tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

nonrd_pickmode.c (158408B)


      1 /*
      2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
      3 *
      4 * This source code is subject to the terms of the BSD 2 Clause License and
      5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
      6 * was not distributed with this source code in the LICENSE file, you can
      7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
      8 * Media Patent License 1.0 was not distributed with this source code in the
      9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
     10 
     11 */
     12 
     13 #include <assert.h>
     14 #include <limits.h>
     15 #include <math.h>
     16 #include <stdio.h>
     17 
     18 #include "av1/common/reconinter.h"
     19 #include "av1/common/reconintra.h"
     20 
     21 #include "av1/encoder/encodemv.h"
     22 #include "av1/encoder/intra_mode_search.h"
     23 #include "av1/encoder/model_rd.h"
     24 #include "av1/encoder/motion_search_facade.h"
     25 #include "av1/encoder/nonrd_opt.h"
     26 #include "av1/encoder/palette.h"
     27 #include "av1/encoder/reconinter_enc.h"
     28 #include "av1/encoder/var_based_part.h"
     29 
     30 static inline int early_term_inter_search_with_sse(int early_term_idx,
     31                                                   BLOCK_SIZE bsize,
     32                                                   int64_t this_sse,
     33                                                   int64_t best_sse,
     34                                                   PREDICTION_MODE this_mode) {
     35  // Aggressiveness to terminate inter mode search early is adjusted based on
     36  // speed and block size.
     37  static const double early_term_thresh[4][4] = { { 0.65, 0.65, 0.65, 0.7 },
     38                                                  { 0.6, 0.65, 0.85, 0.9 },
     39                                                  { 0.5, 0.5, 0.55, 0.6 },
     40                                                  { 0.6, 0.75, 0.85, 0.85 } };
     41  static const double early_term_thresh_newmv_nearestmv[4] = { 0.3, 0.3, 0.3,
     42                                                               0.3 };
     43 
     44  const int size_group = size_group_lookup[bsize];
     45  assert(size_group < 4);
     46  assert((early_term_idx > 0) && (early_term_idx < EARLY_TERM_INDICES));
     47  const double threshold =
     48      ((early_term_idx == EARLY_TERM_IDX_4) &&
     49       (this_mode == NEWMV || this_mode == NEARESTMV))
     50          ? early_term_thresh_newmv_nearestmv[size_group]
     51          : early_term_thresh[early_term_idx - 1][size_group];
     52 
     53  // Terminate inter mode search early based on best sse so far.
     54  if ((early_term_idx > 0) && (threshold * this_sse > best_sse)) {
     55    return 1;
     56  }
     57  return 0;
     58 }
     59 
     60 static inline void init_best_pickmode(BEST_PICKMODE *bp) {
     61  bp->best_sse = INT64_MAX;
     62  bp->best_mode = NEARESTMV;
     63  bp->best_ref_frame = LAST_FRAME;
     64  bp->best_second_ref_frame = NONE_FRAME;
     65  bp->best_tx_size = TX_8X8;
     66  bp->tx_type = DCT_DCT;
     67  bp->best_pred_filter = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
     68  bp->best_mode_skip_txfm = 0;
     69  bp->best_mode_initial_skip_flag = 0;
     70  bp->best_pred = NULL;
     71  bp->best_motion_mode = SIMPLE_TRANSLATION;
     72  bp->num_proj_ref = 0;
     73  av1_zero(bp->wm_params);
     74  av1_zero(bp->pmi);
     75 }
     76 
     77 // Copy best inter mode parameters to best_pickmode
     78 static inline void update_search_state_nonrd(
     79    InterModeSearchStateNonrd *search_state, MB_MODE_INFO *const mi,
     80    TxfmSearchInfo *txfm_info, RD_STATS *nonskip_rdc, PICK_MODE_CONTEXT *ctx,
     81    PREDICTION_MODE this_best_mode, const int64_t sse_y) {
     82  BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
     83 
     84  best_pickmode->best_sse = sse_y;
     85  best_pickmode->best_mode = this_best_mode;
     86  best_pickmode->best_motion_mode = mi->motion_mode;
     87  best_pickmode->wm_params = mi->wm_params;
     88  best_pickmode->num_proj_ref = mi->num_proj_ref;
     89  best_pickmode->best_pred_filter = mi->interp_filters;
     90  best_pickmode->best_tx_size = mi->tx_size;
     91  best_pickmode->best_ref_frame = mi->ref_frame[0];
     92  best_pickmode->best_second_ref_frame = mi->ref_frame[1];
     93  best_pickmode->best_mode_skip_txfm = search_state->this_rdc.skip_txfm;
     94  best_pickmode->best_mode_initial_skip_flag =
     95      (nonskip_rdc->rate == INT_MAX && search_state->this_rdc.skip_txfm);
     96  if (!best_pickmode->best_mode_skip_txfm) {
     97    memcpy(ctx->blk_skip, txfm_info->blk_skip,
     98           sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
     99  }
    100 }
    101 
    102 static inline int subpel_select(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
    103                                int_mv *mv, MV ref_mv, FULLPEL_MV start_mv,
    104                                bool fullpel_performed_well) {
    105  const int frame_lowmotion = cpi->rc.avg_frame_low_motion;
    106  const int reduce_mv_pel_precision_highmotion =
    107      cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion;
    108 
    109  // Reduce MV precision for higher int MV value & frame-level motion
    110  if (reduce_mv_pel_precision_highmotion >= 3) {
    111    int mv_thresh = 4;
    112    const int is_low_resoln =
    113        (cpi->common.width * cpi->common.height <= 320 * 240);
    114    mv_thresh = (bsize > BLOCK_32X32) ? 2 : (bsize > BLOCK_16X16) ? 4 : 6;
    115    if (frame_lowmotion > 0 && frame_lowmotion < 40) mv_thresh = 12;
    116    mv_thresh = (is_low_resoln) ? mv_thresh >> 1 : mv_thresh;
    117    if (abs(mv->as_fullmv.row) >= mv_thresh ||
    118        abs(mv->as_fullmv.col) >= mv_thresh)
    119      return HALF_PEL;
    120  } else if (reduce_mv_pel_precision_highmotion >= 1) {
    121    int mv_thresh;
    122    const int th_vals[2][3] = { { 4, 8, 10 }, { 4, 6, 8 } };
    123    const int th_idx = reduce_mv_pel_precision_highmotion - 1;
    124    assert(th_idx >= 0 && th_idx < 2);
    125    if (frame_lowmotion > 0 && frame_lowmotion < 40)
    126      mv_thresh = 12;
    127    else
    128      mv_thresh = (bsize >= BLOCK_32X32)   ? th_vals[th_idx][0]
    129                  : (bsize >= BLOCK_16X16) ? th_vals[th_idx][1]
    130                                           : th_vals[th_idx][2];
    131    if (abs(mv->as_fullmv.row) >= (mv_thresh << 1) ||
    132        abs(mv->as_fullmv.col) >= (mv_thresh << 1))
    133      return FULL_PEL;
    134    else if (abs(mv->as_fullmv.row) >= mv_thresh ||
    135             abs(mv->as_fullmv.col) >= mv_thresh)
    136      return HALF_PEL;
    137  }
    138  // Reduce MV precision for relatively static (e.g. background), low-complex
    139  // large areas
    140  if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 2) {
    141    const int qband = x->qindex >> (QINDEX_BITS - 2);
    142    assert(qband < 4);
    143    if (x->content_state_sb.source_sad_nonrd <= kVeryLowSad &&
    144        bsize > BLOCK_16X16 && qband != 0) {
    145      if (x->source_variance < 500)
    146        return FULL_PEL;
    147      else if (x->source_variance < 5000)
    148        return HALF_PEL;
    149    }
    150  } else if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 1) {
    151    if (fullpel_performed_well && ref_mv.row == 0 && ref_mv.col == 0 &&
    152        start_mv.row == 0 && start_mv.col == 0)
    153      return HALF_PEL;
    154  }
    155  return cpi->sf.mv_sf.subpel_force_stop;
    156 }
    157 
    158 static bool use_aggressive_subpel_search_method(MACROBLOCK *x,
    159                                                bool use_adaptive_subpel_search,
    160                                                bool fullpel_performed_well) {
    161  if (!use_adaptive_subpel_search) return false;
    162  const int qband = x->qindex >> (QINDEX_BITS - 2);
    163  assert(qband < 4);
    164  if ((qband > 0) && (fullpel_performed_well ||
    165                      (x->content_state_sb.source_sad_nonrd <= kLowSad) ||
    166                      (x->source_variance < 100)))
    167    return true;
    168  return false;
    169 }
    170 
    171 /*!\brief Runs Motion Estimation for a specific block and specific ref frame.
    172 *
    173 * \ingroup nonrd_mode_search
    174 * \callgraph
    175 * \callergraph
    176 * Finds the best Motion Vector by running Motion Estimation for a specific
    177 * block and a specific reference frame. Exits early if RDCost of Full Pel part
    178 * exceeds best RD Cost fund so far
    179 * \param[in]    cpi                      Top-level encoder structure
    180 * \param[in]    x                        Pointer to structure holding all the
    181 *                                        data for the current macroblock
    182 * \param[in]    bsize                    Current block size
    183 * \param[in]    tmp_mv                   Pointer to best found New MV
    184 * \param[in]    rate_mv                  Pointer to Rate of the best new MV
    185 * \param[in]    best_rd_sofar            RD Cost of the best mode found so far
    186 * \param[in]    use_base_mv              Flag, indicating that tmp_mv holds
    187 *                                        specific MV to start the search with
    188 *
    189 * \return Returns 0 if ME was terminated after Full Pel Search because too
    190 * high RD Cost. Otherwise returns 1. Best New MV is placed into \c tmp_mv.
    191 * Rate estimation for this vector is placed to \c rate_mv
    192 */
    193 static int combined_motion_search(AV1_COMP *cpi, MACROBLOCK *x,
    194                                  BLOCK_SIZE bsize, int_mv *tmp_mv,
    195                                  int *rate_mv, int64_t best_rd_sofar,
    196                                  int use_base_mv) {
    197  MACROBLOCKD *xd = &x->e_mbd;
    198  const AV1_COMMON *cm = &cpi->common;
    199  const SPEED_FEATURES *sf = &cpi->sf;
    200  MB_MODE_INFO *mi = xd->mi[0];
    201  int step_param = (sf->rt_sf.fullpel_search_step_param)
    202                       ? sf->rt_sf.fullpel_search_step_param
    203                       : cpi->mv_search_params.mv_step_param;
    204  FULLPEL_MV start_mv;
    205  const int ref = mi->ref_frame[0];
    206  const MV ref_mv = av1_get_ref_mv(x, mi->ref_mv_idx).as_mv;
    207  MV center_mv;
    208  int dis;
    209  int rv = 0;
    210  int cost_list[5];
    211  int search_subpel = 1;
    212 
    213  start_mv = get_fullmv_from_mv(&ref_mv);
    214 
    215  if (!use_base_mv)
    216    center_mv = ref_mv;
    217  else
    218    center_mv = tmp_mv->as_mv;
    219 
    220  const SEARCH_METHODS search_method =
    221      av1_get_default_mv_search_method(x, &cpi->sf.mv_sf, bsize);
    222  const search_site_config *src_search_sites =
    223      av1_get_search_site_config(cpi, x, search_method);
    224  FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
    225  FULLPEL_MV_STATS best_mv_stats;
    226  av1_make_default_fullpel_ms_params(&full_ms_params, cpi, x, bsize, &center_mv,
    227                                     start_mv, src_search_sites, search_method,
    228                                     /*fine_search_interval=*/0);
    229 
    230  const unsigned int full_var_rd = av1_full_pixel_search(
    231      start_mv, &full_ms_params, step_param, cond_cost_list(cpi, cost_list),
    232      &tmp_mv->as_fullmv, &best_mv_stats, NULL);
    233 
    234  // calculate the bit cost on motion vector
    235  MV mvp_full = get_mv_from_fullmv(&tmp_mv->as_fullmv);
    236 
    237  *rate_mv = av1_mv_bit_cost(&mvp_full, &ref_mv, x->mv_costs->nmv_joint_cost,
    238                             x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
    239 
    240  // TODO(kyslov) Account for Rate Mode!
    241  rv = !(RDCOST(x->rdmult, (*rate_mv), 0) > best_rd_sofar);
    242 
    243  if (rv && search_subpel) {
    244    SUBPEL_MOTION_SEARCH_PARAMS ms_params;
    245    av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv,
    246                                      cost_list);
    247    const bool fullpel_performed_well =
    248        (bsize == BLOCK_64X64 && full_var_rd * 40 < 62267 * 7) ||
    249        (bsize == BLOCK_32X32 && full_var_rd * 8 < 42380) ||
    250        (bsize == BLOCK_16X16 && full_var_rd * 8 < 10127);
    251    if (sf->rt_sf.reduce_mv_pel_precision_highmotion ||
    252        sf->rt_sf.reduce_mv_pel_precision_lowcomplex)
    253      ms_params.forced_stop = subpel_select(cpi, x, bsize, tmp_mv, ref_mv,
    254                                            start_mv, fullpel_performed_well);
    255 
    256    MV subpel_start_mv = get_mv_from_fullmv(&tmp_mv->as_fullmv);
    257    assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, subpel_start_mv));
    258    // adaptively downgrade subpel search method based on block properties
    259    if (use_aggressive_subpel_search_method(
    260            x, sf->rt_sf.use_adaptive_subpel_search, fullpel_performed_well))
    261      av1_find_best_sub_pixel_tree_pruned_more(
    262          xd, cm, &ms_params, subpel_start_mv, &best_mv_stats, &tmp_mv->as_mv,
    263          &dis, &x->pred_sse[ref], NULL);
    264    else
    265      cpi->mv_search_params.find_fractional_mv_step(
    266          xd, cm, &ms_params, subpel_start_mv, &best_mv_stats, &tmp_mv->as_mv,
    267          &dis, &x->pred_sse[ref], NULL);
    268    *rate_mv =
    269        av1_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->mv_costs->nmv_joint_cost,
    270                        x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
    271  }
    272  // The final MV can not be equal to the reference MV as this will trigger an
    273  // assert later. This can happen if both NEAREST and NEAR modes were skipped.
    274  rv = (tmp_mv->as_mv.col != ref_mv.col || tmp_mv->as_mv.row != ref_mv.row);
    275  return rv;
    276 }
    277 
    278 /*!\brief Searches for the best New Motion Vector.
    279 *
    280 * \ingroup nonrd_mode_search
    281 * \callgraph
    282 * \callergraph
    283 * Finds the best Motion Vector by doing Motion Estimation. Uses reduced
    284 * complexity ME for non-LAST frames or calls \c combined_motion_search
    285 * for LAST reference frame
    286 * \param[in]    cpi                      Top-level encoder structure
    287 * \param[in]    x                        Pointer to structure holding all the
    288 *                                        data for the current macroblock
    289 * \param[in]    frame_mv                 Array that holds MVs for all modes
    290 *                                        and ref frames
    291 * \param[in]    ref_frame                Reference frame for which to find
    292 *                                        the best New MVs
    293 * \param[in]    gf_temporal_ref          Flag, indicating temporal reference
    294 *                                        for GOLDEN frame
    295 * \param[in]    bsize                    Current block size
    296 * \param[in]    mi_row                   Row index in 4x4 units
    297 * \param[in]    mi_col                   Column index in 4x4 units
    298 * \param[in]    rate_mv                  Pointer to Rate of the best new MV
    299 * \param[in]    best_rdc                 Pointer to the RD Cost for the best
    300 *                                        mode found so far
    301 *
    302 * \return Returns -1 if the search was not done, otherwise returns 0.
    303 * Best New MV is placed into \c frame_mv array, Rate estimation for this
    304 * vector is placed to \c rate_mv
    305 */
    306 static int search_new_mv(AV1_COMP *cpi, MACROBLOCK *x,
    307                         int_mv frame_mv[][REF_FRAMES],
    308                         MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
    309                         BLOCK_SIZE bsize, int mi_row, int mi_col, int *rate_mv,
    310                         RD_STATS *best_rdc) {
    311  MACROBLOCKD *const xd = &x->e_mbd;
    312  MB_MODE_INFO *const mi = xd->mi[0];
    313  AV1_COMMON *cm = &cpi->common;
    314  int_mv *this_ref_frm_newmv = &frame_mv[NEWMV][ref_frame];
    315  unsigned int y_sad_zero;
    316  if (ref_frame > LAST_FRAME && cpi->oxcf.rc_cfg.mode == AOM_CBR &&
    317      (cpi->ref_frame_flags & AOM_LAST_FLAG) && gf_temporal_ref) {
    318    int tmp_sad;
    319    int dis;
    320 
    321    if (bsize < BLOCK_16X16) return -1;
    322 
    323    int me_search_size_col = block_size_wide[bsize] >> 1;
    324    int me_search_size_row = block_size_high[bsize] >> 1;
    325    MV ref_mv = av1_get_ref_mv(x, 0).as_mv;
    326    tmp_sad = av1_int_pro_motion_estimation(
    327        cpi, x, bsize, mi_row, mi_col, &ref_mv, &y_sad_zero, me_search_size_col,
    328        me_search_size_row);
    329 
    330    if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
    331 
    332    this_ref_frm_newmv->as_int = mi->mv[0].as_int;
    333    int_mv best_mv = mi->mv[0];
    334    best_mv.as_mv.row >>= 3;
    335    best_mv.as_mv.col >>= 3;
    336    this_ref_frm_newmv->as_mv.row >>= 3;
    337    this_ref_frm_newmv->as_mv.col >>= 3;
    338 
    339    SUBPEL_MOTION_SEARCH_PARAMS ms_params;
    340    av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv, NULL);
    341    if (cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion ||
    342        cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex) {
    343      FULLPEL_MV start_mv = { .row = 0, .col = 0 };
    344      ms_params.forced_stop =
    345          subpel_select(cpi, x, bsize, &best_mv, ref_mv, start_mv, false);
    346    }
    347    MV start_mv = get_mv_from_fullmv(&best_mv.as_fullmv);
    348    assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, start_mv));
    349    cpi->mv_search_params.find_fractional_mv_step(
    350        xd, cm, &ms_params, start_mv, NULL, &best_mv.as_mv, &dis,
    351        &x->pred_sse[ref_frame], NULL);
    352    this_ref_frm_newmv->as_int = best_mv.as_int;
    353 
    354    // When NEWMV is same as ref_mv from the drl, it is preferred to code the
    355    // MV as NEARESTMV or NEARMV. In this case, NEWMV needs to be skipped to
    356    // avoid an assert failure at a later stage. The scenario can occur if
    357    // NEARESTMV was not evaluated for ALTREF.
    358    if (this_ref_frm_newmv->as_mv.col == ref_mv.col &&
    359        this_ref_frm_newmv->as_mv.row == ref_mv.row)
    360      return -1;
    361 
    362    *rate_mv = av1_mv_bit_cost(&this_ref_frm_newmv->as_mv, &ref_mv,
    363                               x->mv_costs->nmv_joint_cost,
    364                               x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
    365  } else if (!combined_motion_search(cpi, x, bsize, &frame_mv[NEWMV][ref_frame],
    366                                     rate_mv, best_rdc->rdcost, 0)) {
    367    return -1;
    368  }
    369 
    370  return 0;
    371 }
    372 
    373 static void estimate_single_ref_frame_costs(const AV1_COMMON *cm,
    374                                            const MACROBLOCKD *xd,
    375                                            const ModeCosts *mode_costs,
    376                                            int segment_id, BLOCK_SIZE bsize,
    377                                            unsigned int *ref_costs_single) {
    378  int seg_ref_active =
    379      segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
    380  if (seg_ref_active) {
    381    memset(ref_costs_single, 0, REF_FRAMES * sizeof(*ref_costs_single));
    382  } else {
    383    int intra_inter_ctx = av1_get_intra_inter_context(xd);
    384    ref_costs_single[INTRA_FRAME] =
    385        mode_costs->intra_inter_cost[intra_inter_ctx][0];
    386    unsigned int base_cost = mode_costs->intra_inter_cost[intra_inter_ctx][1];
    387    if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT &&
    388        is_comp_ref_allowed(bsize)) {
    389      const int comp_ref_type_ctx = av1_get_comp_reference_type_context(xd);
    390      base_cost += mode_costs->comp_ref_type_cost[comp_ref_type_ctx][1];
    391    }
    392    ref_costs_single[LAST_FRAME] = base_cost;
    393    ref_costs_single[GOLDEN_FRAME] = base_cost;
    394    ref_costs_single[ALTREF_FRAME] = base_cost;
    395    // add cost for last, golden, altref
    396    ref_costs_single[LAST_FRAME] += mode_costs->single_ref_cost[0][0][0];
    397    ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][0][1];
    398    ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][1][0];
    399    ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][0][1];
    400    ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][2][0];
    401  }
    402 }
    403 
    404 static inline void set_force_skip_flag(const AV1_COMP *const cpi,
    405                                       MACROBLOCK *const x, unsigned int sse,
    406                                       int *force_skip) {
    407  if (x->txfm_search_params.tx_mode_search_type == TX_MODE_SELECT &&
    408      cpi->sf.rt_sf.tx_size_level_based_on_qstep &&
    409      cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
    410    const int qstep = x->plane[AOM_PLANE_Y].dequant_QTX[1] >> (x->e_mbd.bd - 5);
    411    const unsigned int qstep_sq = qstep * qstep;
    412    // If the sse is low for low source variance blocks, mark those as
    413    // transform skip.
    414    // Note: Though qstep_sq is based on ac qstep, the threshold is kept
    415    // low so that reliable early estimate of tx skip can be obtained
    416    // through its comparison with sse.
    417    if (sse < qstep_sq && x->source_variance < qstep_sq &&
    418        x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
    419        x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 0)
    420      *force_skip = 1;
    421  }
    422 }
    423 
    424 #define CAP_TX_SIZE_FOR_BSIZE_GT32(tx_mode_search_type, bsize) \
    425  (((tx_mode_search_type) != ONLY_4X4 && (bsize) > BLOCK_32X32) ? true : false)
    426 #define TX_SIZE_FOR_BSIZE_GT32 (TX_16X16)
    427 
    428 static TX_SIZE calculate_tx_size(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
    429                                 MACROBLOCK *const x, unsigned int var,
    430                                 unsigned int sse, int *force_skip) {
    431  MACROBLOCKD *const xd = &x->e_mbd;
    432  TX_SIZE tx_size;
    433  const TxfmSearchParams *txfm_params = &x->txfm_search_params;
    434  if (txfm_params->tx_mode_search_type == TX_MODE_SELECT) {
    435    int multiplier = 8;
    436    unsigned int var_thresh = 0;
    437    unsigned int is_high_var = 1;
    438    // Use quantizer based thresholds to determine transform size.
    439    if (cpi->sf.rt_sf.tx_size_level_based_on_qstep) {
    440      const int qband = x->qindex >> (QINDEX_BITS - 2);
    441      const int mult[4] = { 8, 7, 6, 5 };
    442      assert(qband < 4);
    443      multiplier = mult[qband];
    444      const int qstep = x->plane[AOM_PLANE_Y].dequant_QTX[1] >> (xd->bd - 5);
    445      const unsigned int qstep_sq = qstep * qstep;
    446      var_thresh = qstep_sq * 2;
    447      if (cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
    448        // If the sse is low for low source variance blocks, mark those as
    449        // transform skip.
