength_bits; j++) {
+        cm->cdef_strengths[j] = best_lev0[j];
+        cm->cdef_uv_strengths[j] = best_lev1[j];
+      }
+    }
+  }
+  nb_strengths = 1 << nb_strength_bits;
+
+  cm->cdef_bits = nb_strength_bits;
+  cm->nb_cdef_strengths = nb_strengths;
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    int best_gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    best_gi = 0;
+    for (gi = 0; gi < cm->nb_cdef_strengths; gi++) {
+      uint64_t curr = mse[0][i][cm->cdef_strengths[gi]];
+      if (nplanes >= 3) curr += mse[1][i][cm->cdef_uv_strengths[gi]];
+      if (curr < best_mse) {
+        best_gi = gi;
+        best_mse = curr;
+      }
+    }
+    selected_strength[i] = best_gi;
+    cm->mi_grid_visible[sb_index[i]]->mbmi.cdef_strength = best_gi;
+  }
+  cm->cdef_dering_damping = dering_damping;
+  cm->cdef_clpf_damping = clpf_damping;
+  aom_free(mse[0]);
+  aom_free(mse[1]);
+  for (pli = 0; pli < nplanes; pli++) {
+    aom_free(src[pli]);
+    aom_free(ref_coeff[pli]);
+  }
+  aom_free(sb_index);
+  aom_free(selected_strength);
+}

third_party/aom/av1/encoder/picklpf.c

@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./aom_scale_rtcd.h"
+
+#include "aom_dsp/psnr.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/picklpf.h"
+
+int av1_get_max_filter_level(const AV1_COMP *cpi) {
+  if (cpi->oxcf.pass == 2) {
+    return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
+                                                 : MAX_LOOP_FILTER;
+  } else {
+    return MAX_LOOP_FILTER;
+  }
+}
+
+static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd,
+                                AV1_COMP *const cpi, int filt_level,
+                                int partial_frame) {
+  AV1_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+
+#if CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_CB4X4
+  av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level, 1,
+                        partial_frame);
+#else
+  if (cpi->num_workers > 1)
+    av1_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+                             filt_level, 1, partial_frame, cpi->workers,
+                             cpi->num_workers, &cpi->lf_row_sync);
+  else
+    av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+                          1, partial_frame);
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    filt_err = aom_highbd_get_y_sse(sd, cm->frame_to_show);
+  } else {
+    filt_err = aom_get_y_sse(sd, cm->frame_to_show);
+  }
+#else
+  filt_err = aom_get_y_sse(sd, cm->frame_to_show);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // Re-instate the unfiltered frame
+  aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
+
+  return filt_err;
+}
+
+int av1_search_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                            int partial_frame, double *best_cost_ret) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const struct loopfilter *const lf = &cm->lf;
+  const int min_filter_level = 0;
+  const int max_filter_level = av1_get_max_filter_level(cpi);
+  int filt_direction = 0;
+  int64_t best_err;
+  int filt_best;
+  MACROBLOCK *x = &cpi->td.mb;
+
+  // Start the search at the previous frame filter level unless it is now out of
+  // range.
+  int filt_mid = clamp(lf->filter_level, min_filter_level, max_filter_level);
+  int filter_step = filt_mid < 16 ? 4 : filt_mid / 4;
+  // Sum squared error at each filter level
+  int64_t ss_err[MAX_LOOP_FILTER + 1];
+
+  // Set each entry to -1
+  memset(ss_err, 0xFF, sizeof(ss_err));
+
+  //  Make a copy of the unfiltered / processed recon buffer
+  aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
+
+  best_err = try_filter_frame(sd, cpi, filt_mid, partial_frame);
+  filt_best = filt_mid;
+  ss_err[filt_mid] = best_err;
+
+  while (filter_step > 0) {
+    const int filt_high = AOMMIN(filt_mid + filter_step, max_filter_level);
+    const int filt_low = AOMMAX(filt_mid - filter_step, min_filter_level);
+
+    // Bias against raising loop filter in favor of lowering it.
+    int64_t bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
+
+    if ((cpi->oxcf.pass == 2) && (cpi->twopass.section_intra_rating < 20))
+      bias = (bias * cpi->twopass.section_intra_rating) / 20;
+
+    // yx, bias less for large block size
+    if (cm->tx_mode != ONLY_4X4) bias >>= 1;
+
+    if (filt_direction <= 0 && filt_low != filt_mid) {
+      // Get Low filter error score
+      if (ss_err[filt_low] < 0) {
+        ss_err[filt_low] = try_filter_frame(sd, cpi, filt_low, partial_frame);
+      }
+      // If value is close to the best so far then bias towards a lower loop
+      // filter value.
+      if (ss_err[filt_low] < (best_err + bias)) {
+        // Was it actually better than the previous best?
+        if (ss_err[filt_low] < best_err) {
+          best_err = ss_err[filt_low];
+        }
+        filt_best = filt_low;
+      }
+    }
+
+    // Now look at filt_high
+    if (filt_direction >= 0 && filt_high != filt_mid) {
+      if (ss_err[filt_high] < 0) {
+        ss_err[filt_high] = try_filter_frame(sd, cpi, filt_high, partial_frame);
+      }
+      // If value is significantly better than previous best, bias added against
+      // raising filter value
+      if (ss_err[filt_high] < (best_err - bias)) {
+        best_err = ss_err[filt_high];
+        filt_best = filt_high;
+      }
+    }
+
+    // Half the step distance if the best filter value was the same as last time
+    if (filt_best == filt_mid) {
+      filter_step /= 2;
+      filt_direction = 0;
+    } else {
+      filt_direction = (filt_best < filt_mid) ? -1 : 1;
+      filt_mid = filt_best;
+    }
+  }
+
+  // Update best error
+  best_err = ss_err[filt_best];
+
+  if (best_cost_ret)
+    *best_cost_ret = RDCOST_DBL(x->rdmult, x->rddiv, 0, best_err);
+  return filt_best;
+}
+
+void av1_pick_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                           LPF_PICK_METHOD method) {
+  AV1_COMMON *const cm = &cpi->common;
+  struct loopfilter *const lf = &cm->lf;
+
+  lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0 : cpi->oxcf.sharpness;
+
+  if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
+    lf->filter_level = 0;
+  } else if (method >= LPF_PICK_FROM_Q) {
+    const int min_filter_level = 0;
+    const int max_filter_level = av1_get_max_filter_level(cpi);
+    const int q = av1_ac_quant(cm->base_qindex, 0, cm->bit_depth);
+// These values were determined by linear fitting the result of the
+// searched level, filt_guess = q * 0.316206 + 3.87252
+#if CONFIG_HIGHBITDEPTH
+    int filt_guess;
+    switch (cm->bit_depth) {
+      case AOM_BITS_8:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+        break;
+      case AOM_BITS_10:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 4060632, 20);
+        break;
+      case AOM_BITS_12:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 16242526, 22);
+        break;
+      default:
+        assert(0 &&
+               "bit_depth should be AOM_BITS_8, AOM_BITS_10 "
+               "or AOM_BITS_12");
+        return;
+    }
+#else
+    int filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+#endif  // CONFIG_HIGHBITDEPTH
+    if (cm->frame_type == KEY_FRAME) filt_guess -= 4;
+    lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
+  } else {
+    lf->filter_level = av1_search_filter_level(
+        sd, cpi, method == LPF_PICK_FROM_SUBIMAGE, NULL);
+  }
+
+#if CONFIG_EXT_TILE
+  // TODO(any): 0 loopfilter level is only necessary if individual tile
+  // decoding is required. We need to communicate this requirement to this
+  // code and force loop filter level 0 only if required.
+  if (cm->tile_encoding_mode) lf->filter_level = 0;
+#endif  // CONFIG_EXT_TILE
+}

third_party/aom/av1/encoder/picklpf.h

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_PICKLPF_H_
+#define AV1_ENCODER_PICKLPF_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+struct yv12_buffer_config;
+struct AV1_COMP;
+int av1_get_max_filter_level(const AV1_COMP *cpi);
+int av1_search_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                            int partial_frame, double *err);
+void av1_pick_filter_level(const struct yv12_buffer_config *sd,
+                           struct AV1_COMP *cpi, LPF_PICK_METHOD method);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_PICKLPF_H_

third_party/aom/av1/encoder/pickrst.h

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_ENCODER_PICKRST_H_
+#define AV1_ENCODER_PICKRST_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+struct yv12_buffer_config;
+struct AV1_COMP;
+
+void av1_pick_filter_restoration(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                                 LPF_PICK_METHOD method);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_PICKRST_H_

third_party/aom/av1/encoder/pvq_encoder.c

@@ -0,0 +1,988 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include "aom_dsp/entcode.h"
+#include "aom_dsp/entenc.h"
+#include "av1/common/blockd.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/partition.h"
+#include "av1/common/pvq_state.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/pvq_encoder.h"
+#include "aom_ports/system_state.h"
+
+/*Shift to ensure that the upper bound (i.e. for the max blocksize) of the
+   dot-product of the 1st band of chroma with the luma ref doesn't overflow.*/
+#define OD_CFL_FLIP_SHIFT (OD_LIMIT_BSIZE_MAX + 0)
+
+void aom_write_symbol_pvq(aom_writer *w, int symb, aom_cdf_prob *cdf,
+    int nsymbs) {
+  if (cdf[0] == 0)
+    aom_cdf_init_q15_1D(cdf, nsymbs, CDF_SIZE(nsymbs));
+  aom_write_symbol(w, symb, cdf, nsymbs);
+}
+
+static void aom_encode_pvq_codeword(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const od_coeff *in, int n, int k) {
+  int i;
+  aom_encode_band_pvq_splits(w, adapt, in, n, k, 0);
+  for (i = 0; i < n; i++) if (in[i]) aom_write_bit(w, in[i] < 0);
+}
+
+/* Computes 1/sqrt(i) using a table for small values. */
+static double od_rsqrt_table(int i) {
+  static double table[16] = {
+    1.000000, 0.707107, 0.577350, 0.500000,
+    0.447214, 0.408248, 0.377964, 0.353553,
+    0.333333, 0.316228, 0.301511, 0.288675,
+    0.277350, 0.267261, 0.258199, 0.250000};
+  if (i <= 16) return table[i-1];
+  else return 1./sqrt(i);
+}
+
+/*Computes 1/sqrt(start+2*i+1) using a lookup table containing the results
+   where 0 <= i < table_size.*/
+static double od_custom_rsqrt_dynamic_table(const double* table,
+ const int table_size, const double start, const int i) {
+  if (i < table_size) return table[i];
+  else return od_rsqrt_table((int)(start + 2*i + 1));
+}
+
+/*Fills tables used in od_custom_rsqrt_dynamic_table for a given start.*/
+static void od_fill_dynamic_rsqrt_table(double *table, const int table_size,
+ const double start) {
+  int i;
+  for (i = 0; i < table_size; i++)
+    table[i] = od_rsqrt_table((int)(start + 2*i + 1));
+}
+
+/** Find the codepoint on the given PSphere closest to the desired
+ * vector. Double-precision PVQ search just to make sure our tests
+ * aren't limited by numerical accuracy.
+ *
+ * @param [in]      xcoeff  input vector to quantize (x in the math doc)
+ * @param [in]      n       number of dimensions
+ * @param [in]      k       number of pulses
+ * @param [out]     ypulse  optimal codevector found (y in the math doc)
+ * @param [out]     g2      multiplier for the distortion (typically squared
+ *                          gain units)
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @param [in]      prev_k  number of pulses already in ypulse that we should
+ *                          reuse for the search (or 0 for a new search)
+ * @return                  cosine distance between x and y (between 0 and 1)
+ */
+double pvq_search_rdo_double_c(const od_val16 *xcoeff, int n, int k,
+ od_coeff *ypulse, double g2, double pvq_norm_lambda, int prev_k) {
+  int i, j;
+  double xy;
+  double yy;
+  /* TODO - This blows our 8kB stack space budget and should be fixed when
+   converting PVQ to fixed point. */
+  double x[MAXN];
+  double xx;
+  double lambda;
+  double norm_1;
+  int rdo_pulses;
+  double delta_rate;
+  xx = xy = yy = 0;
+  for (j = 0; j < n; j++) {
+    x[j] = fabs((float)xcoeff[j]);
+    xx += x[j]*x[j];
+  }
+  norm_1 = 1./sqrt(1e-30 + xx);
+  lambda = pvq_norm_lambda/(1e-30 + g2);
+  i = 0;
+  if (prev_k > 0 && prev_k <= k) {
+    /* We reuse pulses from a previous search so we don't have to search them
+       again. */
+    for (j = 0; j < n; j++) {
+      ypulse[j] = abs(ypulse[j]);
+      xy += x[j]*ypulse[j];
+      yy += ypulse[j]*ypulse[j];
+      i += ypulse[j];
+    }
+  }
+  else if (k > 2) {
+    double l1_norm;
+    double l1_inv;
+    l1_norm = 0;
+    for (j = 0; j < n; j++) l1_norm += x[j];
+    l1_inv = 1./OD_MAXF(l1_norm, 1e-100);
+    for (j = 0; j < n; j++) {
+      double tmp;
+      tmp = k*x[j]*l1_inv;
+      ypulse[j] = OD_MAXI(0, (int)floor(tmp));
+      xy += x[j]*ypulse[j];
+      yy += ypulse[j]*ypulse[j];
+      i += ypulse[j];
+    }
+  }
+  else OD_CLEAR(ypulse, n);
+
+  /* Only use RDO on the last few pulses. This not only saves CPU, but using
+     RDO on all pulses actually makes the results worse for reasons I don't
+     fully understand. */
+  rdo_pulses = 1 + k/4;
+  /* Rough assumption for now, the last position costs about 3 bits more than
+     the first. */
+  delta_rate = 3./n;
+  /* Search one pulse at a time */
+  for (; i < k - rdo_pulses; i++) {
+    int pos;
+    double best_xy;
+    double best_yy;
+    pos = 0;
+    best_xy = -10;
+    best_yy = 1;
+    for (j = 0; j < n; j++) {
+      double tmp_xy;
+      double tmp_yy;
+      tmp_xy = xy + x[j];
+      tmp_yy = yy + 2*ypulse[j] + 1;
+      tmp_xy *= tmp_xy;
+      if (j == 0 || tmp_xy*best_yy > best_xy*tmp_yy) {
+        best_xy = tmp_xy;
+        best_yy = tmp_yy;
+        pos = j;
+      }
+    }
+    xy = xy + x[pos];
+    yy = yy + 2*ypulse[pos] + 1;
+    ypulse[pos]++;
+  }
+  /* Search last pulses with RDO. Distortion is D = (x-y)^2 = x^2 - 2*x*y + y^2
+     and since x^2 and y^2 are constant, we just maximize x*y, plus a
+     lambda*rate term. Note that since x and y aren't normalized here,
+     we need to divide by sqrt(x^2)*sqrt(y^2). */
+  for (; i < k; i++) {
+    double rsqrt_table[4];
+    int rsqrt_table_size = 4;
+    int pos;
+    double best_cost;
+    pos = 0;
+    best_cost = -1e5;
+    /*Fill the small rsqrt lookup table with inputs relative to yy.
+      Specifically, the table of n values is filled with
+       rsqrt(yy + 1), rsqrt(yy + 2 + 1) .. rsqrt(yy + 2*(n-1) + 1).*/
+    od_fill_dynamic_rsqrt_table(rsqrt_table, rsqrt_table_size, yy);
+    for (j = 0; j < n; j++) {
+      double tmp_xy;
+      double tmp_yy;
+      tmp_xy = xy + x[j];
+      /*Calculate rsqrt(yy + 2*ypulse[j] + 1) using an optimized method.*/
+      tmp_yy = od_custom_rsqrt_dynamic_table(rsqrt_table, rsqrt_table_size,
+       yy, ypulse[j]);
+      tmp_xy = 2*tmp_xy*norm_1*tmp_yy - lambda*j*delta_rate;
+      if (j == 0 || tmp_xy > best_cost) {
+        best_cost = tmp_xy;
+        pos = j;
+      }
+    }
+    xy = xy + x[pos];
+    yy = yy + 2*ypulse[pos] + 1;
+    ypulse[pos]++;
+  }
+  for (i = 0; i < n; i++) {
+    if (xcoeff[i] < 0) ypulse[i] = -ypulse[i];
+  }
+  return xy/(1e-100 + sqrt(xx*yy));
+}
+
+/** Encodes the gain so that the return value increases with the
+ * distance |x-ref|, so that we can encode a zero when x=ref. The
+ * value x=0 is not covered because it is only allowed in the noref
+ * case.
+ *
+ * @param [in]      x      quantized gain to encode
+ * @param [in]      ref    quantized gain of the reference
+ * @return                 interleave-encoded quantized gain value
+ */
+static int neg_interleave(int x, int ref) {
+  if (x < ref) return -2*(x - ref) - 1;
+  else if (x < 2*ref) return 2*(x - ref);
+  else return x-1;
+}
+
+int od_vector_is_null(const od_coeff *x, int len) {
+  int i;
+  for (i = 0; i < len; i++) if (x[i]) return 0;
+  return 1;
+}
+
+static double od_pvq_rate(int qg, int icgr, int theta, int ts,
+  const od_adapt_ctx *adapt, const od_coeff *y0, int k, int n, int speed) {
+  double rate;
+  if (k == 0) rate = 0;
+  else if (speed > 0) {
+    int i;
+    int sum;
+    double f;
+    /* Compute "center of mass" of the pulse vector. */
+    sum = 0;
+    for (i = 0; i < n - (theta != -1); i++) sum += i*abs(y0[i]);
+    f = sum/(double)(k*n);
+    /* Estimates the number of bits it will cost to encode K pulses in
+       N dimensions based on hand-tuned fit for bitrate vs K, N and
+       "center of mass". */
+    rate = (1 + .4*f)*n*OD_LOG2(1 + OD_MAXF(0, log(n*2*(1*f + .025))*k/n)) + 3;
+  }
+  else {
+    aom_writer w;
+    od_pvq_codeword_ctx cd;
+    int tell;
+#if CONFIG_DAALA_EC
+    od_ec_enc_init(&w.ec, 1000);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+    OD_COPY(&cd, &adapt->pvq.pvq_codeword_ctx, 1);
+#if CONFIG_DAALA_EC
+    tell = od_ec_enc_tell_frac(&w.ec);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+    aom_encode_pvq_codeword(&w, &cd, y0, n - (theta != -1), k);
+#if CONFIG_DAALA_EC
+    rate = (od_ec_enc_tell_frac(&w.ec)-tell)/8.;
+    od_ec_enc_clear(&w.ec);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  }
+  if (qg > 0 && theta >= 0) {
+    /* Approximate cost of entropy-coding theta */
+    rate += .9*OD_LOG2(ts);
+    if (qg == icgr) rate -= .5;
+  }
+  return rate;
+}
+
+#define MAX_PVQ_ITEMS (20)
+/* This stores the information about a PVQ search candidate, so we can sort
+   based on K. */
+typedef struct {
+  int gain;
+  int k;
+  od_val32 qtheta;
+  int theta;
+  int ts;
+  od_val32 qcg;
+} pvq_search_item;
+
+int items_compare(pvq_search_item *a, pvq_search_item *b) {
+  /* Break ties in K with gain to ensure a stable sort.
+     Otherwise, the order depends on qsort implementation. */
+  return a->k == b->k ? a->gain - b->gain : a->k - b->k;
+}
+
+/** Perform PVQ quantization with prediction, trying several
+ * possible gains and angles. See draft-valin-videocodec-pvq and
+ * http://jmvalin.ca/slides/pvq.pdf for more details.
+ *
+ * @param [out]    out         coefficients after quantization
+ * @param [in]     x0          coefficients before quantization
+ * @param [in]     r0          reference, aka predicted coefficients
+ * @param [in]     n           number of dimensions
+ * @param [in]     q0          quantization step size
+ * @param [out]    y           pulse vector (i.e. selected PVQ codevector)
+ * @param [out]    itheta      angle between input and reference (-1 if noref)
+ * @param [out]    vk          total number of pulses
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [out]    skip_diff   distortion cost of skipping this block
+ *                             (accumulated)
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     pli         plane index
+ * @param [in]     adapt       probability adaptation context
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @param [in]     speed       Make search faster by making approximations
+ * @return         gain        index of the quatized gain
+*/
+static int pvq_theta(od_coeff *out, const od_coeff *x0, const od_coeff *r0,
+    int n, int q0, od_coeff *y, int *itheta, int *vk,
+    od_val16 beta, double *skip_diff, int is_keyframe, int pli,
+    const od_adapt_ctx *adapt, const int16_t *qm, const int16_t *qm_inv,
+    double pvq_norm_lambda, int speed) {
+  od_val32 g;
+  od_val32 gr;
+  od_coeff y_tmp[MAXN + 3];
+  int i;
+  /* Number of pulses. */
+  int k;
+  /* Companded gain of x and reference, normalized to q. */
+  od_val32 cg;
+  od_val32 cgr;
+  int icgr;
+  int qg;
+  /* Best RDO cost (D + lamdba*R) so far. */
+  double best_cost;
+  double dist0;
+  /* Distortion (D) that corresponds to the best RDO cost. */
+  double best_dist;
+  double dist;
+  /* Sign of Householder reflection. */
+  int s;
+  /* Dimension on which Householder reflects. */
+  int m;
+  od_val32 theta;
+  double corr;
+  int best_k;
+  od_val32 best_qtheta;
+  od_val32 gain_offset;
+  int noref;
+  double skip_dist;
+  int cfl_enabled;
+  int skip;
+  double gain_weight;
+  od_val16 x16[MAXN];
+  od_val16 r16[MAXN];
+  int xshift;
+  int rshift;
+  /* Give more weight to gain error when calculating the total distortion. */
+  gain_weight = 1.0;
+  OD_ASSERT(n > 1);
+  corr = 0;
+#if !defined(OD_FLOAT_PVQ)
+  /* Shift needed to make x fit in 16 bits even after rotation.
+     This shift value is not normative (it can be changed without breaking
+     the bitstream) */
+  xshift = OD_MAXI(0, od_vector_log_mag(x0, n) - 15);
+  /* Shift needed to make the reference fit in 15 bits, so that the Householder
+     vector can fit in 16 bits.
+     This shift value *is* normative, and has to match the decoder. */
+  rshift = OD_MAXI(0, od_vector_log_mag(r0, n) - 14);
+#else
+  xshift = 0;
+  rshift = 0;
+#endif
+  for (i = 0; i < n; i++) {
+#if defined(OD_FLOAT_PVQ)
+    /*This is slightly different from the original float PVQ code,
+       where the qm was applied in the accumulation in od_pvq_compute_gain and
+       the vectors were od_coeffs, not od_val16 (i.e. double).*/
+    x16[i] = x0[i]*(double)qm[i]*OD_QM_SCALE_1;
+    r16[i] = r0[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+    x16[i] = OD_SHR_ROUND(x0[i]*qm[i], OD_QM_SHIFT + xshift);
+    r16[i] = OD_SHR_ROUND(r0[i]*qm[i], OD_QM_SHIFT + rshift);
+#endif
+    corr += OD_MULT16_16(x16[i], r16[i]);
+  }
+  cfl_enabled = is_keyframe && pli != 0 && !OD_DISABLE_CFL;
+  cg  = od_pvq_compute_gain(x16, n, q0, &g, beta, xshift);
+  cgr = od_pvq_compute_gain(r16, n, q0, &gr, beta, rshift);
+  if (cfl_enabled) cgr = OD_CGAIN_SCALE;
+  /* gain_offset is meant to make sure one of the quantized gains has
+     exactly the same gain as the reference. */
+#if defined(OD_FLOAT_PVQ)
+  icgr = (int)floor(.5 + cgr);
+#else
+  icgr = OD_SHR_ROUND(cgr, OD_CGAIN_SHIFT);
+#endif
+  gain_offset = cgr - OD_SHL(icgr, OD_CGAIN_SHIFT);
+  /* Start search with null case: gain=0, no pulse. */
+  qg = 0;
+  dist = gain_weight*cg*cg*OD_CGAIN_SCALE_2;
+  best_dist = dist;
+  best_cost = dist + pvq_norm_lambda*od_pvq_rate(0, 0, -1, 0, adapt, NULL, 0,
+    n, speed);
+  noref = 1;
+  best_k = 0;
+  *itheta = -1;
+  OD_CLEAR(y, n);
+  best_qtheta = 0;
+  m = 0;
+  s = 1;
+  corr = corr/(1e-100 + g*(double)gr/OD_SHL(1, xshift + rshift));
+  corr = OD_MAXF(OD_MINF(corr, 1.), -1.);
+  if (is_keyframe) skip_dist = gain_weight*cg*cg*OD_CGAIN_SCALE_2;
+  else {
+    skip_dist = gain_weight*(cg - cgr)*(cg - cgr)
+     + cgr*(double)cg*(2 - 2*corr);
+    skip_dist *= OD_CGAIN_SCALE_2;
+  }
+  if (!is_keyframe) {
+    /* noref, gain=0 isn't allowed, but skip is allowed. */
+    od_val32 scgr;
+    scgr = OD_MAXF(0,gain_offset);
+    if (icgr == 0) {
+      best_dist = gain_weight*(cg - scgr)*(cg - scgr)
+       + scgr*(double)cg*(2 - 2*corr);
+      best_dist *= OD_CGAIN_SCALE_2;
+    }
+    best_cost = best_dist + pvq_norm_lambda*od_pvq_rate(0, icgr, 0, 0, adapt,
+     NULL, 0, n, speed);
+    best_qtheta = 0;
+    *itheta = 0;
+    noref = 0;
+  }
+  dist0 = best_dist;
+  if (n <= OD_MAX_PVQ_SIZE && !od_vector_is_null(r0, n) && corr > 0) {
+    od_val16 xr[MAXN];
+    int gain_bound;
+    int prev_k;
+    pvq_search_item items[MAX_PVQ_ITEMS];
+    int idx;
+    int nitems;
+    double cos_dist;
+    idx = 0;
+    gain_bound = OD_SHR(cg - gain_offset, OD_CGAIN_SHIFT);
+    /* Perform theta search only if prediction is useful. */
+    theta = OD_ROUND32(OD_THETA_SCALE*acos(corr));
+    m = od_compute_householder(r16, n, gr, &s, rshift);
+    od_apply_householder(xr, x16, r16, n);
+    prev_k = 0;
+    for (i = m; i < n - 1; i++) xr[i] = xr[i + 1];
+    /* Compute all candidate PVQ searches within a reasonable range of gain
+       and theta. */
+    for (i = OD_MAXI(1, gain_bound - 1); i <= gain_bound + 1; i++) {
+      int j;
+      od_val32 qcg;
+      int ts;
+      int theta_lower;
+      int theta_upper;
+      /* Quantized companded gain */
+      qcg = OD_SHL(i, OD_CGAIN_SHIFT) + gain_offset;
+      /* Set angular resolution (in ra) to match the encoded gain */
+      ts = od_pvq_compute_max_theta(qcg, beta);
+      theta_lower = OD_MAXI(0, (int)floor(.5 +
+       theta*OD_THETA_SCALE_1*2/M_PI*ts) - 2);
+      theta_upper = OD_MINI(ts - 1, (int)ceil(theta*OD_THETA_SCALE_1*2/M_PI*ts));
+      /* Include the angles within a reasonable range. */
+      for (j = theta_lower; j <= theta_upper; j++) {
+        od_val32 qtheta;
+        qtheta = od_pvq_compute_theta(j, ts);
+        k = od_pvq_compute_k(qcg, j, 0, n, beta);
+        items[idx].gain = i;
+        items[idx].theta = j;
+        items[idx].k = k;
+        items[idx].qcg = qcg;
+        items[idx].qtheta = qtheta;
+        items[idx].ts = ts;
+        idx++;
+        OD_ASSERT(idx < MAX_PVQ_ITEMS);
+      }
+    }
+    nitems = idx;
+    cos_dist = 0;
+    /* Sort PVQ search candidates in ascending order of pulses K so that
+       we can reuse all the previously searched pulses across searches. */
+    qsort(items, nitems, sizeof(items[0]),
+     (int (*)(const void *, const void *))items_compare);
+    /* Search for the best gain/theta in order. */
+    for (idx = 0; idx < nitems; idx++) {
+      int j;
+      od_val32 qcg;
+      int ts;
+      double cost;
+      double dist_theta;
+      double sin_prod;
+      od_val32 qtheta;
+      /* Quantized companded gain */
+      qcg = items[idx].qcg;
+      i = items[idx].gain;
+      j = items[idx].theta;
+      /* Set angular resolution (in ra) to match the encoded gain */
+      ts = items[idx].ts;
+      /* Search for the best angle within a reasonable range. */
+      qtheta = items[idx].qtheta;
+      k = items[idx].k;
+      /* Compute the minimal possible distortion by not taking the PVQ
+         cos_dist into account. */
+      dist_theta = 2 - 2.*od_pvq_cos(theta - qtheta)*OD_TRIG_SCALE_1;
+      dist = gain_weight*(qcg - cg)*(qcg - cg) + qcg*(double)cg*dist_theta;
+      dist *= OD_CGAIN_SCALE_2;
+      /* If we have no hope of beating skip (including a 1-bit worst-case
+         penalty), stop now. */
+      if (dist > dist0 + 1.0*pvq_norm_lambda && k != 0) continue;
+      sin_prod = od_pvq_sin(theta)*OD_TRIG_SCALE_1*od_pvq_sin(qtheta)*
+       OD_TRIG_SCALE_1;
+      /* PVQ search, using a gain of qcg*cg*sin(theta)*sin(qtheta) since
+         that's the factor by which cos_dist is multiplied to get the
+         distortion metric. */
+      if (k == 0) {
+        cos_dist = 0;
+        OD_CLEAR(y_tmp, n-1);
+      }
+      else if (k != prev_k) {
+        cos_dist = pvq_search_rdo_double(xr, n - 1, k, y_tmp,
+         qcg*(double)cg*sin_prod*OD_CGAIN_SCALE_2, pvq_norm_lambda, prev_k);
+      }
+      prev_k = k;
+      /* See Jmspeex' Journal of Dubious Theoretical Results. */
+      dist_theta = 2 - 2.*od_pvq_cos(theta - qtheta)*OD_TRIG_SCALE_1
+       + sin_prod*(2 - 2*cos_dist);
+      dist = gain_weight*(qcg - cg)*(qcg - cg) + qcg*(double)cg*dist_theta;
+      dist *= OD_CGAIN_SCALE_2;
+      /* Do approximate RDO. */
+      cost = dist + pvq_norm_lambda*od_pvq_rate(i, icgr, j, ts, adapt, y_tmp,
+       k, n, speed);
+      if (cost < best_cost) {
+        best_cost = cost;
+        best_dist = dist;
+        qg = i;
+        best_k = k;
+        best_qtheta = qtheta;
+        *itheta = j;
+        noref = 0;
+        OD_COPY(y, y_tmp, n - 1);
+      }
+    }
+  }
+  /* Don't bother with no-reference version if there's a reasonable
+     correlation. */
+  if (n <= OD_MAX_PVQ_SIZE && (corr < .5
+        || cg < (od_val32)(OD_SHL(2, OD_CGAIN_SHIFT)))) {
+    int gain_bound;
+    int prev_k;
+    gain_bound = OD_SHR(cg, OD_CGAIN_SHIFT);
+    prev_k = 0;
+    /* Search for the best gain (haven't determined reasonable range yet). */
+    for (i = OD_MAXI(1, gain_bound); i <= gain_bound + 1; i++) {
+      double cos_dist;
+      double cost;
+      od_val32 qcg;
+      qcg = OD_SHL(i, OD_CGAIN_SHIFT);
+      k = od_pvq_compute_k(qcg, -1, 1, n, beta);
+      /* Compute the minimal possible distortion by not taking the PVQ
+         cos_dist into account. */
+      dist = gain_weight*(qcg - cg)*(qcg - cg);
+      dist *= OD_CGAIN_SCALE_2;
+      if (dist > dist0 && k != 0) continue;
+      cos_dist = pvq_search_rdo_double(x16, n, k, y_tmp,
+       qcg*(double)cg*OD_CGAIN_SCALE_2, pvq_norm_lambda, prev_k);
+      prev_k = k;
+      /* See Jmspeex' Journal of Dubious Theoretical Results. */
+      dist = gain_weight*(qcg - cg)*(qcg - cg)
+       + qcg*(double)cg*(2 - 2*cos_dist);
+      dist *= OD_CGAIN_SCALE_2;
+      /* Do approximate RDO. */
+      cost = dist + pvq_norm_lambda*od_pvq_rate(i, 0, -1, 0, adapt, y_tmp, k,
+       n, speed);
+      if (cost <= best_cost) {
+        best_cost = cost;
+        best_dist = dist;
+        qg = i;
+        noref = 1;
+        best_k = k;
+        *itheta = -1;
+        OD_COPY(y, y_tmp, n);
+      }
+    }
+  }
+  k = best_k;
+  theta = best_qtheta;
+  skip = 0;
+  if (noref) {
+    if (qg == 0) skip = OD_PVQ_SKIP_ZERO;
+  }
+  else {
+    if (!is_keyframe && qg == 0) {
+      skip = (icgr ? OD_PVQ_SKIP_ZERO : OD_PVQ_SKIP_COPY);
+    }
+    if (qg == icgr && *itheta == 0 && !cfl_enabled) skip = OD_PVQ_SKIP_COPY;
+  }
+  /* Synthesize like the decoder would. */
+  if (skip) {
+    if (skip == OD_PVQ_SKIP_COPY) OD_COPY(out, r0, n);
+    else OD_CLEAR(out, n);
+  }
+  else {
+    if (noref) gain_offset = 0;
+    g = od_gain_expand(OD_SHL(qg, OD_CGAIN_SHIFT) + gain_offset, q0, beta);
+    od_pvq_synthesis_partial(out, y, r16, n, noref, g, theta, m, s,
+     qm_inv);
+  }
+  *vk = k;
+  *skip_diff += skip_dist - best_dist;
+  /* Encode gain differently depending on whether we use prediction or not.
+     Special encoding on inter frames where qg=0 is allowed for noref=0
+     but not noref=1.*/
+  if (is_keyframe) return noref ? qg : neg_interleave(qg, icgr);
+  else return noref ? qg - 1 : neg_interleave(qg + 1, icgr + 1);
+}
+
+/** Encodes a single vector of integers (eg, a partition within a
+ *  coefficient block) using PVQ
+ *
+ * @param [in,out] w          multi-symbol entropy encoder
+ * @param [in]     qg         quantized gain
+ * @param [in]     theta      quantized post-prediction theta
+ * @param [in]     in         coefficient vector to code
+ * @param [in]     n          number of coefficients in partition
+ * @param [in]     k          number of pulses in partition
+ * @param [in,out] model      entropy encoder state
+ * @param [in,out] adapt      adaptation context
+ * @param [in,out] exg        ExQ16 expectation of gain value
+ * @param [in,out] ext        ExQ16 expectation of theta value
+ * @param [in]     cdf_ctx    selects which cdf context to use
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     code_skip  whether the "skip rest" flag is allowed
+ * @param [in]     skip_rest  when set, we skip all higher bands
+ * @param [in]     encode_flip whether we need to encode the CfL flip flag now
+ * @param [in]     flip       value of the CfL flip flag
+ */
+void pvq_encode_partition(aom_writer *w,
+                                 int qg,
+                                 int theta,
+                                 const od_coeff *in,
+                                 int n,
+                                 int k,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 int cdf_ctx,
+                                 int is_keyframe,
+                                 int code_skip,
+                                 int skip_rest,
+                                 int encode_flip,
+                                 int flip) {
+  int noref;
+  int id;
+  noref = (theta == -1);
+  id = (qg > 0) + 2*OD_MINI(theta + 1,3) + 8*code_skip*skip_rest;
+  if (is_keyframe) {
+    OD_ASSERT(id != 8);
+    if (id >= 8) id--;
+  }
+  else {
+    OD_ASSERT(id != 10);
+    if (id >= 10) id--;
+  }
+  /* Jointly code gain, theta and noref for small values. Then we handle
+     larger gain and theta values. For noref, theta = -1. */
+  aom_write_symbol_pvq(w, id, &adapt->pvq.pvq_gaintheta_cdf[cdf_ctx][0],
+   8 + 7*code_skip);
+  if (encode_flip) {
+    /* We could eventually do some smarter entropy coding here, but it would
+       have to be good enough to overcome the overhead of the entropy coder.
+       An early attempt using a "toogle" flag with simple adaptation wasn't
+       worth the trouble. */
+    aom_write_bit(w, flip);
+  }
+  if (qg > 0) {
+    int tmp;
+    tmp = *exg;
+    generic_encode(w, &model[!noref], qg - 1, &tmp, 2);
+    OD_IIR_DIADIC(*exg, qg << 16, 2);
+  }
+  if (theta > 1) {
+    int tmp;
+    tmp = *ext;
+    generic_encode(w, &model[2], theta - 2, &tmp, 2);
+    OD_IIR_DIADIC(*ext, theta << 16, 2);
+  }
+  aom_encode_pvq_codeword(w, &adapt->pvq.pvq_codeword_ctx, in,
+   n - (theta != -1), k);
+}
+
+/** Quantizes a scalar with rate-distortion optimization (RDO)
+ * @param [in] x      unquantized value
+ * @param [in] q      quantization step size
+ * @param [in] delta0 rate increase for encoding a 1 instead of a 0
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @retval quantized value
+ */
+int od_rdo_quant(od_coeff x, int q, double delta0, double pvq_norm_lambda) {
+  int n;
+  /* Optimal quantization threshold is 1/2 + lambda*delta_rate/2. See
+     Jmspeex' Journal of Dubious Theoretical Results for details. */
+  n = OD_DIV_R0(abs(x), q);
+  if ((double)abs(x)/q < (double)n/2 + pvq_norm_lambda*delta0/(2*n)) {
+    return 0;
+  }
+  else {
+    return OD_DIV_R0(x, q);
+  }
+}
+
+/** Encode a coefficient block (excepting DC) using PVQ
+ *
+ * @param [in,out] enc     daala encoder context
+ * @param [in]     ref     'reference' (prediction) vector
+ * @param [in]     in      coefficient block to quantize and encode
+ * @param [out]    out     quantized coefficient block
+ * @param [in]     q0      scale/quantizer
+ * @param [in]     pli     plane index
+ * @param [in]     bs      log of the block size minus two
+ * @param [in]     beta    per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     qm      QM with magnitude compensation
+ * @param [in]     qm_inv  Inverse of QM with magnitude compensation
+ * @param [in]     speed   Make search faster by making approximations
+ * @param [in]     pvq_info If null, conisdered as RDO search mode
+ * @return         Returns block skip info indicating whether DC/AC are coded.
+ *                 bit0: DC is coded, bit1: AC is coded (1 means coded)
+ *
+ */
+PVQ_SKIP_TYPE od_pvq_encode(daala_enc_ctx *enc,
+                   od_coeff *ref,
+                   const od_coeff *in,
+                   od_coeff *out,
+                   int q_dc,
+                   int q_ac,
+                   int pli,
+                   int bs,
+                   const od_val16 *beta,
+                   int is_keyframe,
+                   const int16_t *qm,
+                   const int16_t *qm_inv,
+                   int speed,
+                   PVQ_INFO *pvq_info){
+  int theta[PVQ_MAX_PARTITIONS];
+  int qg[PVQ_MAX_PARTITIONS];
+  int k[PVQ_MAX_PARTITIONS];
+  od_coeff y[OD_TXSIZE_MAX*OD_TXSIZE_MAX];
+  int *exg;
+  int *ext;
+  int nb_bands;
+  int i;
+  const int *off;
+  int size[PVQ_MAX_PARTITIONS];
+  generic_encoder *model;
+  double skip_diff;
+  int tell;
+  uint16_t *skip_cdf;
+  od_rollback_buffer buf;
+  int dc_quant;
+  int flip;
+  int cfl_encoded;
+  int skip_rest;
+  int skip_dir;
+  int skip_theta_value;
+  const unsigned char *pvq_qm;
+  double dc_rate;
+  int use_masking;
+  PVQ_SKIP_TYPE ac_dc_coded;
+
+  aom_clear_system_state();
+
+  use_masking = enc->use_activity_masking;
+
+  if (use_masking)
+    pvq_qm = &enc->state.pvq_qm_q4[pli][0];
+  else
+    pvq_qm = 0;
+
+  exg = &enc->state.adapt->pvq.pvq_exg[pli][bs][0];
+  ext = enc->state.adapt->pvq.pvq_ext + bs*PVQ_MAX_PARTITIONS;
+  skip_cdf = enc->state.adapt->skip_cdf[2*bs + (pli != 0)];
+  model = enc->state.adapt->pvq.pvq_param_model;
+  nb_bands = OD_BAND_OFFSETS[bs][0];
+  off = &OD_BAND_OFFSETS[bs][1];
+
+  if (use_masking)
+    dc_quant = OD_MAXI(1, q_dc * pvq_qm[od_qm_get_index(bs, 0)] >> 4);
+  else
+    dc_quant = OD_MAXI(1, q_dc);
+
+  tell = 0;
+  for (i = 0; i < nb_bands; i++) size[i] = off[i+1] - off[i];
+  skip_diff = 0;
+  flip = 0;
+  /*If we are coding a chroma block of a keyframe, we are doing CfL.*/
+  if (pli != 0 && is_keyframe) {
+    od_val32 xy;
+    xy = 0;
+    /*Compute the dot-product of the first band of chroma with the luma ref.*/
+    for (i = off[0]; i < off[1]; i++) {
+#if defined(OD_FLOAT_PVQ)
+      xy += ref[i]*(double)qm[i]*OD_QM_SCALE_1*
+       (double)in[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+      od_val32 rq;
+      od_val32 inq;
+      rq = ref[i]*qm[i];
+      inq = in[i]*qm[i];
+      xy += OD_SHR(rq*(int64_t)inq, OD_SHL(OD_QM_SHIFT + OD_CFL_FLIP_SHIFT,
+       1));
+#endif
+    }
+    /*If cos(theta) < 0, then |theta| > pi/2 and we should negate the ref.*/
+    if (xy < 0) {
+      flip = 1;
+      for(i = off[0]; i < off[nb_bands]; i++) ref[i] = -ref[i];
+    }
+  }
+  for (i = 0; i < nb_bands; i++) {
+    int q;
+
+    if (use_masking)
+      q = OD_MAXI(1, q_ac * pvq_qm[od_qm_get_index(bs, i + 1)] >> 4);
+    else
+      q = OD_MAXI(1, q_ac);
+
+    qg[i] = pvq_theta(out + off[i], in + off[i], ref + off[i], size[i],
+     q, y + off[i], &theta[i], &k[i], beta[i], &skip_diff, is_keyframe,
+     pli, enc->state.adapt, qm + off[i], qm_inv + off[i],
+     enc->pvq_norm_lambda, speed);
+  }
+  od_encode_checkpoint(enc, &buf);
+  if (is_keyframe) out[0] = 0;
+  else {
+    int n;
+    n = OD_DIV_R0(abs(in[0] - ref[0]), dc_quant);
+    if (n == 0) {
+      out[0] = 0;
+    } else {
+      int tell2;
+      od_rollback_buffer dc_buf;
+
+      dc_rate = -OD_LOG2((double)(skip_cdf[3] - skip_cdf[2])/
+       (double)(skip_cdf[2] - skip_cdf[1]));
+      dc_rate += 1;
+
+#if CONFIG_DAALA_EC
+      tell2 = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+      od_encode_checkpoint(enc, &dc_buf);
+      generic_encode(&enc->w, &enc->state.adapt->model_dc[pli],
+       n - 1, &enc->state.adapt->ex_dc[pli][bs][0], 2);
+#if CONFIG_DAALA_EC
+      tell2 = od_ec_enc_tell_frac(&enc->w.ec) - tell2;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+      dc_rate += tell2/8.0;
+      od_encode_rollback(enc, &dc_buf);
+
+      out[0] = od_rdo_quant(in[0] - ref[0], dc_quant, dc_rate,
+       enc->pvq_norm_lambda);
+    }
+  }
+#if CONFIG_DAALA_EC
+  tell = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  /* Code as if we're not skipping. */
+  aom_write_symbol(&enc->w, 2 + (out[0] != 0), skip_cdf, 4);
+  ac_dc_coded = AC_CODED + (out[0] != 0);
+  cfl_encoded = 0;
+  skip_rest = 1;
+  skip_theta_value = is_keyframe ? -1 : 0;
+  for (i = 1; i < nb_bands; i++) {
+    if (theta[i] != skip_theta_value || qg[i]) skip_rest = 0;
+  }
+  skip_dir = 0;
+  if (nb_bands > 1) {
+    for (i = 0; i < 3; i++) {
+      int j;
+      int tmp;
+      tmp = 1;
+      // ToDo(yaowu): figure out better stop condition without gcc warning.
+      for (j = i + 1; j < nb_bands && j < PVQ_MAX_PARTITIONS; j += 3) {
+        if (theta[j] != skip_theta_value || qg[j]) tmp = 0;
+      }
+      skip_dir |= tmp << i;
+    }
+  }
+  if (theta[0] == skip_theta_value && qg[0] == 0 && skip_rest) nb_bands = 0;
+
+  /* NOTE: There was no other better place to put this function. */
+  if (pvq_info)
+    av1_store_pvq_enc_info(pvq_info, qg, theta, k, y, nb_bands, off, size,
+      skip_rest, skip_dir, bs);
+
+  for (i = 0; i < nb_bands; i++) {
+    int encode_flip;
+    /* Encode CFL flip bit just after the first time it's used. */
+    encode_flip = pli != 0 && is_keyframe && theta[i] != -1 && !cfl_encoded;
+    if (i == 0 || (!skip_rest && !(skip_dir & (1 << ((i - 1)%3))))) {
+      pvq_encode_partition(&enc->w, qg[i], theta[i], y + off[i],
+       size[i], k[i], model, enc->state.adapt, exg + i, ext + i,
+       (pli != 0)*OD_TXSIZES*PVQ_MAX_PARTITIONS + bs*PVQ_MAX_PARTITIONS + i,
+       is_keyframe, i == 0 && (i < nb_bands - 1), skip_rest, encode_flip, flip);
+    }
+    if (i == 0 && !skip_rest && bs > 0) {
+      aom_write_symbol(&enc->w, skip_dir,
+       &enc->state.adapt->pvq.pvq_skip_dir_cdf[(pli != 0) + 2*(bs - 1)][0], 7);
+    }
+    if (encode_flip) cfl_encoded = 1;
+  }
+#if CONFIG_DAALA_EC
+  tell = od_ec_enc_tell_frac(&enc->w.ec) - tell;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  /* Account for the rate of skipping the AC, based on the same DC decision
+     we made when trying to not skip AC. */
+  {
+    double skip_rate;
+    if (out[0] != 0) {
+      skip_rate = -OD_LOG2((skip_cdf[1] - skip_cdf[0])/
+     (double)skip_cdf[3]);
+    }
+    else {
+      skip_rate = -OD_LOG2(skip_cdf[0]/
+     (double)skip_cdf[3]);
+    }
+    tell -= (int)floor(.5+8*skip_rate);
+  }
+  if (nb_bands == 0 || skip_diff <= enc->pvq_norm_lambda/8*tell) {
+    if (is_keyframe) out[0] = 0;
+    else {
+      int n;
+      n = OD_DIV_R0(abs(in[0] - ref[0]), dc_quant);
+      if (n == 0) {
+        out[0] = 0;
+      } else {
+        int tell2;
+        od_rollback_buffer dc_buf;
+
+        dc_rate = -OD_LOG2((double)(skip_cdf[1] - skip_cdf[0])/
+         (double)skip_cdf[0]);
+        dc_rate += 1;
+
+#if CONFIG_DAALA_EC
+        tell2 = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+        od_encode_checkpoint(enc, &dc_buf);
+        generic_encode(&enc->w, &enc->state.adapt->model_dc[pli],
+         n - 1, &enc->state.adapt->ex_dc[pli][bs][0], 2);
+#if CONFIG_DAALA_EC
+        tell2 = od_ec_enc_tell_frac(&enc->w.ec) - tell2;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+        dc_rate += tell2/8.0;
+        od_encode_rollback(enc, &dc_buf);
+
+        out[0] = od_rdo_quant(in[0] - ref[0], dc_quant, dc_rate,
+         enc->pvq_norm_lambda);
+      }
+    }
+    /* We decide to skip, roll back everything as it was before. */
+    od_encode_rollback(enc, &buf);
+    aom_write_symbol(&enc->w, out[0] != 0, skip_cdf, 4);
+    ac_dc_coded = (out[0] != 0);
+    if (is_keyframe) for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = 0;
+    else for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = ref[i];
+  }
+  if (pvq_info)
+    pvq_info->ac_dc_coded = ac_dc_coded;
+  return ac_dc_coded;
+}

third_party/aom/av1/encoder/pvq_encoder.h

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_pvq_encoder_H)
+# define _pvq_encoder_H (1)
+# include "aom_dsp/bitwriter.h"
+# include "aom_dsp/entenc.h"
+# include "av1/common/blockd.h"
+# include "av1/common/pvq.h"
+# include "av1/encoder/encint.h"
+
+void aom_write_symbol_pvq(aom_writer *w, int symb, aom_cdf_prob *cdf,
+    int nsymbs);
+
+void aom_encode_band_pvq_splits(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const int *y, int n, int k, int level);
+
+void aom_laplace_encode_special(aom_writer *w, int x, unsigned decay);
+
+void pvq_encode_partition(aom_writer *w,
+                                 int qg,
+                                 int theta,
+                                 const od_coeff *in,
+                                 int n,
+                                 int k,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 int cdf_ctx,
+                                 int is_keyframe,
+                                 int code_skip,
+                                 int skip_rest,
+                                 int encode_flip,
+                                 int flip);
+
+PVQ_SKIP_TYPE od_pvq_encode(daala_enc_ctx *enc, od_coeff *ref,
+    const od_coeff *in, od_coeff *out, int q_dc, int q_ac, int pli, int bs,
+    const od_val16 *beta, int is_keyframe,
+    const int16_t *qm, const int16_t *qm_inv, int speed,
+    PVQ_INFO *pvq_info);
+
+#endif

third_party/aom/av1/encoder/ransac.h

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RANSAC_H_
+#define AV1_ENCODER_RANSAC_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <math.h>
+#include <memory.h>
+
+#include "av1/common/warped_motion.h"
+
+typedef int (*RansacFunc)(int *matched_points, int npoints,
+                          int *num_inliers_by_motion, double *params_by_motion,
+                          int num_motions);
+
+/* Each of these functions fits a motion model from a set of
+   corresponding points in 2 frames using RANSAC. */
+int ransac_homography(int *matched_points, int npoints,
+                      int *num_inliers_by_motion, double *params_by_motion,
+                      int num_motions);
+int ransac_affine(int *matched_points, int npoints, int *num_inliers_by_motion,
+                  double *params_by_motion, int num_motions);
+int ransac_hortrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_motions);
+int ransac_vertrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_motions);
+int ransac_rotzoom(int *matched_points, int npoints, int *num_inliers_by_motion,
+                   double *params_by_motion, int num_motions);
+int ransac_translation(int *matched_points, int npoints,
+                       int *num_inliers_by_motion, double *params_by_motion,
+                       int num_motions);
+#endif  // AV1_ENCODER_RANSAC_H_