    450        // Note: Though qstep_sq is based on ac qstep, the threshold is kept
    451        // low so that reliable early estimate of tx skip can be obtained
    452        // through its comparison with sse.
    453        if (sse < qstep_sq && x->source_variance < qstep_sq &&
    454            x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
    455            x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 0)
    456          *force_skip = 1;
    457        // Further lower transform size based on aq mode only if residual
    458        // variance is high.
    459        is_high_var = (var >= var_thresh);
    460      }
    461    }
    462    // Choose larger transform size for blocks where dc component is dominant or
    463    // the ac component is low.
    464    if (sse > ((var * multiplier) >> 2) || (var < var_thresh))
    465      tx_size =
    466          AOMMIN(max_txsize_lookup[bsize],
    467                 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
    468    else
    469      tx_size = TX_8X8;
    470 
    471    if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ &&
    472        cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) && is_high_var)
    473      tx_size = TX_8X8;
    474    else if (tx_size > TX_16X16)
    475      tx_size = TX_16X16;
    476  } else {
    477    tx_size =
    478        AOMMIN(max_txsize_lookup[bsize],
    479               tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
    480  }
    481 
    482  if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize))
    483    tx_size = TX_SIZE_FOR_BSIZE_GT32;
    484 
    485  return AOMMIN(tx_size, TX_16X16);
    486 }
    487 
    488 static void block_variance(const uint8_t *src, int src_stride,
    489                           const uint8_t *ref, int ref_stride, int w, int h,
    490                           unsigned int *sse, int *sum, int block_size,
    491                           uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
    492  int k = 0;
    493  *sse = 0;
    494  *sum = 0;
    495 
    496  // This function is called for block sizes >= BLOCK_32x32. As per the design
    497  // the aom_get_var_sse_sum_8x8_quad() processes four 8x8 blocks (in a 8x32)
    498  // per call. Hence the width and height of the block need to be at least 8 and
    499  // 32 samples respectively.
    500  assert(w >= 32);
    501  assert(h >= 8);
    502  for (int row = 0; row < h; row += block_size) {
    503    for (int col = 0; col < w; col += 32) {
    504      aom_get_var_sse_sum_8x8_quad(src + src_stride * row + col, src_stride,
    505                                   ref + ref_stride * row + col, ref_stride,
    506                                   &sse8x8[k], &sum8x8[k], sse, sum,
    507                                   &var8x8[k]);
    508      k += 4;
    509    }
    510  }
    511 }
    512 
    513 static void block_variance_16x16_dual(const uint8_t *src, int src_stride,
    514                                      const uint8_t *ref, int ref_stride, int w,
    515                                      int h, unsigned int *sse, int *sum,
    516                                      int block_size, uint32_t *sse16x16,
    517                                      uint32_t *var16x16) {
    518  int k = 0;
    519  *sse = 0;
    520  *sum = 0;
    521  // This function is called for block sizes >= BLOCK_32x32. As per the design
    522  // the aom_get_var_sse_sum_16x16_dual() processes four 16x16 blocks (in a
    523  // 16x32) per call. Hence the width and height of the block need to be at
    524  // least 16 and 32 samples respectively.
    525  assert(w >= 32);
    526  assert(h >= 16);
    527  for (int row = 0; row < h; row += block_size) {
    528    for (int col = 0; col < w; col += 32) {
    529      aom_get_var_sse_sum_16x16_dual(src + src_stride * row + col, src_stride,
    530                                     ref + ref_stride * row + col, ref_stride,
    531                                     &sse16x16[k], sse, sum, &var16x16[k]);
    532      k += 2;
    533    }
    534  }
    535 }
    536 
    537 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
    538                               unsigned int *sse_i, int *sum_i,
    539                               unsigned int *var_o, unsigned int *sse_o,
    540                               int *sum_o) {
    541  const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
    542  const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
    543  const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
    544  int row, col, k = 0;
    545 
    546  for (row = 0; row < nh; row += 2) {
    547    for (col = 0; col < nw; col += 2) {
    548      sse_o[k] = sse_i[row * nw + col] + sse_i[row * nw + col + 1] +
    549                 sse_i[(row + 1) * nw + col] + sse_i[(row + 1) * nw + col + 1];
    550      sum_o[k] = sum_i[row * nw + col] + sum_i[row * nw + col + 1] +
    551                 sum_i[(row + 1) * nw + col] + sum_i[(row + 1) * nw + col + 1];
    552      var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
    553                                       (b_width_log2_lookup[unit_size] +
    554                                        b_height_log2_lookup[unit_size] + 6));
    555      k++;
    556    }
    557  }
    558 }
    559 
    560 // Adjust the ac_thr according to speed, width, height and normalized sum
    561 static int ac_thr_factor(int speed, int width, int height, int norm_sum) {
    562  if (speed >= 8 && norm_sum < 5) {
    563    if (width <= 640 && height <= 480)
    564      return 4;
    565    else
    566      return 2;
    567  }
    568  return 1;
    569 }
    570 
    571 // Sets early_term flag based on chroma planes prediction
    572 static inline void set_early_term_based_on_uv_plane(
    573    AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, MACROBLOCKD *xd, int mi_row,
    574    int mi_col, int *early_term, int num_blk, const unsigned int *sse_tx,
    575    const unsigned int *var_tx, int sum, unsigned int var, unsigned int sse) {
    576  AV1_COMMON *const cm = &cpi->common;
    577  struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
    578  const uint32_t dc_quant = p->dequant_QTX[0];
    579  const uint32_t ac_quant = p->dequant_QTX[1];
    580  int64_t dc_thr = dc_quant * dc_quant >> 6;
    581  int64_t ac_thr = ac_quant * ac_quant >> 6;
    582  const int bw = b_width_log2_lookup[bsize];
    583  const int bh = b_height_log2_lookup[bsize];
    584  int ac_test = 1;
    585  int dc_test = 1;
    586  const int norm_sum = abs(sum) >> (bw + bh);
    587 
    588 #if CONFIG_AV1_TEMPORAL_DENOISING
    589  if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
    590      cpi->oxcf.speed > 5)
    591    ac_thr = av1_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
    592                                     norm_sum, cpi->svc.temporal_layer_id);
    593  else
    594    ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
    595 #else
    596  ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
    597 
    598 #endif
    599 
    600  if (cpi->sf.rt_sf.increase_source_sad_thresh) {
    601    dc_thr = dc_thr << 1;
    602    ac_thr = ac_thr << 2;
    603  }
    604 
    605  if (cpi->common.width * cpi->common.height >= 1280 * 720 &&
    606      cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN &&
    607      x->content_state_sb.source_sad_nonrd > kLowSad &&
    608      (sse >> (bw + bh)) > 1000) {
    609    dc_thr = dc_thr >> 4;
    610    ac_thr = ac_thr >> 4;
    611  }
    612 
    613  for (int k = 0; k < num_blk; k++) {
    614    // Check if all ac coefficients can be quantized to zero.
    615    if (!(var_tx[k] < ac_thr || var == 0)) {
    616      ac_test = 0;
    617      break;
    618    }
    619    // Check if dc coefficient can be quantized to zero.
    620    if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
    621      dc_test = 0;
    622      break;
    623    }
    624  }
    625 
    626  // Check if chroma can be skipped based on ac and dc test flags.
    627  if (ac_test && dc_test) {
    628    int skip_uv[2] = { 0 };
    629    unsigned int var_uv[2];
    630    unsigned int sse_uv[2];
    631    // Transform skipping test in UV planes.
    632    for (int plane = AOM_PLANE_U; plane <= AOM_PLANE_V; plane++) {
    633      int j = plane - 1;
    634      skip_uv[j] = 1;
    635      if (x->color_sensitivity[COLOR_SENS_IDX(plane)]) {
    636        skip_uv[j] = 0;
    637        struct macroblock_plane *const puv = &x->plane[plane];
    638        struct macroblockd_plane *const puvd = &xd->plane[plane];
    639        const BLOCK_SIZE uv_bsize = get_plane_block_size(
    640            bsize, puvd->subsampling_x, puvd->subsampling_y);
    641        // Adjust these thresholds for UV.
    642        const int shift_ac = cpi->sf.rt_sf.increase_source_sad_thresh ? 5 : 3;
    643        const int shift_dc = cpi->sf.rt_sf.increase_source_sad_thresh ? 4 : 3;
    644        const int64_t uv_dc_thr =
    645            (puv->dequant_QTX[0] * puv->dequant_QTX[0]) >> shift_dc;
    646        const int64_t uv_ac_thr =
    647            (puv->dequant_QTX[1] * puv->dequant_QTX[1]) >> shift_ac;
    648        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
    649                                      plane, plane);
    650        var_uv[j] = cpi->ppi->fn_ptr[uv_bsize].vf(puv->src.buf, puv->src.stride,
    651                                                  puvd->dst.buf,
    652                                                  puvd->dst.stride, &sse_uv[j]);
    653        if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
    654            (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
    655          skip_uv[j] = 1;
    656        else
    657          break;
    658      }
    659    }
    660    if (skip_uv[0] & skip_uv[1]) {
    661      *early_term = 1;
    662    }
    663  }
    664 }
    665 
    666 static inline void calc_rate_dist_block_param(AV1_COMP *cpi, MACROBLOCK *x,
    667                                              RD_STATS *rd_stats,
    668                                              int calculate_rd, int *early_term,
    669                                              BLOCK_SIZE bsize,
    670                                              unsigned int sse) {
    671  if (calculate_rd) {
    672    if (!*early_term) {
    673      const int bw = block_size_wide[bsize];
    674      const int bh = block_size_high[bsize];
    675 
    676      model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, rd_stats->sse, bw * bh,
    677                            &rd_stats->rate, &rd_stats->dist);
    678    }
    679 
    680    if (*early_term) {
    681      rd_stats->rate = 0;
    682      rd_stats->dist = sse << 4;
    683    }
    684  }
    685 }
    686 
    687 static void model_skip_for_sb_y_large_64(AV1_COMP *cpi, BLOCK_SIZE bsize,
    688                                         int mi_row, int mi_col, MACROBLOCK *x,
    689                                         MACROBLOCKD *xd, RD_STATS *rd_stats,
    690                                         int *early_term, int calculate_rd,
    691                                         int64_t best_sse,
    692                                         unsigned int *var_output,
    693                                         unsigned int var_prune_threshold) {
    694  // Note our transform coeffs are 8 times an orthogonal transform.
    695  // Hence quantizer step is also 8 times. To get effective quantizer
    696  // we need to divide by 8 before sending to modeling function.
    697  unsigned int sse;
    698  struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
    699  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
    700  int test_skip = 1;
    701  unsigned int var;
    702  int sum;
    703  const int bw = b_width_log2_lookup[bsize];
    704  const int bh = b_height_log2_lookup[bsize];
    705  unsigned int sse16x16[64] = { 0 };
    706  unsigned int var16x16[64] = { 0 };
    707  assert(xd->mi[0]->tx_size == TX_16X16);
    708  assert(bsize > BLOCK_32X32);
    709 
    710  // Calculate variance for whole partition, and also save 16x16 blocks'
    711  // variance to be used in following transform skipping test.
    712  block_variance_16x16_dual(p->src.buf, p->src.stride, pd->dst.buf,
    713                            pd->dst.stride, 4 << bw, 4 << bh, &sse, &sum, 16,
    714                            sse16x16, var16x16);
    715 
    716  var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
    717  if (var_output) {
    718    *var_output = var;
    719    if (*var_output > var_prune_threshold) {
    720      return;
    721    }
    722  }
    723 
    724  rd_stats->sse = sse;
    725  // Skipping test
    726  *early_term = 0;
    727  set_force_skip_flag(cpi, x, sse, early_term);
    728  // The code below for setting skip flag assumes transform size of at least
    729  // 8x8, so force this lower limit on transform.
    730  MB_MODE_INFO *const mi = xd->mi[0];
    731  if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
    732      early_term_inter_search_with_sse(
    733          cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
    734          mi->mode))
    735    test_skip = 0;
    736 
    737  if (*early_term) test_skip = 0;
    738 
    739  // Evaluate if the partition block is a skippable block in Y plane.
    740  if (test_skip) {
    741    const unsigned int *sse_tx = sse16x16;
    742    const unsigned int *var_tx = var16x16;
    743    const unsigned int num_block = (1 << (bw + bh - 2)) >> 2;
    744    set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
    745                                     early_term, num_block, sse_tx, var_tx, sum,
    746                                     var, sse);
    747  }
    748  calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
    749                             sse);
    750 }
    751 
    752 static void model_skip_for_sb_y_large(AV1_COMP *cpi, BLOCK_SIZE bsize,
    753                                      int mi_row, int mi_col, MACROBLOCK *x,
    754                                      MACROBLOCKD *xd, RD_STATS *rd_stats,
    755                                      int *early_term, int calculate_rd,
    756                                      int64_t best_sse,
    757                                      unsigned int *var_output,
    758                                      unsigned int var_prune_threshold) {
    759  if (x->force_zeromv_skip_for_blk) {
    760    *early_term = 1;
    761    rd_stats->rate = 0;
    762    rd_stats->dist = 0;
    763    rd_stats->sse = 0;
    764    return;
    765  }
    766 
    767  // For block sizes greater than 32x32, the transform size is always 16x16.
    768  // This function avoids calling calculate_variance() for tx_size 16x16 cases
    769  // by directly populating variance at tx_size level from
    770  // block_variance_16x16_dual() function.
    771  const TxfmSearchParams *txfm_params = &x->txfm_search_params;
    772  if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize)) {
    773    xd->mi[0]->tx_size = TX_SIZE_FOR_BSIZE_GT32;
    774    model_skip_for_sb_y_large_64(cpi, bsize, mi_row, mi_col, x, xd, rd_stats,
    775                                 early_term, calculate_rd, best_sse, var_output,
    776                                 var_prune_threshold);
    777    return;
    778  }
    779 
    780  // Note our transform coeffs are 8 times an orthogonal transform.
    781  // Hence quantizer step is also 8 times. To get effective quantizer
    782  // we need to divide by 8 before sending to modeling function.
    783  unsigned int sse;
    784  struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
    785  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
    786  int test_skip = 1;
    787  unsigned int var;
    788  int sum;
    789 
    790  const int bw = b_width_log2_lookup[bsize];
    791  const int bh = b_height_log2_lookup[bsize];
    792  unsigned int sse8x8[256] = { 0 };
    793  int sum8x8[256] = { 0 };
    794  unsigned int var8x8[256] = { 0 };
    795  TX_SIZE tx_size;
    796 
    797  // Calculate variance for whole partition, and also save 8x8 blocks' variance
    798  // to be used in following transform skipping test.
    799  block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
    800                 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
    801  var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
    802  if (var_output) {
    803    *var_output = var;
    804    if (*var_output > var_prune_threshold) {
    805      return;
    806    }
    807  }
    808 
    809  rd_stats->sse = sse;
    810  // Skipping test
    811  *early_term = 0;
    812  tx_size = calculate_tx_size(cpi, bsize, x, var, sse, early_term);
    813  assert(tx_size <= TX_16X16);
    814  // The code below for setting skip flag assumes transform size of at least
    815  // 8x8, so force this lower limit on transform.
    816  if (tx_size < TX_8X8) tx_size = TX_8X8;
    817  xd->mi[0]->tx_size = tx_size;
    818 
    819  MB_MODE_INFO *const mi = xd->mi[0];
    820  if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
    821      early_term_inter_search_with_sse(
    822          cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
    823          mi->mode))
    824    test_skip = 0;
    825 
    826  if (*early_term) test_skip = 0;
    827 
    828  // Evaluate if the partition block is a skippable block in Y plane.
    829  if (test_skip) {
    830    unsigned int sse16x16[64] = { 0 };
    831    int sum16x16[64] = { 0 };
    832    unsigned int var16x16[64] = { 0 };
    833    const unsigned int *sse_tx = sse8x8;
    834    const unsigned int *var_tx = var8x8;
    835    unsigned int num_blks = 1 << (bw + bh - 2);
    836 
    837    if (tx_size >= TX_16X16) {
    838      calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
    839                         sum16x16);
    840      sse_tx = sse16x16;
    841      var_tx = var16x16;
    842      num_blks = num_blks >> 2;
    843    }
    844    set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
    845                                     early_term, num_blks, sse_tx, var_tx, sum,
    846                                     var, sse);
    847  }
    848  calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
    849                             sse);
    850 }
    851 
    852 static void model_rd_for_sb_y(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
    853                              MACROBLOCK *x, MACROBLOCKD *xd,
    854                              RD_STATS *rd_stats, unsigned int *var_out,
    855                              int calculate_rd, int *early_term) {
    856  if (x->force_zeromv_skip_for_blk && early_term != NULL) {
    857    *early_term = 1;
    858    rd_stats->rate = 0;
    859    rd_stats->dist = 0;
    860    rd_stats->sse = 0;
    861  }
    862 
    863  // Note our transform coeffs are 8 times an orthogonal transform.
    864  // Hence quantizer step is also 8 times. To get effective quantizer
    865  // we need to divide by 8 before sending to modeling function.
    866  const int ref = xd->mi[0]->ref_frame[0];
    867 
    868  assert(bsize < BLOCK_SIZES_ALL);
    869 
    870  struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
    871  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
    872  unsigned int sse;
    873  int rate;
    874  int64_t dist;
    875 
    876  unsigned int var = cpi->ppi->fn_ptr[bsize].vf(
    877      p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse);
    878  int force_skip = 0;
    879  xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, x, var, sse, &force_skip);
    880  if (var_out) {
    881    *var_out = var;
    882  }
    883 
    884  if (calculate_rd && (!force_skip || ref == INTRA_FRAME)) {
    885    const int bwide = block_size_wide[bsize];
    886    const int bhigh = block_size_high[bsize];
    887    model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, sse, bwide * bhigh, &rate,
    888                          &dist);
    889  } else {
    890    rate = INT_MAX;  // this will be overwritten later with av1_block_yrd
    891    dist = INT_MAX;
    892  }
    893  rd_stats->sse = sse;
    894  x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
    895 
    896  if (force_skip && ref > INTRA_FRAME) {
    897    rate = 0;
    898    dist = (int64_t)sse << 4;
    899  }
    900 
    901  assert(rate >= 0);
    902 
    903  rd_stats->skip_txfm = (rate == 0);
    904  rate = AOMMIN(rate, INT_MAX);
    905  rd_stats->rate = rate;
    906  rd_stats->dist = dist;
    907 }
    908 
    909 static inline int get_drl_cost(PREDICTION_MODE this_mode, int ref_mv_idx,
    910                               const MB_MODE_INFO_EXT *mbmi_ext,
    911                               const int (*const drl_mode_cost0)[2],
    912                               int8_t ref_frame_type) {
    913  int cost = 0;
    914  if (this_mode == NEWMV || this_mode == NEW_NEWMV) {
    915    for (int idx = 0; idx < 2; ++idx) {
    916      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
    917        uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
    918        cost += drl_mode_cost0[drl_ctx][ref_mv_idx != idx];
    919        if (ref_mv_idx == idx) return cost;
    920      }
    921    }
    922    return cost;
    923  }
    924 
    925  if (have_nearmv_in_inter_mode(this_mode)) {
    926    for (int idx = 1; idx < 3; ++idx) {
    927      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
    928        uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
    929        cost += drl_mode_cost0[drl_ctx][ref_mv_idx != (idx - 1)];
    930        if (ref_mv_idx == (idx - 1)) return cost;
    931      }
    932    }
    933    return cost;
    934  }
    935  return cost;
    936 }
    937 
    938 static int cost_mv_ref(const ModeCosts *const mode_costs, PREDICTION_MODE mode,
    939                       int16_t mode_context) {
    940  if (is_inter_compound_mode(mode)) {
    941    return mode_costs
    942        ->inter_compound_mode_cost[mode_context][INTER_COMPOUND_OFFSET(mode)];
    943  }
    944 
    945  int mode_cost = 0;
    946  int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
    947 
    948  assert(is_inter_mode(mode));
    949 
    950  if (mode == NEWMV) {
    951    mode_cost = mode_costs->newmv_mode_cost[mode_ctx][0];
    952    return mode_cost;
    953  } else {
    954    mode_cost = mode_costs->newmv_mode_cost[mode_ctx][1];
    955    mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
    956 
    957    if (mode == GLOBALMV) {
    958      mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][0];
    959      return mode_cost;
    960    } else {
    961      mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][1];
    962      mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
    963      mode_cost += mode_costs->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
    964      return mode_cost;
    965    }
    966  }
    967 }
    968 
    969 static void newmv_diff_bias(MACROBLOCKD *xd, PREDICTION_MODE this_mode,
    970                            RD_STATS *this_rdc, BLOCK_SIZE bsize, int mv_row,
    971                            int mv_col, int speed, uint32_t spatial_variance,
    972                            CONTENT_STATE_SB content_state_sb) {
    973  // Bias against MVs associated with NEWMV mode that are very different from
    974  // top/left neighbors.
    975  if (this_mode == NEWMV) {
    976    int al_mv_average_row;
    977    int al_mv_average_col;
    978    int row_diff, col_diff;
    979    int above_mv_valid = 0;
    980    int left_mv_valid = 0;
    981    int above_row = INVALID_MV_ROW_COL, above_col = INVALID_MV_ROW_COL;
    982    int left_row = INVALID_MV_ROW_COL, left_col = INVALID_MV_ROW_COL;
    983    if (bsize >= BLOCK_64X64 && content_state_sb.source_sad_nonrd != kHighSad &&
    984        spatial_variance < 300 &&
    985        (mv_row > 16 || mv_row < -16 || mv_col > 16 || mv_col < -16)) {
    986      this_rdc->rdcost = this_rdc->rdcost << 2;
    987      return;
    988    }
    989    if (xd->above_mbmi) {
    990      above_mv_valid = xd->above_mbmi->mv[0].as_int != INVALID_MV;
    991      above_row = xd->above_mbmi->mv[0].as_mv.row;
    992      above_col = xd->above_mbmi->mv[0].as_mv.col;
    993    }
    994    if (xd->left_mbmi) {
    995      left_mv_valid = xd->left_mbmi->mv[0].as_int != INVALID_MV;
    996      left_row = xd->left_mbmi->mv[0].as_mv.row;
    997      left_col = xd->left_mbmi->mv[0].as_mv.col;
    998    }
    999    if (above_mv_valid && left_mv_valid) {
   1000      al_mv_average_row = (above_row + left_row + 1) >> 1;
   1001      al_mv_average_col = (above_col + left_col + 1) >> 1;
   1002    } else if (above_mv_valid) {
   1003      al_mv_average_row = above_row;
   1004      al_mv_average_col = above_col;
   1005    } else if (left_mv_valid) {
   1006      al_mv_average_row = left_row;
   1007      al_mv_average_col = left_col;
   1008    } else {
   1009      al_mv_average_row = al_mv_average_col = 0;
   1010    }
   1011    row_diff = al_mv_average_row - mv_row;
   1012    col_diff = al_mv_average_col - mv_col;
   1013    if (row_diff > 80 || row_diff < -80 || col_diff > 80 || col_diff < -80) {
   1014      if (bsize >= BLOCK_32X32)
   1015        this_rdc->rdcost = this_rdc->rdcost << 1;
   1016      else
   1017        this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
   1018    }
   1019  } else {
   1020    // Bias for speed >= 8 for low spatial variance.