third_party/aom/av1/encoder/ratectrl.h

@@ -0,0 +1,284 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RATECTRL_H_
+#define AV1_ENCODER_RATECTRL_H_
+
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Bits Per MB at different Q (Multiplied by 512)
+#define BPER_MB_NORMBITS 9
+
+#define MIN_GF_INTERVAL 4
+#define MAX_GF_INTERVAL 16
+#define FIXED_GF_INTERVAL 8  // Used in some testing modes only
+
+#if CONFIG_EXT_REFS
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_LOW = 1,
+  INTER_HIGH = 2,
+  GF_ARF_LOW = 3,
+  GF_ARF_STD = 4,
+  KF_STD = 5,
+  RATE_FACTOR_LEVELS = 6
+} RATE_FACTOR_LEVEL;
+#else
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_HIGH = 1,
+  GF_ARF_LOW = 2,
+  GF_ARF_STD = 3,
+  KF_STD = 4,
+  RATE_FACTOR_LEVELS = 5
+} RATE_FACTOR_LEVEL;
+#endif  // CONFIG_EXT_REFS
+
+// Internal frame scaling level.
+typedef enum {
+  UNSCALED = 0,     // Frame is unscaled.
+  SCALE_STEP1 = 1,  // First-level down-scaling.
+  FRAME_SCALE_STEPS
+} FRAME_SCALE_LEVEL;
+
+// Frame dimensions multiplier wrt the native frame size, in 1/16ths,
+// specified for the scale-up case.
+// e.g. 24 => 16/24 = 2/3 of native size. The restriction to 1/16th is
+// intended to match the capabilities of the normative scaling filters,
+// giving precedence to the up-scaling accuracy.
+static const int frame_scale_factor[FRAME_SCALE_STEPS] = { 16, 24 };
+
+// Multiplier of the target rate to be used as threshold for triggering scaling.
+static const double rate_thresh_mult[FRAME_SCALE_STEPS] = { 1.0, 2.0 };
+
+// Scale dependent Rate Correction Factor multipliers. Compensates for the
+// greater number of bits per pixel generated in down-scaled frames.
+static const double rcf_mult[FRAME_SCALE_STEPS] = { 1.0, 2.0 };
+
+typedef struct {
+  // Rate targetting variables
+  int base_frame_target;  // A baseline frame target before adjustment
+                          // for previous under or over shoot.
+  int this_frame_target;  // Actual frame target after rc adjustment.
+  int projected_frame_size;
+  int sb64_target_rate;
+  int last_q[FRAME_TYPES];  // Separate values for Intra/Inter
+  int last_boosted_qindex;  // Last boosted GF/KF/ARF q
+  int last_kf_qindex;       // Q index of the last key frame coded.
+
+  int gfu_boost;
+  int last_boost;
+  int kf_boost;
+
+  double rate_correction_factors[RATE_FACTOR_LEVELS];
+
+  int frames_since_golden;
+  int frames_till_gf_update_due;
+  int min_gf_interval;
+  int max_gf_interval;
+  int static_scene_max_gf_interval;
+  int baseline_gf_interval;
+  int constrained_gf_group;
+  int frames_to_key;
+  int frames_since_key;
+  int this_key_frame_forced;
+  int next_key_frame_forced;
+  int source_alt_ref_pending;
+  int source_alt_ref_active;
+  int is_src_frame_alt_ref;
+
+#if CONFIG_EXT_REFS
+  // Length of the bi-predictive frame group interval
+  int bipred_group_interval;
+
+  // NOTE: Different types of frames may have different bits allocated
+  //       accordingly, aiming to achieve the overall optimal RD performance.
+  int is_bwd_ref_frame;
+  int is_last_bipred_frame;
+  int is_bipred_frame;
+  int is_src_frame_ext_arf;
+#endif  // CONFIG_EXT_REFS
+
+  int avg_frame_bandwidth;  // Average frame size target for clip
+  int min_frame_bandwidth;  // Minimum allocation used for any frame
+  int max_frame_bandwidth;  // Maximum burst rate allowed for a frame.
+
+  int ni_av_qi;
+  int ni_tot_qi;
+  int ni_frames;
+  int avg_frame_qindex[FRAME_TYPES];
+  double tot_q;
+  double avg_q;
+
+  int64_t buffer_level;
+  int64_t bits_off_target;
+  int64_t vbr_bits_off_target;
+  int64_t vbr_bits_off_target_fast;
+
+  int decimation_factor;
+  int decimation_count;
+
+  int rolling_target_bits;
+  int rolling_actual_bits;
+
+  int long_rolling_target_bits;
+  int long_rolling_actual_bits;
+
+  int rate_error_estimate;
+
+  int64_t total_actual_bits;
+  int64_t total_target_bits;
+  int64_t total_target_vs_actual;
+
+  int worst_quality;
+  int best_quality;
+
+  int64_t starting_buffer_level;
+  int64_t optimal_buffer_level;
+  int64_t maximum_buffer_size;
+
+  // rate control history for last frame(1) and the frame before(2).
+  // -1: undershot
+  //  1: overshoot
+  //  0: not initialized.
+  int rc_1_frame;
+  int rc_2_frame;
+  int q_1_frame;
+  int q_2_frame;
+
+  // Auto frame-scaling variables.
+  FRAME_SCALE_LEVEL frame_size_selector;
+  FRAME_SCALE_LEVEL next_frame_size_selector;
+  int frame_width[FRAME_SCALE_STEPS];
+  int frame_height[FRAME_SCALE_STEPS];
+  int rf_level_maxq[RATE_FACTOR_LEVELS];
+} RATE_CONTROL;
+
+struct AV1_COMP;
+struct AV1EncoderConfig;
+
+void av1_rc_init(const struct AV1EncoderConfig *oxcf, int pass,
+                 RATE_CONTROL *rc);
+
+int av1_estimate_bits_at_q(FRAME_TYPE frame_kind, int q, int mbs,
+                           double correction_factor, aom_bit_depth_t bit_depth);
+
+double av1_convert_qindex_to_q(int qindex, aom_bit_depth_t bit_depth);
+
+void av1_rc_init_minq_luts(void);
+
+int av1_rc_get_default_min_gf_interval(int width, int height, double framerate);
+// Note av1_rc_get_default_max_gf_interval() requires the min_gf_interval to
+// be passed in to ensure that the max_gf_interval returned is at least as bis
+// as that.
+int av1_rc_get_default_max_gf_interval(double framerate, int min_frame_rate);
+
+// Generally at the high level, the following flow is expected
+// to be enforced for rate control:
+// First call per frame, one of:
+//   av1_rc_get_one_pass_vbr_params()
+//   av1_rc_get_one_pass_cbr_params()
+//   av1_rc_get_first_pass_params()
+//   av1_rc_get_second_pass_params()
+// depending on the usage to set the rate control encode parameters desired.
+//
+// Then, call encode_frame_to_data_rate() to perform the
+// actual encode. This function will in turn call encode_frame()
+// one or more times, followed by one of:
+//   av1_rc_postencode_update()
+//   av1_rc_postencode_update_drop_frame()
+//
+// The majority of rate control parameters are only expected
+// to be set in the av1_rc_get_..._params() functions and
+// updated during the av1_rc_postencode_update...() functions.
+// The only exceptions are av1_rc_drop_frame() and
+// av1_rc_update_rate_correction_factors() functions.
+
+// Functions to set parameters for encoding before the actual
+// encode_frame_to_data_rate() function.
+void av1_rc_get_one_pass_vbr_params(struct AV1_COMP *cpi);
+void av1_rc_get_one_pass_cbr_params(struct AV1_COMP *cpi);
+
+// Post encode update of the rate control parameters based
+// on bytes used
+void av1_rc_postencode_update(struct AV1_COMP *cpi, uint64_t bytes_used);
+// Post encode update of the rate control parameters for dropped frames
+void av1_rc_postencode_update_drop_frame(struct AV1_COMP *cpi);
+
+// Updates rate correction factors
+// Changes only the rate correction factors in the rate control structure.
+void av1_rc_update_rate_correction_factors(struct AV1_COMP *cpi);
+
+// Decide if we should drop this frame: For 1-pass CBR.
+// Changes only the decimation count in the rate control structure
+int av1_rc_drop_frame(struct AV1_COMP *cpi);
+
+// Computes frame size bounds.
+void av1_rc_compute_frame_size_bounds(const struct AV1_COMP *cpi,
+                                      int this_frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit);
+
+// Picks q and q bounds given the target for bits
+int av1_rc_pick_q_and_bounds(const struct AV1_COMP *cpi, int *bottom_index,
+                             int *top_index);
+
+// Estimates q to achieve a target bits per frame
+int av1_rc_regulate_q(const struct AV1_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality);
+
+// Estimates bits per mb for a given qindex and correction factor.
+int av1_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor, aom_bit_depth_t bit_depth);
+
+// Clamping utilities for bitrate targets for iframes and pframes.
+int av1_rc_clamp_iframe_target_size(const struct AV1_COMP *const cpi,
+                                    int target);
+int av1_rc_clamp_pframe_target_size(const struct AV1_COMP *const cpi,
+                                    int target);
+// Utility to set frame_target into the RATE_CONTROL structure
+// This function is called only from the av1_rc_get_..._params() functions.
+void av1_rc_set_frame_target(struct AV1_COMP *cpi, int target);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a target q value
+int av1_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       aom_bit_depth_t bit_depth);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a value that should equate to the given rate ratio.
+int av1_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               aom_bit_depth_t bit_depth);
+
+int av1_frame_type_qdelta(const struct AV1_COMP *cpi, int rf_level, int q);
+
+void av1_rc_update_framerate(struct AV1_COMP *cpi);
+
+void av1_rc_set_gf_interval_range(const struct AV1_COMP *const cpi,
+                                  RATE_CONTROL *const rc);
+
+void av1_set_target_rate(struct AV1_COMP *cpi);
+
+int av1_resize_one_pass_cbr(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_RATECTRL_H_

third_party/aom/av1/encoder/ratectrl_xiph.h

@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2001-2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_ratectrl_xiph_H)
+#define _ratectrl_xiph_H (1)
+
+#include "av1/encoder/ratectrl.h"
+#include "aom/internal/aom_codec_internal.h"
+
+/*Frame types.*/
+#define OD_I_FRAME (0)
+#define OD_P_FRAME (1)
+#define OD_GOLDEN_P_FRAME (2)
+#define OD_ALTREF_P_FRAME (3)
+
+#define OD_FRAME_NSUBTYPES (OD_ALTREF_P_FRAME + 1)
+
+/* Periodic boost (in between golden frames) strength - lower is more */
+#define OD_PERIODIC_BOOST_DIV (10)
+
+/* Constants for frame QP modulation <- tweak these
+ * Adjusts how the rate control system decides the quantizers per frame
+ * (sub)type */
+#define OD_MQP_I (0.98)
+#define OD_MQP_P (1.06)
+#define OD_MQP_GP (0.99)
+#define OD_MQP_AP (0.92)
+#define OD_DQP_I (-2)
+#define OD_DQP_P (0)
+#define OD_DQP_GP (-2)
+#define OD_DQP_AP (-2)
+
+/*Fractional_coded_quantizer ~=
+   log2(quantizer / (1 << OD_COEFF_SHIFT))*6.307 + 6.235*/
+/*Base/scale factor for linear quantizer to fractional coded quantizer
+   conversion (6.307 * 2^12) */
+#define OD_LOG_QUANTIZER_BASE_Q12 (0x0064EB)
+/*Inverse of above scale factor.*/
+#define OD_LOG_QUANTIZER_EXP_Q12 (0x000289)
+/*Offset for linear quantizer to fractional coded quantizer
+   conversion (6.235 * 2^45) */
+#define OD_LOG_QUANTIZER_OFFSET_Q45 (0x0000C7851EB851ECLL)
+
+#define OD_RC_2PASS_MAGIC (0x53015641) /* [A, V, 1, S] in little endian */
+#define OD_RC_2PASS_SUMMARY_SZ (4 + 1 + (4 + 4 + 8) * OD_FRAME_NSUBTYPES)
+#define OD_RC_2PASS_PACKET_SZ (1 + 4)
+#define OD_RC_2PASS_MIN (OD_RC_2PASS_PACKET_SZ + OD_RC_2PASS_SUMMARY_SZ)
+#define OD_RC_2PASS_VERSION (1)
+
+/*A 2nd order low-pass Bessel follower.
+  We use this for rate control because it has fast reaction time, but is
+   critically damped.*/
+typedef struct od_iir_bessel2 {
+  int32_t c[2];
+  int64_t g;
+  int32_t x[2];
+  int32_t y[2];
+} od_iir_bessel2;
+
+/* The 2-pass metrics associated with a single frame. */
+typedef struct od_frame_metrics {
+  /*The log base 2 of the scale factor for this frame in Q24 format.*/
+  int64_t log_scale;
+  /*The frame type from pass 1.*/
+  unsigned frame_type : 1;
+} od_frame_metrics;
+
+/*Rate control setup and working state information.*/
+typedef struct od_rc_state {
+  /* Image format */
+  int frame_width;
+  int frame_height;
+  int bit_depth;
+
+  /* Framerate */
+  double framerate;
+  /* Keyframe rate */
+  int keyframe_rate;
+  /* Golden frame period */
+  int goldenframe_rate;
+  /* Altref frame period */
+  int altref_rate;
+  /*The target bit-rate in bits per second.*/
+  int64_t target_bitrate;
+  /* Quality level for non-bitrate-targeting */
+  int quality;
+  /* Copied from oxcf->frame_periodic_boost */
+  int periodic_boosts;
+  /* Max Q */
+  int maxq;
+  /* Min Q */
+  int minq;
+  /* Quantizer to use for the first pass */
+  int firstpass_quant;
+
+  /* 2-pass metrics */
+  od_frame_metrics cur_metrics;
+
+  /* 2-pass state */
+  int64_t scale_sum[OD_FRAME_NSUBTYPES];
+  int nframes[OD_FRAME_NSUBTYPES];
+
+  /* 2-pass bytestream reader/writer context */
+  uint8_t *twopass_buffer;
+  int twopass_buffer_bytes;
+
+  /* Pass 1 stats packet storage */
+  uint8_t firstpass_buffer[OD_RC_2PASS_SUMMARY_SZ];
+
+  /* Every state packet from the first pass in a single buffer */
+  uint8_t *twopass_allframes_buf;
+  size_t twopass_allframes_buf_size;
+
+  /* Actual returned quantizer */
+  int target_quantizer;
+  /*The full-precision, unmodulated quantizer upon which
+    our modulated quantizers are based.*/
+  int base_quantizer;
+
+  /* Increments by 1 for each frame. */
+  int64_t cur_frame;
+
+  /* End of input flag */
+  int end_of_input;
+  /* Closed GOP flag */
+  int closed_gop;
+  /*The number of frames over which to distribute the reservoir usage.*/
+  int reservoir_frame_delay;
+  /*Will we drop frames to meet bitrate target?*/
+  unsigned char drop_frames;
+  /*Do we respect the maximum reservoir fullness?*/
+  unsigned char cap_overflow;
+  /*Can the reservoir go negative?*/
+  unsigned char cap_underflow;
+  /*Two-pass mode state.
+    0 => 1-pass encoding.
+    1 => 1st pass of 2-pass encoding.
+    2 => 2nd pass of 2-pass encoding.*/
+  int twopass_state;
+  /*The log of the number of pixels in a frame in Q57 format.*/
+  int64_t log_npixels;
+  /*The target average bits per frame.*/
+  int64_t bits_per_frame;
+  /*The current bit reservoir fullness (bits available to be used).*/
+  int64_t reservoir_fullness;
+  /*The target buffer fullness.
+    This is where we'd like to be by the last keyframe the appears in the next
+     buf_delay frames.*/
+  int64_t reservoir_target;
+  /*The maximum buffer fullness (total size of the buffer).*/
+  int64_t reservoir_max;
+  /*The log of estimated scale factor for the rate model in Q57 format.*/
+  int64_t log_scale[OD_FRAME_NSUBTYPES];
+  /*The exponent used in the rate model in Q8 format.*/
+  unsigned exp[OD_FRAME_NSUBTYPES];
+  /*The log of an estimated scale factor used to obtain the real framerate, for
+     VFR sources or, e.g., 12 fps content doubled to 24 fps, etc.*/
+  int64_t log_drop_scale[OD_FRAME_NSUBTYPES];
+  /*The total drop count from the previous frame.*/
+  uint32_t prev_drop_count[OD_FRAME_NSUBTYPES];
+  /*Second-order lowpass filters to track scale and VFR/drops.*/
+  od_iir_bessel2 scalefilter[OD_FRAME_NSUBTYPES];
+  od_iir_bessel2 vfrfilter[OD_FRAME_NSUBTYPES];
+  int frame_count[OD_FRAME_NSUBTYPES];
+  int inter_p_delay;
+  int inter_delay_target;
+  /*The total accumulated estimation bias.*/
+  int64_t rate_bias;
+} od_rc_state;
+
+int od_enc_rc_init(od_rc_state *rc, int64_t bitrate, int delay_ms);
+
+int od_enc_rc_select_quantizers_and_lambdas(od_rc_state *rc,
+                                            int is_golden_frame,
+                                            int is_altref_frame, int frame_type,
+                                            int *bottom_idx, int *top_idx);
+
+/* Returns 1 if the frame should be dropped */
+int od_enc_rc_update_state(od_rc_state *rc, int64_t bits, int is_golden_frame,
+                           int is_altref_frame, int frame_type, int droppable);
+
+int od_frame_type(od_rc_state *rc, int64_t coding_frame_count, int *is_golden,
+                  int *is_altref, int64_t *ip_count);
+
+int od_enc_rc_resize(od_rc_state *rc);
+
+int od_enc_rc_2pass_out(od_rc_state *rc, struct aom_codec_pkt_list *pkt_list,
+                        int summary);
+
+int od_enc_rc_2pass_in(od_rc_state *rc);
+
+#endif

third_party/aom/av1/encoder/rd.h

@@ -0,0 +1,505 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RD_H_
+#define AV1_ENCODER_RD_H_
+
+#include <limits.h>
+
+#if CONFIG_ANS
+#include "aom_dsp/ans.h"
+#endif  // CONFIG_ANS
+#include "av1/common/blockd.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/cost.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDDIV_BITS 7
+#define RD_EPB_SHIFT 6
+
+#define RDCOST(RM, DM, R, D) \
+  (ROUND_POWER_OF_TWO(((int64_t)R) * (RM), AV1_PROB_COST_SHIFT) + (D << DM))
+
+#define RDCOST_DBL(RM, DM, R, D)                                   \
+  (((((double)(R)) * (RM)) / (double)(1 << AV1_PROB_COST_SHIFT)) + \
+   ((double)(D) * (1 << (DM))))
+
+#define QIDX_SKIP_THRESH 115
+
+#define MV_COST_WEIGHT 108
+#define MV_COST_WEIGHT_SUB 120
+
+#define INVALID_MV 0x80008000
+
+#if CONFIG_EXT_REFS
+#define MAX_REFS 15
+#else
+#define MAX_REFS 6
+#endif  // CONFIG_EXT_REFS
+
+#define RD_THRESH_MAX_FACT 64
+#define RD_THRESH_INC 1
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION av1_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+  THR_NEARESTMV,
+#if CONFIG_EXT_REFS
+  THR_NEARESTL2,
+  THR_NEARESTL3,
+  THR_NEARESTB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEARESTA,
+  THR_NEARESTG,
+
+  THR_DC,
+
+  THR_NEWMV,
+#if CONFIG_EXT_REFS
+  THR_NEWL2,
+  THR_NEWL3,
+  THR_NEWB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEWA,
+  THR_NEWG,
+
+  THR_NEARMV,
+#if CONFIG_EXT_REFS
+  THR_NEARL2,
+  THR_NEARL3,
+  THR_NEARB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEARA,
+  THR_NEARG,
+
+  THR_ZEROMV,
+#if CONFIG_EXT_REFS
+  THR_ZEROL2,
+  THR_ZEROL3,
+  THR_ZEROB,
+#endif  // CONFIG_EXT_REFS
+  THR_ZEROG,
+  THR_ZEROA,
+
+#if CONFIG_EXT_INTER
+
+  THR_COMP_NEAREST_NEARESTLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTL2A,
+  THR_COMP_NEAREST_NEARESTL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTGA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTLB,
+  THR_COMP_NEAREST_NEARESTL2B,
+  THR_COMP_NEAREST_NEARESTL3B,
+  THR_COMP_NEAREST_NEARESTGB,
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  THR_COMP_NEARESTLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARESTL2A,
+  THR_COMP_NEARESTL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEARESTGA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARESTLB,
+  THR_COMP_NEARESTL2B,
+  THR_COMP_NEARESTL3B,
+  THR_COMP_NEARESTGB,
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  THR_TM,
+
+#if CONFIG_ALT_INTRA
+  THR_SMOOTH,
+#endif  // CONFIG_ALT_INTRA
+
+#if CONFIG_EXT_INTER
+
+  THR_COMP_NEAR_NEARESTLA,
+  THR_COMP_NEAREST_NEARLA,
+  THR_COMP_NEAR_NEARLA,
+  THR_COMP_NEW_NEARESTLA,
+  THR_COMP_NEAREST_NEWLA,
+  THR_COMP_NEW_NEARLA,
+  THR_COMP_NEAR_NEWLA,
+  THR_COMP_NEW_NEWLA,
+  THR_COMP_ZERO_ZEROLA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAR_NEARESTL2A,
+  THR_COMP_NEAREST_NEARL2A,
+  THR_COMP_NEAR_NEARL2A,
+  THR_COMP_NEW_NEARESTL2A,
+  THR_COMP_NEAREST_NEWL2A,
+  THR_COMP_NEW_NEARL2A,
+  THR_COMP_NEAR_NEWL2A,
+  THR_COMP_NEW_NEWL2A,
+  THR_COMP_ZERO_ZEROL2A,
+
+  THR_COMP_NEAR_NEARESTL3A,
+  THR_COMP_NEAREST_NEARL3A,
+  THR_COMP_NEAR_NEARL3A,
+  THR_COMP_NEW_NEARESTL3A,
+  THR_COMP_NEAREST_NEWL3A,
+  THR_COMP_NEW_NEARL3A,
+  THR_COMP_NEAR_NEWL3A,
+  THR_COMP_NEW_NEWL3A,
+  THR_COMP_ZERO_ZEROL3A,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_NEAR_NEARESTGA,
+  THR_COMP_NEAREST_NEARGA,
+  THR_COMP_NEAR_NEARGA,
+  THR_COMP_NEW_NEARESTGA,
+  THR_COMP_NEAREST_NEWGA,
+  THR_COMP_NEW_NEARGA,
+  THR_COMP_NEAR_NEWGA,
+  THR_COMP_NEW_NEWGA,
+  THR_COMP_ZERO_ZEROGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAR_NEARESTLB,
+  THR_COMP_NEAREST_NEARLB,
+  THR_COMP_NEAR_NEARLB,
+  THR_COMP_NEW_NEARESTLB,
+  THR_COMP_NEAREST_NEWLB,
+  THR_COMP_NEW_NEARLB,
+  THR_COMP_NEAR_NEWLB,
+  THR_COMP_NEW_NEWLB,
+  THR_COMP_ZERO_ZEROLB,
+
+  THR_COMP_NEAR_NEARESTL2B,
+  THR_COMP_NEAREST_NEARL2B,
+  THR_COMP_NEAR_NEARL2B,
+  THR_COMP_NEW_NEARESTL2B,
+  THR_COMP_NEAREST_NEWL2B,
+  THR_COMP_NEW_NEARL2B,
+  THR_COMP_NEAR_NEWL2B,
+  THR_COMP_NEW_NEWL2B,
+  THR_COMP_ZERO_ZEROL2B,
+
+  THR_COMP_NEAR_NEARESTL3B,
+  THR_COMP_NEAREST_NEARL3B,
+  THR_COMP_NEAR_NEARL3B,
+  THR_COMP_NEW_NEARESTL3B,
+  THR_COMP_NEAREST_NEWL3B,
+  THR_COMP_NEW_NEARL3B,
+  THR_COMP_NEAR_NEWL3B,
+  THR_COMP_NEW_NEWL3B,
+  THR_COMP_ZERO_ZEROL3B,
+
+  THR_COMP_NEAR_NEARESTGB,
+  THR_COMP_NEAREST_NEARGB,
+  THR_COMP_NEAR_NEARGB,
+  THR_COMP_NEW_NEARESTGB,
+  THR_COMP_NEAREST_NEWGB,
+  THR_COMP_NEW_NEARGB,
+  THR_COMP_NEAR_NEWGB,
+  THR_COMP_NEW_NEWGB,
+  THR_COMP_ZERO_ZEROGB,
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  THR_COMP_NEARLA,
+  THR_COMP_NEWLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARL2A,
+  THR_COMP_NEWL2A,
+  THR_COMP_NEARL3A,
+  THR_COMP_NEWL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEARGA,
+  THR_COMP_NEWGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARLB,
+  THR_COMP_NEWLB,
+  THR_COMP_NEARL2B,
+  THR_COMP_NEWL2B,
+  THR_COMP_NEARL3B,
+  THR_COMP_NEWL3B,
+  THR_COMP_NEARGB,
+  THR_COMP_NEWGB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_ZEROLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_ZEROL2A,
+  THR_COMP_ZEROL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_ZEROGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_ZEROLB,
+  THR_COMP_ZEROL2B,
+  THR_COMP_ZEROL3B,
+  THR_COMP_ZEROGB,
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  THR_H_PRED,
+  THR_V_PRED,
+  THR_D135_PRED,
+  THR_D207_PRED,
+  THR_D153_PRED,
+  THR_D63_PRED,
+  THR_D117_PRED,
+  THR_D45_PRED,
+
+#if CONFIG_EXT_INTER
+  THR_COMP_INTERINTRA_ZEROL,
+  THR_COMP_INTERINTRA_NEARESTL,
+  THR_COMP_INTERINTRA_NEARL,
+  THR_COMP_INTERINTRA_NEWL,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_INTERINTRA_ZEROL2,
+  THR_COMP_INTERINTRA_NEARESTL2,
+  THR_COMP_INTERINTRA_NEARL2,
+  THR_COMP_INTERINTRA_NEWL2,
+
+  THR_COMP_INTERINTRA_ZEROL3,
+  THR_COMP_INTERINTRA_NEARESTL3,
+  THR_COMP_INTERINTRA_NEARL3,
+  THR_COMP_INTERINTRA_NEWL3,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_INTERINTRA_ZEROG,
+  THR_COMP_INTERINTRA_NEARESTG,
+  THR_COMP_INTERINTRA_NEARG,
+  THR_COMP_INTERINTRA_NEWG,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_INTERINTRA_ZEROB,
+  THR_COMP_INTERINTRA_NEARESTB,
+  THR_COMP_INTERINTRA_NEARB,
+  THR_COMP_INTERINTRA_NEWB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_INTERINTRA_ZEROA,
+  THR_COMP_INTERINTRA_NEARESTA,
+  THR_COMP_INTERINTRA_NEARA,
+  THR_COMP_INTERINTRA_NEWA,
+#endif  // CONFIG_EXT_INTER
+  MAX_MODES
+} THR_MODES;
+
+typedef enum {
+  THR_LAST,
+#if CONFIG_EXT_REFS
+  THR_LAST2,
+  THR_LAST3,
+  THR_BWDR,
+#endif  // CONFIG_EXT_REFS
+  THR_GOLD,
+  THR_ALTR,
+
+  THR_COMP_LA,
+#if CONFIG_EXT_REFS
+  THR_COMP_L2A,
+  THR_COMP_L3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_GA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_LB,
+  THR_COMP_L2B,
+  THR_COMP_L3B,
+  THR_COMP_GB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+  // Thresh_mult is used to set a threshold for the rd score. A higher value
+  // means that we will accept the best mode so far more often. This number
+  // is used in combination with the current block size, and thresh_freq_fact
+  // to pick a threshold.
+  int thresh_mult[MAX_MODES];
+  int thresh_mult_sub8x8[MAX_REFS];
+
+  int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+
+  int64_t prediction_type_threshes[TOTAL_REFS_PER_FRAME][REFERENCE_MODES];
+
+  int RDMULT;
+  int RDDIV;
+} RD_OPT;
+
+static INLINE void av1_init_rd_stats(RD_STATS *rd_stats) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats->rate = 0;
+  rd_stats->dist = 0;
+  rd_stats->rdcost = 0;
+  rd_stats->sse = 0;
+  rd_stats->skip = 1;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats->txb_coeff_cost[plane] = 0;
+#if CONFIG_VAR_TX
+    {
+      int r, c;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+          rd_stats->txb_coeff_cost_map[plane][r][c] = 0;
+    }
+#endif
+  }
+#endif
+}
+
+static INLINE void av1_invalid_rd_stats(RD_STATS *rd_stats) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats->rate = INT_MAX;
+  rd_stats->dist = INT64_MAX;
+  rd_stats->rdcost = INT64_MAX;
+  rd_stats->sse = INT64_MAX;
+  rd_stats->skip = 0;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats->txb_coeff_cost[plane] = INT_MAX;
+#if CONFIG_VAR_TX
+    {
+      int r, c;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+          rd_stats->txb_coeff_cost_map[plane][r][c] = INT_MAX;
+    }
+#endif
+  }
+#endif
+}
+
+static INLINE void av1_merge_rd_stats(RD_STATS *rd_stats_dst,
+                                      const RD_STATS *rd_stats_src) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats_dst->rate += rd_stats_src->rate;
+  rd_stats_dst->dist += rd_stats_src->dist;
+  rd_stats_dst->sse += rd_stats_src->sse;
+  rd_stats_dst->skip &= rd_stats_src->skip;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats_dst->txb_coeff_cost[plane] += rd_stats_src->txb_coeff_cost[plane];
+#if CONFIG_VAR_TX
+    {
+      // TODO(angiebird): optimize this part
+      int r, c;
+      int ref_txb_coeff_cost = 0;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+          rd_stats_dst->txb_coeff_cost_map[plane][r][c] +=
+              rd_stats_src->txb_coeff_cost_map[plane][r][c];
+          ref_txb_coeff_cost += rd_stats_dst->txb_coeff_cost_map[plane][r][c];
+        }
+      assert(ref_txb_coeff_cost == rd_stats_dst->txb_coeff_cost[plane]);
+    }
+#endif
+  }
+#endif
+}
+
+struct TileInfo;
+struct TileDataEnc;
+struct AV1_COMP;
+struct macroblock;
+
+int av1_compute_rd_mult(const struct AV1_COMP *cpi, int qindex);
+
+void av1_initialize_rd_consts(struct AV1_COMP *cpi);
+
+void av1_initialize_me_consts(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                              int qindex);
+
+void av1_model_rd_from_var_lapndz(int64_t var, unsigned int n,
+                                  unsigned int qstep, int *rate, int64_t *dist);
+
+int av1_get_switchable_rate(const struct AV1_COMP *cpi, const MACROBLOCKD *xd);
+
+int av1_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block,
+                            int stride);
+
+int16_t *av1_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block,
+                                       int16_t *base);
+
+YV12_BUFFER_CONFIG *av1_get_scaled_ref_frame(const struct AV1_COMP *cpi,
+                                             int ref_frame);
+
+void av1_init_me_luts(void);
+
+#if CONFIG_REF_MV
+void av1_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, int ref,
+                    int ref_mv_idx);
+#endif
+
+void av1_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
+                              ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]);
+
+void av1_set_rd_speed_thresholds(struct AV1_COMP *cpi);
+
+void av1_set_rd_speed_thresholds_sub8x8(struct AV1_COMP *cpi);
+
+void av1_update_rd_thresh_fact(const AV1_COMMON *const cm,
+                               int (*fact)[MAX_MODES], int rd_thresh, int bsize,
+                               int best_mode_index);
+
+void av1_fill_token_costs(av1_coeff_cost *c,
+                          av1_coeff_probs_model (*p)[PLANE_TYPES]);
+
+static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
+                                      int thresh_fact) {
+  return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
+}
+
+void av1_mv_pred(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride, int ref_frame,
+                 BLOCK_SIZE block_size);
+
+static INLINE void set_error_per_bit(MACROBLOCK *x, int rdmult) {
+  x->errorperbit = rdmult >> RD_EPB_SHIFT;
+  x->errorperbit += (x->errorperbit == 0);
+}
+
+void av1_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv);
+
+int av1_get_intra_cost_penalty(int qindex, int qdelta,
+                               aom_bit_depth_t bit_depth);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_RD_H_

third_party/aom/av1/encoder/rdopt.h

@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RDOPT_H_
+#define AV1_ENCODER_RDOPT_H_
+
+#include "av1/common/blockd.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct TileInfo;
+struct AV1_COMP;
+struct macroblock;
+struct RD_STATS;
+
+#if CONFIG_RD_DEBUG
+static INLINE void av1_update_txb_coeff_cost(RD_STATS *rd_stats, int plane,
+                                             TX_SIZE tx_size, int blk_row,
+                                             int blk_col, int txb_coeff_cost) {
+  (void)blk_row;
+  (void)blk_col;
+  (void)tx_size;
+  rd_stats->txb_coeff_cost[plane] += txb_coeff_cost;
+
+#if CONFIG_VAR_TX
+  {
+    const int txb_h = tx_size_high_unit[tx_size];
+    const int txb_w = tx_size_wide_unit[tx_size];
+    int idx, idy;
+    for (idy = 0; idy < txb_h; ++idy)
+      for (idx = 0; idx < txb_w; ++idx)
+        rd_stats->txb_coeff_cost_map[plane][blk_row + idy][blk_col + idx] = 0;
+
+    rd_stats->txb_coeff_cost_map[plane][blk_row][blk_col] = txb_coeff_cost;
+  }
+  assert(blk_row < TXB_COEFF_COST_MAP_SIZE);
+  assert(blk_col < TXB_COEFF_COST_MAP_SIZE);
+#endif
+}
+#endif
+
+typedef enum OUTPUT_STATUS {
+  OUTPUT_HAS_PREDICTED_PIXELS,
+  OUTPUT_HAS_DECODED_PIXELS
+} OUTPUT_STATUS;
+
+void av1_dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                    BLOCK_SIZE plane_bsize, int block, int blk_row, int blk_col,
+                    TX_SIZE tx_size, int64_t *out_dist, int64_t *out_sse,
+                    OUTPUT_STATUS output_status);
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_cost_coeffs(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                    int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
+                    const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                    int use_fast_coef_costing);
+#endif
+void av1_rd_pick_intra_mode_sb(const struct AV1_COMP *cpi, struct macroblock *x,
+                               struct RD_STATS *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd);
+
+unsigned int av1_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs);
+#if CONFIG_HIGHBITDEPTH
+unsigned int av1_high_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                                const struct buf_2d *ref,
+                                                BLOCK_SIZE bs, int bd);
+#endif
+
+void av1_rd_pick_inter_mode_sb(const struct AV1_COMP *cpi,
+                               struct TileDataEnc *tile_data,
+                               struct macroblock *x, int mi_row, int mi_col,
+                               struct RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                               int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far);
+
+void av1_rd_pick_inter_mode_sb_seg_skip(
+    const struct AV1_COMP *cpi, struct TileDataEnc *tile_data,
+    struct macroblock *x, int mi_row, int mi_col, struct RD_STATS *rd_cost,
+    BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far);
+
+int av1_internal_image_edge(const struct AV1_COMP *cpi);
+int av1_active_h_edge(const struct AV1_COMP *cpi, int mi_row, int mi_step);
+int av1_active_v_edge(const struct AV1_COMP *cpi, int mi_col, int mi_step);
+int av1_active_edge_sb(const struct AV1_COMP *cpi, int mi_row, int mi_col);
+
+void av1_rd_pick_inter_mode_sub8x8(const struct AV1_COMP *cpi,
+                                   struct TileDataEnc *tile_data,
+                                   struct macroblock *x, int mi_row, int mi_col,
+                                   struct RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                                   int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far);
+
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+void av1_check_ncobmc_rd(const struct AV1_COMP *cpi, struct macroblock *x,
+                         int mi_row, int mi_col);
+#endif  // CONFIG_MOTION_VAR && CONFIG_NCOBMC
+
+#if CONFIG_SUPERTX
+#if CONFIG_VAR_TX
+void av1_tx_block_rd_b(const AV1_COMP *cpi, MACROBLOCK *x, TX_SIZE tx_size,
+                       int blk_row, int blk_col, int plane, int block,
+                       int plane_bsize, const ENTROPY_CONTEXT *a,
+                       const ENTROPY_CONTEXT *l, RD_STATS *rd_stats);
+#endif
+
+void av1_txfm_rd_in_plane_supertx(MACROBLOCK *x, const AV1_COMP *cpi, int *rate,
+                                  int64_t *distortion, int *skippable,
+                                  int64_t *sse, int64_t ref_best_rd, int plane,
+                                  BLOCK_SIZE bsize, TX_SIZE tx_size,
+                                  int use_fast_coef_casting);
+#endif  // CONFIG_SUPERTX
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+int av1_tx_type_cost(const AV1_COMP *cpi, const MACROBLOCKD *xd,
+                     BLOCK_SIZE bsize, int plane, TX_SIZE tx_size,
+                     TX_TYPE tx_type);
+
+#endif  // AV1_ENCODER_RDOPT_H_

third_party/aom/av1/encoder/segmentation.c

@@ -0,0 +1,394 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/pred_common.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/subexp.h"
+
+void av1_enable_segmentation(struct segmentation *seg) {
+  seg->enabled = 1;
+  seg->update_map = 1;
+  seg->update_data = 1;
+}
+
+void av1_disable_segmentation(struct segmentation *seg) {
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+}
+
+void av1_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta) {
+  seg->abs_delta = abs_delta;
+
+  memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data));
+}
+void av1_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] &= ~(1 << feature_id);
+}
+
+void av1_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id) {
+  seg->feature_data[segment_id][feature_id] = 0;
+}
+
+// Based on set of segment counts calculate a probability tree
+static void calc_segtree_probs(unsigned *segcounts,
+                               aom_prob *segment_tree_probs,
+                               const aom_prob *cur_tree_probs,
+                               const int probwt) {
+  // Work out probabilities of each segment
+  const unsigned cc[4] = { segcounts[0] + segcounts[1],
+                           segcounts[2] + segcounts[3],
+                           segcounts[4] + segcounts[5],
+                           segcounts[6] + segcounts[7] };
+  const unsigned ccc[2] = { cc[0] + cc[1], cc[2] + cc[3] };
+  int i;
+
+  segment_tree_probs[0] = get_binary_prob(ccc[0], ccc[1]);
+  segment_tree_probs[1] = get_binary_prob(cc[0], cc[1]);
+  segment_tree_probs[2] = get_binary_prob(cc[2], cc[3]);
+  segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]);
+  segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]);
+  segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]);
+  segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]);
+
+  for (i = 0; i < 7; i++) {
+    const unsigned *ct =
+        i == 0 ? ccc : i < 3 ? cc + (i & 2) : segcounts + (i - 3) * 2;
+    av1_prob_diff_update_savings_search(ct, cur_tree_probs[i],
+                                        &segment_tree_probs[i],
+                                        DIFF_UPDATE_PROB, probwt);
+  }
+}
+
+// Based on set of segment counts and probabilities calculate a cost estimate
+static int cost_segmap(unsigned *segcounts, aom_prob *probs) {
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+  const int c0123 = c01 + c23;
+  const int c4567 = c45 + c67;
+
+  // Cost the top node of the tree
+  int cost = c0123 * av1_cost_zero(probs[0]) + c4567 * av1_cost_one(probs[0]);
+
+  // Cost subsequent levels
+  if (c0123 > 0) {
+    cost += c01 * av1_cost_zero(probs[1]) + c23 * av1_cost_one(probs[1]);
+
+    if (c01 > 0)
+      cost += segcounts[0] * av1_cost_zero(probs[3]) +
+              segcounts[1] * av1_cost_one(probs[3]);
+    if (c23 > 0)
+      cost += segcounts[2] * av1_cost_zero(probs[4]) +
+              segcounts[3] * av1_cost_one(probs[4]);
+  }
+
+  if (c4567 > 0) {
+    cost += c45 * av1_cost_zero(probs[2]) + c67 * av1_cost_one(probs[2]);
+
+    if (c45 > 0)
+      cost += segcounts[4] * av1_cost_zero(probs[5]) +
+              segcounts[5] * av1_cost_one(probs[5]);
+    if (c67 > 0)
+      cost += segcounts[6] * av1_cost_zero(probs[6]) +
+              segcounts[7] * av1_cost_one(probs[6]);
+  }
+
+  return cost;
+}
+
+static void count_segs(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                       const TileInfo *tile, MODE_INFO **mi,
+                       unsigned *no_pred_segcounts,
+                       unsigned (*temporal_predictor_count)[2],
+                       unsigned *t_unpred_seg_counts, int bw, int bh,
+                       int mi_row, int mi_col) {
+  int segment_id;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  xd->mi = mi;
+  segment_id = xd->mi[0]->mbmi.segment_id;
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  // Count the number of hits on each segment with no prediction
+  no_pred_segcounts[segment_id]++;
+
+  // Temporal prediction not allowed on key frames
+  if (cm->frame_type != KEY_FRAME) {
+    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+    // Test to see if the segment id matches the predicted value.
+    const int pred_segment_id =
+        get_segment_id(cm, cm->last_frame_seg_map, bsize, mi_row, mi_col);
+    const int pred_flag = pred_segment_id == segment_id;
+    const int pred_context = av1_get_pred_context_seg_id(xd);
+
+    // Store the prediction status for this mb and update counts
+    // as appropriate
+    xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
+    temporal_predictor_count[pred_context][pred_flag]++;
+
+    // Update the "unpredicted" segment count
+    if (!pred_flag) t_unpred_seg_counts[segment_id]++;
+  }
+}
+
+static void count_segs_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                          const TileInfo *tile, MODE_INFO **mi,
+                          unsigned *no_pred_segcounts,
+                          unsigned (*temporal_predictor_count)[2],
+                          unsigned *t_unpred_seg_counts, int mi_row, int mi_col,
+                          BLOCK_SIZE bsize) {
+  const int mis = cm->mi_stride;
+  const int bs = mi_size_wide[bsize], hbs = bs / 2;
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPE partition;
+#else
+  int bw, bh;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+#if CONFIG_EXT_PARTITION_TYPES
+  if (bsize == BLOCK_8X8)
+    partition = PARTITION_NONE;
+  else
+    partition = get_partition(cm, mi_row, mi_col, bsize);
+  switch (partition) {
+    case PARTITION_NONE:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+      break;
+    case PARTITION_HORZ:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+                 mi_row + hbs, mi_col);
+      break;
+    case PARTITION_VERT:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+                 mi_col + hbs);
+      break;
+    case PARTITION_HORZ_A:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row, mi_col + hbs);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+                 mi_row + hbs, mi_col);
+      break;
+    case PARTITION_HORZ_B:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col);
+      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col + hbs);
+      break;
+    case PARTITION_VERT_A:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+                 mi_col + hbs);
+      break;
+    case PARTITION_VERT_B:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row, mi_col + hbs);
+      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col + hbs);
+      break;
+    case PARTITION_SPLIT: {
+      const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+      int n;
+
+      assert(num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type] < bs &&
+             num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type] < bs);
+
+      for (n = 0; n < 4; n++) {
+        const int mi_dc = hbs * (n & 1);
+        const int mi_dr = hbs * (n >> 1);
+
+        count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
+                      temporal_predictor_count, t_unpred_seg_counts,
+                      mi_row + mi_dr, mi_col + mi_dc, subsize);
+      }
+    } break;
+    default: assert(0);
+  }
+#else
+  bw = mi_size_wide[mi[0]->mbmi.sb_type];
+  bh = mi_size_high[mi[0]->mbmi.sb_type];
+
+  if (bw == bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+  } else if (bw == bs && bh < bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+               mi_row + hbs, mi_col);
+  } else if (bw < bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+               mi_col + hbs);
+  } else {
+    const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+    int n;
+
+    assert(bw < bs && bh < bs);
+
+    for (n = 0; n < 4; n++) {
+      const int mi_dc = hbs * (n & 1);
+      const int mi_dr = hbs * (n >> 1);
+
+      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
+                    temporal_predictor_count, t_unpred_seg_counts,
+                    mi_row + mi_dr, mi_col + mi_dc, subsize);
+    }
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+void av1_choose_segmap_coding_method(AV1_COMMON *cm, MACROBLOCKD *xd) {
+  struct segmentation *seg = &cm->seg;
+  struct segmentation_probs *segp = &cm->fc->seg;
+
+  int no_pred_cost;
+  int t_pred_cost = INT_MAX;
+
+  int i, tile_col, tile_row, mi_row, mi_col;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+
+  unsigned(*temporal_predictor_count)[2] = cm->counts.seg.pred;
+  unsigned *no_pred_segcounts = cm->counts.seg.tree_total;
+  unsigned *t_unpred_seg_counts = cm->counts.seg.tree_mispred;
+
+  aom_prob no_pred_tree[SEG_TREE_PROBS];
+  aom_prob t_pred_tree[SEG_TREE_PROBS];
+  aom_prob t_nopred_prob[PREDICTION_PROBS];
+
+  (void)xd;
+
+  // We are about to recompute all the segment counts, so zero the accumulators.
+  av1_zero(cm->counts.seg);
+
+  // First of all generate stats regarding how well the last segment map
+  // predicts this one
+  for (tile_row = 0; tile_row < cm->tile_rows; tile_row++) {
+    TileInfo tile_info;
+    av1_tile_set_row(&tile_info, cm, tile_row);
+    for (tile_col = 0; tile_col < cm->tile_cols; tile_col++) {
+      MODE_INFO **mi_ptr;
+      av1_tile_set_col(&tile_info, cm, tile_col);
+#if CONFIG_TILE_GROUPS && CONFIG_DEPENDENT_HORZTILES
+      av1_tile_set_tg_boundary(&tile_info, cm, tile_row, tile_col);
+#endif
+      mi_ptr = cm->mi_grid_visible + tile_info.mi_row_start * cm->mi_stride +
+               tile_info.mi_col_start;
+      for (mi_row = tile_info.mi_row_start; mi_row < tile_info.mi_row_end;
+           mi_row += cm->mib_size, mi_ptr += cm->mib_size * cm->mi_stride) {
+        MODE_INFO **mi = mi_ptr;
+        for (mi_col = tile_info.mi_col_start; mi_col < tile_info.mi_col_end;
+             mi_col += cm->mib_size, mi += cm->mib_size) {
+          count_segs_sb(cm, xd, &tile_info, mi, no_pred_segcounts,
+                        temporal_predictor_count, t_unpred_seg_counts, mi_row,
+                        mi_col, cm->sb_size);
+        }
+      }
+    }
+  }
+
+  // Work out probability tree for coding segments without prediction
+  // and the cost.
+  calc_segtree_probs(no_pred_segcounts, no_pred_tree, segp->tree_probs, probwt);
+  no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree);
+
+  // Key frames cannot use temporal prediction
+  if (!frame_is_intra_only(cm) && !cm->error_resilient_mode) {
+    // Work out probability tree for coding those segments not
+    // predicted using the temporal method and the cost.
+    calc_segtree_probs(t_unpred_seg_counts, t_pred_tree, segp->tree_probs,
+                       probwt);
+    t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree);
+
+    // Add in the cost of the signaling for each prediction context.
+    for (i = 0; i < PREDICTION_PROBS; i++) {
+      const int count0 = temporal_predictor_count[i][0];
+      const int count1 = temporal_predictor_count[i][1];
+
+      t_nopred_prob[i] = get_binary_prob(count0, count1);
+      av1_prob_diff_update_savings_search(
+          temporal_predictor_count[i], segp->pred_probs[i], &t_nopred_prob[i],
+          DIFF_UPDATE_PROB, probwt);
+
+      // Add in the predictor signaling cost
+      t_pred_cost += count0 * av1_cost_zero(t_nopred_prob[i]) +
+                     count1 * av1_cost_one(t_nopred_prob[i]);
+    }
+  }
+
+  // Now choose which coding method to use.
+  if (t_pred_cost < no_pred_cost) {
+    assert(!cm->error_resilient_mode);
+    seg->temporal_update = 1;
+  } else {
+    seg->temporal_update = 0;
+  }
+}
+
+void av1_reset_segment_features(AV1_COMMON *cm) {
+  struct segmentation *seg = &cm->seg;
+
+  // Set up default state for MB feature flags
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+  av1_clearall_segfeatures(seg);
+}

third_party/aom/av1/encoder/segmentation.h

@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SEGMENTATION_H_
+#define AV1_ENCODER_SEGMENTATION_H_
+
+#include "av1/common/blockd.h"
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_enable_segmentation(struct segmentation *seg);
+void av1_disable_segmentation(struct segmentation *seg);
+
+void av1_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id);
+void av1_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id);
+
+// The values given for each segment can be either deltas (from the default
+// value chosen for the frame) or absolute values.
+//
+// Valid range for abs values is (0-127 for MB_LVL_ALT_Q), (0-63 for
+// SEGMENT_ALT_LF)
+// Valid range for delta values are (+/-127 for MB_LVL_ALT_Q), (+/-63 for
+// SEGMENT_ALT_LF)
+//
+// abs_delta = SEGMENT_DELTADATA (deltas) abs_delta = SEGMENT_ABSDATA (use
+// the absolute values given).
+void av1_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta);
+
+void av1_choose_segmap_coding_method(AV1_COMMON *cm, MACROBLOCKD *xd);
+
+void av1_reset_segment_features(AV1_COMMON *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SEGMENTATION_H_

third_party/aom/av1/encoder/speed_features.c

@@ -0,0 +1,506 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/speed_features.h"
+#include "av1/encoder/rdopt.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#define MAX_MESH_SPEED 5  // Max speed setting for mesh motion method
+static MESH_PATTERN
+    good_quality_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
+      { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
+      { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
+      { { 64, 8 }, { 14, 2 }, { 7, 1 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+    };
+static unsigned char good_quality_max_mesh_pct[MAX_MESH_SPEED + 1] = {
+  50, 25, 15, 5, 1, 1
+};
+
+#if CONFIG_INTRABC
+// TODO(aconverse@google.com): These settings are pretty relaxed, tune them for
+// each speed setting
+static MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+};
+static uint8_t intrabc_max_mesh_pct[MAX_MESH_SPEED + 1] = { 100, 100, 100,
+                                                            25,  25,  10 };
+#endif
+
+// Intra only frames, golden frames (except alt ref overlays) and
+// alt ref frames tend to be coded at a higher than ambient quality
+static int frame_is_boosted(const AV1_COMP *cpi) {
+  return frame_is_kf_gf_arf(cpi);
+}
+
+// Sets a partition size down to which the auto partition code will always
+// search (can go lower), based on the image dimensions. The logic here
+// is that the extent to which ringing artefacts are offensive, depends
+// partly on the screen area that over which they propogate. Propogation is
+// limited by transform block size but the screen area take up by a given block
+// size will be larger for a small image format stretched to full screen.
+static BLOCK_SIZE set_partition_min_limit(AV1_COMMON *const cm) {
+  unsigned int screen_area = (cm->width * cm->height);
+
+  // Select block size based on image format size.
+  if (screen_area < 1280 * 720) {
+    // Formats smaller in area than 720P
+    return BLOCK_4X4;
+  } else if (screen_area < 1920 * 1080) {
+    // Format >= 720P and < 1080P
+    return BLOCK_8X8;
+  } else {
+    // Formats 1080P and up
+    return BLOCK_16X16;
+  }
+}
+
+static void set_good_speed_feature_framesize_dependent(AV1_COMP *cpi,
+                                                       SPEED_FEATURES *sf,
+                                                       int speed) {
+  AV1_COMMON *const cm = &cpi->common;
+
+  if (speed >= 1) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask =
+          cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 21);
+    }
+  }
+
+  if (speed >= 2) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask =
+          cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
+      sf->adaptive_pred_interp_filter = 0;
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+      sf->partition_search_breakout_rate_thr = 120;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+      sf->partition_search_breakout_dist_thr = (1 << 22);
+      sf->partition_search_breakout_rate_thr = 100;
+    }
+    sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
+  }
+
+  if (speed >= 3) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = DISABLE_ALL_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+      sf->partition_search_breakout_rate_thr = 200;
+    } 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->partition_search_breakout_dist_thr = (1 << 23);
+      sf->partition_search_breakout_rate_thr = 120;
+    }
+  }
+
+  // If this is a two pass clip that fits the criteria for animated or
+  // graphics content then reset disable_split_mask for speeds 1-4.
+  // Also if the image edge is internal to the coded area.
+  if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
+      ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+       (av1_internal_image_edge(cpi)))) {
+    sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+  }
+
+  if (speed >= 4) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 26);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+    }
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+  }
+}
+
+static void set_good_speed_feature(AV1_COMP *cpi, AV1_COMMON *cm,
+                                   SPEED_FEATURES *sf, int speed) {
+  const int boosted = frame_is_boosted(cpi);
+
+  if (speed >= 1) {
+    sf->tx_type_search.fast_intra_tx_type_search = 1;
+    sf->tx_type_search.fast_inter_tx_type_search = 1;
+  }
+
+  if (speed >= 2) {
+    if ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+        av1_internal_image_edge(cpi)) {
+      sf->use_square_partition_only = !frame_is_boosted(cpi);
+    } else {
+      sf->use_square_partition_only = !frame_is_intra_only(cm);
+    }
+
+    sf->less_rectangular_check = 1;
+
+    sf->use_rd_breakout = 1;
+    sf->adaptive_motion_search = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 1;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->mode_skip_start = 10;
+    sf->adaptive_pred_interp_filter = 1;
+
+    sf->recode_loop = ALLOW_RECODE_KFARFGF;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_64X64] = INTRA_DC_H_V;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+
+    sf->tx_size_search_breakout = 1;
+    sf->partition_search_breakout_rate_thr = 80;
+    sf->tx_type_search.prune_mode = PRUNE_ONE;
+    // Use transform domain distortion.
+    // Note var-tx expt always uses pixel domain distortion.
+    sf->use_transform_domain_distortion = 1;
+#if CONFIG_EXT_INTER
+    sf->disable_wedge_search_var_thresh = 100;
+    sf->fast_wedge_sign_estimate = 1;
+#endif  // CONFIG_EXT_INTER
+  }
+
+  if (speed >= 3) {
+    sf->tx_size_search_method =
+        frame_is_boosted(cpi) ? USE_FULL_RD : USE_LARGESTALL;
+    sf->mode_search_skip_flags =
+        (cm->frame_type == KEY_FRAME)
+            ? 0
+            : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
+                  FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->allow_partition_search_skip = 1;
+    sf->use_upsampled_references = 0;
+    sf->adaptive_rd_thresh = 2;
+#if CONFIG_EXT_TX
+    sf->tx_type_search.prune_mode = PRUNE_TWO;
+#endif
+  }
+
+  if (speed >= 4) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->tx_size_search_method =
+        frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
+    sf->adaptive_pred_interp_filter = 0;
+    sf->adaptive_mode_search = 1;
+    sf->cb_partition_search = !boosted;
+    sf->cb_pred_filter_search = 1;
+    sf->alt_ref_search_fp = 1;
+    sf->recode_loop = ALLOW_RECODE_KFMAXBW;
+    sf->adaptive_rd_thresh = 3;
+    sf->mode_skip_start = 6;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_64X64] = INTRA_DC;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
+    sf->adaptive_interp_filter_search = 1;
+  }
+
+  if (speed >= 5) {
+    sf->use_square_partition_only = 1;
+    sf->tx_size_search_method = USE_LARGESTALL;
+    sf->mv.search_method = BIGDIA;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
+    sf->adaptive_rd_thresh = 4;
+    if (cm->frame_type != KEY_FRAME)
+      sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
+    sf->disable_filter_search_var_thresh = 200;
+    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
+    sf->use_fast_coef_costing = 1;
+    sf->partition_search_breakout_rate_thr = 300;
+  }
+
+  if (speed >= 6) {
+    int i;
+    sf->optimize_coefficients = 0;
+    sf->mv.search_method = HEX;
+    sf->disable_filter_search_var_thresh = 500;
+    for (i = 0; i < TX_SIZES; ++i) {
+      sf->intra_y_mode_mask[i] = INTRA_DC;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->partition_search_breakout_rate_thr = 500;
+    sf->mv.reduce_first_step_size = 1;
+    sf->simple_model_rd_from_var = 1;
+  }
+  if (speed >= 7) {
+    const int is_keyframe = cm->frame_type == KEY_FRAME;
+    const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
+    sf->default_max_partition_size = BLOCK_32X32;
+    sf->default_min_partition_size = BLOCK_8X8;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->frame_parameter_update = 0;
+    sf->mv.search_method = FAST_HEX;
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
+#if CONFIG_EXT_PARTITION
+    sf->inter_mode_mask[BLOCK_64X128] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_128X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_128X128] = INTER_NEAREST;
+#endif  // CONFIG_EXT_PARTITION
+    sf->partition_search_type = REFERENCE_PARTITION;
+    sf->default_min_partition_size = BLOCK_8X8;
+    sf->reuse_inter_pred_sby = 1;
+    sf->force_frame_boost =
+        is_keyframe ||
+        (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
+    sf->max_delta_qindex = is_keyframe ? 20 : 15;
+    sf->coeff_prob_appx_step = 4;
+    sf->mode_search_skip_flags |= FLAG_SKIP_INTRA_DIRMISMATCH;
+  }
+}
+
+void av1_set_speed_features_framesize_dependent(AV1_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  AV1_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+// Limit memory usage for high resolutions
+#if CONFIG_EXT_REFS
+  // TODO(zoeliu): Temporary solution to resolve the insufficient RAM issue for
+  //               ext-refs. Need to work with @yunqingwang to have a more
+  //               effective solution.
+  if (AOMMIN(cm->width, cm->height) > 720) {
+    // Turn off the use of upsampled references for HD resolution
+    sf->use_upsampled_references = 0;
+  } else if ((AOMMIN(cm->width, cm->height) > 540) &&
+             (oxcf->profile != PROFILE_0)) {
+    sf->use_upsampled_references = 0;
+  }
+#else
+  if (AOMMIN(cm->width, cm->height) > 1080) {
+    sf->use_upsampled_references = 0;
+  } else if ((AOMMIN(cm->width, cm->height) > 720) &&
+             (oxcf->profile != PROFILE_0)) {
+    sf->use_upsampled_references = 0;
+  }
+#endif  // CONFIG_EXT_REFS
+
+  if (oxcf->mode == GOOD) {
+    set_good_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
+  }
+
+  if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
+    sf->adaptive_pred_interp_filter = 0;
+  }
+
+  // Check for masked out split cases.
+  for (i = 0; i < MAX_REFS; ++i) {
+    if (sf->disable_split_mask & (1 << i)) {
+      rd->thresh_mult_sub8x8[i] = INT_MAX;
+    }
+  }
+
+  // This is only used in motion vector unit test.
+  if (cpi->oxcf.motion_vector_unit_test == 1)
+    cpi->find_fractional_mv_step = av1_return_max_sub_pixel_mv;
+  else if (cpi->oxcf.motion_vector_unit_test == 2)
+    cpi->find_fractional_mv_step = av1_return_min_sub_pixel_mv;
+}
+
+void av1_set_speed_features_framesize_independent(AV1_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  int i;
+
+  // best quality defaults
+  sf->frame_parameter_update = 1;
+  sf->mv.search_method = NSTEP;
+  sf->recode_loop = ALLOW_RECODE;
+  sf->mv.subpel_search_method = SUBPEL_TREE;
+  sf->mv.subpel_iters_per_step = 2;
+  sf->mv.subpel_force_stop = 0;
+  sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
+  sf->mv.reduce_first_step_size = 0;
+  sf->coeff_prob_appx_step = 1;
+  sf->mv.auto_mv_step_size = 0;
+  sf->mv.fullpel_search_step_param = 6;
+  sf->comp_inter_joint_search_thresh = BLOCK_4X4;
+  sf->adaptive_rd_thresh = 0;
+  sf->tx_size_search_method = USE_FULL_RD;
+  sf->adaptive_motion_search = 0;
+  sf->adaptive_pred_interp_filter = 0;
+  sf->adaptive_mode_search = 0;
+  sf->cb_pred_filter_search = 0;
+  sf->cb_partition_search = 0;
+  sf->alt_ref_search_fp = 0;
+  sf->partition_search_type = SEARCH_PARTITION;
+  sf->tx_type_search.prune_mode = NO_PRUNE;
+  sf->tx_type_search.fast_intra_tx_type_search = 0;
+  sf->tx_type_search.fast_inter_tx_type_search = 0;
+  sf->less_rectangular_check = 0;
+  sf->use_square_partition_only = 0;
+  sf->auto_min_max_partition_size = NOT_IN_USE;
+  sf->rd_auto_partition_min_limit = BLOCK_4X4;
+  sf->default_max_partition_size = BLOCK_LARGEST;
+  sf->default_min_partition_size = BLOCK_4X4;
+  sf->adjust_partitioning_from_last_frame = 0;
+  sf->last_partitioning_redo_frequency = 4;
+  sf->disable_split_mask = 0;
+  sf->mode_search_skip_flags = 0;
+  sf->force_frame_boost = 0;
+  sf->max_delta_qindex = 0;
+  sf->disable_filter_search_var_thresh = 0;
+  sf->adaptive_interp_filter_search = 0;
+  sf->allow_partition_search_skip = 0;
+  sf->use_upsampled_references = 1;
+#if CONFIG_EXT_INTER
+  sf->disable_wedge_search_var_thresh = 0;
+  sf->fast_wedge_sign_estimate = 0;
+#endif  // CONFIG_EXT_INTER
+
+  for (i = 0; i < TX_SIZES; i++) {
+    sf->intra_y_mode_mask[i] = INTRA_ALL;
+    sf->intra_uv_mode_mask[i] = INTRA_ALL;
+  }
+  sf->use_rd_breakout = 0;
+  sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
+  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;
+  for (i = 0; i < BLOCK_SIZES; ++i) sf->inter_mode_mask[i] = INTER_ALL;
+  sf->max_intra_bsize = BLOCK_LARGEST;
+  sf->reuse_inter_pred_sby = 0;
+  // This setting only takes effect when partition_search_type is set
+  // to FIXED_PARTITION.
+  sf->always_this_block_size = BLOCK_16X16;
+  sf->search_type_check_frequency = 50;
+  // Recode loop tolerance %.
+  sf->recode_tolerance = 25;
+  sf->default_interp_filter = SWITCHABLE;
+  sf->tx_size_search_breakout = 0;
+  sf->partition_search_breakout_dist_thr = 0;
+  sf->partition_search_breakout_rate_thr = 0;
+  sf->simple_model_rd_from_var = 0;
+
+  // Set this at the appropriate speed levels
+  sf->use_transform_domain_distortion = 0;
+
+  if (oxcf->mode == GOOD
+#if CONFIG_XIPHRC
+      || oxcf->pass == 1
+#endif
+      )
+    set_good_speed_feature(cpi, cm, sf, oxcf->speed);
+
+  // sf->partition_search_breakout_dist_thr is set assuming max 64x64
+  // blocks. Normalise this if the blocks are bigger.
+  if (MAX_SB_SIZE_LOG2 > 6) {
+    sf->partition_search_breakout_dist_thr <<= 2 * (MAX_SB_SIZE_LOG2 - 6);
+  }
+
+  cpi->full_search_sad = av1_full_search_sad;
+  cpi->diamond_search_sad = av1_diamond_search_sad;
+
+  sf->allow_exhaustive_searches = 1;
+  int speed = (oxcf->speed > MAX_MESH_SPEED) ? MAX_MESH_SPEED : oxcf->speed;
+  if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+    sf->exhaustive_searches_thresh = (1 << 24);
+  else
+    sf->exhaustive_searches_thresh = (1 << 25);
+  sf->max_exaustive_pct = good_quality_max_mesh_pct[speed];
+  if (speed > 0)
+    sf->exhaustive_searches_thresh = sf->exhaustive_searches_thresh << 1;
+
+  for (i = 0; i < MAX_MESH_STEP; ++i) {
+    sf->mesh_patterns[i].range = good_quality_mesh_patterns[speed][i].range;
+    sf->mesh_patterns[i].interval =
+        good_quality_mesh_patterns[speed][i].interval;
+  }
+#if CONFIG_INTRABC
+  if ((frame_is_intra_only(cm) && cm->allow_screen_content_tools) &&
+      (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
+       cpi->oxcf.content == AOM_CONTENT_SCREEN)) {
+    for (i = 0; i < MAX_MESH_STEP; ++i) {
+      sf->mesh_patterns[i].range = intrabc_mesh_patterns[speed][i].range;
+      sf->mesh_patterns[i].interval = intrabc_mesh_patterns[speed][i].interval;
+    }
+    sf->max_exaustive_pct = intrabc_max_mesh_pct[speed];
+  }
+#endif  // CONFIG_INTRABC
+
+#if !CONFIG_XIPHRC
+  // Slow quant, dct and trellis not worthwhile for first pass
+  // so make sure they are always turned off.
+  if (oxcf->pass == 1) sf->optimize_coefficients = 0;
+#endif
+
+  // No recode for 1 pass.
+  if (oxcf->pass == 0) {
+    sf->recode_loop = DISALLOW_RECODE;
+    sf->optimize_coefficients = 0;
+  }
+
+  if (sf->mv.subpel_search_method == SUBPEL_TREE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned_more;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned_evenmore;
+  }
+
+#if !CONFIG_AOM_QM
+  x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
+#else
+  // FIXME: trellis not very efficient for quantisation matrices
+  x->optimize = 0;
+#endif
+
+  x->min_partition_size = sf->default_min_partition_size;
+  x->max_partition_size = sf->default_max_partition_size;
+
+  if (!cpi->oxcf.frame_periodic_boost) {
+    sf->max_delta_qindex = 0;
+  }
+
+  // This is only used in motion vector unit test.
+  if (cpi->oxcf.motion_vector_unit_test == 1)
+    cpi->find_fractional_mv_step = av1_return_max_sub_pixel_mv;
+  else if (cpi->oxcf.motion_vector_unit_test == 2)
+    cpi->find_fractional_mv_step = av1_return_min_sub_pixel_mv;
+}