   1021    if (speed >= 8 && spatial_variance < 150 &&
   1022        (mv_row > 64 || mv_row < -64 || mv_col > 64 || mv_col < -64))
   1023      this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
   1024  }
   1025 }
   1026 
   1027 static inline void update_thresh_freq_fact(AV1_COMP *cpi, MACROBLOCK *x,
   1028                                           BLOCK_SIZE bsize,
   1029                                           MV_REFERENCE_FRAME ref_frame,
   1030                                           THR_MODES best_mode_idx,
   1031                                           PREDICTION_MODE mode) {
   1032  const THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
   1033  const BLOCK_SIZE min_size = AOMMAX(bsize - 3, BLOCK_4X4);
   1034  const BLOCK_SIZE max_size = AOMMIN(bsize + 6, BLOCK_128X128);
   1035  for (BLOCK_SIZE bs = min_size; bs <= max_size; bs += 3) {
   1036    int *freq_fact = &x->thresh_freq_fact[bs][thr_mode_idx];
   1037    if (thr_mode_idx == best_mode_idx) {
   1038      *freq_fact -= (*freq_fact >> 4);
   1039    } else {
   1040      *freq_fact =
   1041          AOMMIN(*freq_fact + RD_THRESH_INC,
   1042                 cpi->sf.inter_sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
   1043    }
   1044  }
   1045 }
   1046 
   1047 #if CONFIG_AV1_TEMPORAL_DENOISING
   1048 static void av1_pickmode_ctx_den_update(
   1049    AV1_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
   1050    unsigned int ref_frame_cost[REF_FRAMES],
   1051    int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES], int reuse_inter_pred,
   1052    BEST_PICKMODE *bp) {
   1053  ctx_den->zero_last_cost_orig = zero_last_cost_orig;
   1054  ctx_den->ref_frame_cost = ref_frame_cost;
   1055  ctx_den->frame_mv = frame_mv;
   1056  ctx_den->reuse_inter_pred = reuse_inter_pred;
   1057  ctx_den->best_tx_size = bp->best_tx_size;
   1058  ctx_den->best_mode = bp->best_mode;
   1059  ctx_den->best_ref_frame = bp->best_ref_frame;
   1060  ctx_den->best_pred_filter = bp->best_pred_filter;
   1061  ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
   1062 }
   1063 
   1064 static void recheck_zeromv_after_denoising(
   1065    AV1_COMP *cpi, MB_MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
   1066    AV1_DENOISER_DECISION decision, AV1_PICKMODE_CTX_DEN *ctx_den,
   1067    struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_STATS *best_rdc,
   1068    BEST_PICKMODE *best_pickmode, BLOCK_SIZE bsize, int mi_row, int mi_col) {
   1069  // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
   1070  // denoised result. Only do this under noise conditions, and if rdcost of
   1071  // ZEROMV on original source is not significantly higher than rdcost of best
   1072  // mode.
   1073  if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
   1074      ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
   1075      ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
   1076       (ctx_den->best_ref_frame == GOLDEN_FRAME &&
   1077        cpi->svc.number_spatial_layers == 1 &&
   1078        decision == FILTER_ZEROMV_BLOCK))) {
   1079    // Check if we should pick ZEROMV on denoised signal.
   1080    AV1_COMMON *const cm = &cpi->common;
   1081    RD_STATS this_rdc;
   1082    const ModeCosts *mode_costs = &x->mode_costs;
   1083    TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   1084    MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
   1085 
   1086    mi->mode = GLOBALMV;
   1087    mi->ref_frame[0] = LAST_FRAME;
   1088    mi->ref_frame[1] = NONE_FRAME;
   1089    set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
   1090    mi->mv[0].as_int = 0;
   1091    mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
   1092    xd->plane[AOM_PLANE_Y].pre[0] = yv12_mb[LAST_FRAME][AOM_PLANE_Y];
   1093    av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
   1094    unsigned int var;
   1095    model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc, &var, 1, NULL);
   1096 
   1097    const int16_t mode_ctx =
   1098        av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
   1099    this_rdc.rate += cost_mv_ref(mode_costs, GLOBALMV, mode_ctx);
   1100 
   1101    this_rdc.rate += ctx_den->ref_frame_cost[LAST_FRAME];
   1102    this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
   1103    txfm_info->skip_txfm = this_rdc.skip_txfm;
   1104    // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
   1105    // is higher than best_ref mode (on original source).
   1106    if (this_rdc.rdcost > best_rdc->rdcost) {
   1107      this_rdc = *best_rdc;
   1108      mi->mode = best_pickmode->best_mode;
   1109      mi->ref_frame[0] = best_pickmode->best_ref_frame;
   1110      set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
   1111      mi->interp_filters = best_pickmode->best_pred_filter;
   1112      if (best_pickmode->best_ref_frame == INTRA_FRAME) {
   1113        mi->mv[0].as_int = INVALID_MV;
   1114      } else {
   1115        mi->mv[0].as_int = ctx_den
   1116                               ->frame_mv[best_pickmode->best_mode]
   1117                                         [best_pickmode->best_ref_frame]
   1118                               .as_int;
   1119        if (ctx_den->reuse_inter_pred) {
   1120          xd->plane[AOM_PLANE_Y].pre[0] = yv12_mb[GOLDEN_FRAME][AOM_PLANE_Y];
   1121          av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
   1122        }
   1123      }
   1124      mi->tx_size = best_pickmode->best_tx_size;
   1125      txfm_info->skip_txfm = best_pickmode->best_mode_skip_txfm;
   1126    } else {
   1127      ctx_den->best_ref_frame = LAST_FRAME;
   1128      *best_rdc = this_rdc;
   1129    }
   1130  }
   1131 }
   1132 #endif  // CONFIG_AV1_TEMPORAL_DENOISING
   1133 
   1134 /*!\brief Searches for the best interpolation filter
   1135 *
   1136 * \ingroup nonrd_mode_search
   1137 * \callgraph
   1138 * \callergraph
   1139 * Iterates through subset of possible interpolation filters (EIGHTTAP_REGULAR,
   1140 * EIGTHTAP_SMOOTH, MULTITAP_SHARP, depending on FILTER_SEARCH_SIZE) and selects
   1141 * the one that gives lowest RD cost. RD cost is calculated using curvfit model.
   1142 * Support for dual filters (different filters in the x & y directions) is
   1143 * allowed if sf.interp_sf.disable_dual_filter = 0.
   1144 *
   1145 * \param[in]    cpi                  Top-level encoder structure
   1146 * \param[in]    x                    Pointer to structure holding all the
   1147 *                                    data for the current macroblock
   1148 * \param[in]    this_rdc             Pointer to calculated RD Cost
   1149 * \param[in]    inter_pred_params_sr Pointer to structure holding parameters of
   1150                                      inter prediction for single reference
   1151 * \param[in]    mi_row               Row index in 4x4 units
   1152 * \param[in]    mi_col               Column index in 4x4 units
   1153 * \param[in]    tmp_buffer           Pointer to a temporary buffer for
   1154 *                                    prediction re-use
   1155 * \param[in]    bsize                Current block size
   1156 * \param[in]    reuse_inter_pred     Flag, indicating prediction re-use
   1157 * \param[out]   this_mode_pred       Pointer to store prediction buffer
   1158 *                                    for prediction re-use
   1159 * \param[out]   this_early_term      Flag, indicating that transform can be
   1160 *                                    skipped
   1161 * \param[out]   var                  The residue variance of the current
   1162 *                                    predictor.
   1163 * \param[in]    use_model_yrd_large  Flag, indicating special logic to handle
   1164 *                                    large blocks
   1165 * \param[in]    best_sse             Best sse so far.
   1166 * \param[in]    is_single_pred       Flag, indicating single mode.
   1167 *
   1168 * \remark Nothing is returned. Instead, calculated RD cost is placed to
   1169 * \c this_rdc and best filter is placed to \c mi->interp_filters. In case
   1170 * \c reuse_inter_pred flag is set, this function also outputs
   1171 * \c this_mode_pred. Also \c this_early_temp is set if transform can be
   1172 * skipped
   1173 */
   1174 static void search_filter_ref(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
   1175                              InterPredParams *inter_pred_params_sr, int mi_row,
   1176                              int mi_col, PRED_BUFFER *tmp_buffer,
   1177                              BLOCK_SIZE bsize, int reuse_inter_pred,
   1178                              PRED_BUFFER **this_mode_pred,
   1179                              int *this_early_term, unsigned int *var,
   1180                              int use_model_yrd_large, int64_t best_sse,
   1181                              int is_single_pred) {
   1182  AV1_COMMON *const cm = &cpi->common;
   1183  MACROBLOCKD *const xd = &x->e_mbd;
   1184  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
   1185  MB_MODE_INFO *const mi = xd->mi[0];
   1186  const int bw = block_size_wide[bsize];
   1187  int dim_factor =
   1188      (cpi->sf.interp_sf.disable_dual_filter == 0) ? FILTER_SEARCH_SIZE : 1;
   1189  RD_STATS pf_rd_stats[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
   1190  TX_SIZE pf_tx_size[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
   1191  PRED_BUFFER *current_pred = *this_mode_pred;
   1192  int best_skip = 0;
   1193  int best_early_term = 0;
   1194  int64_t best_cost = INT64_MAX;
   1195  int best_filter_index = -1;
   1196 
   1197  SubpelParams subpel_params;
   1198  // Initialize inter prediction params at mode level for single reference
   1199  // mode.
   1200  if (is_single_pred)
   1201    init_inter_mode_params(&mi->mv[0].as_mv, inter_pred_params_sr,
   1202                           &subpel_params, xd->block_ref_scale_factors[0],
   1203                           pd->pre->width, pd->pre->height);
   1204  for (int filter_idx = 0; filter_idx < FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE;
   1205       ++filter_idx) {
   1206    int64_t cost;
   1207    if (cpi->sf.interp_sf.disable_dual_filter &&
   1208        filters_ref_set[filter_idx].as_filters.x_filter !=
   1209            filters_ref_set[filter_idx].as_filters.y_filter)
   1210      continue;
   1211 
   1212    mi->interp_filters.as_int = filters_ref_set[filter_idx].as_int;
   1213    if (is_single_pred)
   1214      av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
   1215                                            &subpel_params);
   1216    else
   1217      av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   1218                                    AOM_PLANE_Y, AOM_PLANE_Y);
   1219    unsigned int curr_var = UINT_MAX;
   1220    if (use_model_yrd_large)
   1221      model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
   1222                                &pf_rd_stats[filter_idx], this_early_term, 1,
   1223                                best_sse, &curr_var, UINT_MAX);
   1224    else
   1225      model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[filter_idx], &curr_var,
   1226                        1, NULL);
   1227    pf_rd_stats[filter_idx].rate += av1_get_switchable_rate(
   1228        x, xd, cm->features.interp_filter, cm->seq_params->enable_dual_filter);
   1229    cost = RDCOST(x->rdmult, pf_rd_stats[filter_idx].rate,
   1230                  pf_rd_stats[filter_idx].dist);
   1231    pf_tx_size[filter_idx] = mi->tx_size;
   1232    if (cost < best_cost) {
   1233      *var = curr_var;
   1234      best_filter_index = filter_idx;
   1235      best_cost = cost;
   1236      best_skip = pf_rd_stats[filter_idx].skip_txfm;
   1237      best_early_term = *this_early_term;
   1238      if (reuse_inter_pred) {
   1239        if (*this_mode_pred != current_pred) {
   1240          free_pred_buffer(*this_mode_pred);
   1241          *this_mode_pred = current_pred;
   1242        }
   1243        current_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
   1244        pd->dst.buf = current_pred->data;
   1245        pd->dst.stride = bw;
   1246      }
   1247    }
   1248  }
   1249  assert(best_filter_index >= 0 &&
   1250         best_filter_index < dim_factor * FILTER_SEARCH_SIZE);
   1251  if (reuse_inter_pred && *this_mode_pred != current_pred)
   1252    free_pred_buffer(current_pred);
   1253 
   1254  mi->interp_filters.as_int = filters_ref_set[best_filter_index].as_int;
   1255  mi->tx_size = pf_tx_size[best_filter_index];
   1256  this_rdc->rate = pf_rd_stats[best_filter_index].rate;
   1257  this_rdc->dist = pf_rd_stats[best_filter_index].dist;
   1258  this_rdc->sse = pf_rd_stats[best_filter_index].sse;
   1259  this_rdc->skip_txfm = (best_skip || best_early_term);
   1260  *this_early_term = best_early_term;
   1261  if (reuse_inter_pred) {
   1262    pd->dst.buf = (*this_mode_pred)->data;
   1263    pd->dst.stride = (*this_mode_pred)->stride;
   1264  } else if (best_filter_index < dim_factor * FILTER_SEARCH_SIZE - 1) {
   1265    if (is_single_pred)
   1266      av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
   1267                                            &subpel_params);
   1268    else
   1269      av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   1270                                    AOM_PLANE_Y, AOM_PLANE_Y);
   1271  }
   1272 }
   1273 #if !CONFIG_REALTIME_ONLY
   1274 
   1275 static inline int is_warped_mode_allowed(const AV1_COMP *cpi,
   1276                                         MACROBLOCK *const x,
   1277                                         const MB_MODE_INFO *mbmi) {
   1278  const FeatureFlags *const features = &cpi->common.features;
   1279  const MACROBLOCKD *xd = &x->e_mbd;
   1280 
   1281  if (cpi->sf.inter_sf.extra_prune_warped) return 0;
   1282  if (has_second_ref(mbmi)) return 0;
   1283  MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
   1284 
   1285  if (features->switchable_motion_mode) {
   1286    // Determine which motion modes to search if more than SIMPLE_TRANSLATION
   1287    // is allowed.
   1288    last_motion_mode_allowed = motion_mode_allowed(
   1289        xd->global_motion, xd, mbmi, features->allow_warped_motion);
   1290  }
   1291 
   1292  if (last_motion_mode_allowed == WARPED_CAUSAL) {
   1293    return 1;
   1294  }
   1295 
   1296  return 0;
   1297 }
   1298 
   1299 static void calc_num_proj_ref(AV1_COMP *cpi, MACROBLOCK *x, MB_MODE_INFO *mi) {
   1300  AV1_COMMON *const cm = &cpi->common;
   1301  MACROBLOCKD *const xd = &x->e_mbd;
   1302  const FeatureFlags *const features = &cm->features;
   1303 
   1304  mi->num_proj_ref = 1;
   1305  WARP_SAMPLE_INFO *const warp_sample_info =
   1306      &x->warp_sample_info[mi->ref_frame[0]];
   1307  int *pts0 = warp_sample_info->pts;
   1308  int *pts_inref0 = warp_sample_info->pts_inref;
   1309  MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
   1310 
   1311  if (features->switchable_motion_mode) {
   1312    // Determine which motion modes to search if more than SIMPLE_TRANSLATION
   1313    // is allowed.
   1314    last_motion_mode_allowed = motion_mode_allowed(
   1315        xd->global_motion, xd, mi, features->allow_warped_motion);
   1316  }
   1317 
   1318  if (last_motion_mode_allowed == WARPED_CAUSAL) {
   1319    if (warp_sample_info->num < 0) {
   1320      warp_sample_info->num = av1_findSamples(cm, xd, pts0, pts_inref0);
   1321    }
   1322    mi->num_proj_ref = warp_sample_info->num;
   1323  }
   1324 }
   1325 
   1326 static void search_motion_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
   1327                               int mi_row, int mi_col, BLOCK_SIZE bsize,
   1328                               int *this_early_term, int use_model_yrd_large,
   1329                               int *rate_mv, int64_t best_sse) {
   1330  AV1_COMMON *const cm = &cpi->common;
   1331  MACROBLOCKD *const xd = &x->e_mbd;
   1332  const FeatureFlags *const features = &cm->features;
   1333  MB_MODE_INFO *const mi = xd->mi[0];
   1334  RD_STATS pf_rd_stats[MOTION_MODE_SEARCH_SIZE] = { 0 };
   1335  int best_skip = 0;
   1336  int best_early_term = 0;
   1337  int64_t best_cost = INT64_MAX;
   1338  int best_mode_index = -1;
   1339  const int interp_filter = features->interp_filter;
   1340 
   1341  const MOTION_MODE motion_modes[MOTION_MODE_SEARCH_SIZE] = {
   1342    SIMPLE_TRANSLATION, WARPED_CAUSAL
   1343  };
   1344  int mode_search_size = is_warped_mode_allowed(cpi, x, mi) ? 2 : 1;
   1345 
   1346  WARP_SAMPLE_INFO *const warp_sample_info =
   1347      &x->warp_sample_info[mi->ref_frame[0]];
   1348  int *pts0 = warp_sample_info->pts;
   1349  int *pts_inref0 = warp_sample_info->pts_inref;
   1350 
   1351  const int total_samples = mi->num_proj_ref;
   1352  if (total_samples == 0) {
   1353    // Do not search WARPED_CAUSAL if there are no samples to use to determine
   1354    // warped parameters.
   1355    mode_search_size = 1;
   1356  }
   1357 
   1358  const MB_MODE_INFO base_mbmi = *mi;
   1359  MB_MODE_INFO best_mbmi;
   1360 
   1361  for (int mode_index = 0; mode_index < mode_search_size; ++mode_index) {
   1362    int64_t cost = INT64_MAX;
   1363    MOTION_MODE motion_mode = motion_modes[mode_index];
   1364    *mi = base_mbmi;
   1365    mi->motion_mode = motion_mode;
   1366    if (motion_mode == SIMPLE_TRANSLATION) {
   1367      mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
   1368 
   1369      av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   1370                                    AOM_PLANE_Y, AOM_PLANE_Y);
   1371      if (use_model_yrd_large)
   1372        model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
   1373                                  &pf_rd_stats[mode_index], this_early_term, 1,
   1374                                  best_sse, NULL, UINT_MAX);
   1375      else
   1376        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[mode_index], NULL, 1,
   1377                          NULL);
   1378      pf_rd_stats[mode_index].rate +=
   1379          av1_get_switchable_rate(x, xd, cm->features.interp_filter,
   1380                                  cm->seq_params->enable_dual_filter);
   1381      cost = RDCOST(x->rdmult, pf_rd_stats[mode_index].rate,
   1382                    pf_rd_stats[mode_index].dist);
   1383    } else if (motion_mode == WARPED_CAUSAL) {
   1384      int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
   1385      const ModeCosts *mode_costs = &x->mode_costs;
   1386      mi->wm_params.wmtype = DEFAULT_WMTYPE;
   1387      mi->interp_filters =
   1388          av1_broadcast_interp_filter(av1_unswitchable_filter(interp_filter));
   1389 
   1390      memcpy(pts, pts0, total_samples * 2 * sizeof(*pts0));
   1391      memcpy(pts_inref, pts_inref0, total_samples * 2 * sizeof(*pts_inref0));
   1392      // Select the samples according to motion vector difference
   1393      if (mi->num_proj_ref > 1) {
   1394        mi->num_proj_ref = av1_selectSamples(&mi->mv[0].as_mv, pts, pts_inref,
   1395                                             mi->num_proj_ref, bsize);
   1396      }
   1397 
   1398      // Compute the warped motion parameters with a least squares fit
   1399      //  using the collected samples
   1400      if (!av1_find_projection(mi->num_proj_ref, pts, pts_inref, bsize,
   1401                               mi->mv[0].as_mv.row, mi->mv[0].as_mv.col,
   1402                               &mi->wm_params, mi_row, mi_col)) {
   1403        if (mi->mode == NEWMV) {
   1404          const int_mv mv0 = mi->mv[0];
   1405          const WarpedMotionParams wm_params0 = mi->wm_params;
   1406          const int num_proj_ref0 = mi->num_proj_ref;
   1407 
   1408          const int_mv ref_mv = av1_get_ref_mv(x, 0);
   1409          SUBPEL_MOTION_SEARCH_PARAMS ms_params;
   1410          av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize,
   1411                                            &ref_mv.as_mv, NULL);
   1412 
   1413          // Refine MV in a small range.
   1414          av1_refine_warped_mv(xd, cm, &ms_params, bsize, pts0, pts_inref0,
   1415                               total_samples, cpi->sf.mv_sf.warp_search_method,
   1416                               cpi->sf.mv_sf.warp_search_iters);
   1417          if (mi->mv[0].as_int == ref_mv.as_int) {
   1418            continue;
   1419          }
   1420 
   1421          if (mv0.as_int != mi->mv[0].as_int) {
   1422            // Keep the refined MV and WM parameters.
   1423            int tmp_rate_mv = av1_mv_bit_cost(
   1424                &mi->mv[0].as_mv, &ref_mv.as_mv, x->mv_costs->nmv_joint_cost,
   1425                x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
   1426            *rate_mv = tmp_rate_mv;
   1427          } else {
   1428            // Restore the old MV and WM parameters.
   1429            mi->mv[0] = mv0;
   1430            mi->wm_params = wm_params0;
   1431            mi->num_proj_ref = num_proj_ref0;
   1432          }
   1433        }
   1434        // Build the warped predictor
   1435        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   1436                                      AOM_PLANE_Y, av1_num_planes(cm) - 1);
   1437        if (use_model_yrd_large)
   1438          model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
   1439                                    &pf_rd_stats[mode_index], this_early_term,
   1440                                    1, best_sse, NULL, UINT_MAX);
   1441        else
   1442          model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[mode_index], NULL,
   1443                            1, NULL);
   1444 
   1445        pf_rd_stats[mode_index].rate +=
   1446            mode_costs->motion_mode_cost[bsize][mi->motion_mode];
   1447        cost = RDCOST(x->rdmult, pf_rd_stats[mode_index].rate,
   1448                      pf_rd_stats[mode_index].dist);
   1449      } else {
   1450        cost = INT64_MAX;
   1451      }
   1452    }
   1453    if (cost < best_cost) {
   1454      best_mode_index = mode_index;
   1455      best_cost = cost;
   1456      best_skip = pf_rd_stats[mode_index].skip_txfm;
   1457      best_early_term = *this_early_term;
   1458      best_mbmi = *mi;
   1459    }
   1460  }
   1461  assert(best_mode_index >= 0 && best_mode_index < FILTER_SEARCH_SIZE);
   1462 
   1463  *mi = best_mbmi;
   1464  this_rdc->rate = pf_rd_stats[best_mode_index].rate;
   1465  this_rdc->dist = pf_rd_stats[best_mode_index].dist;
   1466  this_rdc->sse = pf_rd_stats[best_mode_index].sse;
   1467  this_rdc->skip_txfm = (best_skip || best_early_term);
   1468  *this_early_term = best_early_term;
   1469  if (best_mode_index < FILTER_SEARCH_SIZE - 1) {
   1470    av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   1471                                  AOM_PLANE_Y, AOM_PLANE_Y);
   1472  }
   1473 }
   1474 #endif  // !CONFIG_REALTIME_ONLY
   1475 
   1476 #define COLLECT_NON_SQR_STAT 0
   1477 
   1478 #if COLLECT_NONRD_PICK_MODE_STAT
   1479 
   1480 static inline void print_stage_time(const char *stage_name, int64_t stage_time,
   1481                                    int64_t total_time) {
   1482  printf("    %s: %ld (%f%%)\n", stage_name, stage_time,
   1483         100 * stage_time / (float)total_time);
   1484 }
   1485 
   1486 static void print_time(const mode_search_stat_nonrd *const ms_stat,
   1487                       BLOCK_SIZE bsize, int mi_rows, int mi_cols, int mi_row,
   1488                       int mi_col) {
   1489  if ((mi_row + mi_size_high[bsize] >= mi_rows) &&
   1490      (mi_col + mi_size_wide[bsize] >= mi_cols)) {
   1491    int64_t total_time = 0l;
   1492    int32_t total_blocks = 0;
   1493    for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
   1494      total_time += ms_stat->total_block_times[bs];
   1495      total_blocks += ms_stat->num_blocks[bs];
   1496    }
   1497 
   1498    printf("\n");
   1499    for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
   1500      if (ms_stat->num_blocks[bs] == 0) {
   1501        continue;
   1502      }
   1503      if (!COLLECT_NON_SQR_STAT && block_size_wide[bs] != block_size_high[bs]) {
   1504        continue;
   1505      }
   1506 
   1507      printf("BLOCK_%dX%d Num %d, Time: %ld (%f%%), Avg_time %f:\n",
   1508             block_size_wide[bs], block_size_high[bs], ms_stat->num_blocks[bs],
   1509             ms_stat->total_block_times[bs],
   1510             100 * ms_stat->total_block_times[bs] / (float)total_time,
   1511             (float)ms_stat->total_block_times[bs] / ms_stat->num_blocks[bs]);
   1512      for (int j = 0; j < MB_MODE_COUNT; j++) {
   1513        if (ms_stat->nonskipped_search_times[bs][j] == 0) {
   1514          continue;
   1515        }
   1516 
   1517        int64_t total_mode_time = ms_stat->nonskipped_search_times[bs][j];
   1518        printf("  Mode %d, %d/%d tps %f\n", j,
   1519               ms_stat->num_nonskipped_searches[bs][j],
   1520               ms_stat->num_searches[bs][j],
   1521               ms_stat->num_nonskipped_searches[bs][j] > 0
   1522                   ? (float)ms_stat->nonskipped_search_times[bs][j] /
   1523                         ms_stat->num_nonskipped_searches[bs][j]
   1524                   : 0l);
   1525        if (j >= INTER_MODE_START) {
   1526          total_mode_time = ms_stat->ms_time[bs][j] + ms_stat->ifs_time[bs][j] +
   1527                            ms_stat->model_rd_time[bs][j] +
   1528                            ms_stat->txfm_time[bs][j];
   1529          print_stage_time("Motion Search Time", ms_stat->ms_time[bs][j],
   1530                           total_time);
   1531          print_stage_time("Filter Search Time", ms_stat->ifs_time[bs][j],
   1532                           total_time);
   1533          print_stage_time("Model    RD   Time", ms_stat->model_rd_time[bs][j],
   1534                           total_time);
   1535          print_stage_time("Tranfm Search Time", ms_stat->txfm_time[bs][j],
   1536                           total_time);
   1537        }
   1538        print_stage_time("Total  Mode   Time", total_mode_time, total_time);
   1539      }
   1540      printf("\n");
   1541    }
   1542    printf("Total time = %ld. Total blocks = %d\n", total_time, total_blocks);
   1543  }
   1544 }
   1545 #endif  // COLLECT_NONRD_PICK_MODE_STAT
   1546 
   1547 static bool should_prune_intra_modes_using_neighbors(
   1548    const MACROBLOCKD *xd, bool enable_intra_mode_pruning_using_neighbors,
   1549    PREDICTION_MODE this_mode, PREDICTION_MODE above_mode,
   1550    PREDICTION_MODE left_mode) {
   1551  if (!enable_intra_mode_pruning_using_neighbors) return false;
   1552 
   1553  // Avoid pruning of DC_PRED as it is the most probable mode to win as per the
   1554  // statistics generated for nonrd intra mode evaluations.
   1555  if (this_mode == DC_PRED) return false;
   1556 
   1557  // Enable the pruning for current mode only if it is not the winner mode of
   1558  // both the neighboring blocks (left/top).
   1559  return xd->up_available && this_mode != above_mode && xd->left_available &&
   1560         this_mode != left_mode;
   1561 }
   1562 
   1563 void av1_nonrd_pick_intra_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_cost,
   1564                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
   1565  AV1_COMMON *const cm = &cpi->common;
   1566  MACROBLOCKD *const xd = &x->e_mbd;
   1567  MB_MODE_INFO *const mi = xd->mi[0];
   1568  RD_STATS this_rdc, best_rdc;
   1569  struct estimate_block_intra_args args;
   1570  init_estimate_block_intra_args(&args, cpi, x);
   1571  const TxfmSearchParams *txfm_params = &x->txfm_search_params;
   1572  mi->tx_size =
   1573      AOMMIN(max_txsize_lookup[bsize],
   1574             tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
   1575  assert(IMPLIES(xd->lossless[mi->segment_id], mi->tx_size == TX_4X4));
   1576  const BLOCK_SIZE tx_bsize = txsize_to_bsize[mi->tx_size];
   1577 
   1578  // If the current block size is the same as the transform block size, enable
   1579  // mode pruning based on the best SAD so far.
   1580  if (cpi->sf.rt_sf.prune_intra_mode_using_best_sad_so_far && bsize == tx_bsize)
   1581    args.prune_mode_based_on_sad = true;
   1582 
   1583  int *bmode_costs;
   1584  PREDICTION_MODE best_mode = DC_PRED;
   1585  const MB_MODE_INFO *above_mi = xd->above_mbmi;
   1586  const MB_MODE_INFO *left_mi = xd->left_mbmi;
   1587  const PREDICTION_MODE A = av1_above_block_mode(above_mi);
   1588  const PREDICTION_MODE L = av1_left_block_mode(left_mi);
   1589  const int above_ctx = intra_mode_context[A];
   1590  const int left_ctx = intra_mode_context[L];
   1591  const unsigned int source_variance = x->source_variance;
   1592  bmode_costs = x->mode_costs.y_mode_costs[above_ctx][left_ctx];
   1593  const int mi_row = xd->mi_row;
   1594  const int mi_col = xd->mi_col;
   1595  // Use this flag to signal large flat blocks that may need special
   1596  // treatment: in the current case H/V/SMOOTH may not be skipped if
   1597  // DC has nonzero distortion and skippable is set. This is to remove
   1598  // visual artifacts observed for screen in realtime mode.
   1599  const bool flat_blocks_screen =
   1600      cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   1601      cpi->oxcf.mode == REALTIME && x->source_variance == 0 &&
   1602      bsize >= BLOCK_32X32;
   1603  av1_invalid_rd_stats(&best_rdc);
   1604  av1_invalid_rd_stats(&this_rdc);
   1605 
   1606  init_mbmi_nonrd(mi, DC_PRED, INTRA_FRAME, NONE_FRAME, cm);
   1607  mi->mv[0].as_int = mi->mv[1].as_int = INVALID_MV;
   1608 
   1609  bool allow_skip_nondc = true;
   1610  // Change the limit of this loop to add other intra prediction
   1611  // mode tests.
   1612  for (int mode_index = 0; mode_index < RTC_INTRA_MODES; ++mode_index) {
   1613    PREDICTION_MODE this_mode = intra_mode_list[mode_index];
   1614 
   1615    // Force DC for spatially flat block for large bsize, on top-left corner.
   1616    // This removed potential artifact observed in gray scale image for high Q.
   1617    if (x->source_variance == 0 && mi_col == 0 && mi_row == 0 &&
   1618        bsize >= BLOCK_32X32 && this_mode > 0)
   1619      continue;
   1620 
   1621    // As per the statistics generated for intra mode evaluation in the nonrd
   1622    // path, it is found that the probability of H_PRED mode being the winner is
   1623    // very low when the best mode so far is V_PRED (out of DC_PRED and V_PRED).
   1624    // If V_PRED is the winner mode out of DC_PRED and V_PRED, it could imply
   1625    // the presence of a vertically dominant pattern. Hence, H_PRED mode is not
   1626    // evaluated.
   1627    if (cpi->sf.rt_sf.prune_h_pred_using_best_mode_so_far &&
   1628        this_mode == H_PRED && best_mode == V_PRED && allow_skip_nondc)
   1629      continue;
   1630 
   1631    if (should_prune_intra_modes_using_neighbors(
   1632            xd, cpi->sf.rt_sf.enable_intra_mode_pruning_using_neighbors,
   1633            this_mode, A, L)) {
   1634      // Prune V_PRED and H_PRED if source variance of the block is less than
   1635      // or equal to 50. The source variance threshold is obtained empirically.
   1636      if ((this_mode == V_PRED || this_mode == H_PRED) &&
   1637          source_variance <= 50 && allow_skip_nondc)
   1638        continue;
   1639 
   1640      // As per the statistics, probability of SMOOTH_PRED being the winner is
   1641      // low when best mode so far is DC_PRED (out of DC_PRED, V_PRED and
   1642      // H_PRED). Hence, SMOOTH_PRED mode is not evaluated.
   1643      if (best_mode == DC_PRED && this_mode == SMOOTH_PRED && allow_skip_nondc)
   1644        continue;
   1645    }
   1646 
   1647    this_rdc.dist = this_rdc.rate = 0;
   1648    args.mode = this_mode;
   1649    args.skippable = 1;
   1650    args.rdc = &this_rdc;
   1651    mi->mode = this_mode;
   1652    av1_foreach_transformed_block_in_plane(xd, bsize, AOM_PLANE_Y,
   1653                                           av1_estimate_block_intra, &args);
   1654 
   1655    if (this_rdc.rate == INT_MAX) continue;
   1656 
   1657    const int skip_ctx = av1_get_skip_txfm_context(xd);
   1658    if (args.skippable) {
   1659      this_rdc.rate = x->mode_costs.skip_txfm_cost[skip_ctx][1];
   1660    } else {
   1661      this_rdc.rate += x->mode_costs.skip_txfm_cost[skip_ctx][0];
   1662    }
   1663    this_rdc.rate += bmode_costs[this_mode];
   1664    this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
   1665    if (this_rdc.rdcost < best_rdc.rdcost) {
   1666      best_rdc = this_rdc;
   1667      best_mode = this_mode;
   1668      if (!this_rdc.skip_txfm) {
   1669        if (flat_blocks_screen && args.skippable && best_rdc.dist < 20000) {
   1670          memcpy(ctx->blk_skip, x->txfm_search_info.blk_skip,
   1671                 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
   1672        } else {
   1673          memset(ctx->blk_skip, 0,
   1674                 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
   1675        }
   1676      }
   1677    }
   1678    if (this_mode == DC_PRED) {
   1679      if (flat_blocks_screen && args.skippable && this_rdc.dist > 0)
   1680        allow_skip_nondc = false;
   1681    }
   1682  }
   1683 
   1684  const unsigned int thresh_sad =
   1685      cpi->sf.rt_sf.prune_palette_search_nonrd > 1 ? 100 : 20;
   1686  const unsigned int best_sad_norm =
   1687      args.best_sad >>
   1688      (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
   1689 
   1690  // Try palette if it's enabled.
   1691  bool try_palette =
   1692      cpi->oxcf.tool_cfg.enable_palette &&
   1693      av1_allow_palette(cpi->common.features.allow_screen_content_tools,
   1694                        mi->bsize);
   1695  if (cpi->sf.rt_sf.prune_palette_search_nonrd > 0) {
   1696    bool prune =
   1697        (!args.prune_mode_based_on_sad || best_sad_norm > thresh_sad) &&
   1698        bsize <= BLOCK_16X16 && x->source_variance > 200;
   1699    try_palette &= prune;
   1700  }
   1701  if (try_palette) {
   1702    const TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   1703    const unsigned int intra_ref_frame_cost = 0;
   1704    x->color_palette_thresh = (best_sad_norm < 500) ? 32 : 64;
   1705 
   1706    // Search palette mode for Luma plane in intra frame.
   1707    av1_search_palette_mode_luma(cpi, x, bsize, intra_ref_frame_cost, ctx,
   1708                                 &this_rdc, best_rdc.rdcost);
   1709    // Update best mode data.
   1710    if (this_rdc.rdcost < best_rdc.rdcost) {
   1711      best_mode = DC_PRED;
   1712      mi->mv[0].as_int = INVALID_MV;
   1713      mi->mv[1].as_int = INVALID_MV;
   1714      best_rdc.rate = this_rdc.rate;
   1715      best_rdc.dist = this_rdc.dist;
   1716      best_rdc.rdcost = this_rdc.rdcost;
   1717      if (!this_rdc.skip_txfm) {
   1718        memcpy(ctx->blk_skip, txfm_info->blk_skip,
   1719               sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
   1720      }
   1721      if (xd->tx_type_map[0] != DCT_DCT)
   1722        av1_copy_array(ctx->tx_type_map, xd->tx_type_map, ctx->num_4x4_blk);
   1723    } else {
   1724      av1_zero(mi->palette_mode_info);
   1725    }
   1726  }
   1727 
   1728  mi->mode = best_mode;
   1729  // Keep DC for UV since mode test is based on Y channel only.
   1730  mi->uv_mode = UV_DC_PRED;
   1731  *rd_cost = best_rdc;
   1732 
   1733  // For lossless: always force the skip flags off.
   1734  // Even though the blk_skip is set to 0 above in the rdcost comparison,
   1735  // do it here again in case the above logic changes.
   1736  if (is_lossless_requested(&cpi->oxcf.rc_cfg)) {
   1737    x->txfm_search_info.skip_txfm = 0;
   1738    memset(ctx->blk_skip, 0,
   1739           sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
   1740  }
   1741 
   1742 #if CONFIG_INTERNAL_STATS
   1743  store_coding_context_nonrd(x, ctx, mi->mode);
   1744 #else
   1745  store_coding_context_nonrd(x, ctx);
   1746 #endif  // CONFIG_INTERNAL_STATS
   1747 }
   1748 
   1749 static inline int is_same_gf_and_last_scale(AV1_COMMON *cm) {
   1750  struct scale_factors *const sf_last = get_ref_scale_factors(cm, LAST_FRAME);
   1751  struct scale_factors *const sf_golden =
   1752      get_ref_scale_factors(cm, GOLDEN_FRAME);
   1753  return ((sf_last->x_scale_fp == sf_golden->x_scale_fp) &&
   1754          (sf_last->y_scale_fp == sf_golden->y_scale_fp));
   1755 }
   1756 
   1757 static inline void get_ref_frame_use_mask(AV1_COMP *cpi, MACROBLOCK *x,
   1758                                          MB_MODE_INFO *mi, int mi_row,
   1759                                          int mi_col, BLOCK_SIZE bsize,
   1760                                          int gf_temporal_ref,
   1761                                          int use_ref_frame[],
   1762                                          int *force_skip_low_temp_var) {
   1763  AV1_COMMON *const cm = &cpi->common;
   1764  const struct segmentation *const seg = &cm->seg;
   1765  const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
   1766 
   1767  // When the ref_frame_config is used to set the reference frame structure
   1768  // then the usage of alt_ref is determined by the ref_frame_flags
   1769  // (and not the speed feature use_nonrd_altref_frame).
   1770  int use_alt_ref_frame = cpi->ppi->rtc_ref.set_ref_frame_config ||
   1771                          cpi->sf.rt_sf.use_nonrd_altref_frame;
   1772 
   1773  int use_golden_ref_frame = 1;
   1774  int use_last_ref_frame = 1;
   1775 
   1776  // When the ref_frame_config is used to set the reference frame structure:
   1777  // check if LAST is used as a reference. And only remove golden and altref
   1778  // references below if last is used as a reference.
   1779  if (cpi->ppi->rtc_ref.set_ref_frame_config)
   1780    use_last_ref_frame =
   1781        cpi->ref_frame_flags & AOM_LAST_FLAG ? use_last_ref_frame : 0;
   1782 
   1783  // frame_since_golden is not used when user sets the referene structure.
   1784  if (!cpi->ppi->rtc_ref.set_ref_frame_config && use_last_ref_frame &&
   1785      cpi->rc.frames_since_golden == 0 && gf_temporal_ref) {
   1786    use_golden_ref_frame = 0;
   1787  }
   1788 
   1789  if (use_last_ref_frame && cpi->sf.rt_sf.short_circuit_low_temp_var &&
   1790      x->nonrd_prune_ref_frame_search) {
   1791    if (is_small_sb)
   1792      *force_skip_low_temp_var = av1_get_force_skip_low_temp_var_small_sb(
   1793          &x->part_search_info.variance_low[0], mi_row, mi_col, bsize);
   1794    else
   1795      *force_skip_low_temp_var = av1_get_force_skip_low_temp_var(
   1796          &x->part_search_info.variance_low[0], mi_row, mi_col, bsize);
   1797    // If force_skip_low_temp_var is set, skip golden reference.
   1798    if (*force_skip_low_temp_var) {
   1799      use_golden_ref_frame = 0;
   1800      use_alt_ref_frame = 0;
   1801    }
   1802  }
   1803 
   1804  if (use_last_ref_frame &&
   1805      (x->nonrd_prune_ref_frame_search > 2 || x->force_zeromv_skip_for_blk ||
   1806       (x->nonrd_prune_ref_frame_search > 1 && bsize > BLOCK_64X64))) {
   1807    use_golden_ref_frame = 0;
   1808    use_alt_ref_frame = 0;
   1809  }
   1810 
   1811  if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
   1812      get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
   1813    use_ref_frame[GOLDEN_FRAME] = 1;
   1814    use_ref_frame[ALTREF_FRAME] = 0;
   1815    return;
   1816  } else if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
   1817             get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) ==
   1818                 ALTREF_FRAME) {
   1819    use_ref_frame[GOLDEN_FRAME] = 0;
   1820    use_ref_frame[ALTREF_FRAME] = 1;
   1821    return;
   1822  }
   1823 
   1824  // Skip golden/altref reference if color is set, on flat blocks with motion.
   1825  // For screen: always skip golden/alt (if color_sensitivity_sb_g/alt is set)
   1826  // except when x->nonrd_prune_ref_frame_search = 0. This latter flag
   1827  // may be set in the variance partition when golden is a much better
   1828  // reference than last, in which case it may not be worth skipping
   1829  // golden/altref completely.
   1830  // Condition on use_last_ref to make sure there remains at least one
   1831  // reference.
   1832  if (use_last_ref_frame &&
   1833      ((cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   1834        x->nonrd_prune_ref_frame_search != 0) ||
   1835       (x->source_variance < 200 &&
   1836        x->content_state_sb.source_sad_nonrd >= kLowSad))) {
   1837    if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
   1838        x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
   1839      use_golden_ref_frame = 0;
   1840    if (x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
   1841        x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
   1842      use_alt_ref_frame = 0;
   1843  }
   1844 
   1845  // For non-screen: if golden and altref are not being selected as references
   1846  // (use_golden_ref_frame/use_alt_ref_frame = 0) check to allow golden back
   1847  // based on the sad of nearest/nearmv of LAST ref. If this block sad is large,
   1848  // keep golden as reference. Only do this for the agrressive pruning mode and
   1849  // avoid it when color is set for golden reference.
   1850  if (cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN &&
   1851      (cpi->ref_frame_flags & AOM_LAST_FLAG) && !use_golden_ref_frame &&
   1852      !use_alt_ref_frame && x->pred_mv_sad[LAST_FRAME] != INT_MAX &&
   1853      x->nonrd_prune_ref_frame_search > 2 &&
   1854      x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
   1855      x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) {
   1856    int thr = (cm->width * cm->height > RESOLUTION_288P) ? 100 : 150;
   1857    int pred = x->pred_mv_sad[LAST_FRAME] >>
   1858               (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
   1859    if (pred > thr) use_golden_ref_frame = 1;
   1860  }
   1861 
   1862  use_alt_ref_frame =
   1863      cpi->ref_frame_flags & AOM_ALT_FLAG ? use_alt_ref_frame : 0;
   1864  use_golden_ref_frame =
   1865      cpi->ref_frame_flags & AOM_GOLD_FLAG ? use_golden_ref_frame : 0;
   1866 
   1867  // For spatial layers: enable golden ref if it is set by user and
   1868  // corresponds to the lower spatial layer.
   1869  if (cpi->svc.spatial_layer_id > 0 && (cpi->ref_frame_flags & AOM_GOLD_FLAG) &&
   1870      x->content_state_sb.source_sad_nonrd < kHighSad) {
   1871    const int buffslot_golden =
   1872        cpi->ppi->rtc_ref.ref_idx[GOLDEN_FRAME - LAST_FRAME];
   1873    if (cpi->ppi->rtc_ref.buffer_time_index[buffslot_golden] ==
   1874        cpi->svc.current_superframe)
   1875      use_golden_ref_frame = 1;
   1876  }
   1877 
   1878  use_ref_frame[ALTREF_FRAME] = use_alt_ref_frame;
   1879  use_ref_frame[GOLDEN_FRAME] = use_golden_ref_frame;
   1880  use_ref_frame[LAST_FRAME] = use_last_ref_frame;
   1881  // Keep this assert on, as only 3 references are used in nonrd_pickmode
   1882  // (LAST, GOLDEN, ALTREF), and if all 3 are not set by user then this
   1883  // frame must be an intra-only frame and hence should never enter the
   1884  // pickmode here for inter frames.
   1885  assert(use_last_ref_frame || use_golden_ref_frame || use_alt_ref_frame);
   1886 }
   1887 
   1888 static inline int is_filter_search_enabled_blk(AV1_COMP *cpi, MACROBLOCK *x,
   1889                                               int mi_row, int mi_col,
   1890                                               BLOCK_SIZE bsize, int segment_id,
   1891                                               int cb_pred_filter_search,
   1892                                               InterpFilter *filt_select) {
   1893  const AV1_COMMON *const cm = &cpi->common;
   1894  // filt search disabled
   1895  if (!cpi->sf.rt_sf.use_nonrd_filter_search) return 0;
   1896  // filt search purely based on mode properties
   1897  if (!cb_pred_filter_search) return 1;
   1898  MACROBLOCKD *const xd = &x->e_mbd;
   1899  int enable_interp_search = 0;
   1900  if (!(xd->left_mbmi && xd->above_mbmi)) {
   1901    // neighbors info unavailable
   1902    enable_interp_search = 2;
   1903  } else if (!(is_inter_block(xd->left_mbmi) &&
   1904               is_inter_block(xd->above_mbmi))) {
   1905    // neighbor is INTRA
   1906    enable_interp_search = 2;
   1907  } else if (xd->left_mbmi->interp_filters.as_int !=
   1908             xd->above_mbmi->interp_filters.as_int) {
   1909    // filters are different
   1910    enable_interp_search = 2;
   1911  } else if ((cb_pred_filter_search == 1) &&
   1912             (xd->left_mbmi->interp_filters.as_filters.x_filter !=
   1913              EIGHTTAP_REGULAR)) {
   1914    // not regular
   1915    enable_interp_search = 2;
   1916  } else {
   1917    // enable prediction based on chessboard pattern
   1918    if (xd->left_mbmi->interp_filters.as_filters.x_filter == EIGHTTAP_SMOOTH)
   1919      *filt_select = EIGHTTAP_SMOOTH;
   1920    const int bsl = mi_size_wide_log2[bsize];
   1921    enable_interp_search =
   1922        (bool)((((mi_row + mi_col) >> bsl) +
   1923                get_chessboard_index(cm->current_frame.frame_number)) &
   1924               0x1);
   1925    if (cyclic_refresh_segment_id_boosted(segment_id)) enable_interp_search = 1;
   1926  }
   1927  return enable_interp_search;
   1928 }
   1929 
   1930 static inline int skip_mode_by_threshold(PREDICTION_MODE mode,
   1931                                         MV_REFERENCE_FRAME ref_frame,
   1932                                         int_mv mv, int frames_since_golden,
   1933                                         const int *const rd_threshes,
   1934                                         const int *const rd_thresh_freq_fact,
   1935                                         int64_t best_cost, int best_skip,
   1936                                         int extra_shift) {
   1937  int skip_this_mode = 0;
   1938  const THR_MODES mode_index = mode_idx[ref_frame][INTER_OFFSET(mode)];
   1939  int64_t mode_rd_thresh =
   1940      best_skip ? ((int64_t)rd_threshes[mode_index]) << (extra_shift + 1)
   1941                : ((int64_t)rd_threshes[mode_index]) << extra_shift;
   1942 
   1943  // Increase mode_rd_thresh value for non-LAST for improved encoding
   1944  // speed
   1945  if (ref_frame != LAST_FRAME) {
   1946    mode_rd_thresh = mode_rd_thresh << 1;
   1947    if (ref_frame == GOLDEN_FRAME && frames_since_golden > 4)
   1948      mode_rd_thresh = mode_rd_thresh << (extra_shift + 1);
   1949  }
   1950 
   1951  if (rd_less_than_thresh(best_cost, mode_rd_thresh,
   1952                          rd_thresh_freq_fact[mode_index]))
   1953    if (mv.as_int != 0) skip_this_mode = 1;
   1954 
   1955  return skip_this_mode;
   1956 }
   1957 
   1958 static inline int skip_mode_by_low_temp(
   1959    PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame, BLOCK_SIZE bsize,
   1960    CONTENT_STATE_SB content_state_sb, int_mv mv, int force_skip_low_temp_var) {
   1961  // Skip non-zeromv mode search for non-LAST frame if force_skip_low_temp_var
   1962  // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
   1963  // later.