third_party/aom/av1/encoder/speed_features.h

@@ -0,0 +1,484 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SPEED_FEATURES_H_
+#define AV1_ENCODER_SPEED_FEATURES_H_
+
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
+              (1 << D135_PRED) | (1 << D117_PRED) | (1 << D153_PRED) |
+              (1 << D207_PRED) | (1 << D63_PRED) |
+#if CONFIG_ALT_INTRA
+              (1 << SMOOTH_PRED) |
+#endif  // CONFIG_ALT_INTRA
+              (1 << TM_PRED),
+  INTRA_DC = (1 << DC_PRED),
+  INTRA_DC_TM = (1 << DC_PRED) | (1 << TM_PRED),
+  INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
+  INTRA_DC_TM_H_V =
+      (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) | (1 << H_PRED)
+};
+
+#if CONFIG_EXT_INTER
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV) |
+              (1 << NEAREST_NEARESTMV) | (1 << NEAR_NEARMV) |
+              (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) | (1 << NEW_NEWMV) |
+              (1 << NEAREST_NEWMV) | (1 << NEAR_NEWMV) | (1 << NEW_NEARMV) |
+              (1 << NEW_NEARESTMV) | (1 << ZERO_ZEROMV),
+  INTER_NEAREST = (1 << NEARESTMV) | (1 << NEAREST_NEARESTMV) |
+                  (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) |
+                  (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV) |
+                      (1 << NEAREST_NEARESTMV) | (1 << NEW_NEWMV) |
+                      (1 << NEAR_NEARESTMV) | (1 << NEAREST_NEARMV) |
+                      (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV) |
+                      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) |
+                       (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) |
+                       (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) |
+                       (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV),
+  INTER_NEAREST_NEW_ZERO =
+      (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) | (1 << NEW_NEWMV) |
+      (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) | (1 << NEW_NEARESTMV) |
+      (1 << NEAREST_NEWMV) | (1 << NEW_NEARMV) | (1 << NEAR_NEWMV),
+  INTER_NEAREST_NEAR_NEW =
+      (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << NEW_NEWMV) | (1 << NEAREST_NEARMV) |
+      (1 << NEAR_NEARESTMV) | (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV) |
+      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) | (1 << NEAR_NEARMV),
+  INTER_NEAREST_NEAR_ZERO =
+      (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) | (1 << NEAREST_NEARMV) |
+      (1 << NEAR_NEARESTMV) | (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV) |
+      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) | (1 << NEAR_NEARMV),
+};
+#else
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST = (1 << NEARESTMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
+  INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
+};
+#endif  // CONFIG_EXT_INTER
+
+enum {
+  DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
+                            (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
+
+  DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
+
+  DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
+
+  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) | (1 << THR_GOLD)
+};
+
+typedef enum {
+  DIAMOND = 0,
+  NSTEP = 1,
+  HEX = 2,
+  BIGDIA = 3,
+  SQUARE = 4,
+  FAST_HEX = 5,
+  FAST_DIAMOND = 6
+} SEARCH_METHODS;
+
+typedef enum {
+  // No recode.
+  DISALLOW_RECODE = 0,
+  // Allow recode for KF and exceeding maximum frame bandwidth.
+  ALLOW_RECODE_KFMAXBW = 1,
+  // Allow recode only for KF/ARF/GF frames.
+  ALLOW_RECODE_KFARFGF = 2,
+  // Allow recode for all frames based on bitrate constraints.
+  ALLOW_RECODE = 3,
+} RECODE_LOOP_TYPE;
+
+typedef enum {
+  SUBPEL_TREE = 0,
+  SUBPEL_TREE_PRUNED = 1,           // Prunes 1/2-pel searches
+  SUBPEL_TREE_PRUNED_MORE = 2,      // Prunes 1/2-pel searches more aggressively
+  SUBPEL_TREE_PRUNED_EVENMORE = 3,  // Prunes 1/2- and 1/4-pel searches
+  // Other methods to come
+} SUBPEL_SEARCH_METHODS;
+
+typedef enum {
+  NO_MOTION_THRESHOLD = 0,
+  LOW_MOTION_THRESHOLD = 7
+} MOTION_THRESHOLD;
+
+typedef enum {
+  USE_FULL_RD = 0,
+  USE_LARGESTALL,
+  USE_TX_8X8
+} TX_SIZE_SEARCH_METHOD;
+
+typedef enum {
+  NOT_IN_USE = 0,
+  RELAXED_NEIGHBORING_MIN_MAX = 1
+} AUTO_MIN_MAX_MODE;
+
+typedef enum {
+  // Try the full image with different values.
+  LPF_PICK_FROM_FULL_IMAGE,
+  // Try a small portion of the image with different values.
+  LPF_PICK_FROM_SUBIMAGE,
+  // Estimate the level based on quantizer and frame type
+  LPF_PICK_FROM_Q,
+  // Pick 0 to disable LPF if LPF was enabled last frame
+  LPF_PICK_MINIMAL_LPF
+} LPF_PICK_METHOD;
+
+typedef enum {
+  // Terminate search early based on distortion so far compared to
+  // qp step, distortion in the neighborhood of the frame, etc.
+  FLAG_EARLY_TERMINATE = 1 << 0,
+
+  // Skips comp inter modes if the best so far is an intra mode.
+  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
+
+  // Skips oblique intra modes if the best so far is an inter mode.
+  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
+
+  // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
+  // intra so far is not one of the neighboring directions.
+  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
+
+  // Skips intra modes other than DC_PRED if the source variance is small
+  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
+} MODE_SEARCH_SKIP_LOGIC;
+
+typedef enum {
+  FLAG_SKIP_EIGHTTAP_REGULAR = 1 << EIGHTTAP_REGULAR,
+  FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
+  FLAG_SKIP_MULTITAP_SHARP = 1 << MULTITAP_SHARP,
+} INTERP_FILTER_MASK;
+
+typedef enum {
+  NO_PRUNE = 0,
+  // eliminates one tx type in vertical and horizontal direction
+  PRUNE_ONE = 1,
+#if CONFIG_EXT_TX
+  // eliminates two tx types in each direction
+  PRUNE_TWO = 2,
+#endif
+} TX_TYPE_PRUNE_MODE;
+
+typedef struct {
+  TX_TYPE_PRUNE_MODE prune_mode;
+  int fast_intra_tx_type_search;
+  int fast_inter_tx_type_search;
+} TX_TYPE_SEARCH;
+
+typedef enum {
+  // Search partitions using RD criterion
+  SEARCH_PARTITION,
+
+  // Always use a fixed size partition
+  FIXED_PARTITION,
+
+  REFERENCE_PARTITION,
+
+  // Use an arbitrary partitioning scheme based on source variance within
+  // a 64X64 SB
+  VAR_BASED_PARTITION,
+
+  // Use non-fixed partitions based on source variance
+  SOURCE_VAR_BASED_PARTITION
+} PARTITION_SEARCH_TYPE;
+
+typedef enum {
+  // Does a dry run to see if any of the contexts need to be updated or not,
+  // before the final run.
+  TWO_LOOP = 0,
+
+  // No dry run, also only half the coef contexts and bands are updated.
+  // The rest are not updated at all.
+  ONE_LOOP_REDUCED = 1
+} FAST_COEFF_UPDATE;
+
+typedef struct MV_SPEED_FEATURES {
+  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
+  SEARCH_METHODS search_method;
+
+  // This parameter controls which step in the n-step process we start at.
+  // It's changed adaptively based on circumstances.
+  int reduce_first_step_size;
+
+  // If this is set to 1, we limit the motion search range to 2 times the
+  // largest motion vector found in the last frame.
+  int auto_mv_step_size;
+
+  // Subpel_search_method can only be subpel_tree which does a subpixel
+  // logarithmic search that keeps stepping at 1/2 pixel units until
+  // you stop getting a gain, and then goes on to 1/4 and repeats
+  // the same process. Along the way it skips many diagonals.
+  SUBPEL_SEARCH_METHODS subpel_search_method;
+
+  // Maximum number of steps in logarithmic subpel search before giving up.
+  int subpel_iters_per_step;
+
+  // Control when to stop subpel search
+  int subpel_force_stop;
+
+  // This variable sets the step_param used in full pel motion search.
+  int fullpel_search_step_param;
+} MV_SPEED_FEATURES;
+
+#define MAX_MESH_STEP 4
+
+typedef struct MESH_PATTERN {
+  int range;
+  int interval;
+} MESH_PATTERN;
+
+typedef struct SPEED_FEATURES {
+  MV_SPEED_FEATURES mv;
+
+  // Frame level coding parameter update
+  int frame_parameter_update;
+
+  RECODE_LOOP_TYPE recode_loop;
+
+  // Trellis (dynamic programming) optimization of quantized values (+1, 0).
+  int optimize_coefficients;
+
+  // Always set to 0. If on it enables 0 cost background transmission
+  // (except for the initial transmission of the segmentation). The feature is
+  // disabled because the addition of very large block sizes make the
+  // backgrounds very to cheap to encode, and the segmentation we have
+  // adds overhead.
+  int static_segmentation;
+
+  // If 1 we iterate finding a best reference for 2 ref frames together - via
+  // a log search that iterates 4 times (check around mv for last for best
+  // error of combined predictor then check around mv for alt). If 0 we
+  // we just use the best motion vector found for each frame by itself.
+  BLOCK_SIZE comp_inter_joint_search_thresh;
+
+  // This variable is used to cap the maximum number of times we skip testing a
+  // mode to be evaluated. A high value means we will be faster.
+  int adaptive_rd_thresh;
+
+  // Coefficient probability model approximation step size
+  int coeff_prob_appx_step;
+
+  // The threshold is to determine how slow the motino is, it is used when
+  // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
+  MOTION_THRESHOLD lf_motion_threshold;
+
+  // Determine which method we use to determine transform size. We can choose
+  // between options like full rd, largest for prediction size, largest
+  // for intra and model coefs for the rest.
+  TX_SIZE_SEARCH_METHOD tx_size_search_method;
+
+  // After looking at the first set of modes (set by index here), skip
+  // checking modes for reference frames that don't match the reference frame
+  // of the best so far.
+  int mode_skip_start;
+
+  PARTITION_SEARCH_TYPE partition_search_type;
+
+  TX_TYPE_SEARCH tx_type_search;
+
+  // Used if partition_search_type = FIXED_SIZE_PARTITION
+  BLOCK_SIZE always_this_block_size;
+
+  // Skip rectangular partition test when partition type none gives better
+  // rd than partition type split.
+  int less_rectangular_check;
+
+  // Disable testing non square partitions. (eg 16x32)
+  int use_square_partition_only;
+
+  // Sets min and max partition sizes for this superblock based on the
+  // same superblock in last encoded frame, and the left and above neighbor.
+  AUTO_MIN_MAX_MODE auto_min_max_partition_size;
+  // Ensures the rd based auto partition search will always
+  // go down at least to the specified level.
+  BLOCK_SIZE rd_auto_partition_min_limit;
+
+  // Min and max partition size we enable (block_size) as per auto
+  // min max, but also used by adjust partitioning, and pick_partitioning.
+  BLOCK_SIZE default_min_partition_size;
+  BLOCK_SIZE default_max_partition_size;
+
+  // Whether or not we allow partitions one smaller or one greater than the last
+  // frame's partitioning. Only used if use_lastframe_partitioning is set.
+  int adjust_partitioning_from_last_frame;
+
+  // How frequently we re do the partitioning from scratch. Only used if
+  // use_lastframe_partitioning is set.
+  int last_partitioning_redo_frequency;
+
+  // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
+  // it always, to allow it for only Last frame and Intra, disable it for all
+  // inter modes or to enable it always.
+  int disable_split_mask;
+
+  // TODO(jingning): combine the related motion search speed features
+  // This allows us to use motion search at other sizes as a starting
+  // point for this motion search and limits the search range around it.
+  int adaptive_motion_search;
+
+  // Flag for allowing some use of exhaustive searches;
+  int allow_exhaustive_searches;
+
+  // Threshold for allowing exhaistive motion search.
+  int exhaustive_searches_thresh;
+
+  // Maximum number of exhaustive searches for a frame.
+  int max_exaustive_pct;
+
+  // Pattern to be used for any exhaustive mesh searches.
+  MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
+
+  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
+  // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
+  int adaptive_pred_interp_filter;
+
+  // Adaptive prediction mode search
+  int adaptive_mode_search;
+
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
+  int cb_partition_search;
+
+  int alt_ref_search_fp;
+
+  // Use finer quantizer in every other few frames that run variable block
+  // partition type search.
+  int force_frame_boost;
+
+  // Maximally allowed base quantization index fluctuation.
+  int max_delta_qindex;
+
+  // Implements various heuristics to skip searching modes
+  // The heuristics selected are based on  flags
+  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
+  unsigned int mode_search_skip_flags;
+
+  // A source variance threshold below which filter search is disabled
+  // Choose a very large value (UINT_MAX) to use 8-tap always
+  unsigned int disable_filter_search_var_thresh;
+
+#if CONFIG_EXT_INTER
+  // A source variance threshold below which wedge search is disabled
+  unsigned int disable_wedge_search_var_thresh;
+
+  // Whether fast wedge sign estimate is used
+  int fast_wedge_sign_estimate;
+#endif  // CONFIG_EXT_INTER
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // transform size separately.
+  int intra_y_mode_mask[TX_SIZES];
+  int intra_uv_mode_mask[TX_SIZES];
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // prediction block size separately.
+  int intra_y_mode_bsize_mask[BLOCK_SIZES];
+
+  // This variable enables an early break out of mode testing if the model for
+  // rd built from the prediction signal indicates a value that's much
+  // higher than the best rd we've seen so far.
+  int use_rd_breakout;
+
+  // This feature controls how the loop filter level is determined.
+  LPF_PICK_METHOD lpf_pick;
+
+  // This feature limits the number of coefficients updates we actually do
+  // by only looking at counts from 1/2 the bands.
+  FAST_COEFF_UPDATE use_fast_coef_updates;
+
+  // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
+  // modes are used in order from LSB to MSB for each BLOCK_SIZE.
+  int inter_mode_mask[BLOCK_SIZES];
+
+  // This feature controls whether we do the expensive context update and
+  // calculation in the rd coefficient costing loop.
+  int use_fast_coef_costing;
+
+  // This feature controls the tolerence vs target used in deciding whether to
+  // recode a frame. It has no meaning if recode is disabled.
+  int recode_tolerance;
+
+  // This variable controls the maximum block size where intra blocks can be
+  // used in inter frames.
+  // TODO(aconverse): Fold this into one of the other many mode skips
+  BLOCK_SIZE max_intra_bsize;
+
+  // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
+  // FIXED_PARTITION search type should be used.
+  int search_type_check_frequency;
+
+  // When partition is pre-set, the inter prediction result from pick_inter_mode
+  // can be reused in final block encoding process. It is enabled only for real-
+  // time mode speed 6.
+  int reuse_inter_pred_sby;
+
+  // default interp filter choice
+  InterpFilter default_interp_filter;
+
+  // Early termination in transform size search, which only applies while
+  // tx_size_search_method is USE_FULL_RD.
+  int tx_size_search_breakout;
+
+  // adaptive interp_filter search to allow skip of certain filter types.
+  int adaptive_interp_filter_search;
+
+  // mask for skip evaluation of certain interp_filter type.
+  INTERP_FILTER_MASK interp_filter_search_mask;
+
+  // Partition search early breakout thresholds.
+  int64_t partition_search_breakout_dist_thr;
+  int partition_search_breakout_rate_thr;
+
+  // Allow skipping partition search for still image frame
+  int allow_partition_search_skip;
+
+  // Fast approximation of av1_model_rd_from_var_lapndz
+  int simple_model_rd_from_var;
+
+  // Do sub-pixel search in up-sampled reference frames
+  int use_upsampled_references;
+
+  // Whether to compute distortion in the image domain (slower but
+  // more accurate), or in the transform domain (faster but less acurate).
+  int use_transform_domain_distortion;
+} SPEED_FEATURES;
+
+struct AV1_COMP;
+
+void av1_set_speed_features_framesize_independent(struct AV1_COMP *cpi);
+void av1_set_speed_features_framesize_dependent(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SPEED_FEATURES_H_

third_party/aom/av1/encoder/subexp.c

@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "aom_dsp/bitwriter.h"
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/subexp.h"
+
+static const uint8_t update_bits[255] = {
+  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  6,  6,  6,
+  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  8,  8,  8,  8,  8,  8,
+  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  8,  8,  8,  8,  8,  8,  8,  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 0,
+};
+#define MIN_DELP_BITS 5
+
+static int recenter_nonneg(int v, int m) {
+  if (v > (m << 1))
+    return v;
+  else if (v >= m)
+    return ((v - m) << 1);
+  else
+    return ((m - v) << 1) - 1;
+}
+
+static int remap_prob(int v, int m) {
+  int i;
+  static const uint8_t map_table[MAX_PROB - 1] = {
+    // generated by:
+    //   map_table[j] = split_index(j, MAX_PROB - 1, MODULUS_PARAM);
+    20,  21,  22,  23,  24,  25,  0,   26,  27,  28,  29,  30,  31,  32,  33,
+    34,  35,  36,  37,  1,   38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
+    48,  49,  2,   50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,
+    3,   62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  4,   74,
+    75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  5,   86,  87,  88,
+    89,  90,  91,  92,  93,  94,  95,  96,  97,  6,   98,  99,  100, 101, 102,
+    103, 104, 105, 106, 107, 108, 109, 7,   110, 111, 112, 113, 114, 115, 116,
+    117, 118, 119, 120, 121, 8,   122, 123, 124, 125, 126, 127, 128, 129, 130,
+    131, 132, 133, 9,   134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144,
+    145, 10,  146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 11,
+    158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 12,  170, 171,
+    172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 13,  182, 183, 184, 185,
+    186, 187, 188, 189, 190, 191, 192, 193, 14,  194, 195, 196, 197, 198, 199,
+    200, 201, 202, 203, 204, 205, 15,  206, 207, 208, 209, 210, 211, 212, 213,
+    214, 215, 216, 217, 16,  218, 219, 220, 221, 222, 223, 224, 225, 226, 227,
+    228, 229, 17,  230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241,
+    18,  242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 19,
+  };
+  v--;
+  m--;
+  if ((m << 1) <= MAX_PROB)
+    i = recenter_nonneg(v, m) - 1;
+  else
+    i = recenter_nonneg(MAX_PROB - 1 - v, MAX_PROB - 1 - m) - 1;
+
+  i = map_table[i];
+  return i;
+}
+
+static int prob_diff_update_cost(aom_prob newp, aom_prob oldp) {
+  int delp = remap_prob(newp, oldp);
+  return update_bits[delp] << AV1_PROB_COST_SHIFT;
+}
+
+static void encode_uniform(aom_writer *w, int v) {
+  const int l = 8;
+  const int m = (1 << l) - 190;
+  if (v < m) {
+    aom_write_literal(w, v, l - 1);
+  } else {
+    aom_write_literal(w, m + ((v - m) >> 1), l - 1);
+    aom_write_literal(w, (v - m) & 1, 1);
+  }
+}
+
+static INLINE int write_bit_gte(aom_writer *w, int word, int test) {
+  aom_write_literal(w, word >= test, 1);
+  return word >= test;
+}
+
+static void encode_term_subexp(aom_writer *w, int word) {
+  if (!write_bit_gte(w, word, 16)) {
+    aom_write_literal(w, word, 4);
+  } else if (!write_bit_gte(w, word, 32)) {
+    aom_write_literal(w, word - 16, 4);
+  } else if (!write_bit_gte(w, word, 64)) {
+    aom_write_literal(w, word - 32, 5);
+  } else {
+    encode_uniform(w, word - 64);
+  }
+}
+
+void av1_write_prob_diff_update(aom_writer *w, aom_prob newp, aom_prob oldp) {
+  const int delp = remap_prob(newp, oldp);
+  encode_term_subexp(w, delp);
+}
+
+int av1_prob_diff_update_savings_search(const unsigned int *ct, aom_prob oldp,
+                                        aom_prob *bestp, aom_prob upd,
+                                        int probwt) {
+  const uint32_t old_b = cost_branch256(ct, oldp);
+  int bestsavings = 0;
+  aom_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+
+  if (old_b > (uint32_t)upd_cost + (MIN_DELP_BITS << AV1_PROB_COST_SHIFT)) {
+    for (newp = *bestp; newp != oldp; newp += step) {
+      const int new_b = cost_branch256(ct, newp);
+      const int update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      const int savings = (int)((int64_t)old_b - new_b - update_b * probwt);
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int av1_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const aom_prob oldp,
+                                              aom_prob *bestp, aom_prob upd,
+                                              int stepsize, int probwt) {
+  int i, old_b, new_b, update_b, savings, bestsavings;
+  int newp;
+  const int step_sign = *bestp > oldp ? -1 : 1;
+  const int step = stepsize * step_sign;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  const aom_prob *newplist, *oldplist;
+  aom_prob bestnewp;
+  oldplist = av1_pareto8_full[oldp - 1];
+  old_b = cost_branch256(ct + 2 * PIVOT_NODE, oldp);
+  for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+    old_b += cost_branch256(ct + 2 * i, oldplist[i - UNCONSTRAINED_NODES]);
+
+  bestsavings = 0;
+  bestnewp = oldp;
+
+  assert(stepsize > 0);
+
+  if (old_b > upd_cost + (MIN_DELP_BITS << AV1_PROB_COST_SHIFT)) {
+    for (newp = *bestp; (newp - oldp) * step_sign < 0; newp += step) {
+      if (newp < 1 || newp > 255) continue;
+      newplist = av1_pareto8_full[newp - 1];
+      new_b = cost_branch256(ct + 2 * PIVOT_NODE, newp);
+      for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i - UNCONSTRAINED_NODES]);
+      update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      savings = old_b - new_b - update_b * probwt;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+static int get_cost(unsigned int ct[][2], aom_prob p, int n) {
+  int i, p0 = p;
+  unsigned int total_ct[2] = { 0, 0 };
+  int cost = 0;
+
+  for (i = 0; i <= n; ++i) {
+    cost += cost_branch256(ct[i], p);
+    total_ct[0] += ct[i][0];
+    total_ct[1] += ct[i][1];
+    if (i < n)
+      p = av1_merge_probs(p0, total_ct, COEF_COUNT_SAT, COEF_MAX_UPDATE_FACTOR);
+  }
+  return cost;
+}
+
+int av1_prob_update_search_subframe(unsigned int ct[][2], aom_prob oldp,
+                                    aom_prob *bestp, aom_prob upd, int n) {
+  const int old_b = get_cost(ct, oldp, n);
+  int bestsavings = 0;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  aom_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+
+  for (newp = *bestp; newp != oldp; newp += step) {
+    const int new_b = get_cost(ct, newp, n);
+    const int update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+    const int savings = old_b - new_b - update_b;
+    if (savings > bestsavings) {
+      bestsavings = savings;
+      bestnewp = newp;
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int av1_prob_update_search_model_subframe(
+    unsigned int ct[ENTROPY_NODES][COEF_PROBS_BUFS][2], const aom_prob *oldp,
+    aom_prob *bestp, aom_prob upd, int stepsize, int n) {
+  int i, old_b, new_b, update_b, savings, bestsavings;
+  int newp;
+  const int step_sign = *bestp > oldp[PIVOT_NODE] ? -1 : 1;
+  const int step = stepsize * step_sign;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  aom_prob bestnewp, newplist[ENTROPY_NODES], oldplist[ENTROPY_NODES];
+  av1_model_to_full_probs(oldp, oldplist);
+  memcpy(newplist, oldp, sizeof(aom_prob) * UNCONSTRAINED_NODES);
+  for (i = UNCONSTRAINED_NODES, old_b = 0; i < ENTROPY_NODES; ++i)
+    old_b += get_cost(ct[i], oldplist[i], n);
+  old_b += get_cost(ct[PIVOT_NODE], oldplist[PIVOT_NODE], n);
+
+  bestsavings = 0;
+  bestnewp = oldp[PIVOT_NODE];
+
+  assert(stepsize > 0);
+
+  for (newp = *bestp; (newp - oldp[PIVOT_NODE]) * step_sign < 0; newp += step) {
+    if (newp < 1 || newp > 255) continue;
+    newplist[PIVOT_NODE] = newp;
+    av1_model_to_full_probs(newplist, newplist);
+    for (i = UNCONSTRAINED_NODES, new_b = 0; i < ENTROPY_NODES; ++i)
+      new_b += get_cost(ct[i], newplist[i], n);
+    new_b += get_cost(ct[PIVOT_NODE], newplist[PIVOT_NODE], n);
+    update_b = prob_diff_update_cost(newp, oldp[PIVOT_NODE]) + upd_cost;
+    savings = old_b - new_b - update_b;
+    if (savings > bestsavings) {
+      bestsavings = savings;
+      bestnewp = newp;
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+void av1_cond_prob_diff_update(aom_writer *w, aom_prob *oldp,
+                               const unsigned int ct[2], int probwt) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  aom_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings =
+      av1_prob_diff_update_savings_search(ct, *oldp, &newp, upd, probwt);
+  assert(newp >= 1);
+  if (savings > 0) {
+    aom_write(w, 1, upd);
+    av1_write_prob_diff_update(w, newp, *oldp);
+    *oldp = newp;
+  } else {
+    aom_write(w, 0, upd);
+  }
+}
+
+int av1_cond_prob_diff_update_savings(aom_prob *oldp, const unsigned int ct[2],
+                                      int probwt) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  aom_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings =
+      av1_prob_diff_update_savings_search(ct, *oldp, &newp, upd, probwt);
+  return savings;
+}

third_party/aom/av1/encoder/subexp.h

@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SUBEXP_H_
+#define AV1_ENCODER_SUBEXP_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "aom_dsp/bitwriter.h"
+#include "aom_dsp/prob.h"
+
+void av1_write_prob_diff_update(aom_writer *w, aom_prob newp, aom_prob oldpm);
+
+void av1_cond_prob_diff_update(aom_writer *w, aom_prob *oldp,
+                               const unsigned int ct[2], int probwt);
+
+int av1_prob_diff_update_savings_search(const unsigned int *ct, aom_prob oldp,
+                                        aom_prob *bestp, aom_prob upd,
+                                        int probwt);
+
+int av1_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const aom_prob oldp,
+                                              aom_prob *bestp, aom_prob upd,
+                                              int stepsize, int probwt);
+
+int av1_cond_prob_diff_update_savings(aom_prob *oldp, const unsigned int ct[2],
+                                      int probwt);
+#if CONFIG_SUBFRAME_PROB_UPDATE
+int av1_prob_update_search_subframe(unsigned int ct[][2], aom_prob oldp,
+                                    aom_prob *bestp, aom_prob upd, int n);
+int av1_prob_update_search_model_subframe(
+    unsigned int ct[ENTROPY_NODES][COEF_PROBS_BUFS][2], const aom_prob *oldp,
+    aom_prob *bestp, aom_prob upd, int stepsize, int n);
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SUBEXP_H_

third_party/aom/av1/encoder/temporal_filter.c

@@ -0,0 +1,719 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#include "av1/common/alloccommon.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/odintrin.h"
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/temporal_filter.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_scale/aom_scale.h"
+
+static void temporal_filter_predictors_mb_c(
+    MACROBLOCKD *xd, uint8_t *y_mb_ptr, uint8_t *u_mb_ptr, uint8_t *v_mb_ptr,
+    int stride, int uv_block_width, int uv_block_height, int mv_row, int mv_col,
+    uint8_t *pred, struct scale_factors *scale, int x, int y) {
+  const int which_mv = 0;
+  const MV mv = { mv_row, mv_col };
+  enum mv_precision mv_precision_uv;
+  int uv_stride;
+  // TODO(angiebird): change plane setting accordingly
+  ConvolveParams conv_params = get_conv_params(which_mv, 0);
+
+#if USE_TEMPORALFILTER_12TAP
+#if CONFIG_DUAL_FILTER
+  const InterpFilter interp_filter[4] = { TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP };
+#else
+  const InterpFilter interp_filter = TEMPORALFILTER_12TAP;
+#endif
+  (void)xd;
+#else
+  const InterpFilter interp_filter = xd->mi[0]->mbmi.interp_filter;
+#endif  // USE_TEMPORALFILTER_12TAP
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+  WarpTypesAllowed warp_types;
+  memset(&warp_types, 0, sizeof(WarpTypesAllowed));
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+  if (uv_block_width == 8) {
+    uv_stride = (stride + 1) >> 1;
+    mv_precision_uv = MV_PRECISION_Q4;
+  } else {
+    uv_stride = stride;
+    mv_precision_uv = MV_PRECISION_Q3;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    av1_highbd_build_inter_predictor(y_mb_ptr, stride, &pred[0], 16, &mv, scale,
+                                     16, 16, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     0, MV_PRECISION_Q3, x, y, xd);
+
+    av1_highbd_build_inter_predictor(u_mb_ptr, uv_stride, &pred[256],
+                                     uv_block_width, &mv, scale, uv_block_width,
+                                     uv_block_height, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     1, mv_precision_uv, x, y, xd);
+
+    av1_highbd_build_inter_predictor(v_mb_ptr, uv_stride, &pred[512],
+                                     uv_block_width, &mv, scale, uv_block_width,
+                                     uv_block_height, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     2, mv_precision_uv, x, y, xd);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  av1_build_inter_predictor(y_mb_ptr, stride, &pred[0], 16, &mv, scale, 16, 16,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 0, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            MV_PRECISION_Q3, x, y, xd);
+
+  av1_build_inter_predictor(u_mb_ptr, uv_stride, &pred[256], uv_block_width,
+                            &mv, scale, uv_block_width, uv_block_height,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 1, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            mv_precision_uv, x, y, xd);
+
+  av1_build_inter_predictor(v_mb_ptr, uv_stride, &pred[512], uv_block_width,
+                            &mv, scale, uv_block_width, uv_block_height,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 2, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            mv_precision_uv, x, y, xd);
+}
+
+void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride,
+                                 uint8_t *frame2, unsigned int block_width,
+                                 unsigned int block_height, int strength,
+                                 int filter_weight, unsigned int *accumulator,
+                                 uint16_t *count) {
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+
+      // non-local mean approach
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height && col >= 0 &&
+              col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                       frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx) modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16) modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_temporal_filter_apply_c(
+    uint8_t *frame1_8, unsigned int stride, uint8_t *frame2_8,
+    unsigned int block_width, unsigned int block_height, int strength,
+    int filter_weight, unsigned int *accumulator, uint16_t *count) {
+  uint16_t *frame1 = CONVERT_TO_SHORTPTR(frame1_8);
+  uint16_t *frame2 = CONVERT_TO_SHORTPTR(frame2_8);
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+
+      // non-local mean approach
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height && col >= 0 &&
+              col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                       frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx) modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16) modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static int temporal_filter_find_matching_mb_c(AV1_COMP *cpi,
+                                              uint8_t *arf_frame_buf,
+                                              uint8_t *frame_ptr_buf,
+                                              int stride) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  int bestsme = INT_MAX;
+  int distortion;
+  unsigned int sse;
+  int cost_list[5];
+  MvLimits tmp_mv_limits = x->mv_limits;
+
+  MV best_ref_mv1 = { 0, 0 };
+  MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
+
+  // Save input state
+  struct buf_2d src = x->plane[0].src;
+  struct buf_2d pre = xd->plane[0].pre[0];
+
+  best_ref_mv1_full.col = best_ref_mv1.col >> 3;
+  best_ref_mv1_full.row = best_ref_mv1.row >> 3;
+
+  // Setup frame pointers
+  x->plane[0].src.buf = arf_frame_buf;
+  x->plane[0].src.stride = stride;
+  xd->plane[0].pre[0].buf = frame_ptr_buf;
+  xd->plane[0].pre[0].stride = stride;
+
+  step_param = mv_sf->reduce_first_step_size;
+  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  av1_set_mv_search_range(&x->mv_limits, &best_ref_mv1);
+
+#if CONFIG_REF_MV
+  x->mvcost = x->mv_cost_stack[0];
+  x->nmvjointcost = x->nmv_vec_cost[0];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
+#endif
+
+  // Ignore mv costing by sending NULL pointer instead of cost arrays
+  av1_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
+                 cond_cost_list(cpi, cost_list), &cpi->fn_ptr[BLOCK_16X16], 0,
+                 &best_ref_mv1);
+
+  x->mv_limits = tmp_mv_limits;
+
+  // Ignore mv costing by sending NULL pointer instead of cost array
+  bestsme = cpi->find_fractional_mv_step(
+      x, &best_ref_mv1, cpi->common.allow_high_precision_mv, x->errorperbit,
+      &cpi->fn_ptr[BLOCK_16X16], 0, mv_sf->subpel_iters_per_step,
+      cond_cost_list(cpi, cost_list), NULL, NULL, &distortion, &sse, NULL, 0, 0,
+      0);
+
+  x->e_mbd.mi[0]->bmi[0].as_mv[0] = x->best_mv;
+
+  // Restore input state
+  x->plane[0].src = src;
+  xd->plane[0].pre[0] = pre;
+
+  return bestsme;
+}
+
+static void temporal_filter_iterate_c(AV1_COMP *cpi,
+                                      YV12_BUFFER_CONFIG **frames,
+                                      int frame_count, int alt_ref_index,
+                                      int strength,
+                                      struct scale_factors *scale) {
+  int byte;
+  int frame;
+  int mb_col, mb_row;
+  unsigned int filter_weight;
+  int mb_cols = (frames[alt_ref_index]->y_crop_width + 15) >> 4;
+  int mb_rows = (frames[alt_ref_index]->y_crop_height + 15) >> 4;
+  int mb_y_offset = 0;
+  int mb_uv_offset = 0;
+  DECLARE_ALIGNED(16, unsigned int, accumulator[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint16_t, count[16 * 16 * 3]);
+  MACROBLOCKD *mbd = &cpi->td.mb.e_mbd;
+  YV12_BUFFER_CONFIG *f = frames[alt_ref_index];
+  uint8_t *dst1, *dst2;
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, predictor16[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint8_t, predictor8[16 * 16 * 3]);
+  uint8_t *predictor;
+#else
+  DECLARE_ALIGNED(16, uint8_t, predictor[16 * 16 * 3]);
+#endif
+  const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width = 16 >> mbd->plane[1].subsampling_x;
+
+  // Save input state
+  uint8_t *input_buffer[MAX_MB_PLANE];
+  int i;
+#if CONFIG_HIGHBITDEPTH
+  if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    predictor = CONVERT_TO_BYTEPTR(predictor16);
+  } else {
+    predictor = predictor8;
+  }
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; i++) input_buffer[i] = mbd->plane[i].pre[0].buf;
+
+  for (mb_row = 0; mb_row < mb_rows; mb_row++) {
+    // Source frames are extended to 16 pixels. This is different than
+    //  L/A/G reference frames that have a border of 32 (AV1ENCBORDERINPIXELS)
+    // A 6/8 tap filter is used for motion search.  This requires 2 pixels
+    //  before and 3 pixels after.  So the largest Y mv on a border would
+    //  then be 16 - AOM_INTERP_EXTEND. The UV blocks are half the size of the
+    //  Y and therefore only extended by 8.  The largest mv that a UV block
+    //  can support is 8 - AOM_INTERP_EXTEND.  A UV mv is half of a Y mv.
+    //  (16 - AOM_INTERP_EXTEND) >> 1 which is greater than
+    //  8 - AOM_INTERP_EXTEND.
+    // To keep the mv in play for both Y and UV planes the max that it
+    //  can be on a border is therefore 16 - (2*AOM_INTERP_EXTEND+1).
+    cpi->td.mb.mv_limits.row_min =
+        -((mb_row * 16) + (17 - 2 * AOM_INTERP_EXTEND));
+    cpi->td.mb.mv_limits.row_max =
+        ((mb_rows - 1 - mb_row) * 16) + (17 - 2 * AOM_INTERP_EXTEND);
+
+    for (mb_col = 0; mb_col < mb_cols; mb_col++) {
+      int j, k;
+      int stride;
+
+      memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0]));
+      memset(count, 0, 16 * 16 * 3 * sizeof(count[0]));
+
+      cpi->td.mb.mv_limits.col_min =
+          -((mb_col * 16) + (17 - 2 * AOM_INTERP_EXTEND));
+      cpi->td.mb.mv_limits.col_max =
+          ((mb_cols - 1 - mb_col) * 16) + (17 - 2 * AOM_INTERP_EXTEND);
+
+      for (frame = 0; frame < frame_count; frame++) {
+        const int thresh_low = 10000;
+        const int thresh_high = 20000;
+
+        if (frames[frame] == NULL) continue;
+
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0;
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0;
+
+        if (frame == alt_ref_index) {
+          filter_weight = 2;
+        } else {
+          // Find best match in this frame by MC
+          int err = temporal_filter_find_matching_mb_c(
+              cpi, frames[alt_ref_index]->y_buffer + mb_y_offset,
+              frames[frame]->y_buffer + mb_y_offset, frames[frame]->y_stride);
+
+          // Assign higher weight to matching MB if it's error
+          // score is lower. If not applying MC default behavior
+          // is to weight all MBs equal.
+          filter_weight = err < thresh_low ? 2 : err < thresh_high ? 1 : 0;
+        }
+
+        if (filter_weight != 0) {
+          // Construct the predictors
+          temporal_filter_predictors_mb_c(
+              mbd, frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->u_buffer + mb_uv_offset,
+              frames[frame]->v_buffer + mb_uv_offset, frames[frame]->y_stride,
+              mb_uv_width, mb_uv_height, mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.col, predictor, scale,
+              mb_col * 16, mb_row * 16);
+
+#if CONFIG_HIGHBITDEPTH
+          if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            int adj_strength = strength + 2 * (mbd->bd - 8);
+            // Apply the filter (YUV)
+            av1_highbd_temporal_filter_apply(
+                f->y_buffer + mb_y_offset, f->y_stride, predictor, 16, 16,
+                adj_strength, filter_weight, accumulator, count);
+            av1_highbd_temporal_filter_apply(
+                f->u_buffer + mb_uv_offset, f->uv_stride, predictor + 256,
+                mb_uv_width, mb_uv_height, adj_strength, filter_weight,
+                accumulator + 256, count + 256);
+            av1_highbd_temporal_filter_apply(
+                f->v_buffer + mb_uv_offset, f->uv_stride, predictor + 512,
+                mb_uv_width, mb_uv_height, adj_strength, filter_weight,
+                accumulator + 512, count + 512);
+          } else {
+            // Apply the filter (YUV)
+            av1_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                        predictor, 16, 16, strength,
+                                        filter_weight, accumulator, count);
+            av1_temporal_filter_apply_c(
+                f->u_buffer + mb_uv_offset, f->uv_stride, predictor + 256,
+                mb_uv_width, mb_uv_height, strength, filter_weight,
+                accumulator + 256, count + 256);
+            av1_temporal_filter_apply_c(
+                f->v_buffer + mb_uv_offset, f->uv_stride, predictor + 512,
+                mb_uv_width, mb_uv_height, strength, filter_weight,
+                accumulator + 512, count + 512);
+          }
+#else
+          // Apply the filter (YUV)
+          av1_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                      predictor, 16, 16, strength,
+                                      filter_weight, accumulator, count);
+          av1_temporal_filter_apply_c(f->u_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 256, mb_uv_width,
+                                      mb_uv_height, strength, filter_weight,
+                                      accumulator + 256, count + 256);
+          av1_temporal_filter_apply_c(f->v_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 512, mb_uv_width,
+                                      mb_uv_height, strength, filter_weight,
+                                      accumulator + 512, count + 512);
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+
+#if CONFIG_HIGHBITDEPTH
+      if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        uint16_t *dst1_16;
+        uint16_t *dst2_16;
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            dst1_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        dst2_16 = CONVERT_TO_SHORTPTR(dst2);
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            dst1_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // V
+            dst2_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - mb_uv_width;
+        }
+      } else {
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            dst1[byte] =
+                (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            dst1[byte] =
+                (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // V
+            dst2[byte] =
+                (uint8_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - mb_uv_width;
+        }
+      }
+#else
+      // Normalize filter output to produce AltRef frame
+      dst1 = cpi->alt_ref_buffer.y_buffer;
+      stride = cpi->alt_ref_buffer.y_stride;
+      byte = mb_y_offset;
+      for (i = 0, k = 0; i < 16; i++) {
+        for (j = 0; j < 16; j++, k++) {
+          dst1[byte] =
+              (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - 16;
+      }
+
+      dst1 = cpi->alt_ref_buffer.u_buffer;
+      dst2 = cpi->alt_ref_buffer.v_buffer;
+      stride = cpi->alt_ref_buffer.uv_stride;
+      byte = mb_uv_offset;
+      for (i = 0, k = 256; i < mb_uv_height; i++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
+          int m = k + 256;
+
+          // U
+          dst1[byte] =
+              (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+          // V
+          dst2[byte] =
+              (uint8_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - mb_uv_width;
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+      mb_y_offset += 16;
+      mb_uv_offset += mb_uv_width;
+    }
+    mb_y_offset += 16 * (f->y_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
+  }
+
+  // Restore input state
+  for (i = 0; i < MAX_MB_PLANE; i++) mbd->plane[i].pre[0].buf = input_buffer[i];
+}
+
+// Apply buffer limits and context specific adjustments to arnr filter.
+static void adjust_arnr_filter(AV1_COMP *cpi, int distance, int group_boost,
+                               int *arnr_frames, int *arnr_strength) {
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frames_after_arf =
+      av1_lookahead_depth(cpi->lookahead) - distance - 1;
+  int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+  int frames_bwd;
+  int q, frames, strength;
+
+  // Define the forward and backwards filter limits for this arnr group.
+  if (frames_fwd > frames_after_arf) frames_fwd = frames_after_arf;
+  if (frames_fwd > distance) frames_fwd = distance;
+
+  frames_bwd = frames_fwd;
+
+  // For even length filter there is one more frame backward
+  // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+  if (frames_bwd < distance) frames_bwd += (oxcf->arnr_max_frames + 1) & 0x1;
+
+  // Set the baseline active filter size.
+  frames = frames_bwd + 1 + frames_fwd;
+
+  // Adjust the strength based on active max q.
+  if (cpi->common.current_video_frame > 1)
+    q = ((int)av1_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],
+                                      cpi->common.bit_depth));
+  else
+    q = ((int)av1_convert_qindex_to_q(cpi->rc.avg_frame_qindex[KEY_FRAME],
+                                      cpi->common.bit_depth));
+  if (q > 16) {
+    strength = oxcf->arnr_strength;
+  } else {
+    strength = oxcf->arnr_strength - ((16 - q) / 2);
+    if (strength < 0) strength = 0;
+  }
+
+  // Adjust number of frames in filter and strength based on gf boost level.
+  if (frames > group_boost / 150) {
+    frames = group_boost / 150;
+    frames += !(frames & 1);
+  }
+
+  if (strength > group_boost / 300) {
+    strength = group_boost / 300;
+  }
+
+  // Adjustments for second level arf in multi arf case.
+  if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) {
+      strength >>= 1;
+    }
+  }
+
+  *arnr_frames = frames;
+  *arnr_strength = strength;
+}
+
+void av1_temporal_filter(AV1_COMP *cpi, int distance) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int frame;
+  int frames_to_blur;
+  int start_frame;
+  int strength;
+  int frames_to_blur_backward;
+  int frames_to_blur_forward;
+  struct scale_factors sf;
+  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS] = { NULL };
+#if CONFIG_EXT_REFS
+  const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+#endif
+
+  // Apply context specific adjustments to the arnr filter parameters.
+  adjust_arnr_filter(cpi, distance, rc->gfu_boost, &frames_to_blur, &strength);
+// TODO(weitinglin): Currently, we enforce the filtering strength on
+//                   extra ARFs' to be zeros. We should investigate in which
+//                   case it is more beneficial to use non-zero strength
+//                   filtering.
+#if CONFIG_EXT_REFS
+  if (gf_group->rf_level[gf_group->index] == GF_ARF_LOW) {
+    strength = 0;
+    frames_to_blur = 1;
+  }
+#endif
+
+#if CONFIG_EXT_REFS
+  if (strength == 0 && frames_to_blur == 1) {
+    cpi->is_arf_filter_off[gf_group->arf_update_idx[gf_group->index]] = 1;
+  } else {
+    cpi->is_arf_filter_off[gf_group->arf_update_idx[gf_group->index]] = 0;
+  }
+#endif
+
+  frames_to_blur_backward = (frames_to_blur / 2);
+  frames_to_blur_forward = ((frames_to_blur - 1) / 2);
+  start_frame = distance + frames_to_blur_forward;
+
+  // Setup frame pointers, NULL indicates frame not included in filter.
+  for (frame = 0; frame < frames_to_blur; ++frame) {
+    const int which_buffer = start_frame - frame;
+    struct lookahead_entry *buf =
+        av1_lookahead_peek(cpi->lookahead, which_buffer);
+    frames[frames_to_blur - 1 - frame] = &buf->img;
+  }
+
+  if (frames_to_blur > 0) {
+// Setup scaling factors. Scaling on each of the arnr frames is not
+// supported.
+// ARF is produced at the native frame size and resized when coded.
+#if CONFIG_HIGHBITDEPTH
+    av1_setup_scale_factors_for_frame(
+        &sf, frames[0]->y_crop_width, frames[0]->y_crop_height,
+        frames[0]->y_crop_width, frames[0]->y_crop_height,
+        cpi->common.use_highbitdepth);
+#else
+    av1_setup_scale_factors_for_frame(
+        &sf, frames[0]->y_crop_width, frames[0]->y_crop_height,
+        frames[0]->y_crop_width, frames[0]->y_crop_height);
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+
+  temporal_filter_iterate_c(cpi, frames, frames_to_blur,
+                            frames_to_blur_backward, strength, &sf);
+}

third_party/aom/av1/encoder/temporal_filter.h

@@ -0,0 +1,25 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TEMPORAL_FILTER_H_
+#define AV1_ENCODER_TEMPORAL_FILTER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_temporal_filter(AV1_COMP *cpi, int distance);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TEMPORAL_FILTER_H_