   1964  if (force_skip_low_temp_var && ref_frame != LAST_FRAME && mv.as_int != 0) {
   1965    return 1;
   1966  }
   1967 
   1968  if (content_state_sb.source_sad_nonrd != kHighSad && bsize >= BLOCK_64X64 &&
   1969      force_skip_low_temp_var && mode == NEWMV) {
   1970    return 1;
   1971  }
   1972  return 0;
   1973 }
   1974 
   1975 static inline int skip_mode_by_bsize_and_ref_frame(
   1976    PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame, BLOCK_SIZE bsize,
   1977    int extra_prune, unsigned int sse_zeromv_norm, int more_prune,
   1978    int skip_nearmv) {
   1979  const unsigned int thresh_skip_golden = 500;
   1980 
   1981  if (ref_frame != LAST_FRAME && sse_zeromv_norm < thresh_skip_golden &&
   1982      mode == NEWMV)
   1983    return 1;
   1984 
   1985  if ((bsize == BLOCK_128X128 && mode == NEWMV) ||
   1986      (skip_nearmv && mode == NEARMV))
   1987    return 1;
   1988 
   1989  // Skip testing non-LAST if this flag is set.
   1990  if (extra_prune) {
   1991    if (extra_prune > 1 && ref_frame != LAST_FRAME &&
   1992        (bsize > BLOCK_16X16 && mode == NEWMV))
   1993      return 1;
   1994 
   1995    if (ref_frame != LAST_FRAME && mode == NEARMV) return 1;
   1996 
   1997    if (more_prune && bsize >= BLOCK_32X32 && mode == NEARMV) return 1;
   1998  }
   1999  return 0;
   2000 }
   2001 
   2002 static void set_block_source_sad(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
   2003                                 struct buf_2d *yv12_mb) {
   2004  struct macroblock_plane *const p = &x->plane[0];
   2005  const int y_sad = cpi->ppi->fn_ptr[bsize].sdf(p->src.buf, p->src.stride,
   2006                                                yv12_mb->buf, yv12_mb->stride);
   2007  if (y_sad == 0) x->block_is_zero_sad = 1;
   2008 }
   2009 
   2010 static void set_color_sensitivity(AV1_COMP *cpi, MACROBLOCK *x,
   2011                                  BLOCK_SIZE bsize, int y_sad,
   2012                                  unsigned int source_variance,
   2013                                  struct buf_2d yv12_mb[MAX_MB_PLANE]) {
   2014  const int subsampling_x = cpi->common.seq_params->subsampling_x;
   2015  const int subsampling_y = cpi->common.seq_params->subsampling_y;
   2016  const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
   2017  const int high_res = cpi->common.width * cpi->common.height >= 640 * 360;
   2018  if (bsize == cpi->common.seq_params->sb_size &&
   2019      !x->force_color_check_block_level) {
   2020    // At superblock level color_sensitivity is already set to 0, 1, or 2.
   2021    // 2 is middle/uncertain level. To avoid additional sad
   2022    // computations when bsize = sb_size force level 2 to 1 (certain color)
   2023    // for motion areas. Avoid this shortcut if x->force_color_check_block_level
   2024    // is set.
   2025    if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 2) {
   2026      x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] =
   2027          source_sad_nonrd >= kMedSad ? 1 : 0;
   2028    }
   2029    if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 2) {
   2030      x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] =
   2031          source_sad_nonrd >= kMedSad ? 1 : 0;
   2032    }
   2033    return;
   2034  }
   2035  // Divide factor for comparing uv_sad to y_sad.
   2036  int shift = 3;
   2037  // Threshold for the block spatial source variance.
   2038  unsigned int source_var_thr = 50;
   2039  // Thresholds for normalized uv_sad, the first one is used for
   2040  // low source_varaince.
   2041  int norm_uv_sad_thresh = 100;
   2042  int norm_uv_sad_thresh2 = 40;
   2043  if (source_sad_nonrd >= kMedSad && x->source_variance > 0 && high_res)
   2044    shift = 4;
   2045  if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
   2046    if (cpi->rc.high_source_sad) shift = 6;
   2047    if (source_sad_nonrd > kMedSad) {
   2048      source_var_thr = 1200;
   2049      norm_uv_sad_thresh = 10;
   2050    }
   2051    if (cpi->rc.percent_blocks_with_motion > 90 &&
   2052        cpi->rc.frame_source_sad > 10000 && source_sad_nonrd > kLowSad) {
   2053      // Aggressive setting for color_sensitivity for this content.
   2054      shift = 10;
   2055      norm_uv_sad_thresh = 0;
   2056      norm_uv_sad_thresh2 = 0;
   2057    }
   2058  }
   2059  NOISE_LEVEL noise_level = kLow;
   2060  int norm_sad =
   2061      y_sad >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
   2062  unsigned int thresh_spatial = (cpi->common.width > 1920) ? 5000 : 1000;
   2063  // If the spatial source variance is high and the normalized y_sad
   2064  // is low, then y-channel is likely good for mode estimation, so keep
   2065  // color_sensitivity off. For low noise content for now, since there is
   2066  // some bdrate regression for noisy color clip.
   2067  if (cpi->noise_estimate.enabled)
   2068    noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);
   2069  if (noise_level == kLow && source_variance > thresh_spatial &&
   2070      cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN && norm_sad < 50) {
   2071    x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] = 0;
   2072    x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] = 0;
   2073    return;
   2074  }
   2075  const int num_planes = av1_num_planes(&cpi->common);
   2076 
   2077  for (int plane = AOM_PLANE_U; plane < num_planes; ++plane) {
   2078    // Always check if level = 2. If level = 0 check again for
   2079    // motion areas for higher resolns, where color artifacts
   2080    // are more noticeable. Always check if
   2081    // x->force_color_check_block_level is set.
   2082    if (x->color_sensitivity[COLOR_SENS_IDX(plane)] == 2 ||
   2083        x->force_color_check_block_level ||
   2084        (x->color_sensitivity[COLOR_SENS_IDX(plane)] == 0 &&
   2085         source_sad_nonrd >= kMedSad && high_res)) {
   2086      struct macroblock_plane *const p = &x->plane[plane];
   2087      const BLOCK_SIZE bs =
   2088          get_plane_block_size(bsize, subsampling_x, subsampling_y);
   2089 
   2090      const int uv_sad = cpi->ppi->fn_ptr[bs].sdf(
   2091          p->src.buf, p->src.stride, yv12_mb[plane].buf, yv12_mb[plane].stride);
   2092 
   2093      const int norm_uv_sad =
   2094          uv_sad >> (b_width_log2_lookup[bs] + b_height_log2_lookup[bs]);
   2095      x->color_sensitivity[COLOR_SENS_IDX(plane)] =
   2096          uv_sad > (y_sad >> shift) && norm_uv_sad > norm_uv_sad_thresh2;
   2097      if (source_variance < source_var_thr && norm_uv_sad > norm_uv_sad_thresh)
   2098        x->color_sensitivity[COLOR_SENS_IDX(plane)] = 1;
   2099    }
   2100  }
   2101 }
   2102 
   2103 static void setup_compound_prediction(const AV1_COMMON *cm, MACROBLOCK *x,
   2104                                      struct buf_2d yv12_mb[8][MAX_MB_PLANE],
   2105                                      const int *use_ref_frame_mask,
   2106                                      const MV_REFERENCE_FRAME *rf,
   2107                                      int *ref_mv_idx) {
   2108  MACROBLOCKD *const xd = &x->e_mbd;
   2109  MB_MODE_INFO *const mbmi = xd->mi[0];
   2110  MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
   2111  MV_REFERENCE_FRAME ref_frame_comp;
   2112  if (!use_ref_frame_mask[rf[1]]) {
   2113    // Need to setup pred_block, if it hasn't been done in find_predictors.
   2114    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, rf[1]);
   2115    const int num_planes = av1_num_planes(cm);
   2116    if (yv12 != NULL) {
   2117      const struct scale_factors *const sf =
   2118          get_ref_scale_factors_const(cm, rf[1]);
   2119      av1_setup_pred_block(xd, yv12_mb[rf[1]], yv12, sf, sf, num_planes);
   2120    }
   2121  }
   2122  ref_frame_comp = av1_ref_frame_type(rf);
   2123  mbmi_ext->mode_context[ref_frame_comp] = 0;
   2124  mbmi_ext->ref_mv_count[ref_frame_comp] = UINT8_MAX;
   2125  av1_find_mv_refs(cm, xd, mbmi, ref_frame_comp, mbmi_ext->ref_mv_count,
   2126                   xd->ref_mv_stack, xd->weight, NULL, mbmi_ext->global_mvs,
   2127                   mbmi_ext->mode_context);
   2128  av1_copy_usable_ref_mv_stack_and_weight(xd, mbmi_ext, ref_frame_comp);
   2129  *ref_mv_idx = mbmi->ref_mv_idx + 1;
   2130 }
   2131 
   2132 static void set_compound_mode(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
   2133                              MV_REFERENCE_FRAME ref_frame2, int ref_mv_idx,
   2134                              int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],
   2135                              PREDICTION_MODE this_mode) {
   2136  MACROBLOCKD *const xd = &x->e_mbd;
   2137  MB_MODE_INFO *const mi = xd->mi[0];
   2138  mi->ref_frame[0] = ref_frame;
   2139  mi->ref_frame[1] = ref_frame2;
   2140  mi->compound_idx = 1;
   2141  mi->comp_group_idx = 0;
   2142  mi->interinter_comp.type = COMPOUND_AVERAGE;
   2143  MV_REFERENCE_FRAME ref_frame_comp = av1_ref_frame_type(mi->ref_frame);
   2144  if (this_mode == GLOBAL_GLOBALMV) {
   2145    frame_mv[this_mode][ref_frame].as_int = 0;
   2146    frame_mv[this_mode][ref_frame2].as_int = 0;
   2147  } else if (this_mode == NEAREST_NEARESTMV) {
   2148    frame_mv[this_mode][ref_frame].as_int =
   2149        xd->ref_mv_stack[ref_frame_comp][0].this_mv.as_int;
   2150    frame_mv[this_mode][ref_frame2].as_int =
   2151        xd->ref_mv_stack[ref_frame_comp][0].comp_mv.as_int;
   2152  } else if (this_mode == NEAR_NEARMV) {
   2153    frame_mv[this_mode][ref_frame].as_int =
   2154        xd->ref_mv_stack[ref_frame_comp][ref_mv_idx].this_mv.as_int;
   2155    frame_mv[this_mode][ref_frame2].as_int =
   2156        xd->ref_mv_stack[ref_frame_comp][ref_mv_idx].comp_mv.as_int;
   2157  }
   2158 }
   2159 
   2160 // Prune compound mode if the single mode variance is lower than a fixed
   2161 // percentage of the median value.
   2162 static bool skip_comp_based_on_var(
   2163    const unsigned int (*single_vars)[REF_FRAMES], BLOCK_SIZE bsize) {
   2164  unsigned int best_var = UINT_MAX;
   2165  for (int cur_mode_idx = 0; cur_mode_idx < RTC_INTER_MODES; cur_mode_idx++) {
   2166    for (int ref_idx = 0; ref_idx < REF_FRAMES; ref_idx++) {
   2167      best_var = AOMMIN(best_var, single_vars[cur_mode_idx][ref_idx]);
   2168    }
   2169  }
   2170  const unsigned int thresh_64 = (unsigned int)(0.57356805f * 8659);
   2171  const unsigned int thresh_32 = (unsigned int)(0.23964763f * 4281);
   2172 
   2173  // Currently, the thresh for 128 and 16 are not well-tuned. We are using the
   2174  // results from 64 and 32 as an heuristic.
   2175  switch (bsize) {
   2176    case BLOCK_128X128: return best_var < 4 * thresh_64;
   2177    case BLOCK_64X64: return best_var < thresh_64;
   2178    case BLOCK_32X32: return best_var < thresh_32;
   2179    case BLOCK_16X16: return best_var < thresh_32 / 4;
   2180    default: return false;
   2181  }
   2182 }
   2183 
   2184 static AOM_FORCE_INLINE void fill_single_inter_mode_costs(
   2185    int (*single_inter_mode_costs)[REF_FRAMES], int num_inter_modes,
   2186    const REF_MODE *reference_mode_set, const ModeCosts *mode_costs,
   2187    const int16_t *mode_context) {
   2188  bool ref_frame_used[REF_FRAMES] = { false };
   2189  for (int idx = 0; idx < num_inter_modes; idx++) {
   2190    ref_frame_used[reference_mode_set[idx].ref_frame] = true;
   2191  }
   2192 
   2193  for (int this_ref_frame = LAST_FRAME; this_ref_frame < REF_FRAMES;
   2194       this_ref_frame++) {
   2195    if (!ref_frame_used[this_ref_frame]) {
   2196      continue;
   2197    }
   2198 
   2199    const MV_REFERENCE_FRAME rf[2] = { this_ref_frame, NONE_FRAME };
   2200    const int16_t mode_ctx = av1_mode_context_analyzer(mode_context, rf);
   2201    for (PREDICTION_MODE this_mode = NEARESTMV; this_mode <= NEWMV;
   2202         this_mode++) {
   2203      single_inter_mode_costs[INTER_OFFSET(this_mode)][this_ref_frame] =
   2204          cost_mv_ref(mode_costs, this_mode, mode_ctx);
   2205    }
   2206  }
   2207 }
   2208 
   2209 static inline bool is_globalmv_better(
   2210    PREDICTION_MODE this_mode, MV_REFERENCE_FRAME ref_frame, int rate_mv,
   2211    const ModeCosts *mode_costs,
   2212    const int (*single_inter_mode_costs)[REF_FRAMES],
   2213    const MB_MODE_INFO_EXT *mbmi_ext) {
   2214  const int globalmv_mode_cost =
   2215      single_inter_mode_costs[INTER_OFFSET(GLOBALMV)][ref_frame];
   2216  int this_mode_cost =
   2217      rate_mv + single_inter_mode_costs[INTER_OFFSET(this_mode)][ref_frame];
   2218  if (this_mode == NEWMV || this_mode == NEARMV) {
   2219    const MV_REFERENCE_FRAME rf[2] = { ref_frame, NONE_FRAME };
   2220    this_mode_cost += get_drl_cost(
   2221        NEWMV, 0, mbmi_ext, mode_costs->drl_mode_cost0, av1_ref_frame_type(rf));
   2222  }
   2223  return this_mode_cost > globalmv_mode_cost;
   2224 }
   2225 
   2226 // Set up the mv/ref_frames etc based on the comp_index. Returns 1 if it
   2227 // succeeds, 0 if it fails.
   2228 static inline int setup_compound_params_from_comp_idx(
   2229    const AV1_COMP *cpi, MACROBLOCK *x, struct buf_2d yv12_mb[8][MAX_MB_PLANE],
   2230    PREDICTION_MODE *this_mode, MV_REFERENCE_FRAME *ref_frame,
   2231    MV_REFERENCE_FRAME *ref_frame2, int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],
   2232    const int *use_ref_frame_mask, int comp_index,
   2233    bool comp_use_zero_zeromv_only, MV_REFERENCE_FRAME *last_comp_ref_frame,
   2234    BLOCK_SIZE bsize) {
   2235  const MV_REFERENCE_FRAME *rf = comp_ref_mode_set[comp_index].ref_frame;
   2236  int skip_gf = 0;
   2237  int skip_alt = 0;
   2238  *this_mode = comp_ref_mode_set[comp_index].pred_mode;
   2239  *ref_frame = rf[0];
   2240  *ref_frame2 = rf[1];
   2241  assert(*ref_frame == LAST_FRAME);
   2242  assert(*this_mode == GLOBAL_GLOBALMV || *this_mode == NEAREST_NEARESTMV);
   2243  if (x->source_variance < 50 && bsize > BLOCK_16X16) {
   2244    if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
   2245        x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
   2246      skip_gf = 1;
   2247    if (x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
   2248        x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
   2249      skip_alt = 1;
   2250  }
   2251  if (comp_use_zero_zeromv_only && *this_mode != GLOBAL_GLOBALMV) {
   2252    return 0;
   2253  }
   2254  if (*ref_frame2 == GOLDEN_FRAME &&
   2255      (cpi->sf.rt_sf.ref_frame_comp_nonrd[0] == 0 || skip_gf ||
   2256       !(cpi->ref_frame_flags & AOM_GOLD_FLAG))) {
   2257    return 0;
   2258  } else if (*ref_frame2 == LAST2_FRAME &&
   2259             (cpi->sf.rt_sf.ref_frame_comp_nonrd[1] == 0 ||
   2260              !(cpi->ref_frame_flags & AOM_LAST2_FLAG))) {
   2261    return 0;
   2262  } else if (*ref_frame2 == ALTREF_FRAME &&
   2263             (cpi->sf.rt_sf.ref_frame_comp_nonrd[2] == 0 || skip_alt ||
   2264              !(cpi->ref_frame_flags & AOM_ALT_FLAG))) {
   2265    return 0;
   2266  }
   2267  int ref_mv_idx = 0;
   2268  if (*last_comp_ref_frame != rf[1]) {
   2269    // Only needs to be done once per reference pair.
   2270    setup_compound_prediction(&cpi->common, x, yv12_mb, use_ref_frame_mask, rf,
   2271                              &ref_mv_idx);
   2272    *last_comp_ref_frame = rf[1];
   2273  }
   2274  set_compound_mode(x, *ref_frame, *ref_frame2, ref_mv_idx, frame_mv,
   2275                    *this_mode);
   2276  if (*this_mode != GLOBAL_GLOBALMV &&
   2277      frame_mv[*this_mode][*ref_frame].as_int == 0 &&
   2278      frame_mv[*this_mode][*ref_frame2].as_int == 0) {
   2279    return 0;
   2280  }
   2281 
   2282  return 1;
   2283 }
   2284 
   2285 static inline bool previous_mode_performed_poorly(
   2286    PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame,
   2287    const unsigned int (*vars)[REF_FRAMES],
   2288    const int64_t (*uv_dist)[REF_FRAMES]) {
   2289  unsigned int best_var = UINT_MAX;
   2290  int64_t best_uv_dist = INT64_MAX;
   2291  for (int midx = 0; midx < RTC_INTER_MODES; midx++) {
   2292    best_var = AOMMIN(best_var, vars[midx][ref_frame]);
   2293    best_uv_dist = AOMMIN(best_uv_dist, uv_dist[midx][ref_frame]);
   2294  }
   2295  assert(best_var != UINT_MAX && "Invalid variance data.");
   2296  const float mult = 1.125f;
   2297  bool var_bad = mult * best_var < vars[INTER_OFFSET(mode)][ref_frame];
   2298  if (uv_dist[INTER_OFFSET(mode)][ref_frame] < INT64_MAX &&
   2299      best_uv_dist != uv_dist[INTER_OFFSET(mode)][ref_frame]) {
   2300    // If we have chroma info, then take it into account
   2301    var_bad &= mult * best_uv_dist < uv_dist[INTER_OFFSET(mode)][ref_frame];
   2302  }
   2303  return var_bad;
   2304 }
   2305 
   2306 static inline bool prune_compoundmode_with_singlemode_var(
   2307    PREDICTION_MODE compound_mode, MV_REFERENCE_FRAME ref_frame,
   2308    MV_REFERENCE_FRAME ref_frame2, const int_mv (*frame_mv)[REF_FRAMES],
   2309    const uint8_t (*mode_checked)[REF_FRAMES],
   2310    const unsigned int (*vars)[REF_FRAMES],
   2311    const int64_t (*uv_dist)[REF_FRAMES]) {
   2312  const PREDICTION_MODE single_mode0 = compound_ref0_mode(compound_mode);
   2313  const PREDICTION_MODE single_mode1 = compound_ref1_mode(compound_mode);
   2314 
   2315  bool first_ref_valid = false, second_ref_valid = false;
   2316  bool first_ref_bad = false, second_ref_bad = false;
   2317  if (mode_checked[single_mode0][ref_frame] &&
   2318      frame_mv[single_mode0][ref_frame].as_int ==
   2319          frame_mv[compound_mode][ref_frame].as_int &&
   2320      vars[INTER_OFFSET(single_mode0)][ref_frame] < UINT_MAX) {
   2321    first_ref_valid = true;
   2322    first_ref_bad =
   2323        previous_mode_performed_poorly(single_mode0, ref_frame, vars, uv_dist);
   2324  }
   2325  if (mode_checked[single_mode1][ref_frame2] &&
   2326      frame_mv[single_mode1][ref_frame2].as_int ==
   2327          frame_mv[compound_mode][ref_frame2].as_int &&
   2328      vars[INTER_OFFSET(single_mode1)][ref_frame2] < UINT_MAX) {
   2329    second_ref_valid = true;
   2330    second_ref_bad =
   2331        previous_mode_performed_poorly(single_mode1, ref_frame2, vars, uv_dist);
   2332  }
   2333  if (first_ref_valid && second_ref_valid) {
   2334    return first_ref_bad && second_ref_bad;
   2335  } else if (first_ref_valid || second_ref_valid) {
   2336    return first_ref_bad || second_ref_bad;
   2337  }
   2338  return false;
   2339 }
   2340 
   2341 // Function to setup parameters used for inter mode evaluation in non-rd.