third_party/aom/av1/encoder/tokenize.c

@@ -0,0 +1,887 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/entropy.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/scan.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/tokenize.h"
+
+static const TOKENVALUE dct_cat_lt_10_value_tokens[] = {
+  { 9, 63 }, { 9, 61 }, { 9, 59 }, { 9, 57 }, { 9, 55 }, { 9, 53 }, { 9, 51 },
+  { 9, 49 }, { 9, 47 }, { 9, 45 }, { 9, 43 }, { 9, 41 }, { 9, 39 }, { 9, 37 },
+  { 9, 35 }, { 9, 33 }, { 9, 31 }, { 9, 29 }, { 9, 27 }, { 9, 25 }, { 9, 23 },
+  { 9, 21 }, { 9, 19 }, { 9, 17 }, { 9, 15 }, { 9, 13 }, { 9, 11 }, { 9, 9 },
+  { 9, 7 },  { 9, 5 },  { 9, 3 },  { 9, 1 },  { 8, 31 }, { 8, 29 }, { 8, 27 },
+  { 8, 25 }, { 8, 23 }, { 8, 21 }, { 8, 19 }, { 8, 17 }, { 8, 15 }, { 8, 13 },
+  { 8, 11 }, { 8, 9 },  { 8, 7 },  { 8, 5 },  { 8, 3 },  { 8, 1 },  { 7, 15 },
+  { 7, 13 }, { 7, 11 }, { 7, 9 },  { 7, 7 },  { 7, 5 },  { 7, 3 },  { 7, 1 },
+  { 6, 7 },  { 6, 5 },  { 6, 3 },  { 6, 1 },  { 5, 3 },  { 5, 1 },  { 4, 1 },
+  { 3, 1 },  { 2, 1 },  { 1, 1 },  { 0, 0 },  { 1, 0 },  { 2, 0 },  { 3, 0 },
+  { 4, 0 },  { 5, 0 },  { 5, 2 },  { 6, 0 },  { 6, 2 },  { 6, 4 },  { 6, 6 },
+  { 7, 0 },  { 7, 2 },  { 7, 4 },  { 7, 6 },  { 7, 8 },  { 7, 10 }, { 7, 12 },
+  { 7, 14 }, { 8, 0 },  { 8, 2 },  { 8, 4 },  { 8, 6 },  { 8, 8 },  { 8, 10 },
+  { 8, 12 }, { 8, 14 }, { 8, 16 }, { 8, 18 }, { 8, 20 }, { 8, 22 }, { 8, 24 },
+  { 8, 26 }, { 8, 28 }, { 8, 30 }, { 9, 0 },  { 9, 2 },  { 9, 4 },  { 9, 6 },
+  { 9, 8 },  { 9, 10 }, { 9, 12 }, { 9, 14 }, { 9, 16 }, { 9, 18 }, { 9, 20 },
+  { 9, 22 }, { 9, 24 }, { 9, 26 }, { 9, 28 }, { 9, 30 }, { 9, 32 }, { 9, 34 },
+  { 9, 36 }, { 9, 38 }, { 9, 40 }, { 9, 42 }, { 9, 44 }, { 9, 46 }, { 9, 48 },
+  { 9, 50 }, { 9, 52 }, { 9, 54 }, { 9, 56 }, { 9, 58 }, { 9, 60 }, { 9, 62 }
+};
+const TOKENVALUE *av1_dct_cat_lt_10_value_tokens =
+    dct_cat_lt_10_value_tokens +
+    (sizeof(dct_cat_lt_10_value_tokens) / sizeof(*dct_cat_lt_10_value_tokens)) /
+        2;
+// The corresponding costs of the extrabits for the tokens in the above table
+// are stored in the table below. The values are obtained from looking up the
+// entry for the specified extrabits in the table corresponding to the token
+// (as defined in cost element av1_extra_bits)
+// e.g. {9, 63} maps to cat5_cost[63 >> 1], {1, 1} maps to sign_cost[1 >> 1]
+static const int dct_cat_lt_10_value_cost[] = {
+  3773, 3750, 3704, 3681, 3623, 3600, 3554, 3531, 3432, 3409, 3363, 3340, 3282,
+  3259, 3213, 3190, 3136, 3113, 3067, 3044, 2986, 2963, 2917, 2894, 2795, 2772,
+  2726, 2703, 2645, 2622, 2576, 2553, 3197, 3116, 3058, 2977, 2881, 2800, 2742,
+  2661, 2615, 2534, 2476, 2395, 2299, 2218, 2160, 2079, 2566, 2427, 2334, 2195,
+  2023, 1884, 1791, 1652, 1893, 1696, 1453, 1256, 1229, 864,  512,  512,  512,
+  512,  0,    512,  512,  512,  512,  864,  1229, 1256, 1453, 1696, 1893, 1652,
+  1791, 1884, 2023, 2195, 2334, 2427, 2566, 2079, 2160, 2218, 2299, 2395, 2476,
+  2534, 2615, 2661, 2742, 2800, 2881, 2977, 3058, 3116, 3197, 2553, 2576, 2622,
+  2645, 2703, 2726, 2772, 2795, 2894, 2917, 2963, 2986, 3044, 3067, 3113, 3136,
+  3190, 3213, 3259, 3282, 3340, 3363, 3409, 3432, 3531, 3554, 3600, 3623, 3681,
+  3704, 3750, 3773,
+};
+const int *av1_dct_cat_lt_10_value_cost =
+    dct_cat_lt_10_value_cost +
+    (sizeof(dct_cat_lt_10_value_cost) / sizeof(*dct_cat_lt_10_value_cost)) / 2;
+
+// Array indices are identical to previously-existing CONTEXT_NODE indices
+/* clang-format off */
+const aom_tree_index av1_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  -EOB_TOKEN, 2,                       // 0  = EOB
+  -ZERO_TOKEN, 4,                      // 1  = ZERO
+  -ONE_TOKEN, 6,                       // 2  = ONE
+  8, 12,                               // 3  = LOW_VAL
+  -TWO_TOKEN, 10,                      // 4  = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 5  = THREE
+  14, 16,                              // 6  = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 7  = CAT_ONE
+  18, 20,                              // 8  = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 9  = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 10 = CAT_FIVE
+};
+/* clang-format on */
+
+static const int16_t zero_cost[] = { 0 };
+static const int16_t sign_cost[1] = { 512 };
+static const int16_t cat1_cost[1 << 1] = { 864, 1229 };
+static const int16_t cat2_cost[1 << 2] = { 1256, 1453, 1696, 1893 };
+static const int16_t cat3_cost[1 << 3] = { 1652, 1791, 1884, 2023,
+                                           2195, 2334, 2427, 2566 };
+static const int16_t cat4_cost[1 << 4] = { 2079, 2160, 2218, 2299, 2395, 2476,
+                                           2534, 2615, 2661, 2742, 2800, 2881,
+                                           2977, 3058, 3116, 3197 };
+static const int16_t cat5_cost[1 << 5] = {
+  2553, 2576, 2622, 2645, 2703, 2726, 2772, 2795, 2894, 2917, 2963,
+  2986, 3044, 3067, 3113, 3136, 3190, 3213, 3259, 3282, 3340, 3363,
+  3409, 3432, 3531, 3554, 3600, 3623, 3681, 3704, 3750, 3773
+};
+const int16_t av1_cat6_low_cost[256] = {
+  3378, 3390, 3401, 3413, 3435, 3447, 3458, 3470, 3517, 3529, 3540, 3552, 3574,
+  3586, 3597, 3609, 3671, 3683, 3694, 3706, 3728, 3740, 3751, 3763, 3810, 3822,
+  3833, 3845, 3867, 3879, 3890, 3902, 3973, 3985, 3996, 4008, 4030, 4042, 4053,
+  4065, 4112, 4124, 4135, 4147, 4169, 4181, 4192, 4204, 4266, 4278, 4289, 4301,
+  4323, 4335, 4346, 4358, 4405, 4417, 4428, 4440, 4462, 4474, 4485, 4497, 4253,
+  4265, 4276, 4288, 4310, 4322, 4333, 4345, 4392, 4404, 4415, 4427, 4449, 4461,
+  4472, 4484, 4546, 4558, 4569, 4581, 4603, 4615, 4626, 4638, 4685, 4697, 4708,
+  4720, 4742, 4754, 4765, 4777, 4848, 4860, 4871, 4883, 4905, 4917, 4928, 4940,
+  4987, 4999, 5010, 5022, 5044, 5056, 5067, 5079, 5141, 5153, 5164, 5176, 5198,
+  5210, 5221, 5233, 5280, 5292, 5303, 5315, 5337, 5349, 5360, 5372, 4988, 5000,
+  5011, 5023, 5045, 5057, 5068, 5080, 5127, 5139, 5150, 5162, 5184, 5196, 5207,
+  5219, 5281, 5293, 5304, 5316, 5338, 5350, 5361, 5373, 5420, 5432, 5443, 5455,
+  5477, 5489, 5500, 5512, 5583, 5595, 5606, 5618, 5640, 5652, 5663, 5675, 5722,
+  5734, 5745, 5757, 5779, 5791, 5802, 5814, 5876, 5888, 5899, 5911, 5933, 5945,
+  5956, 5968, 6015, 6027, 6038, 6050, 6072, 6084, 6095, 6107, 5863, 5875, 5886,
+  5898, 5920, 5932, 5943, 5955, 6002, 6014, 6025, 6037, 6059, 6071, 6082, 6094,
+  6156, 6168, 6179, 6191, 6213, 6225, 6236, 6248, 6295, 6307, 6318, 6330, 6352,
+  6364, 6375, 6387, 6458, 6470, 6481, 6493, 6515, 6527, 6538, 6550, 6597, 6609,
+  6620, 6632, 6654, 6666, 6677, 6689, 6751, 6763, 6774, 6786, 6808, 6820, 6831,
+  6843, 6890, 6902, 6913, 6925, 6947, 6959, 6970, 6982
+};
+const int av1_cat6_high_cost[CAT6_HIGH_COST_ENTRIES] = {
+  100,   2263,  2739,  4902,  3160,  5323,  5799,  7962,  3678,  5841,  6317,
+  8480,  6738,  8901,  9377,  11540, 3678,  5841,  6317,  8480,  6738,  8901,
+  9377,  11540, 7256,  9419,  9895,  12058, 10316, 12479, 12955, 15118, 3678,
+  5841,  6317,  8480,  6738,  8901,  9377,  11540, 7256,  9419,  9895,  12058,
+  10316, 12479, 12955, 15118, 7256,  9419,  9895,  12058, 10316, 12479, 12955,
+  15118, 10834, 12997, 13473, 15636, 13894, 16057, 16533, 18696,
+#if CONFIG_HIGHBITDEPTH
+  4193,  6356,  6832,  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410,
+  12573, 10831, 12994, 13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994,
+  13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,
+  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151,
+  14409, 16572, 17048, 19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048,
+  19211, 14927, 17090, 17566, 19729, 17987, 20150, 20626, 22789, 4193,  6356,
+  6832,  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831,
+  12994, 13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633,
+  11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410,
+  12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572,
+  17048, 19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927,
+  17090, 17566, 19729, 17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088,
+  11346, 13509, 13985, 16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563,
+  19726, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605,
+  18081, 20244, 18502, 20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924,
+  17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+  15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659,
+  23822, 22080, 24243, 24719, 26882, 4193,  6356,  6832,  8995,  7253,  9416,
+  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 7771,
+  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151,
+  14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573, 10831, 12994, 13470,
+  15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 11349, 13512,
+  13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090, 17566, 19729, 17987,
+  20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503,
+  16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665,
+  21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442,
+  17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244,
+  18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719,
+  26882, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027,
+  14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924,
+  17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665,
+  21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379,
+  14542, 15018, 17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759,
+  19017, 21180, 21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+  23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120,
+  18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595,
+  24758, 25234, 27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397,
+  23113, 25276, 25752, 27915, 26173, 28336, 28812, 30975, 4193,  6356,  6832,
+  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994,
+  13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349,
+  13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573,
+  10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048,
+  19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090,
+  17566, 19729, 17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346,
+  13509, 13985, 16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087,
+  17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442,
+  17605, 18081, 20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822,
+  22080, 24243, 24719, 26882, 8286,  10449, 10925, 13088, 11346, 13509, 13985,
+  16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027,
+  14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502,
+  20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+  15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243,
+  24719, 26882, 12379, 14542, 15018, 17181, 15439, 17602, 18078, 20241, 15957,
+  18120, 18596, 20759, 19017, 21180, 21656, 23819, 15957, 18120, 18596, 20759,
+  19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234,
+  27397, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698,
+  22174, 24337, 22595, 24758, 25234, 27397, 19535, 21698, 22174, 24337, 22595,
+  24758, 25234, 27397, 23113, 25276, 25752, 27915, 26173, 28336, 28812, 30975,
+  8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027, 14503,
+  16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924, 17087,
+  17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 11864,
+  14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244,
+  18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665, 21141,
+  23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379, 14542,
+  15018, 17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017,
+  21180, 21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819,
+  19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596,
+  20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758,
+  25234, 27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113,
+  25276, 25752, 27915, 26173, 28336, 28812, 30975, 12379, 14542, 15018, 17181,
+  15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+  23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698,
+  22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596, 20759, 19017,
+  21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397,
+  19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113, 25276, 25752,
+  27915, 26173, 28336, 28812, 30975, 16472, 18635, 19111, 21274, 19532, 21695,
+  22171, 24334, 20050, 22213, 22689, 24852, 23110, 25273, 25749, 27912, 20050,
+  22213, 22689, 24852, 23110, 25273, 25749, 27912, 23628, 25791, 26267, 28430,
+  26688, 28851, 29327, 31490, 20050, 22213, 22689, 24852, 23110, 25273, 25749,
+  27912, 23628, 25791, 26267, 28430, 26688, 28851, 29327, 31490, 23628, 25791,
+  26267, 28430, 26688, 28851, 29327, 31490, 27206, 29369, 29845, 32008, 30266,
+  32429, 32905, 35068
+#endif
+};
+
+const uint8_t av1_cat6_skipped_bits_discount[8] = {
+  0, 3, 6, 9, 12, 18, 24, 30
+};
+
+#if CONFIG_NEW_MULTISYMBOL
+const av1_extra_bit av1_extra_bits[ENTROPY_TOKENS] = {
+  { 0, 0, 0, zero_cost },                        // ZERO_TOKEN
+  { 0, 0, 1, sign_cost },                        // ONE_TOKEN
+  { 0, 0, 2, sign_cost },                        // TWO_TOKEN
+  { 0, 0, 3, sign_cost },                        // THREE_TOKEN
+  { 0, 0, 4, sign_cost },                        // FOUR_TOKEN
+  { av1_cat1_cdf, 1, CAT1_MIN_VAL, cat1_cost },  // CATEGORY1_TOKEN
+  { av1_cat2_cdf, 2, CAT2_MIN_VAL, cat2_cost },  // CATEGORY2_TOKEN
+  { av1_cat3_cdf, 3, CAT3_MIN_VAL, cat3_cost },  // CATEGORY3_TOKEN
+  { av1_cat4_cdf, 4, CAT4_MIN_VAL, cat4_cost },  // CATEGORY4_TOKEN
+  { av1_cat5_cdf, 5, CAT5_MIN_VAL, cat5_cost },  // CATEGORY5_TOKEN
+  { av1_cat6_cdf, 18, CAT6_MIN_VAL, 0 },         // CATEGORY6_TOKEN
+  { 0, 0, 0, zero_cost }                         // EOB_TOKEN
+};
+#else
+const av1_extra_bit av1_extra_bits[ENTROPY_TOKENS] = {
+  { 0, 0, 0, zero_cost },                         // ZERO_TOKEN
+  { 0, 0, 1, sign_cost },                         // ONE_TOKEN
+  { 0, 0, 2, sign_cost },                         // TWO_TOKEN
+  { 0, 0, 3, sign_cost },                         // THREE_TOKEN
+  { 0, 0, 4, sign_cost },                         // FOUR_TOKEN
+  { av1_cat1_prob, 1, CAT1_MIN_VAL, cat1_cost },  // CATEGORY1_TOKEN
+  { av1_cat2_prob, 2, CAT2_MIN_VAL, cat2_cost },  // CATEGORY2_TOKEN
+  { av1_cat3_prob, 3, CAT3_MIN_VAL, cat3_cost },  // CATEGORY3_TOKEN
+  { av1_cat4_prob, 4, CAT4_MIN_VAL, cat4_cost },  // CATEGORY4_TOKEN
+  { av1_cat5_prob, 5, CAT5_MIN_VAL, cat5_cost },  // CATEGORY5_TOKEN
+  { av1_cat6_prob, 18, CAT6_MIN_VAL, 0 },         // CATEGORY6_TOKEN
+  { 0, 0, 0, zero_cost }                          // EOB_TOKEN
+};
+#endif
+
+#if !CONFIG_EC_MULTISYMBOL
+const struct av1_token av1_coef_encodings[ENTROPY_TOKENS] = {
+  { 2, 2 },  { 6, 3 },   { 28, 5 },  { 58, 6 },  { 59, 6 },  { 60, 6 },
+  { 61, 6 }, { 124, 7 }, { 125, 7 }, { 126, 7 }, { 127, 7 }, { 0, 1 }
+};
+#endif  // !CONFIG_EC_MULTISYMBOL
+
+struct tokenize_b_args {
+  const AV1_COMP *cpi;
+  ThreadData *td;
+  TOKENEXTRA **tp;
+  int this_rate;
+};
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+static void cost_coeffs_b(int plane, int block, int blk_row, int blk_col,
+                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *const cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  const PLANE_TYPE type = pd->plane_type;
+  const int ref = is_inter_block(mbmi);
+  const TX_TYPE tx_type = get_tx_type(type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, ref);
+  const int rate = av1_cost_coeffs(cpi, x, plane, block, tx_size, scan_order,
+                                   pd->above_context + blk_col,
+                                   pd->left_context + blk_row, 0);
+  args->this_rate += rate;
+  (void)plane_bsize;
+  av1_set_contexts(xd, pd, plane, tx_size, p->eobs[block] > 0, blk_col,
+                   blk_row);
+}
+
+static void set_entropy_context_b(int plane, int block, int blk_row,
+                                  int blk_col, BLOCK_SIZE plane_bsize,
+                                  TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  (void)plane_bsize;
+  av1_set_contexts(xd, pd, plane, tx_size, p->eobs[block] > 0, blk_col,
+                   blk_row);
+}
+
+#if CONFIG_NEW_TOKENSET
+static INLINE void add_token(TOKENEXTRA **t,
+                             aom_cdf_prob (*tail_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+                             aom_cdf_prob (*head_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+                             int eob_val, int first_val, int32_t extra,
+                             uint8_t token) {
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)->tail_cdf = tail_cdf;
+  (*t)->head_cdf = head_cdf;
+  (*t)->eob_val = eob_val;
+  (*t)->first_val = first_val;
+  (*t)++;
+}
+
+#else  // CONFIG_NEW_TOKENSET
+static INLINE void add_token(
+    TOKENEXTRA **t, const aom_prob *context_tree,
+#if CONFIG_EC_MULTISYMBOL
+    aom_cdf_prob (*token_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+#endif  // CONFIG_EC_MULTISYMBOL
+    int32_t extra, uint8_t token, uint8_t skip_eob_node, unsigned int *counts) {
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)->context_tree = context_tree;
+#if CONFIG_EC_MULTISYMBOL
+  (*t)->token_cdf = token_cdf;
+#endif  // CONFIG_EC_MULTISYMBOL
+  (*t)->skip_eob_node = skip_eob_node;
+  (*t)++;
+  ++counts[token];
+}
+#endif  // CONFIG_NEW_TOKENSET
+#endif  // !CONFIG_PVQ || CONFIG_VAR_TX
+
+#if CONFIG_PALETTE
+void av1_tokenize_palette_sb(const AV1_COMP *cpi,
+                             const struct ThreadData *const td, int plane,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate) {
+  const MACROBLOCK *const x = &td->mb;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const uint8_t *const color_map = xd->plane[plane].color_index_map;
+  const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  const int n = pmi->palette_size[plane];
+  int i, j;
+  int this_rate = 0;
+  uint8_t color_order[PALETTE_MAX_SIZE];
+  const aom_prob(
+      *const probs)[PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] =
+      plane == 0 ? av1_default_palette_y_color_index_prob
+                 : av1_default_palette_uv_color_index_prob;
+  int plane_block_width, rows, cols;
+  av1_get_block_dimensions(bsize, plane, xd, &plane_block_width, NULL, &rows,
+                           &cols);
+  assert(plane == 0 || plane == 1);
+
+#if CONFIG_PALETTE_THROUGHPUT
+  int k;
+  for (k = 1; k < rows + cols - 1; ++k) {
+    for (j = AOMMIN(k, cols - 1); j >= AOMMAX(0, k - rows + 1); --j) {
+      i = k - j;
+#else
+  for (i = 0; i < rows; ++i) {
+    for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+#endif  // CONFIG_PALETTE_THROUGHPUT
+      int color_new_idx;
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, i, j, n, color_order, &color_new_idx);
+      assert(color_new_idx >= 0 && color_new_idx < n);
+      if (dry_run == DRY_RUN_COSTCOEFFS)
+        this_rate += cpi->palette_y_color_cost[n - PALETTE_MIN_SIZE][color_ctx]
+                                              [color_new_idx];
+      (*t)->token = color_new_idx;
+      (*t)->context_tree = probs[n - PALETTE_MIN_SIZE][color_ctx];
+      (*t)->skip_eob_node = 0;
+      ++(*t);
+    }
+  }
+  if (rate) *rate += this_rate;
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_PVQ
+static void add_pvq_block(AV1_COMMON *const cm, MACROBLOCK *const x,
+                          PVQ_INFO *pvq) {
+  PVQ_QUEUE *q = x->pvq_q;
+  if (q->curr_pos >= q->buf_len) {
+    int new_buf_len = 2 * q->buf_len + 1;
+    PVQ_INFO *new_buf;
+    CHECK_MEM_ERROR(cm, new_buf, aom_malloc(new_buf_len * sizeof(PVQ_INFO)));
+    memcpy(new_buf, q->buf, q->buf_len * sizeof(PVQ_INFO));
+    aom_free(q->buf);
+    q->buf = new_buf;
+    q->buf_len = new_buf_len;
+  }
+  OD_COPY(q->buf + q->curr_pos, pvq, 1);
+  ++q->curr_pos;
+}
+
+// NOTE: This does not actually generate tokens, instead we store the encoding
+// decisions made for PVQ in a queue that we will read from when
+// actually writing the bitstream in write_modes_b
+static void tokenize_pvq(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  PVQ_INFO *pvq_info;
+
+  (void)block;
+  (void)blk_row;
+  (void)blk_col;
+  (void)plane_bsize;
+  (void)tx_size;
+
+  assert(block < MAX_PVQ_BLOCKS_IN_SB);
+  pvq_info = &x->pvq[block][plane];
+  add_pvq_block((AV1_COMMON * const)cm, x, pvq_info);
+}
+#endif  // CONFIG_PVQ
+
+static void tokenize_b(int plane, int block, int blk_row, int blk_col,
+                       BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+#if !CONFIG_PVQ
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TOKENEXTRA **tp = args->tp;
+  uint8_t token_cache[MAX_TX_SQUARE];
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int pt; /* near block/prev token context index */
+  int c;
+  TOKENEXTRA *t = *tp; /* store tokens starting here */
+  const int eob = p->eobs[block];
+  const PLANE_TYPE type = pd->plane_type;
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+#if CONFIG_SUPERTX
+  const int segment_id = AOMMIN(mbmi->segment_id, mbmi->segment_id_supertx);
+#else
+  const int segment_id = mbmi->segment_id;
+#endif  // CONFIG_SUEPRTX
+  const int16_t *scan, *nb;
+  const TX_TYPE tx_type = get_tx_type(type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int ref = is_inter_block(mbmi);
+  unsigned int(*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      td->rd_counts.coef_counts[txsize_sqr_map[tx_size]][type][ref];
+#if !CONFIG_NEW_TOKENSET
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  const aom_prob(*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->subframe_stats.coef_probs_buf[cpi->common.coef_probs_update_idx]
+                                        [txsize_sqr_map[tx_size]][type][ref];
+#else
+  aom_prob(*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->common.fc->coef_probs[txsize_sqr_map[tx_size]][type][ref];
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#endif  // !CONFIG_NEW_TOKENSET
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cpi->common.fc;
+#endif
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob(
+      *const coef_head_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_head_cdfs[txsize_sqr_map[tx_size]][type][ref];
+  aom_cdf_prob(
+      *const coef_tail_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_tail_cdfs[txsize_sqr_map[tx_size]][type][ref];
+  unsigned int(*const blockz_count)[2] =
+      td->counts->blockz_count[txsize_sqr_map[tx_size]][type][ref];
+  int eob_val;
+  int first_val = 1;
+#else
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob(*const coef_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_cdfs[txsize_sqr_map[tx_size]][type][ref];
+#endif
+  int skip_eob = 0;
+#endif
+  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
+  unsigned int(*const eob_branch)[COEFF_CONTEXTS] =
+      td->counts->eob_branch[txsize_sqr_map[tx_size]][type][ref];
+  const uint8_t *const band = get_band_translate(tx_size);
+  int16_t token;
+  EXTRABIT extra;
+  (void)plane_bsize;
+  pt = get_entropy_context(tx_size, pd->above_context + blk_col,
+                           pd->left_context + blk_row);
+  scan = scan_order->scan;
+  nb = scan_order->neighbors;
+  c = 0;
+
+#if CONFIG_NEW_TOKENSET
+  if (eob == 0)
+    add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt], 1,
+              1, 0, BLOCK_Z_TOKEN);
+
+  ++blockz_count[pt][eob != 0];
+
+  while (c < eob) {
+    int v = qcoeff[scan[c]];
+    first_val = (c == 0);
+
+    if (!v) {
+      add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt],
+                0, first_val, 0, ZERO_TOKEN);
+      ++counts[band[c]][pt][ZERO_TOKEN];
+      token_cache[scan[c]] = 0;
+    } else {
+      eob_val =
+          (c + 1 == eob) ? (c + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
+
+      av1_get_token_extra(v, &token, &extra);
+
+      add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt],
+                eob_val, first_val, extra, (uint8_t)token);
+
+      if (eob_val != LAST_EOB) {
+        ++counts[band[c]][pt][token];
+        ++eob_branch[band[c]][pt];
+        counts[band[c]][pt][EOB_TOKEN] += eob_val != NO_EOB;
+      }
+
+      token_cache[scan[c]] = av1_pt_energy_class[token];
+    }
+    ++c;
+    pt = get_coef_context(nb, token_cache, AOMMIN(c, eob - 1));
+  }
+#else
+  while (c < eob) {
+    const int v = qcoeff[scan[c]];
+    eob_branch[band[c]][pt] += !skip_eob;
+
+    av1_get_token_extra(v, &token, &extra);
+
+    add_token(&t, coef_probs[band[c]][pt],
+#if CONFIG_EC_MULTISYMBOL
+              &coef_cdfs[band[c]][pt],
+#endif
+              extra, (uint8_t)token, (uint8_t)skip_eob, counts[band[c]][pt]);
+
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+    ++c;
+    pt = get_coef_context(nb, token_cache, c);
+    skip_eob = (token == ZERO_TOKEN);
+  }
+  if (c < seg_eob) {
+    add_token(&t, coef_probs[band[c]][pt],
+#if CONFIG_EC_MULTISYMBOL
+              NULL,
+#endif
+              0, EOB_TOKEN, 0, counts[band[c]][pt]);
+    ++eob_branch[band[c]][pt];
+  }
+#endif  // CONFIG_NEW_TOKENSET
+
+#if CONFIG_COEF_INTERLEAVE
+  t->token = EOSB_TOKEN;
+  t++;
+#endif
+
+  *tp = t;
+
+#if CONFIG_ADAPT_SCAN
+  // Since dqcoeff is not available here, we pass qcoeff into
+  // av1_update_scan_count_facade(). The update behavior should be the same
+  // because av1_update_scan_count_facade() only cares if coefficients are zero
+  // or not.
+  av1_update_scan_count_facade((AV1_COMMON *)cm, td->counts, tx_size, tx_type,
+                               qcoeff, c);
+#endif
+
+  av1_set_contexts(xd, pd, plane, tx_size, c > 0, blk_col, blk_row);
+#else   // !CONFIG_PVQ
+  tokenize_pvq(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+#endif  // !CONFIG_PVQ
+}
+
+struct is_skippable_args {
+  uint16_t *eobs;
+  int *skippable;
+};
+static void is_skippable(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *argv) {
+  struct is_skippable_args *args = argv;
+  (void)plane;
+  (void)plane_bsize;
+  (void)tx_size;
+  (void)blk_row;
+  (void)blk_col;
+  args->skippable[0] &= (!args->eobs[block]);
+}
+
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              av1_foreach_transform_block() and simplify is_skippable().
+int av1_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 1;
+  struct is_skippable_args args = { x->plane[plane].eobs, &result };
+  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
+                                         &args);
+  return result;
+}
+
+#if CONFIG_VAR_TX
+void tokenize_vartx(ThreadData *td, TOKENEXTRA **t, RUN_TYPE dry_run,
+                    TX_SIZE tx_size, BLOCK_SIZE plane_bsize, int blk_row,
+                    int blk_col, int block, int plane, void *arg) {
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  TX_SIZE plane_tx_size;
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (tx_size == plane_tx_size) {
+    plane_bsize = get_plane_block_size(mbmi->sb_type, pd);
+    if (!dry_run)
+      tokenize_b(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+    else if (dry_run == DRY_RUN_NORMAL)
+      set_entropy_context_b(plane, block, blk_row, blk_col, plane_bsize,
+                            tx_size, arg);
+    else if (dry_run == DRY_RUN_COSTCOEFFS)
+      cost_coeffs_b(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+  } else {
+    // Half the block size in transform block unit.
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + ((i >> 1) * bsl);
+      const int offsetc = blk_col + ((i & 0x01) * bsl);
+
+      int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      tokenize_vartx(td, t, dry_run, sub_txs, plane_bsize, offsetr, offsetc,
+                     block, plane, arg);
+      block += step;
+    }
+  }
+}
+
+void av1_tokenize_sb_vartx(const AV1_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                           RUN_TYPE dry_run, int mi_row, int mi_col,
+                           BLOCK_SIZE bsize, int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  TOKENEXTRA *t_backup = *t;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  int plane;
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    if (dry_run) *t = t_backup;
+    return;
+  }
+
+  if (!dry_run)
+    td->counts->skip[ctx][0] += skip_inc;
+  else
+    *t = t_backup;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+    if (!is_chroma_reference(mi_row, mi_col, bsize,
+                             xd->plane[plane].subsampling_x,
+                             xd->plane[plane].subsampling_y)) {
+#if !CONFIG_PVQ
+      if (!dry_run) {
+        (*t)->token = EOSB_TOKEN;
+        (*t)++;
+      }
+#endif
+      continue;
+    }
+#endif
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+    const BLOCK_SIZE plane_bsize =
+        AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#else
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#endif
+    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+    const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+    const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
+    int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
+    int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
+    int idx, idy;
+    int block = 0;
+    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
+    for (idy = 0; idy < mi_height; idy += bh) {
+      for (idx = 0; idx < mi_width; idx += bw) {
+        tokenize_vartx(td, t, dry_run, max_tx_size, plane_bsize, idy, idx,
+                       block, plane, &arg);
+        block += step;
+      }
+    }
+
+    if (!dry_run) {
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+    }
+  }
+  if (rate) *rate += arg.this_rate;
+}
+#endif  // CONFIG_VAR_TX
+
+void av1_tokenize_sb(const AV1_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                     RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                     const int mi_row, const int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+#if CONFIG_COEF_INTERLEAVE
+    td->counts->skip[ctx][0] += skip_inc;
+    av1_foreach_transformed_block_interleave(xd, bsize, tokenize_b, &arg);
+#else
+    int plane;
+
+    td->counts->skip[ctx][0] += skip_inc;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y)) {
+#if !CONFIG_PVQ
+        (*t)->token = EOSB_TOKEN;
+        (*t)++;
+#endif
+        continue;
+      }
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_b,
+                                             &arg);
+#if !CONFIG_PVQ
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+#endif  // !CONFIG_PVQ
+    }
+#endif
+  }
+#if !CONFIG_PVQ
+  else if (dry_run == DRY_RUN_NORMAL) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y))
+        continue;
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                             set_entropy_context_b, &arg);
+    }
+  } else if (dry_run == DRY_RUN_COSTCOEFFS) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y))
+        continue;
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, cost_coeffs_b,
+                                             &arg);
+    }
+  }
+#endif  // !CONFIG_PVQ
+
+  if (rate) *rate += arg.this_rate;
+}
+
+#if CONFIG_SUPERTX
+void av1_tokenize_sb_supertx(const AV1_COMP *cpi, ThreadData *td,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &td->mb.e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  TOKENEXTRA *t_backup = *t;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id_supertx, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    if (dry_run) *t = t_backup;
+    return;
+  }
+
+  if (!dry_run) {
+    int plane;
+    td->counts->skip[ctx][0] += skip_inc;
+
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_b,
+                                             &arg);
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+    }
+  } else if (dry_run == DRY_RUN_NORMAL) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                             set_entropy_context_b, &arg);
+    *t = t_backup;
+  } else if (dry_run == DRY_RUN_COSTCOEFFS) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, cost_coeffs_b,
+                                             &arg);
+  }
+  if (rate) *rate += arg.this_rate;
+}
+#endif  // CONFIG_SUPERTX