   2342 static AOM_FORCE_INLINE void set_params_nonrd_pick_inter_mode(
   2343    AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
   2344    RD_STATS *rd_cost, int *force_skip_low_temp_var, int mi_row, int mi_col,
   2345    int gf_temporal_ref, unsigned char segment_id, BLOCK_SIZE bsize
   2346 #if CONFIG_AV1_TEMPORAL_DENOISING
   2347    ,
   2348    PICK_MODE_CONTEXT *ctx, int denoise_svc_pickmode
   2349 #endif
   2350 ) {
   2351  AV1_COMMON *const cm = &cpi->common;
   2352  MACROBLOCKD *const xd = &x->e_mbd;
   2353  TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   2354  MB_MODE_INFO *const mi = xd->mi[0];
   2355  const ModeCosts *mode_costs = &x->mode_costs;
   2356  int skip_pred_mv = 0;
   2357 
   2358  // Initialize variance and distortion (chroma) for all modes and reference
   2359  // frames
   2360  for (int idx = 0; idx < RTC_INTER_MODES; idx++) {
   2361    for (int ref = 0; ref < REF_FRAMES; ref++) {
   2362      search_state->vars[idx][ref] = UINT_MAX;
   2363      search_state->uv_dist[idx][ref] = INT64_MAX;
   2364    }
   2365  }
   2366 
   2367  // Initialize values of color sensitivity with sb level color sensitivity
   2368  av1_copy(x->color_sensitivity, x->color_sensitivity_sb);
   2369 
   2370  init_best_pickmode(&search_state->best_pickmode);
   2371 
   2372  // Estimate cost for single reference frames
   2373  estimate_single_ref_frame_costs(cm, xd, mode_costs, segment_id, bsize,
   2374                                  search_state->ref_costs_single);
   2375 
   2376  // Reset flag to indicate modes evaluated
   2377  av1_zero(search_state->mode_checked);
   2378 
   2379  txfm_info->skip_txfm = 0;
   2380 
   2381  // Initialize mode decisions
   2382  av1_invalid_rd_stats(&search_state->best_rdc);
   2383  av1_invalid_rd_stats(&search_state->this_rdc);
   2384  av1_invalid_rd_stats(rd_cost);
   2385  for (int ref_idx = 0; ref_idx < REF_FRAMES; ++ref_idx) {
   2386    x->warp_sample_info[ref_idx].num = -1;
   2387  }
   2388 
   2389  mi->bsize = bsize;
   2390  mi->ref_frame[0] = NONE_FRAME;
   2391  mi->ref_frame[1] = NONE_FRAME;
   2392 
   2393 #if CONFIG_AV1_TEMPORAL_DENOISING
   2394  if (cpi->oxcf.noise_sensitivity > 0) {
   2395    // if (cpi->ppi->use_svc) denoise_svc_pickmode =
   2396    // av1_denoise_svc_non_key(cpi);
   2397    if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
   2398      av1_denoiser_reset_frame_stats(ctx);
   2399  }
   2400 #endif
   2401 
   2402  // Populate predicated motion vectors for LAST_FRAME
   2403  if (cpi->ref_frame_flags & AOM_LAST_FLAG) {
   2404    find_predictors(cpi, x, LAST_FRAME, search_state->frame_mv,
   2405                    search_state->yv12_mb, bsize, *force_skip_low_temp_var,
   2406                    x->force_zeromv_skip_for_blk,
   2407                    &search_state->use_scaled_ref_frame[LAST_FRAME]);
   2408  }
   2409  // Update mask to use all reference frame
   2410  get_ref_frame_use_mask(cpi, x, mi, mi_row, mi_col, bsize, gf_temporal_ref,
   2411                         search_state->use_ref_frame_mask,
   2412                         force_skip_low_temp_var);
   2413 
   2414  skip_pred_mv = x->force_zeromv_skip_for_blk ||
   2415                 (x->nonrd_prune_ref_frame_search > 2 &&
   2416                  x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] != 2 &&
   2417                  x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] != 2);
   2418 
   2419  // Populate predicated motion vectors for other single reference frame
   2420  // Start at LAST_FRAME + 1.
   2421  for (MV_REFERENCE_FRAME ref_frame_iter = LAST_FRAME + 1;
   2422       ref_frame_iter <= ALTREF_FRAME; ++ref_frame_iter) {
   2423    if (search_state->use_ref_frame_mask[ref_frame_iter]) {
   2424      find_predictors(cpi, x, ref_frame_iter, search_state->frame_mv,
   2425                      search_state->yv12_mb, bsize, *force_skip_low_temp_var,
   2426                      skip_pred_mv,
   2427                      &search_state->use_scaled_ref_frame[ref_frame_iter]);
   2428    }
   2429  }
   2430 }
   2431 
   2432 // Function to check the inter mode can be skipped based on mode statistics and
   2433 // speed features settings.
   2434 static AOM_FORCE_INLINE bool skip_inter_mode_nonrd(
   2435    AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
   2436    int64_t *thresh_sad_pred, int *force_mv_inter_layer, int *is_single_pred,
   2437    PREDICTION_MODE *this_mode, MV_REFERENCE_FRAME *last_comp_ref_frame,
   2438    MV_REFERENCE_FRAME *ref_frame, MV_REFERENCE_FRAME *ref_frame2, int idx,
   2439    int_mv svc_mv, int force_skip_low_temp_var, unsigned int sse_zeromv_norm,
   2440    int num_inter_modes, unsigned char segment_id, BLOCK_SIZE bsize,
   2441    bool comp_use_zero_zeromv_only, bool check_globalmv) {
   2442  AV1_COMMON *const cm = &cpi->common;
   2443  const struct segmentation *const seg = &cm->seg;
   2444  const SVC *const svc = &cpi->svc;
   2445  MACROBLOCKD *const xd = &x->e_mbd;
   2446  MB_MODE_INFO *const mi = xd->mi[0];
   2447  const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
   2448 
   2449  // Skip compound mode based on reference frame mask and type of the mode and
   2450  // for allowed compound modes, setup ref mv stack and reference frame.
   2451  if (idx >= num_inter_modes) {
   2452    const int comp_index = idx - num_inter_modes;
   2453    if (!setup_compound_params_from_comp_idx(
   2454            cpi, x, search_state->yv12_mb, this_mode, ref_frame, ref_frame2,
   2455            search_state->frame_mv, search_state->use_ref_frame_mask,
   2456            comp_index, comp_use_zero_zeromv_only, last_comp_ref_frame,
   2457            bsize)) {
   2458      return true;
   2459    }
   2460    *is_single_pred = 0;
   2461  } else {
   2462    *this_mode = ref_mode_set[idx].pred_mode;
   2463    *ref_frame = ref_mode_set[idx].ref_frame;
   2464    *ref_frame2 = NONE_FRAME;
   2465  }
   2466 
   2467  if (cpi->sf.rt_sf.skip_newmv_mode_sad_screen && cpi->rc.high_source_sad &&
   2468      x->content_state_sb.source_sad_nonrd >= kMedSad && bsize <= BLOCK_16X16 &&
   2469      !x->sb_me_block && (*ref_frame != LAST_FRAME || *this_mode == NEWMV))
   2470    return true;
   2471 
   2472  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP) &&
   2473      (*this_mode != GLOBALMV || *ref_frame != LAST_FRAME))
   2474    return true;
   2475 
   2476  // If the segment reference frame feature is enabled then do nothing if the
   2477  // current ref frame is not allowed.
   2478  if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
   2479    if (get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)(*ref_frame))
   2480      return true;
   2481    return false;
   2482  }
   2483 
   2484  // Skip the mode if use reference frame mask flag is not set.
   2485  if (!search_state->use_ref_frame_mask[*ref_frame]) return true;
   2486 
   2487  // Don't skip non_last references if LAST is not used a reference.
   2488  if (!(cpi->ref_frame_flags & AOM_LAST_FLAG) &&
   2489      (*ref_frame == GOLDEN_FRAME || *ref_frame == ALTREF_FRAME))
   2490    return false;
   2491 
   2492  // Skip mode for some modes and reference frames when
   2493  // force_zeromv_skip_for_blk flag is true.
   2494  if (x->force_zeromv_skip_for_blk &&
   2495      ((!(*this_mode == NEARESTMV &&
   2496          search_state->frame_mv[*this_mode][*ref_frame].as_int == 0) &&
   2497        *this_mode != GLOBALMV) ||
   2498       *ref_frame != LAST_FRAME))
   2499    return true;
   2500 
   2501  if (x->sb_me_block && *ref_frame == LAST_FRAME) {
   2502    // We want to make sure to test the superblock MV:
   2503    // so don't skip (return false) for NEAREST_LAST or NEAR_LAST if they
   2504    // have this sb MV. And don't skip NEWMV_LAST: this will be set to
   2505    // sb MV in handle_inter_mode_nonrd(), in case NEAREST or NEAR don't
   2506    // have it.
   2507    if (*this_mode == NEARESTMV &&
   2508        search_state->frame_mv[NEARESTMV][LAST_FRAME].as_int ==
   2509            x->sb_me_mv.as_int) {
   2510      return false;
   2511    }
   2512    if (*this_mode == NEARMV &&
   2513        search_state->frame_mv[NEARMV][LAST_FRAME].as_int ==
   2514            x->sb_me_mv.as_int) {
   2515      return false;
   2516    }
   2517    if (*this_mode == NEWMV) {
   2518      return false;
   2519    }
   2520  }
   2521 
   2522  // Skip the single reference mode for which mode check flag is set.
   2523  if (*is_single_pred && search_state->mode_checked[*this_mode][*ref_frame]) {
   2524    return true;
   2525  }
   2526 
   2527  // Skip GLOBALMV mode if check_globalmv flag is not enabled.
   2528  if (!check_globalmv && *this_mode == GLOBALMV) {
   2529    return true;
   2530  }
   2531 
   2532 #if COLLECT_NONRD_PICK_MODE_STAT
   2533  aom_usec_timer_start(&x->ms_stat_nonrd.timer1);
   2534  x->ms_stat_nonrd.num_searches[bsize][*this_mode]++;
   2535 #endif
   2536  mi->mode = *this_mode;
   2537  mi->ref_frame[0] = *ref_frame;
   2538  mi->ref_frame[1] = *ref_frame2;
   2539 
   2540  // Skip compound mode based on variance of previously evaluated single
   2541  // reference modes.
   2542  if (rt_sf->prune_compoundmode_with_singlemode_var && !*is_single_pred &&
   2543      prune_compoundmode_with_singlemode_var(
   2544          *this_mode, *ref_frame, *ref_frame2, search_state->frame_mv,
   2545          search_state->mode_checked, search_state->vars,
   2546          search_state->uv_dist)) {
   2547    return true;
   2548  }
   2549 
   2550  *force_mv_inter_layer = 0;
   2551  if (cpi->ppi->use_svc && svc->spatial_layer_id > 0 &&
   2552      ((*ref_frame == LAST_FRAME && svc->skip_mvsearch_last) ||
   2553       (*ref_frame == GOLDEN_FRAME && svc->skip_mvsearch_gf) ||
   2554       (*ref_frame == ALTREF_FRAME && svc->skip_mvsearch_altref))) {
   2555    // Only test mode if NEARESTMV/NEARMV is (svc_mv.mv.col, svc_mv.mv.row),
   2556    // otherwise set NEWMV to (svc_mv.mv.col, svc_mv.mv.row).
   2557    // Skip newmv and filter search.
   2558    *force_mv_inter_layer = 1;
   2559    if (*this_mode == NEWMV) {
   2560      search_state->frame_mv[*this_mode][*ref_frame] = svc_mv;
   2561    } else if (search_state->frame_mv[*this_mode][*ref_frame].as_int !=
   2562               svc_mv.as_int) {
   2563      return true;
   2564    }
   2565  }
   2566 
   2567  // For screen content: skip mode testing based on source_sad.
   2568  if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   2569      !x->force_zeromv_skip_for_blk) {
   2570    // If source_sad is computed: skip non-zero motion
   2571    // check for stationary (super)blocks. Otherwise if superblock
   2572    // has motion skip the modes with zero motion on last reference
   2573    // for flat blocks, and color is not set.
   2574    // For the latter condition: the same condition should apply
   2575    // to newmv if (0, 0), so this latter condition is repeated
   2576    // below after search_new_mv.
   2577    if (rt_sf->source_metrics_sb_nonrd) {
   2578      if ((search_state->frame_mv[*this_mode][*ref_frame].as_int != 0 &&
   2579           x->content_state_sb.source_sad_nonrd == kZeroSad) ||
   2580          (search_state->frame_mv[*this_mode][*ref_frame].as_int == 0 &&
   2581           x->block_is_zero_sad == 0 && *ref_frame == LAST_FRAME &&
   2582           ((x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
   2583             x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) ||
   2584            cpi->rc.high_source_sad) &&
   2585           x->source_variance == 0))
   2586        return true;
   2587    }
   2588    // Skip NEWMV search for flat blocks.
   2589    if (rt_sf->skip_newmv_flat_blocks_screen && *this_mode == NEWMV &&
   2590        x->source_variance < 100)
   2591      return true;
   2592    // Skip non-LAST for color on flat blocks.
   2593    if (*ref_frame > LAST_FRAME && x->source_variance == 0 &&
   2594        (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
   2595         x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 1))
   2596      return true;
   2597  }
   2598 
   2599  // Skip mode based on block size, reference frame mode and other block
   2600  // properties.
   2601  if (skip_mode_by_bsize_and_ref_frame(
   2602          *this_mode, *ref_frame, bsize, x->nonrd_prune_ref_frame_search,
   2603          sse_zeromv_norm, rt_sf->nonrd_aggressive_skip,
   2604          rt_sf->increase_source_sad_thresh))
   2605    return true;
   2606 
   2607  // Skip mode based on low temporal variance and souce sad.
   2608  if (skip_mode_by_low_temp(*this_mode, *ref_frame, bsize, x->content_state_sb,
   2609                            search_state->frame_mv[*this_mode][*ref_frame],
   2610                            force_skip_low_temp_var))
   2611    return true;
   2612 
   2613  // Disable this drop out case if the ref frame segment level feature is
   2614  // enabled for this segment. This is to prevent the possibility that we
   2615  // end up unable to pick any mode.
   2616  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
   2617    // Check for skipping GOLDEN and ALTREF based pred_mv_sad.
   2618    if (rt_sf->nonrd_prune_ref_frame_search > 0 &&
   2619        x->pred_mv_sad[*ref_frame] != INT_MAX && *ref_frame != LAST_FRAME) {
   2620      if ((int64_t)(x->pred_mv_sad[*ref_frame]) > *thresh_sad_pred) return true;
   2621    }
   2622  }
   2623 
   2624  // Check for skipping NEARMV based on pred_mv_sad.
   2625  if (*this_mode == NEARMV && x->pred_mv1_sad[*ref_frame] != INT_MAX &&
   2626      x->pred_mv1_sad[*ref_frame] > (x->pred_mv0_sad[*ref_frame] << 1))
   2627    return true;
   2628 
   2629  // Skip single reference mode based on rd threshold.
   2630  if (*is_single_pred) {
   2631    if (skip_mode_by_threshold(
   2632            *this_mode, *ref_frame,
   2633            search_state->frame_mv[*this_mode][*ref_frame],
   2634            cpi->rc.frames_since_golden, cpi->rd.threshes[segment_id][bsize],
   2635            x->thresh_freq_fact[bsize], search_state->best_rdc.rdcost,
   2636            search_state->best_pickmode.best_mode_skip_txfm,
   2637            (rt_sf->nonrd_aggressive_skip ? 1 : 0)))
   2638      return true;
   2639  }
   2640  return false;
   2641 }
   2642 
   2643 // Function to perform inter mode evaluation for non-rd
   2644 static AOM_FORCE_INLINE bool handle_inter_mode_nonrd(
   2645    AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
   2646    PICK_MODE_CONTEXT *ctx, PRED_BUFFER **this_mode_pred,
   2647    PRED_BUFFER *tmp_buffer, InterPredParams inter_pred_params_sr,
   2648    int *best_early_term, unsigned int *sse_zeromv_norm, bool *check_globalmv,
   2649 #if CONFIG_AV1_TEMPORAL_DENOISING
   2650    int64_t *zero_last_cost_orig, int denoise_svc_pickmode,
   2651 #endif
   2652    int idx, int force_mv_inter_layer, int is_single_pred, int gf_temporal_ref,
   2653    int use_model_yrd_large, int filter_search_enabled_blk, BLOCK_SIZE bsize,
   2654    PREDICTION_MODE this_mode, InterpFilter filt_select,
   2655    int cb_pred_filter_search, int reuse_inter_pred,
   2656    int *sb_me_has_been_tested) {
   2657  AV1_COMMON *const cm = &cpi->common;
   2658  MACROBLOCKD *const xd = &x->e_mbd;
   2659  MB_MODE_INFO *const mi = xd->mi[0];
   2660  const MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
   2661  const int mi_row = xd->mi_row;
   2662  const int mi_col = xd->mi_col;
   2663  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
   2664  const int bw = block_size_wide[bsize];
   2665  const InterpFilter filter_ref = cm->features.interp_filter;
   2666  const InterpFilter default_interp_filter = EIGHTTAP_REGULAR;
   2667  TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   2668  const ModeCosts *mode_costs = &x->mode_costs;
   2669  const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
   2670  BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
   2671 
   2672  MV_REFERENCE_FRAME ref_frame = mi->ref_frame[0];
   2673  MV_REFERENCE_FRAME ref_frame2 = mi->ref_frame[1];
   2674  int_mv *const this_mv = &search_state->frame_mv[this_mode][ref_frame];
   2675  unsigned int var = UINT_MAX;
   2676  int this_early_term = 0;
   2677  int rate_mv = 0;
   2678  int is_skippable;
   2679  int skip_this_mv = 0;
   2680  unsigned int var_threshold = UINT_MAX;
   2681  PREDICTION_MODE this_best_mode;
   2682  RD_STATS nonskip_rdc;
   2683  av1_invalid_rd_stats(&nonskip_rdc);
   2684 
   2685  if (x->sb_me_block && this_mode == NEWMV && ref_frame == LAST_FRAME) {
   2686    // Set the NEWMV_LAST to the sb MV.
   2687    search_state->frame_mv[NEWMV][LAST_FRAME].as_int = x->sb_me_mv.as_int;
   2688  } else if (this_mode == NEWMV && !force_mv_inter_layer) {
   2689 #if COLLECT_NONRD_PICK_MODE_STAT
   2690    aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
   2691 #endif
   2692    // Find the best motion vector for single/compound mode.
   2693    const bool skip_newmv = search_new_mv(
   2694        cpi, x, search_state->frame_mv, ref_frame, gf_temporal_ref, bsize,
   2695        mi_row, mi_col, &rate_mv, &search_state->best_rdc);
   2696 #if COLLECT_NONRD_PICK_MODE_STAT
   2697    aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
   2698    x->ms_stat_nonrd.ms_time[bsize][this_mode] +=
   2699        aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
   2700 #endif
   2701    // Skip NEWMV mode,
   2702    //   (i). For bsize smaller than 16X16
   2703    //  (ii). Based on sad of the predicted mv w.r.t LAST_FRAME
   2704    // (iii). When motion vector is same as that of reference mv
   2705    if (skip_newmv) {
   2706      return true;
   2707    }
   2708  }
   2709 
   2710  // Check the current motion vector is same as that of previously evaluated
   2711  // motion vectors.
   2712  for (PREDICTION_MODE inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV;
   2713       inter_mv_mode++) {
   2714    if (inter_mv_mode == this_mode) continue;
   2715    if (is_single_pred &&
   2716        search_state->mode_checked[inter_mv_mode][ref_frame] &&
   2717        this_mv->as_int ==
   2718            search_state->frame_mv[inter_mv_mode][ref_frame].as_int) {
   2719      skip_this_mv = 1;
   2720      break;
   2721    }
   2722  }
   2723 
   2724  // Skip single mode if current motion vector is same that of previously
   2725  // evaluated motion vectors.
   2726  if (skip_this_mv && is_single_pred) return true;
   2727 
   2728  // For screen: for spatially flat blocks with non-zero motion,
   2729  // skip newmv if the motion vector is (0, 0)-LAST, and color is not set.
   2730  if (this_mode == NEWMV && cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   2731      cpi->svc.spatial_layer_id == 0 && rt_sf->source_metrics_sb_nonrd) {
   2732    if (this_mv->as_int == 0 && ref_frame == LAST_FRAME &&
   2733        x->block_is_zero_sad == 0 &&
   2734        ((x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
   2735          x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) ||
   2736         cpi->rc.high_source_sad) &&
   2737        x->source_variance == 0)
   2738      return true;
   2739  }
   2740 
   2741  mi->mode = this_mode;
   2742  mi->mv[0].as_int = this_mv->as_int;
   2743  mi->mv[1].as_int = 0;
   2744  if (!is_single_pred)
   2745    mi->mv[1].as_int = search_state->frame_mv[this_mode][ref_frame2].as_int;
   2746 
   2747  // Set buffers to store predicted samples for reuse
   2748  if (reuse_inter_pred) {
   2749    if (!*this_mode_pred) {
   2750      *this_mode_pred = &tmp_buffer[3];
   2751    } else {
   2752      *this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
   2753      pd->dst.buf = (*this_mode_pred)->data;
   2754      pd->dst.stride = bw;
   2755    }
   2756  }
   2757 
   2758  mi->motion_mode = SIMPLE_TRANSLATION;
   2759 #if !CONFIG_REALTIME_ONLY
   2760  if (cpi->oxcf.motion_mode_cfg.allow_warped_motion) {
   2761    calc_num_proj_ref(cpi, x, mi);
   2762  }
   2763 #endif
   2764  // set variance threshold for compound mode pruning
   2765  if (rt_sf->prune_compoundmode_with_singlecompound_var && !is_single_pred &&
   2766      use_model_yrd_large) {
   2767    const PREDICTION_MODE single_mode0 = compound_ref0_mode(this_mode);
   2768    const PREDICTION_MODE single_mode1 = compound_ref1_mode(this_mode);
   2769    var_threshold =
   2770        AOMMIN(var_threshold,
   2771               search_state->vars[INTER_OFFSET(single_mode0)][ref_frame]);
   2772    var_threshold =
   2773        AOMMIN(var_threshold,
   2774               search_state->vars[INTER_OFFSET(single_mode1)][ref_frame2]);
   2775  }
   2776 
   2777  // decide interpolation filter, build prediction signal, get sse
   2778  const bool is_mv_subpel =
   2779      (mi->mv[0].as_mv.row & 0x07) || (mi->mv[0].as_mv.col & 0x07);
   2780  const bool enable_filt_search_this_mode =
   2781      (filter_search_enabled_blk == 2)
   2782          ? true
   2783          : (filter_search_enabled_blk && !force_mv_inter_layer &&
   2784             is_single_pred &&
   2785             (ref_frame == LAST_FRAME || !x->nonrd_prune_ref_frame_search));
   2786  if (is_mv_subpel && enable_filt_search_this_mode) {
   2787 #if COLLECT_NONRD_PICK_MODE_STAT
   2788    aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
   2789 #endif
   2790    search_filter_ref(
   2791        cpi, x, &search_state->this_rdc, &inter_pred_params_sr, mi_row, mi_col,
   2792        tmp_buffer, bsize, reuse_inter_pred, this_mode_pred, &this_early_term,
   2793        &var, use_model_yrd_large, best_pickmode->best_sse, is_single_pred);
   2794 #if COLLECT_NONRD_PICK_MODE_STAT
   2795    aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
   2796    x->ms_stat_nonrd.ifs_time[bsize][this_mode] +=
   2797        aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
   2798 #endif
   2799 #if !CONFIG_REALTIME_ONLY
   2800  } else if (cpi->oxcf.motion_mode_cfg.allow_warped_motion &&
   2801             this_mode == NEWMV) {
   2802    // Find the best motion mode when current mode is NEWMV
   2803    search_motion_mode(cpi, x, &search_state->this_rdc, mi_row, mi_col, bsize,
   2804                       &this_early_term, use_model_yrd_large, &rate_mv,
   2805                       best_pickmode->best_sse);
   2806    if (this_mode == NEWMV) {
   2807      this_mv[0] = mi->mv[0];
   2808    }
   2809 #endif
   2810  } else {
   2811    mi->interp_filters =
   2812        (filter_ref == SWITCHABLE)
   2813            ? av1_broadcast_interp_filter(default_interp_filter)
   2814            : av1_broadcast_interp_filter(filter_ref);
   2815    if (force_mv_inter_layer)
   2816      mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
   2817 
   2818    // If it is sub-pel motion and cb_pred_filter_search is enabled, select
   2819    // the pre-decided filter
   2820    if (is_mv_subpel && cb_pred_filter_search)
   2821      mi->interp_filters = av1_broadcast_interp_filter(filt_select);
   2822 
   2823 #if COLLECT_NONRD_PICK_MODE_STAT
   2824    aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
   2825 #endif
   2826    if (is_single_pred) {
   2827      SubpelParams subpel_params;
   2828      // Initialize inter mode level params for single reference mode.
   2829      init_inter_mode_params(&mi->mv[0].as_mv, &inter_pred_params_sr,
   2830                             &subpel_params, xd->block_ref_scale_factors[0],
   2831                             pd->pre->width, pd->pre->height);
   2832      av1_enc_build_inter_predictor_y_nonrd(xd, &inter_pred_params_sr,
   2833                                            &subpel_params);
   2834    } else {
   2835      av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   2836                                    AOM_PLANE_Y, AOM_PLANE_Y);
   2837    }
   2838 
   2839    if (use_model_yrd_large) {
   2840      model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
   2841                                &search_state->this_rdc, &this_early_term, 0,
   2842                                best_pickmode->best_sse, &var, var_threshold);
   2843    } else {
   2844      model_rd_for_sb_y(cpi, bsize, x, xd, &search_state->this_rdc, &var, 0,
   2845                        &this_early_term);
   2846    }
   2847 #if COLLECT_NONRD_PICK_MODE_STAT
   2848    aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
   2849    x->ms_stat_nonrd.model_rd_time[bsize][this_mode] +=
   2850        aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
   2851 #endif
   2852  }
   2853 
   2854  // update variance for single mode
   2855  if (is_single_pred) {
   2856    search_state->vars[INTER_OFFSET(this_mode)][ref_frame] = var;
   2857    if (this_mv->as_int == 0) {
   2858      search_state->vars[INTER_OFFSET(GLOBALMV)][ref_frame] = var;
   2859    }
   2860  }
   2861  // prune compound mode based on single mode var threshold
   2862  if (!is_single_pred && var > var_threshold) {
   2863    if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
   2864    return true;
   2865  }
   2866 
   2867  if (ref_frame == LAST_FRAME && this_mv->as_int == 0) {
   2868    *sse_zeromv_norm = (unsigned int)(search_state->this_rdc.sse >>
   2869                                      (b_width_log2_lookup[bsize] +
   2870                                       b_height_log2_lookup[bsize]));
   2871  }
   2872 
   2873  // Perform early termination based on sse.
   2874  if (rt_sf->sse_early_term_inter_search &&
   2875      early_term_inter_search_with_sse(rt_sf->sse_early_term_inter_search,
   2876                                       bsize, search_state->this_rdc.sse,
   2877                                       best_pickmode->best_sse, this_mode)) {
   2878    if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
   2879    return true;
   2880  }
   2881 
   2882 #if COLLECT_NONRD_PICK_MODE_STAT
   2883  x->ms_stat_nonrd.num_nonskipped_searches[bsize][this_mode]++;
   2884 #endif
   2885 
   2886  const int skip_ctx = av1_get_skip_txfm_context(xd);
   2887  const int skip_txfm_cost = mode_costs->skip_txfm_cost[skip_ctx][1];
   2888  const int no_skip_txfm_cost = mode_costs->skip_txfm_cost[skip_ctx][0];
   2889  const int64_t sse_y = search_state->this_rdc.sse;
   2890 
   2891  if (this_early_term) {
   2892    search_state->this_rdc.skip_txfm = 1;
   2893    search_state->this_rdc.rate = skip_txfm_cost;
   2894    search_state->this_rdc.dist = search_state->this_rdc.sse << 4;
   2895  } else {
   2896 #if COLLECT_NONRD_PICK_MODE_STAT
   2897    aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
   2898 #endif
   2899    // Calculates RD Cost using Hadamard transform.
   2900    av1_block_yrd(x, &search_state->this_rdc, &is_skippable, bsize,
   2901                  mi->tx_size);
   2902    if (search_state->this_rdc.skip_txfm ||
   2903        RDCOST(x->rdmult, search_state->this_rdc.rate,
   2904               search_state->this_rdc.dist) >=
   2905            RDCOST(x->rdmult, 0, search_state->this_rdc.sse)) {
   2906      if (!search_state->this_rdc.skip_txfm) {
   2907        // Need to store "real" rdc for possible future use if UV rdc
   2908        // disallows tx skip
   2909        nonskip_rdc = search_state->this_rdc;
   2910        nonskip_rdc.rate += no_skip_txfm_cost;
   2911      }
   2912      search_state->this_rdc.rate = skip_txfm_cost;
   2913      search_state->this_rdc.skip_txfm = 1;
   2914      search_state->this_rdc.dist = search_state->this_rdc.sse;
   2915    } else {
   2916      search_state->this_rdc.rate += no_skip_txfm_cost;
   2917    }
   2918 
   2919    // Populate predicted sample for chroma planes based on color sensitivity.
   2920    if ((x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
   2921         x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])) {
   2922      RD_STATS rdc_uv;
   2923      const BLOCK_SIZE uv_bsize =
   2924          get_plane_block_size(bsize, xd->plane[AOM_PLANE_U].subsampling_x,
   2925                               xd->plane[AOM_PLANE_U].subsampling_y);
   2926      if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)]) {
   2927        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   2928                                      AOM_PLANE_U, AOM_PLANE_U);
   2929      }
   2930      if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) {
   2931        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   2932                                      AOM_PLANE_V, AOM_PLANE_V);
   2933      }
   2934      // Compute sse for chroma planes.
   2935      const int64_t sse_uv = av1_model_rd_for_sb_uv(
   2936          cpi, uv_bsize, x, xd, &rdc_uv, AOM_PLANE_U, AOM_PLANE_V);
   2937      if (rdc_uv.dist < x->min_dist_inter_uv)
   2938        x->min_dist_inter_uv = rdc_uv.dist;
   2939      search_state->this_rdc.sse += sse_uv;
   2940      // Restore Y rdc if UV rdc disallows txfm skip
   2941      if (search_state->this_rdc.skip_txfm && !rdc_uv.skip_txfm &&
   2942          nonskip_rdc.rate != INT_MAX)
   2943        search_state->this_rdc = nonskip_rdc;
   2944      if (is_single_pred) {
   2945        search_state->uv_dist[INTER_OFFSET(this_mode)][ref_frame] = rdc_uv.dist;
   2946      }
   2947      search_state->this_rdc.rate += rdc_uv.rate;
   2948      search_state->this_rdc.dist += rdc_uv.dist;
   2949      search_state->this_rdc.skip_txfm =
   2950          search_state->this_rdc.skip_txfm && rdc_uv.skip_txfm;
   2951    }
   2952 #if COLLECT_NONRD_PICK_MODE_STAT
   2953    aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
   2954    x->ms_stat_nonrd.txfm_time[bsize][this_mode] +=
   2955        aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
   2956 #endif
   2957  }
   2958 
   2959  this_best_mode = this_mode;
   2960  // TODO(kyslov) account for UV prediction cost
   2961  search_state->this_rdc.rate += rate_mv;
   2962  if (!is_single_pred) {
   2963    const int16_t mode_ctx =
   2964        av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
   2965    search_state->this_rdc.rate += cost_mv_ref(mode_costs, this_mode, mode_ctx);
   2966  } else {
   2967    // If the current mode has zeromv but is not GLOBALMV, compare the rate
   2968    // cost. If GLOBALMV is cheaper, use GLOBALMV instead.
   2969    if (this_mode != GLOBALMV &&
   2970        this_mv->as_int == search_state->frame_mv[GLOBALMV][ref_frame].as_int) {
   2971      if (is_globalmv_better(this_mode, ref_frame, rate_mv, mode_costs,
   2972                             search_state->single_inter_mode_costs, mbmi_ext)) {
   2973        this_best_mode = GLOBALMV;
   2974      }
   2975    }
   2976 
   2977    search_state->this_rdc.rate +=
   2978        search_state
   2979            ->single_inter_mode_costs[INTER_OFFSET(this_best_mode)][ref_frame];
   2980  }
   2981 
   2982  if (is_single_pred && this_mv->as_int == 0 && var < UINT_MAX) {
   2983    search_state->vars[INTER_OFFSET(GLOBALMV)][ref_frame] = var;
   2984  }
   2985 
   2986  search_state->this_rdc.rate += search_state->ref_costs_single[ref_frame];
   2987 
   2988  search_state->this_rdc.rdcost = RDCOST(x->rdmult, search_state->this_rdc.rate,
   2989                                         search_state->this_rdc.dist);
   2990  if (cpi->oxcf.rc_cfg.mode == AOM_CBR && is_single_pred) {
   2991    newmv_diff_bias(xd, this_best_mode, &search_state->this_rdc, bsize,
   2992                    search_state->frame_mv[this_best_mode][ref_frame].as_mv.row,
   2993                    search_state->frame_mv[this_best_mode][ref_frame].as_mv.col,
   2994                    cpi->speed, x->source_variance, x->content_state_sb);
   2995  }
   2996 
   2997 #if CONFIG_AV1_TEMPORAL_DENOISING
   2998  if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
   2999      cpi->denoiser.denoising_level > kDenLowLow) {
   3000    av1_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
   3001    // Keep track of zero_last cost.
   3002    if (ref_frame == LAST_FRAME && this_mv->as_int == 0)
   3003      *zero_last_cost_orig = search_state->this_rdc.rdcost;
   3004  }
   3005 #else
   3006  (void)(sse_y);
   3007 #endif
   3008 
   3009  search_state->mode_checked[this_mode][ref_frame] = 1;
   3010  search_state->mode_checked[this_best_mode][ref_frame] = 1;
   3011 
   3012  if (*check_globalmv) {
   3013    int32_t abs_mv =
   3014        abs(search_state->frame_mv[this_best_mode][ref_frame].as_mv.row) +
   3015        abs(search_state->frame_mv[this_best_mode][ref_frame].as_mv.col);
   3016    // Early exit check: if the magnitude of this_best_mode's mv is small
   3017    // enough, we skip GLOBALMV check in the next loop iteration.
   3018    if (abs_mv < 2) {
   3019      *check_globalmv = false;
   3020    }
   3021  }
   3022 #if COLLECT_NONRD_PICK_MODE_STAT
   3023  aom_usec_timer_mark(&x->ms_stat_nonrd.timer1);
   3024  x->ms_stat_nonrd.nonskipped_search_times[bsize][this_mode] +=
   3025      aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer1);
   3026 #endif
   3027 
   3028  if (x->sb_me_block && ref_frame == LAST_FRAME &&
   3029      search_state->frame_mv[this_best_mode][ref_frame].as_int ==
   3030          x->sb_me_mv.as_int)
   3031    *sb_me_has_been_tested = 1;
   3032 
   3033  // Copy best mode params to search state
   3034  if (search_state->this_rdc.rdcost < search_state->best_rdc.rdcost) {
   3035    search_state->best_rdc = search_state->this_rdc;
   3036    *best_early_term = this_early_term;
   3037    update_search_state_nonrd(search_state, mi, txfm_info, &nonskip_rdc, ctx,
   3038                              this_best_mode, sse_y);
   3039 
   3040    // This is needed for the compound modes.
   3041    search_state->frame_mv_best[this_best_mode][ref_frame].as_int =
   3042        search_state->frame_mv[this_best_mode][ref_frame].as_int;
   3043    if (ref_frame2 > NONE_FRAME) {
   3044      search_state->frame_mv_best[this_best_mode][ref_frame2].as_int =
   3045          search_state->frame_mv[this_best_mode][ref_frame2].as_int;
   3046    }
   3047 
   3048    if (reuse_inter_pred) {
   3049      free_pred_buffer(best_pickmode->best_pred);
   3050      best_pickmode->best_pred = *this_mode_pred;
   3051    }
   3052  } else {
   3053    if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
   3054  }
   3055 
   3056  if (*best_early_term && (idx > 0 || rt_sf->nonrd_aggressive_skip)) {
   3057    txfm_info->skip_txfm = 1;
   3058    if (!x->sb_me_block || *sb_me_has_been_tested) return false;
   3059  }
   3060  return true;
   3061 }
   3062 
   3063 // Function to perform screen content mode evaluation for non-rd
   3064 static AOM_FORCE_INLINE void handle_screen_content_mode_nonrd(
   3065    AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
   3066    PRED_BUFFER *this_mode_pred, PICK_MODE_CONTEXT *ctx,
   3067    PRED_BUFFER *tmp_buffer, struct buf_2d *orig_dst, int skip_idtx_palette,
   3068    int try_palette, BLOCK_SIZE bsize, int reuse_inter_pred, int mi_col,
   3069    int mi_row) {
   3070  AV1_COMMON *const cm = &cpi->common;
   3071  const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
   3072  MACROBLOCKD *const xd = &x->e_mbd;
   3073  MB_MODE_INFO *const mi = xd->mi[0];
   3074  struct macroblockd_plane *const pd = &xd->plane[0];
   3075  const int bw = block_size_wide[bsize];
   3076  const int bh = block_size_high[bsize];
   3077  TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   3078  BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
   3079 
   3080  // TODO(marpan): Only allow for 8 bit-depth for now, re-enable for 10/12 bit
   3081  // when issue 3359 is fixed.
   3082  if (cm->seq_params->bit_depth == 8 && rt_sf->use_idtx_nonrd &&
   3083      !skip_idtx_palette && !cpi->oxcf.txfm_cfg.use_inter_dct_only &&
   3084      !x->force_zeromv_skip_for_blk &&
   3085      is_inter_mode(best_pickmode->best_mode) &&
   3086      best_pickmode->best_pred != NULL &&
   3087      (!rt_sf->prune_idtx_nonrd ||
   3088       (rt_sf->prune_idtx_nonrd && bsize <= BLOCK_32X32 &&
   3089        best_pickmode->best_mode_skip_txfm != 1 && x->source_variance > 200))) {
   3090    RD_STATS idtx_rdc;
   3091    av1_init_rd_stats(&idtx_rdc);
   3092    int is_skippable;
   3093    this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
   3094    pd->dst.buf = this_mode_pred->data;
   3095    pd->dst.stride = bw;
   3096    const PRED_BUFFER *const best_pred = best_pickmode->best_pred;
   3097    av1_block_yrd_idtx(x, best_pred->data, best_pred->stride, &idtx_rdc,
   3098                       &is_skippable, bsize, mi->tx_size);
   3099    int64_t idx_rdcost_y = RDCOST(x->rdmult, idtx_rdc.rate, idtx_rdc.dist);
   3100    int allow_idtx = 1;
   3101    // Incorporate color into rd cost.
   3102    if ((x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
   3103         x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])) {
   3104      RD_STATS rdc_uv;
   3105      const BLOCK_SIZE uv_bsize =
   3106          get_plane_block_size(bsize, xd->plane[AOM_PLANE_U].subsampling_x,
   3107                               xd->plane[AOM_PLANE_U].subsampling_y);
   3108      if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)]) {
   3109        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   3110                                      AOM_PLANE_U, AOM_PLANE_U);
   3111      }
   3112      if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) {
   3113        av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
   3114                                      AOM_PLANE_V, AOM_PLANE_V);
   3115      }
   3116      av1_model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, AOM_PLANE_U,
   3117                             AOM_PLANE_V);
   3118      if (rdc_uv.dist < x->min_dist_inter_uv)
   3119        x->min_dist_inter_uv = rdc_uv.dist;
   3120      idtx_rdc.rate += rdc_uv.rate;
   3121      idtx_rdc.dist += rdc_uv.dist;
   3122      idtx_rdc.skip_txfm = idtx_rdc.skip_txfm && rdc_uv.skip_txfm;
   3123      if (idx_rdcost_y == 0 && rdc_uv.dist > 0 && x->source_variance < 3000 &&
   3124          x->content_state_sb.source_sad_nonrd > kMedSad)
   3125        allow_idtx = 0;
   3126    }
   3127    int64_t idx_rdcost = RDCOST(x->rdmult, idtx_rdc.rate, idtx_rdc.dist);
   3128    if (allow_idtx && idx_rdcost < search_state->best_rdc.rdcost) {
   3129      best_pickmode->tx_type = IDTX;
   3130      search_state->best_rdc.rdcost = idx_rdcost;
   3131      best_pickmode->best_mode_skip_txfm = idtx_rdc.skip_txfm;
   3132      if (!idtx_rdc.skip_txfm) {
   3133        memcpy(ctx->blk_skip, txfm_info->blk_skip,
   3134               sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
   3135      }
   3136      xd->tx_type_map[0] = best_pickmode->tx_type;
   3137      memset(ctx->tx_type_map, best_pickmode->tx_type, ctx->num_4x4_blk);
   3138      memset(xd->tx_type_map, best_pickmode->tx_type, ctx->num_4x4_blk);
   3139    }
   3140    pd->dst = *orig_dst;
   3141  }
   3142 
   3143  if (!try_palette) return;
   3144  const unsigned int intra_ref_frame_cost =
   3145      search_state->ref_costs_single[INTRA_FRAME];
   3146 
   3147  if (!is_mode_intra(best_pickmode->best_mode)) {
   3148    PRED_BUFFER *const best_pred = best_pickmode->best_pred;
   3149    if (reuse_inter_pred && best_pred != NULL) {
   3150      if (best_pred->data == orig_dst->buf) {
   3151        this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
   3152        aom_convolve_copy(best_pred->data, best_pred->stride,
   3153                          this_mode_pred->data, this_mode_pred->stride, bw, bh);
   3154        best_pickmode->best_pred = this_mode_pred;
   3155      }
   3156    }
   3157    pd->dst = *orig_dst;
   3158  }
   3159  // Search palette mode for Luma plane in inter frame.
   3160  av1_search_palette_mode_luma(cpi, x, bsize, intra_ref_frame_cost, ctx,
   3161                               &search_state->this_rdc,
   3162                               search_state->best_rdc.rdcost);
   3163  // Update best mode data in search_state
   3164  if (search_state->this_rdc.rdcost < search_state->best_rdc.rdcost) {
   3165    best_pickmode->pmi = mi->palette_mode_info;
   3166    best_pickmode->best_mode = DC_PRED;
   3167    mi->mv[0].as_int = INVALID_MV;
   3168    mi->mv[1].as_int = INVALID_MV;
   3169    best_pickmode->best_ref_frame = INTRA_FRAME;
   3170    best_pickmode->best_second_ref_frame = NONE;
   3171    search_state->best_rdc.rate = search_state->this_rdc.rate;
   3172    search_state->best_rdc.dist = search_state->this_rdc.dist;
   3173    search_state->best_rdc.rdcost = search_state->this_rdc.rdcost;
   3174    best_pickmode->best_mode_skip_txfm = search_state->this_rdc.skip_txfm;
   3175    // Keep the skip_txfm off if the color_sensitivity is set.
   3176    if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
   3177        x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])
   3178      search_state->this_rdc.skip_txfm = 0;
   3179    if (!search_state->this_rdc.skip_txfm) {
   3180      memcpy(ctx->blk_skip, txfm_info->blk_skip,
   3181             sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
   3182    }
   3183    if (xd->tx_type_map[0] != DCT_DCT)
   3184      av1_copy_array(ctx->tx_type_map, xd->tx_type_map, ctx->num_4x4_blk);
   3185  }
   3186 }
   3187 
   3188 static inline bool enable_palette(AV1_COMP *cpi, bool is_mode_intra,
   3189                                  BLOCK_SIZE bsize,
   3190                                  unsigned int source_variance,
   3191                                  int force_zeromv_skip, int skip_idtx_palette,
   3192                                  int force_palette_test,
   3193                                  unsigned int best_intra_sad_norm) {
   3194  const unsigned int sad_thresh =
   3195      cpi->sf.rt_sf.prune_palette_search_nonrd > 2
   3196          ? (cpi->oxcf.frm_dim_cfg.width * cpi->oxcf.frm_dim_cfg.height <=
   3197             1280 * 720)
   3198                ? 6
   3199                : 12
   3200          : 10;
   3201  if (!cpi->oxcf.tool_cfg.enable_palette) return false;
   3202  if (!av1_allow_palette(cpi->common.features.allow_screen_content_tools,
   3203                         bsize)) {
   3204    return false;
   3205  }
   3206  if (skip_idtx_palette) return false;
   3207 
   3208  if (cpi->sf.rt_sf.prune_palette_search_nonrd > 1 &&
   3209      ((cpi->rc.high_source_sad && cpi->ppi->rtc_ref.non_reference_frame) ||
   3210       bsize > BLOCK_16X16)) {
   3211    return false;
   3212  }
   3213 
   3214  if (prune_palette_testing_inter(cpi, source_variance) &&
   3215      best_intra_sad_norm < sad_thresh)
   3216    return false;
   3217 
   3218  if ((is_mode_intra || force_palette_test) && source_variance > 0 &&
   3219      !force_zeromv_skip &&
   3220      (cpi->rc.high_source_sad || source_variance > 300)) {
   3221    return true;
   3222  } else {
   3223    return false;
   3224  }
   3225 }
   3226 
   3227 /*!\brief AV1 inter mode selection based on Non-RD optimized model.
   3228 *
   3229 * \ingroup nonrd_mode_search
   3230 * \callgraph
   3231 * Top level function for Non-RD optimized inter mode selection.
   3232 * This finction will loop over subset of inter modes and select the best one
   3233 * based on calculated modelled RD cost. While making decisions which modes to
   3234 * check, this function applies heuristics based on previously checked modes,
   3235 * block residual variance, block size, and other factors to prune certain
   3236 * modes and reference frames. Currently only single reference frame modes
   3237 * are checked. Additional heuristics are applied to decide if intra modes
   3238 *  need to be checked.
   3239 *  *
   3240 * \param[in]    cpi            Top-level encoder structure
   3241 * \param[in]    tile_data      Pointer to struct holding adaptive
   3242                                data/contexts/models for the tile during
   3243                                encoding
   3244 * \param[in]    x              Pointer to structure holding all the data for
   3245                                the current macroblock
   3246 * \param[in]    rd_cost        Struct to keep track of the RD information
   3247 * \param[in]    bsize          Current block size
   3248 * \param[in]    ctx            Structure to hold snapshot of coding context
   3249                                during the mode picking process
   3250 *
   3251 * \remark Nothing is returned. Instead, the MB_MODE_INFO struct inside x
   3252 * is modified to store information about the best mode computed
   3253 * in this function. The rd_cost struct is also updated with the RD stats
   3254 * corresponding to the best mode found.