third_party/aom/av1/encoder/tokenize.h

@@ -0,0 +1,151 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TOKENIZE_H_
+#define AV1_ENCODER_TOKENIZE_H_
+
+#include "av1/common/entropy.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/treewriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define EOSB_TOKEN 127  // Not signalled, encoder only
+
+#if CONFIG_HIGHBITDEPTH
+typedef int32_t EXTRABIT;
+#else
+typedef int16_t EXTRABIT;
+#endif
+
+typedef struct {
+  int16_t token;
+  EXTRABIT extra;
+} TOKENVALUE;
+
+typedef struct {
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob (*tail_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+  aom_cdf_prob (*head_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+  int eob_val;
+  int first_val;
+#elif CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob (*token_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+#endif
+  const aom_prob *context_tree;
+  EXTRABIT extra;
+  uint8_t token;
+  uint8_t skip_eob_node;
+} TOKENEXTRA;
+
+extern const aom_tree_index av1_coef_tree[];
+extern const aom_tree_index av1_coef_con_tree[];
+#if !CONFIG_EC_MULTISYMBOL
+extern const struct av1_token av1_coef_encodings[];
+#endif  // !CONFIG_EC_MULTISYMBOL
+
+int av1_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+struct AV1_COMP;
+struct ThreadData;
+
+typedef enum {
+  OUTPUT_ENABLED = 0,
+  DRY_RUN_NORMAL,
+  DRY_RUN_COSTCOEFFS,
+} RUN_TYPE;
+
+// Note in all the tokenize functions rate if non NULL is incremented
+// with the coefficient token cost only if dry_run = DRY_RUN_COSTCOEFS,
+// otherwise rate is not incremented.
+#if CONFIG_VAR_TX
+void av1_tokenize_sb_vartx(const struct AV1_COMP *cpi, struct ThreadData *td,
+                           TOKENEXTRA **t, RUN_TYPE dry_run, int mi_row,
+                           int mi_col, BLOCK_SIZE bsize, int *rate);
+#endif
+#if CONFIG_PALETTE
+void av1_tokenize_palette_sb(const struct AV1_COMP *cpi,
+                             const struct ThreadData *const td, int plane,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate);
+#endif  // CONFIG_PALETTE
+void av1_tokenize_sb(const struct AV1_COMP *cpi, struct ThreadData *td,
+                     TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                     int *rate, const int mi_row, const int mi_col);
+#if CONFIG_SUPERTX
+void av1_tokenize_sb_supertx(const struct AV1_COMP *cpi, struct ThreadData *td,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate);
+#endif
+
+extern const int16_t *av1_dct_value_cost_ptr;
+/* TODO: The Token field should be broken out into a separate char array to
+ *  improve cache locality, since it's needed for costing when the rest of the
+ *  fields are not.
+ */
+extern const TOKENVALUE *av1_dct_value_tokens_ptr;
+extern const TOKENVALUE *av1_dct_cat_lt_10_value_tokens;
+extern const int *av1_dct_cat_lt_10_value_cost;
+extern const int16_t av1_cat6_low_cost[256];
+#if CONFIG_HIGHBITDEPTH
+#define CAT6_HIGH_COST_ENTRIES 1024
+#else
+#define CAT6_HIGH_COST_ENTRIES 64
+#endif
+extern const int av1_cat6_high_cost[CAT6_HIGH_COST_ENTRIES];
+extern const uint8_t av1_cat6_skipped_bits_discount[8];
+
+static INLINE void av1_get_token_extra(int v, int16_t *token, EXTRABIT *extra) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    *token = CATEGORY6_TOKEN;
+    if (v >= CAT6_MIN_VAL)
+      *extra = 2 * v - 2 * CAT6_MIN_VAL;
+    else
+      *extra = -2 * v - 2 * CAT6_MIN_VAL + 1;
+    return;
+  }
+  *token = av1_dct_cat_lt_10_value_tokens[v].token;
+  *extra = av1_dct_cat_lt_10_value_tokens[v].extra;
+}
+static INLINE int16_t av1_get_token(int v) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) return 10;
+  return av1_dct_cat_lt_10_value_tokens[v].token;
+}
+
+static INLINE int av1_get_token_cost(int v, int16_t *token, int cat6_bits) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    EXTRABIT extrabits;
+    *token = CATEGORY6_TOKEN;
+    extrabits = abs(v) - CAT6_MIN_VAL;
+    return av1_cat6_low_cost[extrabits & 0xff] +
+           av1_cat6_high_cost[extrabits >> 8] -
+           av1_cat6_skipped_bits_discount[18 - cat6_bits];
+  }
+  *token = av1_dct_cat_lt_10_value_tokens[v].token;
+  return av1_dct_cat_lt_10_value_cost[v];
+}
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+static INLINE int get_tx_eob(const struct segmentation *seg, int segment_id,
+                             TX_SIZE tx_size) {
+  const int eob_max = tx_size_2d[tx_size];
+  return segfeature_active(seg, segment_id, SEG_LVL_SKIP) ? 0 : eob_max;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TOKENIZE_H_

third_party/aom/av1/encoder/treewriter.c

@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/treewriter.h"
+
+static void tree2tok(struct av1_token *tokens, const aom_tree_index *tree,
+                     int i, int v, int l) {
+  v += v;
+  ++l;
+
+  do {
+    const aom_tree_index j = tree[i++];
+    if (j <= 0) {
+      tokens[-j].value = v;
+      tokens[-j].len = l;
+    } else {
+      tree2tok(tokens, tree, j, v, l);
+    }
+  } while (++v & 1);
+}
+
+void av1_tokens_from_tree(struct av1_token *tokens,
+                          const aom_tree_index *tree) {
+  tree2tok(tokens, tree, 0, 0, 0);
+}
+
+static unsigned int convert_distribution(unsigned int i, aom_tree tree,
+                                         unsigned int branch_ct[][2],
+                                         const unsigned int num_events[]) {
+  unsigned int left, right;
+
+  if (tree[i] <= 0)
+    left = num_events[-tree[i]];
+  else
+    left = convert_distribution(tree[i], tree, branch_ct, num_events);
+
+  if (tree[i + 1] <= 0)
+    right = num_events[-tree[i + 1]];
+  else
+    right = convert_distribution(tree[i + 1], tree, branch_ct, num_events);
+
+  branch_ct[i >> 1][0] = left;
+  branch_ct[i >> 1][1] = right;
+  return left + right;
+}
+
+void av1_tree_probs_from_distribution(aom_tree tree,
+                                      unsigned int branch_ct[/* n-1 */][2],
+                                      const unsigned int num_events[/* n */]) {
+  convert_distribution(0, tree, branch_ct, num_events);
+}

third_party/aom/av1/encoder/treewriter.h

@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TREEWRITER_H_
+#define AV1_ENCODER_TREEWRITER_H_
+
+#include "aom_dsp/bitwriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_tree_probs_from_distribution(aom_tree tree,
+                                      unsigned int branch_ct[/* n - 1 */][2],
+                                      const unsigned int num_events[/* n */]);
+
+struct av1_token {
+  int value;
+  int len;
+};
+
+void av1_tokens_from_tree(struct av1_token *, const aom_tree_index *);
+
+static INLINE void av1_write_token(aom_writer *w, const aom_tree_index *tree,
+                                   const aom_prob *probs,
+                                   const struct av1_token *token) {
+  aom_write_tree(w, tree, probs, token->value, token->len, 0);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TREEWRITER_H_

third_party/aom/av1/encoder/variance_tree.c

@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/variance_tree.h"
+#include "av1/encoder/encoder.h"
+
+void av1_setup_var_tree(struct AV1Common *cm, ThreadData *td) {
+  int i, j;
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 1024;
+  const int tree_nodes = 1024 + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 256;
+  const int tree_nodes = 256 + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int index = 0;
+  VAR_TREE *this_var;
+  int nodes;
+
+  aom_free(td->var_tree);
+  CHECK_MEM_ERROR(cm, td->var_tree,
+                  aom_calloc(tree_nodes, sizeof(*td->var_tree)));
+
+  this_var = &td->var_tree[0];
+
+  // Sets up all the leaf nodes in the tree.
+  for (index = 0; index < leaf_nodes; ++index) {
+    VAR_TREE *const leaf = &td->var_tree[index];
+    leaf->split[0] = NULL;
+  }
+
+  // Each node has 4 leaf nodes, fill in the child pointers
+  // from leafs to the root.
+  for (nodes = leaf_nodes >> 2; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i, ++index) {
+      VAR_TREE *const node = &td->var_tree[index];
+      for (j = 0; j < 4; j++) node->split[j] = this_var++;
+    }
+  }
+
+  // Set up the root node for the largest superblock size
+  i = MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2;
+  td->var_root[i] = &td->var_tree[tree_nodes - 1];
+  // Set up the root nodes for the rest of the possible superblock sizes
+  while (--i >= 0) {
+    td->var_root[i] = td->var_root[i + 1]->split[0];
+  }
+}
+
+void av1_free_var_tree(ThreadData *td) {
+  aom_free(td->var_tree);
+  td->var_tree = NULL;
+}

third_party/aom/av1/encoder/variance_tree.h

@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_VARIANCE_TREE_H_
+#define AV1_ENCODER_VARIANCE_TREE_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct ThreadData;
+
+typedef struct {
+  int64_t sum_square_error;
+  int64_t sum_error;
+  int log2_count;
+  int variance;
+} VAR;
+
+typedef struct {
+  VAR none;
+  VAR horz[2];
+  VAR vert[2];
+} partition_variance;
+
+typedef struct VAR_TREE {
+  int force_split;
+  partition_variance variances;
+  struct VAR_TREE *split[4];
+  BLOCK_SIZE bsize;
+  const uint8_t *src;
+  const uint8_t *ref;
+  int src_stride;
+  int ref_stride;
+  int width;
+  int height;
+#if CONFIG_HIGHBITDEPTH
+  int highbd;
+#endif  // CONFIG_HIGHBITDEPTH
+} VAR_TREE;
+
+void av1_setup_var_tree(struct AV1Common *cm, struct ThreadData *td);
+void av1_free_var_tree(struct ThreadData *td);
+
+// Set variance values given sum square error, sum error, count.
+static INLINE void fill_variance(int64_t s2, int64_t s, int c, VAR *v) {
+  v->sum_square_error = s2;
+  v->sum_error = s;
+  v->log2_count = c;
+  v->variance =
+      (int)(256 * (v->sum_square_error -
+                   ((v->sum_error * v->sum_error) >> v->log2_count)) >>
+            v->log2_count);
+}
+
+static INLINE void sum_2_variances(const VAR *a, const VAR *b, VAR *r) {
+  assert(a->log2_count == b->log2_count);
+  fill_variance(a->sum_square_error + b->sum_square_error,
+                a->sum_error + b->sum_error, a->log2_count + 1, r);
+}
+
+static INLINE void fill_variance_node(VAR_TREE *vt) {
+  sum_2_variances(&vt->split[0]->variances.none, &vt->split[1]->variances.none,
+                  &vt->variances.horz[0]);
+  sum_2_variances(&vt->split[2]->variances.none, &vt->split[3]->variances.none,
+                  &vt->variances.horz[1]);
+  sum_2_variances(&vt->split[0]->variances.none, &vt->split[2]->variances.none,
+                  &vt->variances.vert[0]);
+  sum_2_variances(&vt->split[1]->variances.none, &vt->split[3]->variances.none,
+                  &vt->variances.vert[1]);
+  sum_2_variances(&vt->variances.vert[0], &vt->variances.vert[1],
+                  &vt->variances.none);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_VARIANCE_TREE_H_ */

third_party/aom/av1/encoder/wedge_utils.c

@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+
+#include "aom_ports/mem.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "av1/common/reconinter.h"
+
+#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS)
+
+/**
+ * Computes SSE of a compound predictor constructed from 2 fundamental
+ * predictors p0 and p1 using blending with mask.
+ *
+ * r1:  Residuals of p1.
+ *      (source - p1)
+ * d:   Difference of p1 and p0.
+ *      (p1 - p0)
+ * m:   The blending mask
+ * N:   Number of pixels
+ *
+ * 'r1', 'd', and 'm' are contiguous.
+ *
+ * Computes:
+ *  Sum((MAX_MASK_VALUE*r1 + mask*d)**2), which is equivalent to:
+ *  Sum((mask*r0 + (MAX_MASK_VALUE-mask)*r1)**2),
+ *    where r0 is (source - p0), and r1 is (source - p1), which is in turn
+ *    is equivalent to:
+ *  Sum((source*MAX_MASK_VALUE - (mask*p0 + (MAX_MASK_VALUE-mask)*p1))**2),
+ *    which is the SSE of the residuals of the compound predictor scaled up by
+ *    MAX_MASK_VALUE**2.
+ *
+ * Note that we clamp the partial term in the loop to 16 bits signed. This is
+ * to facilitate equivalent SIMD implementation. It should have no effect if
+ * residuals are within 16 - WEDGE_WEIGHT_BITS (=10) signed, which always
+ * holds for 8 bit input, and on real input, it should hold practically always,
+ * as residuals are expected to be small.
+ */
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d,
+                                        const uint8_t *m, int N) {
+  uint64_t csse = 0;
+  int i;
+
+  for (i = 0; i < N; i++) {
+    int32_t t = MAX_MASK_VALUE * r1[i] + m[i] * d[i];
+    t = clamp(t, INT16_MIN, INT16_MAX);
+    csse += t * t;
+  }
+  return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS);
+}
+
+/**
+ * Choose the mask sign for a compound predictor.
+ *
+ * ds:    Difference of the squares of the residuals.
+ *        r0**2 - r1**2
+ * m:     The blending mask
+ * N:     Number of pixels
+ * limit: Pre-computed threshold value.
+ *        MAX_MASK_VALUE/2 * (sum(r0**2) - sum(r1**2))
+ *
+ * 'ds' and 'm' are contiguous.
+ *
+ * Returns true if the negated mask has lower SSE compared to the positive
+ * mask. Computation is based on:
+ *  Sum((mask*r0 + (MAX_MASK_VALUE-mask)*r1)**2)
+ *                                     >
+ *                                Sum(((MAX_MASK_VALUE-mask)*r0 + mask*r1)**2)
+ *
+ *  which can be simplified to:
+ *
+ *  Sum(mask*(r0**2 - r1**2)) > MAX_MASK_VALUE/2 * (sum(r0**2) - sum(r1**2))
+ *
+ *  The right hand side does not depend on the mask, and needs to be passed as
+ *  the 'limit' parameter.
+ *
+ *  After pre-computing (r0**2 - r1**2), which is passed in as 'ds', the left
+ *  hand side is simply a scalar product between an int16_t and uint8_t vector.
+ *
+ *  Note that for efficiency, ds is stored on 16 bits. Real input residuals
+ *  being small, this should not cause a noticeable issue.
+ */
+int av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N,
+                                    int64_t limit) {
+  int64_t acc = 0;
+
+  do {
+    acc += *ds++ * *m++;
+  } while (--N);
+
+  return acc > limit;
+}
+
+/**
+ * Compute the element-wise difference of the squares of 2 arrays.
+ *
+ * d: Difference of the squares of the inputs: a**2 - b**2
+ * a: First input array
+ * b: Second input array
+ * N: Number of elements
+ *
+ * 'd', 'a', and 'b' are contiguous.
+ *
+ * The result is saturated to signed 16 bits.
+ */
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a,
+                                       const int16_t *b, int N) {
+  int i;
+
+  for (i = 0; i < N; i++)
+    d[i] = clamp(a[i] * a[i] - b[i] * b[i], INT16_MIN, INT16_MAX);
+}

third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c

@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>
+#include <stdint.h>
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+// Coefficient quantization phase 1
+// param[0-2] : rounding/quan/dequan constants
+static INLINE void quantize_coeff_phase1(__m128i *coeff, const __m128i *param,
+                                         const int shift, const int scale,
+                                         __m128i *qcoeff, __m128i *dquan,
+                                         __m128i *sign) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi32(1);
+
+  *sign = _mm_cmplt_epi32(*coeff, zero);
+  *sign = _mm_or_si128(*sign, one);
+  *coeff = _mm_abs_epi32(*coeff);
+
+  qcoeff[0] = _mm_add_epi32(*coeff, param[0]);
+  qcoeff[1] = _mm_unpackhi_epi32(qcoeff[0], zero);
+  qcoeff[0] = _mm_unpacklo_epi32(qcoeff[0], zero);
+
+  qcoeff[0] = _mm_mul_epi32(qcoeff[0], param[1]);
+  qcoeff[0] = _mm_srli_epi64(qcoeff[0], shift);
+  dquan[0] = _mm_mul_epi32(qcoeff[0], param[2]);
+  dquan[0] = _mm_srli_epi64(dquan[0], scale);
+}
+
+// Coefficient quantization phase 2
+static INLINE void quantize_coeff_phase2(__m128i *qcoeff, __m128i *dquan,
+                                         const __m128i *sign,
+                                         const __m128i *param, const int shift,
+                                         const int scale, tran_low_t *qAddr,
+                                         tran_low_t *dqAddr) {
+  __m128i mask0L = _mm_set_epi32(-1, -1, 0, 0);
+  __m128i mask0H = _mm_set_epi32(0, 0, -1, -1);
+
+  qcoeff[1] = _mm_mul_epi32(qcoeff[1], param[1]);
+  qcoeff[1] = _mm_srli_epi64(qcoeff[1], shift);
+  dquan[1] = _mm_mul_epi32(qcoeff[1], param[2]);
+  dquan[1] = _mm_srli_epi64(dquan[1], scale);
+
+  // combine L&H
+  qcoeff[0] = _mm_shuffle_epi32(qcoeff[0], 0xd8);
+  qcoeff[1] = _mm_shuffle_epi32(qcoeff[1], 0x8d);
+
+  qcoeff[0] = _mm_and_si128(qcoeff[0], mask0H);
+  qcoeff[1] = _mm_and_si128(qcoeff[1], mask0L);
+
+  dquan[0] = _mm_shuffle_epi32(dquan[0], 0xd8);
+  dquan[1] = _mm_shuffle_epi32(dquan[1], 0x8d);
+
+  dquan[0] = _mm_and_si128(dquan[0], mask0H);
+  dquan[1] = _mm_and_si128(dquan[1], mask0L);
+
+  qcoeff[0] = _mm_or_si128(qcoeff[0], qcoeff[1]);
+  dquan[0] = _mm_or_si128(dquan[0], dquan[1]);
+
+  qcoeff[0] = _mm_sign_epi32(qcoeff[0], *sign);
+  dquan[0] = _mm_sign_epi32(dquan[0], *sign);
+
+  _mm_storeu_si128((__m128i *)qAddr, qcoeff[0]);
+  _mm_storeu_si128((__m128i *)dqAddr, dquan[0]);
+}
+
+static INLINE void find_eob(tran_low_t *qcoeff_ptr, const int16_t *iscan,
+                            __m128i *eob) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, iscanIdx;
+  const __m128i q0 = _mm_loadu_si128((__m128i const *)qcoeff_ptr);
+  const __m128i q1 = _mm_loadu_si128((__m128i const *)(qcoeff_ptr + 4));
+  __m128i nz_flag0 = _mm_cmpeq_epi32(q0, zero);
+  __m128i nz_flag1 = _mm_cmpeq_epi32(q1, zero);
+
+  nz_flag0 = _mm_cmpeq_epi32(nz_flag0, zero);
+  nz_flag1 = _mm_cmpeq_epi32(nz_flag1, zero);
+
+  mask = _mm_packs_epi32(nz_flag0, nz_flag1);
+  iscanIdx = _mm_loadu_si128((__m128i const *)iscan);
+  iscanIdx = _mm_sub_epi16(iscanIdx, mask);
+  iscanIdx = _mm_and_si128(iscanIdx, mask);
+  *eob = _mm_max_epi16(*eob, iscanIdx);
+}
+
+static INLINE uint16_t get_accumulated_eob(__m128i *eob) {
+  __m128i eob_shuffled;
+  uint16_t eobValue;
+  eob_shuffled = _mm_shuffle_epi32(*eob, 0xe);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(*eob, 0xe);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(*eob, 0x1);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eobValue = _mm_extract_epi16(*eob, 0);
+  return eobValue;
+}
+
+void av1_highbd_quantize_fp_sse4_1(
+    const tran_low_t *coeff_ptr, intptr_t count, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan, int log_scale) {
+  __m128i coeff[2], qcoeff[2], dequant[2], qparam[3], coeff_sign;
+  __m128i eob = _mm_setzero_si128();
+  const tran_low_t *src = coeff_ptr;
+  tran_low_t *quanAddr = qcoeff_ptr;
+  tran_low_t *dquanAddr = dqcoeff_ptr;
+  const int shift = 16 - log_scale;
+  const int coeff_stride = 4;
+  const int quan_stride = coeff_stride;
+  (void)skip_block;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  memset(quanAddr, 0, count * sizeof(quanAddr[0]));
+  memset(dquanAddr, 0, count * sizeof(dquanAddr[0]));
+
+  if (!skip_block) {
+    coeff[0] = _mm_loadu_si128((__m128i const *)src);
+
+    qparam[0] =
+        _mm_set_epi32(round_ptr[1], round_ptr[1], round_ptr[1], round_ptr[0]);
+    qparam[1] = _mm_set_epi64x(quant_ptr[1], quant_ptr[0]);
+    qparam[2] = _mm_set_epi64x(dequant_ptr[1], dequant_ptr[0]);
+
+    // DC and first 3 AC
+    quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff, dequant,
+                          &coeff_sign);
+
+    // update round/quan/dquan for AC
+    qparam[0] = _mm_unpackhi_epi64(qparam[0], qparam[0]);
+    qparam[1] = _mm_set_epi64x(quant_ptr[1], quant_ptr[1]);
+    qparam[2] = _mm_set_epi64x(dequant_ptr[1], dequant_ptr[1]);
+
+    quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                          log_scale, quanAddr, dquanAddr);
+
+    // next 4 AC
+    coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride));
+    quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff, dequant,
+                          &coeff_sign);
+    quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                          log_scale, quanAddr + quan_stride,
+                          dquanAddr + quan_stride);
+
+    find_eob(quanAddr, iscan, &eob);
+
+    count -= 8;
+
+    // loop for the rest of AC
+    while (count > 0) {
+      src += coeff_stride << 1;
+      quanAddr += quan_stride << 1;
+      dquanAddr += quan_stride << 1;
+      iscan += quan_stride << 1;
+
+      coeff[0] = _mm_loadu_si128((__m128i const *)src);
+      coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride));
+
+      quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff,
+                            dequant, &coeff_sign);
+      quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                            log_scale, quanAddr, dquanAddr);
+
+      quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff,
+                            dequant, &coeff_sign);
+      quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                            log_scale, quanAddr + quan_stride,
+                            dquanAddr + quan_stride);
+
+      find_eob(quanAddr, iscan, &eob);
+
+      count -= 8;
+    }
+    *eob_ptr = get_accumulated_eob(&eob);
+  } else {
+    *eob_ptr = 0;
+  }
+}

third_party/aom/av1/encoder/x86/av1_quantize_sse2.c

@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "aom/aom_integer.h"
+
+void av1_quantize_fp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan_ptr,
+                          const int16_t *iscan_ptr) {
+  __m128i zero;
+  __m128i thr;
+  int16_t nzflag;
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i *)round_ptr);
+        quant = _mm_load_si128((const __m128i *)quant_ptr);
+        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    thr = _mm_srai_epi16(dequant, 1);
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        coeff0 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+                 _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}

third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm

@@ -0,0 +1,204 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%define private_prefix av1
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+%macro QUANTIZE_FP 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, fp_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m1, m5
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, fp_32x32
+  psllw                           m2, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                            r5q, [  r5q+ncoeffq*2]
+  lea                            r3q, [ r3q+ncoeffq*2]
+  lea                            r4q, [r4q+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m8
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; r4[i] = r3[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+  psrlw                           m0, m3, 2
+%else
+  psrlw                           m0, m3, 1
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m8
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+
+  pcmpgtw                         m7, m6,  m0
+  pcmpgtw                        m12, m11, m0
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+
+  or                              r6, r2
+  jz .skip_iter
+
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m14
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; r4[i] = r3[i] * q
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m14
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+  jmp .accumulate_eob
+.skip_iter:
+  mova            [r3q+ncoeffq*2+ 0], m5
+  mova            [r3q+ncoeffq*2+16], m5
+  mova            [r4q+ncoeffq*2+ 0], m5
+  mova            [r4q+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                           [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+  lea                            r0q, [r0q+ncoeffq*2]
+  lea                            r2q, [r2q+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova            [r0q+ncoeffq*2+ 0], m7
+  mova            [r0q+ncoeffq*2+16], m7
+  mova            [r2q+ncoeffq*2+ 0], m7
+  mova            [r2q+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                     word [r3q], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FP fp, 7
+QUANTIZE_FP fp_32x32, 7

third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm

@@ -0,0 +1,219 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(av1_ssim_parms_16x16_sse2) PRIVATE
+sym(av1_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(av1_ssim_parms_8x8_sse2) PRIVATE
+sym(av1_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret

third_party/aom/av1/encoder/x86/dct_sse2.asm

@@ -0,0 +1,87 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+%define private_prefix av1
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+                                ; 00 01 02 03
+                                ; 10 11 12 13
+                                ; 20 21 22 23
+                                ; 30 31 32 33
+  punpcklwd       m0,        m1 ; 00 10 01 11  02 12 03 13
+  punpcklwd       m2,        m3 ; 20 30 21 31  22 32 23 33
+  mova            m1,        m0
+  punpckldq       m0,        m2 ; 00 10 20 30  01 11 21 31
+  punpckhdq       m1,        m2 ; 02 12 22 32  03 13 23 33
+%endmacro
+
+INIT_XMM sse2
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  SWAP            1,         2
+  psrldq          m1,        m0, 8
+  psrldq          m3,        m2, 8
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+
+%if CONFIG_HIGHBITDEPTH
+  ; sign extension
+  mova            m2,             m0
+  mova            m3,             m1
+  punpcklwd       m0,             m0
+  punpcklwd       m1,             m1
+  punpckhwd       m2,             m2
+  punpckhwd       m3,             m3
+  psrad           m0,             16
+  psrad           m1,             16
+  psrad           m2,             16
+  psrad           m3,             16
+  mova            [outputq],      m0
+  mova            [outputq + 16], m2
+  mova            [outputq + 32], m1
+  mova            [outputq + 48], m3
+%else
+  mova            [outputq],      m0
+  mova            [outputq + 16], m1
+%endif
+
+  RET