   3255 */
   3256 void av1_nonrd_pick_inter_mode_sb(AV1_COMP *cpi, TileDataEnc *tile_data,
   3257                                  MACROBLOCK *x, RD_STATS *rd_cost,
   3258                                  BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
   3259  AV1_COMMON *const cm = &cpi->common;
   3260  SVC *const svc = &cpi->svc;
   3261  MACROBLOCKD *const xd = &x->e_mbd;
   3262  MB_MODE_INFO *const mi = xd->mi[0];
   3263  const struct segmentation *const seg = &cm->seg;
   3264  struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
   3265  const MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
   3266  MV_REFERENCE_FRAME ref_frame, ref_frame2;
   3267  const unsigned char segment_id = mi->segment_id;
   3268  int best_early_term = 0;
   3269  int force_skip_low_temp_var = 0;
   3270  unsigned int sse_zeromv_norm = UINT_MAX;
   3271  const int num_inter_modes = NUM_INTER_MODES;
   3272  const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
   3273  bool check_globalmv = rt_sf->check_globalmv_on_single_ref;
   3274  PRED_BUFFER tmp_buffer[4];
   3275  DECLARE_ALIGNED(16, uint8_t, pred_buf[MAX_MB_PLANE * MAX_SB_SQUARE]);
   3276  PRED_BUFFER *this_mode_pred = NULL;
   3277  const int reuse_inter_pred =
   3278      rt_sf->reuse_inter_pred_nonrd && cm->seq_params->bit_depth == AOM_BITS_8;
   3279  InterModeSearchStateNonrd search_state;
   3280  av1_zero(search_state.use_ref_frame_mask);
   3281  av1_zero(search_state.use_scaled_ref_frame);
   3282  BEST_PICKMODE *const best_pickmode = &search_state.best_pickmode;
   3283  (void)tile_data;
   3284 
   3285  const int bh = block_size_high[bsize];
   3286  const int bw = block_size_wide[bsize];
   3287  const int pixels_in_block = bh * bw;
   3288  struct buf_2d orig_dst = pd->dst;
   3289  const TxfmSearchParams *txfm_params = &x->txfm_search_params;
   3290  TxfmSearchInfo *txfm_info = &x->txfm_search_info;
   3291 #if COLLECT_NONRD_PICK_MODE_STAT
   3292  // Mode statistics can be collected only when num_workers is 1
   3293  assert(cpi->mt_info.num_workers <= 1);
   3294  aom_usec_timer_start(&x->ms_stat_nonrd.bsize_timer);
   3295 #endif
   3296  int64_t thresh_sad_pred = INT64_MAX;
   3297  const int mi_row = xd->mi_row;
   3298  const int mi_col = xd->mi_col;
   3299  int_mv svc_mv = { .as_int = 0 };
   3300  int force_mv_inter_layer = 0;
   3301  bool comp_use_zero_zeromv_only = 0;
   3302  int tot_num_comp_modes = NUM_COMP_INTER_MODES_RT;
   3303 #if CONFIG_AV1_TEMPORAL_DENOISING
   3304  const int denoise_recheck_zeromv = 1;
   3305  AV1_PICKMODE_CTX_DEN ctx_den;
   3306  int64_t zero_last_cost_orig = INT64_MAX;
   3307  int denoise_svc_pickmode = 1;
   3308  const int resize_pending = is_frame_resize_pending(cpi);
   3309 #endif
   3310  const ModeCosts *mode_costs = &x->mode_costs;
   3311  struct scale_factors sf_no_scale;
   3312  av1_setup_scale_factors_for_frame(&sf_no_scale, cm->width, cm->height,
   3313                                    cm->width, cm->height);
   3314  if (reuse_inter_pred) {
   3315    for (int buf_idx = 0; buf_idx < 3; buf_idx++) {
   3316      tmp_buffer[buf_idx].data = &pred_buf[pixels_in_block * buf_idx];
   3317      tmp_buffer[buf_idx].stride = bw;
   3318      tmp_buffer[buf_idx].in_use = 0;
   3319    }
   3320    tmp_buffer[3].data = pd->dst.buf;
   3321    tmp_buffer[3].stride = pd->dst.stride;
   3322    tmp_buffer[3].in_use = 0;
   3323  }
   3324 
   3325  const int gf_temporal_ref = is_same_gf_and_last_scale(cm);
   3326 
   3327  // If the lower spatial layer uses an averaging filter for downsampling
   3328  // (phase = 8), the target decimated pixel is shifted by (1/2, 1/2) relative
   3329  // to source, so use subpel motion vector to compensate. The nonzero motion
   3330  // is half pixel shifted to left and top, so (-4, -4). This has more effect
   3331  // on higher resolutions, so condition it on that for now.
   3332  // Exclude quality layers, which have the same resolution and hence no shift.
   3333  if (cpi->ppi->use_svc && svc->spatial_layer_id > 0 &&
   3334      !svc->has_lower_quality_layer &&
   3335      svc->downsample_filter_phase[svc->spatial_layer_id - 1] == 8 &&
   3336      cm->width * cm->height > 640 * 480) {
   3337    svc_mv.as_mv.row = -4;
   3338    svc_mv.as_mv.col = -4;
   3339  }
   3340 
   3341  // Setup parameters used for inter mode evaluation.
   3342  set_params_nonrd_pick_inter_mode(cpi, x, &search_state, rd_cost,
   3343                                   &force_skip_low_temp_var, mi_row, mi_col,
   3344                                   gf_temporal_ref, segment_id, bsize
   3345 #if CONFIG_AV1_TEMPORAL_DENOISING
   3346                                   ,
   3347                                   ctx, denoise_svc_pickmode
   3348 #endif
   3349  );
   3350 
   3351  if (rt_sf->use_comp_ref_nonrd && is_comp_ref_allowed(bsize)) {
   3352    // Only search compound if bsize \gt BLOCK_16X16.
   3353    if (bsize > BLOCK_16X16) {
   3354      comp_use_zero_zeromv_only = rt_sf->check_only_zero_zeromv_on_large_blocks;
   3355    } else {
   3356      tot_num_comp_modes = 0;
   3357    }
   3358  } else {
   3359    tot_num_comp_modes = 0;
   3360  }
   3361 
   3362  // No compound if SEG_LVL_REF_FRAME is set.
   3363  if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
   3364    tot_num_comp_modes = 0;
   3365 
   3366  if (x->pred_mv_sad[LAST_FRAME] != INT_MAX) {
   3367    thresh_sad_pred = ((int64_t)x->pred_mv_sad[LAST_FRAME]) << 1;
   3368    // Increase threshold for less aggressive pruning.
   3369    if (rt_sf->nonrd_prune_ref_frame_search == 1)
   3370      thresh_sad_pred += (x->pred_mv_sad[LAST_FRAME] >> 2);
   3371  }
   3372 
   3373  const int use_model_yrd_large = get_model_rd_flag(cpi, xd, bsize);
   3374 
   3375  // decide block-level interp filter search flags:
   3376  // filter_search_enabled_blk:
   3377  // 0: disabled
   3378  // 1: filter search depends on mode properties
   3379  // 2: filter search forced since prediction is unreliable
   3380  // cb_pred_filter_search 0: disabled cb prediction
   3381  InterpFilter filt_select = EIGHTTAP_REGULAR;
   3382  const int cb_pred_filter_search =
   3383      x->content_state_sb.source_sad_nonrd > kVeryLowSad
   3384          ? cpi->sf.interp_sf.cb_pred_filter_search
   3385          : 0;
   3386  const int filter_search_enabled_blk =
   3387      is_filter_search_enabled_blk(cpi, x, mi_row, mi_col, bsize, segment_id,
   3388                                   cb_pred_filter_search, &filt_select);
   3389 
   3390 #if COLLECT_NONRD_PICK_MODE_STAT
   3391  x->ms_stat_nonrd.num_blocks[bsize]++;
   3392 #endif
   3393  init_mbmi_nonrd(mi, DC_PRED, NONE_FRAME, NONE_FRAME, cm);
   3394  mi->tx_size = AOMMIN(
   3395      AOMMIN(max_txsize_lookup[bsize],
   3396             tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]),
   3397      TX_16X16);
   3398 
   3399  fill_single_inter_mode_costs(search_state.single_inter_mode_costs,
   3400                               num_inter_modes, ref_mode_set, mode_costs,
   3401                               mbmi_ext->mode_context);
   3402 
   3403  MV_REFERENCE_FRAME last_comp_ref_frame = NONE_FRAME;
   3404 
   3405  // Initialize inter prediction params at block level for single reference
   3406  // mode.
   3407  InterPredParams inter_pred_params_sr;
   3408  init_inter_block_params(&inter_pred_params_sr, pd->width, pd->height,
   3409                          mi_row * MI_SIZE, mi_col * MI_SIZE, pd->subsampling_x,
   3410                          pd->subsampling_y, xd->bd, is_cur_buf_hbd(xd),
   3411                          /*is_intrabc=*/0);
   3412  inter_pred_params_sr.conv_params =
   3413      get_conv_params(/*do_average=*/0, AOM_PLANE_Y, xd->bd);
   3414 
   3415  x->block_is_zero_sad = x->content_state_sb.source_sad_nonrd == kZeroSad ||
   3416                         segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
   3417  if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   3418      !x->force_zeromv_skip_for_blk &&
   3419      x->content_state_sb.source_sad_nonrd != kZeroSad &&
   3420      x->source_variance == 0 && bsize < cm->seq_params->sb_size &&
   3421      search_state.yv12_mb[LAST_FRAME][0].width == cm->width &&
   3422      search_state.yv12_mb[LAST_FRAME][0].height == cm->height) {
   3423    set_block_source_sad(cpi, x, bsize, &search_state.yv12_mb[LAST_FRAME][0]);
   3424  }
   3425 
   3426  int sb_me_has_been_tested = 0;
   3427  x->sb_me_block = x->sb_me_partition;
   3428  // Only use this feature (force testing of superblock motion) if coding
   3429  // block size is large.
   3430  if (x->sb_me_block) {
   3431    if (cm->seq_params->sb_size == BLOCK_128X128 && bsize < BLOCK_64X64)
   3432      x->sb_me_block = 0;
   3433    else if (cm->seq_params->sb_size == BLOCK_64X64 && bsize < BLOCK_32X32)
   3434      x->sb_me_block = 0;
   3435  }
   3436 
   3437  x->min_dist_inter_uv = INT64_MAX;
   3438  for (int idx = 0; idx < num_inter_modes + tot_num_comp_modes; ++idx) {
   3439    // If we are at the first compound mode, and the single modes already
   3440    // perform well, then end the search.
   3441    if (rt_sf->skip_compound_based_on_var && idx == num_inter_modes &&
   3442        skip_comp_based_on_var(search_state.vars, bsize)) {
   3443      break;
   3444    }
   3445 
   3446    int is_single_pred = 1;
   3447    PREDICTION_MODE this_mode;
   3448 
   3449    if (idx == 0 && !x->force_zeromv_skip_for_blk) {
   3450      // Set color sensitivity on first tested mode only.
   3451      // Use y-sad already computed in find_predictors: take the sad with motion
   3452      // vector closest to 0; the uv-sad computed below in set_color_sensitivity
   3453      // is for zeromv.
   3454      // For screen: first check if golden reference is being used, if so,
   3455      // force color_sensitivity on (=1) if the color sensitivity for sb_g is 1.
   3456      // The check in set_color_sensitivity() will then follow and check for
   3457      // setting the flag if the level is still 2 or 0.
   3458      if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   3459          search_state.use_ref_frame_mask[GOLDEN_FRAME]) {
   3460        if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1)
   3461          x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] = 1;
   3462        if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
   3463          x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] = 1;
   3464      }
   3465      if (search_state.use_ref_frame_mask[LAST_FRAME] &&
   3466          x->pred_mv0_sad[LAST_FRAME] != INT_MAX) {
   3467        int y_sad = x->pred_mv0_sad[LAST_FRAME];
   3468        if (x->pred_mv1_sad[LAST_FRAME] != INT_MAX &&
   3469            (abs(search_state.frame_mv[NEARMV][LAST_FRAME].as_mv.col) +
   3470             abs(search_state.frame_mv[NEARMV][LAST_FRAME].as_mv.row)) <
   3471                (abs(search_state.frame_mv[NEARESTMV][LAST_FRAME].as_mv.col) +
   3472                 abs(search_state.frame_mv[NEARESTMV][LAST_FRAME].as_mv.row)))
   3473          y_sad = x->pred_mv1_sad[LAST_FRAME];
   3474        set_color_sensitivity(cpi, x, bsize, y_sad, x->source_variance,
   3475                              search_state.yv12_mb[LAST_FRAME]);
   3476      }
   3477    }
   3478 
   3479    // Check the inter mode can be skipped based on mode statistics and speed
   3480    // features settings.
   3481    if (skip_inter_mode_nonrd(cpi, x, &search_state, &thresh_sad_pred,
   3482                              &force_mv_inter_layer, &is_single_pred,
   3483                              &this_mode, &last_comp_ref_frame, &ref_frame,
   3484                              &ref_frame2, idx, svc_mv, force_skip_low_temp_var,
   3485                              sse_zeromv_norm, num_inter_modes, segment_id,
   3486                              bsize, comp_use_zero_zeromv_only, check_globalmv))
   3487      continue;
   3488 
   3489    // Select prediction reference frames.
   3490    for (int plane = 0; plane < MAX_MB_PLANE; plane++) {
   3491      xd->plane[plane].pre[0] = search_state.yv12_mb[ref_frame][plane];
   3492      if (!is_single_pred)
   3493        xd->plane[plane].pre[1] = search_state.yv12_mb[ref_frame2][plane];
   3494    }
   3495 
   3496    mi->ref_frame[0] = ref_frame;
   3497    mi->ref_frame[1] = ref_frame2;
   3498    set_ref_ptrs(cm, xd, ref_frame, ref_frame2);
   3499 
   3500    // Check if the scaled reference frame should be used. This is set in the
   3501    // find_predictors() for each usable reference. If so, set the
   3502    // block_ref_scale_factors[] to no reference scaling.
   3503    if (search_state.use_scaled_ref_frame[ref_frame]) {
   3504      xd->block_ref_scale_factors[0] = &sf_no_scale;
   3505    }
   3506    if (!is_single_pred && search_state.use_scaled_ref_frame[ref_frame2]) {
   3507      xd->block_ref_scale_factors[1] = &sf_no_scale;
   3508    }
   3509 
   3510    // Perform inter mode evaluation for non-rd
   3511    if (!handle_inter_mode_nonrd(
   3512            cpi, x, &search_state, ctx, &this_mode_pred, tmp_buffer,
   3513            inter_pred_params_sr, &best_early_term, &sse_zeromv_norm,
   3514            &check_globalmv,
   3515 #if CONFIG_AV1_TEMPORAL_DENOISING
   3516            &zero_last_cost_orig, denoise_svc_pickmode,
   3517 #endif
   3518            idx, force_mv_inter_layer, is_single_pred, gf_temporal_ref,
   3519            use_model_yrd_large, filter_search_enabled_blk, bsize, this_mode,
   3520            filt_select, cb_pred_filter_search, reuse_inter_pred,
   3521            &sb_me_has_been_tested)) {
   3522      break;
   3523    }
   3524  }
   3525 
   3526  // Restore mode data of best inter mode
   3527  mi->mode = best_pickmode->best_mode;
   3528  mi->motion_mode = best_pickmode->best_motion_mode;
   3529  mi->wm_params = best_pickmode->wm_params;
   3530  mi->num_proj_ref = best_pickmode->num_proj_ref;
   3531  mi->interp_filters = best_pickmode->best_pred_filter;
   3532  mi->tx_size = best_pickmode->best_tx_size;
   3533  memset(mi->inter_tx_size, mi->tx_size, sizeof(mi->inter_tx_size));
   3534  mi->ref_frame[0] = best_pickmode->best_ref_frame;
   3535  mi->mv[0].as_int = search_state
   3536                         .frame_mv_best[best_pickmode->best_mode]
   3537                                       [best_pickmode->best_ref_frame]
   3538                         .as_int;
   3539  mi->mv[1].as_int = 0;
   3540  if (best_pickmode->best_second_ref_frame > INTRA_FRAME) {
   3541    mi->ref_frame[1] = best_pickmode->best_second_ref_frame;
   3542    mi->mv[1].as_int = search_state
   3543                           .frame_mv_best[best_pickmode->best_mode]
   3544                                         [best_pickmode->best_second_ref_frame]
   3545                           .as_int;
   3546  }
   3547  // Perform intra prediction search, if the best SAD is above a certain
   3548  // threshold.
   3549  mi->angle_delta[PLANE_TYPE_Y] = 0;
   3550  mi->angle_delta[PLANE_TYPE_UV] = 0;
   3551  mi->filter_intra_mode_info.use_filter_intra = 0;
   3552 
   3553 #if COLLECT_NONRD_PICK_MODE_STAT
   3554  aom_usec_timer_start(&x->ms_stat_nonrd.timer1);
   3555  x->ms_stat_nonrd.num_searches[bsize][DC_PRED]++;
   3556  x->ms_stat_nonrd.num_nonskipped_searches[bsize][DC_PRED]++;
   3557 #endif
   3558 
   3559  int force_palette_test = 0;
   3560  if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
   3561      x->content_state_sb.source_sad_nonrd != kZeroSad &&
   3562      bsize <= BLOCK_16X16) {
   3563    unsigned int thresh_sse = cpi->rc.high_source_sad ? 15000 : 200000;
   3564    unsigned int thresh_source_var = cpi->rc.high_source_sad ? 50 : 200;
   3565    unsigned int best_sse_inter_motion =
   3566        (unsigned int)(search_state.best_rdc.sse >>
   3567                       (b_width_log2_lookup[bsize] +
   3568                        b_height_log2_lookup[bsize]));
   3569    if (best_sse_inter_motion > thresh_sse &&
   3570        x->source_variance > thresh_source_var)
   3571      force_palette_test = 1;
   3572  }
   3573 
   3574  // For the SEG_LVL_REF_FRAME inter_mode must be selected if reference set is
   3575  // not INTRA_FRAME, so skip all intra mode (and palette below).
   3576  const int inter_forced_on_segment =
   3577      segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
   3578      get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != INTRA_FRAME;
   3579 
   3580  // Evaluate Intra modes in inter frame
   3581  unsigned int best_intra_sad_norm = UINT_MAX;
   3582  if (!x->force_zeromv_skip_for_blk && !inter_forced_on_segment)
   3583    av1_estimate_intra_mode(cpi, x, bsize, best_early_term,
   3584                            search_state.ref_costs_single[INTRA_FRAME],
   3585                            reuse_inter_pred, &orig_dst, tmp_buffer,
   3586                            &this_mode_pred, &search_state.best_rdc,
   3587                            best_pickmode, ctx, &best_intra_sad_norm);
   3588 
   3589  int skip_idtx_palette = (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
   3590                           x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) &&
   3591                          x->content_state_sb.source_sad_nonrd != kZeroSad &&
   3592                          !cpi->rc.high_source_sad &&
   3593                          (cpi->rc.high_motion_content_screen_rtc ||
   3594                           cpi->rc.frame_source_sad < 10000);
   3595 
   3596  bool try_palette = enable_palette(
   3597      cpi, is_mode_intra(best_pickmode->best_mode), bsize, x->source_variance,
   3598      x->force_zeromv_skip_for_blk, skip_idtx_palette, force_palette_test,
   3599      best_intra_sad_norm);
   3600 
   3601  if (try_palette && prune_palette_testing_inter(cpi, x->source_variance))
   3602    x->color_palette_thresh =
   3603        cpi->sf.rt_sf.prune_palette_search_nonrd > 2 ? 20 : 32;
   3604 
   3605  if (!inter_forced_on_segment) {
   3606    // Perform screen content mode evaluation for non-rd
   3607    handle_screen_content_mode_nonrd(cpi, x, &search_state, this_mode_pred, ctx,
   3608                                     tmp_buffer, &orig_dst, skip_idtx_palette,
   3609                                     try_palette, bsize, reuse_inter_pred,
   3610                                     mi_col, mi_row);
   3611  }
   3612 
   3613 #if COLLECT_NONRD_PICK_MODE_STAT
   3614  aom_usec_timer_mark(&x->ms_stat_nonrd.timer1);
   3615  x->ms_stat_nonrd.nonskipped_search_times[bsize][DC_PRED] +=
   3616      aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer1);
   3617 #endif
   3618 
   3619  pd->dst = orig_dst;
   3620  // Best mode is finalized. Restore the mode data to mbmi
   3621  if (try_palette) mi->palette_mode_info = best_pickmode->pmi;
   3622  mi->mode = best_pickmode->best_mode;
   3623  mi->ref_frame[0] = best_pickmode->best_ref_frame;
   3624  mi->ref_frame[1] = best_pickmode->best_second_ref_frame;
   3625  // For lossless: always force the skip flags off.
   3626  if (is_lossless_requested(&cpi->oxcf.rc_cfg)) {
   3627    txfm_info->skip_txfm = 0;
   3628    memset(ctx->blk_skip, 0, sizeof(ctx->blk_skip[0]) * ctx->num_4x4_blk);
   3629  } else {
   3630    txfm_info->skip_txfm = best_pickmode->best_mode_skip_txfm;
   3631  }
   3632  if (has_second_ref(mi)) {
   3633    mi->comp_group_idx = 0;
   3634    mi->compound_idx = 1;
   3635    mi->interinter_comp.type = COMPOUND_AVERAGE;
   3636  }
   3637 
   3638  if (!is_inter_block(mi)) {
   3639    mi->interp_filters = av1_broadcast_interp_filter(SWITCHABLE_FILTERS);
   3640  } else {
   3641    // If inter mode is selected and ref_frame was one that uses the
   3642    // scaled reference frame, then we can't use reuse_inter_pred.
   3643    if (search_state.use_scaled_ref_frame[best_pickmode->best_ref_frame] ||
   3644        (has_second_ref(mi) &&
   3645         search_state
   3646             .use_scaled_ref_frame[best_pickmode->best_second_ref_frame]))
   3647      x->reuse_inter_pred = 0;
   3648  }
   3649 
   3650  // Restore the predicted samples of best mode to final buffer
   3651  if (reuse_inter_pred && best_pickmode->best_pred != NULL) {
   3652    PRED_BUFFER *const best_pred = best_pickmode->best_pred;
   3653    if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
   3654      aom_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
   3655                        pd->dst.stride, bw, bh);
   3656    }
   3657  }
   3658 
   3659 #if CONFIG_AV1_TEMPORAL_DENOISING
   3660  if (cpi->oxcf.noise_sensitivity > 0 && resize_pending == 0 &&
   3661      denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
   3662      cpi->denoiser.reset == 0) {
   3663    AV1_DENOISER_DECISION decision = COPY_BLOCK;
   3664    ctx->sb_skip_denoising = 0;
   3665    av1_pickmode_ctx_den_update(
   3666        &ctx_den, zero_last_cost_orig, search_state.ref_costs_single,
   3667        search_state.frame_mv, reuse_inter_pred, best_pickmode);
   3668    av1_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision,
   3669                         gf_temporal_ref);
   3670    if (denoise_recheck_zeromv)
   3671      recheck_zeromv_after_denoising(
   3672          cpi, mi, x, xd, decision, &ctx_den, search_state.yv12_mb,
   3673          &search_state.best_rdc, best_pickmode, bsize, mi_row, mi_col);
   3674    best_pickmode->best_ref_frame = ctx_den.best_ref_frame;
   3675  }
   3676 #endif
   3677 
   3678  // Update the factors used for RD thresholding for all modes.
   3679  if (cpi->sf.inter_sf.adaptive_rd_thresh && !has_second_ref(mi)) {
   3680    THR_MODES best_mode_idx =
   3681        mode_idx[best_pickmode->best_ref_frame][mode_offset(mi->mode)];
   3682    if (best_pickmode->best_ref_frame == INTRA_FRAME) {
   3683      // Only consider the modes that are included in the intra_mode_list.
   3684      int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
   3685      for (int mode_index = 0; mode_index < intra_modes; mode_index++) {
   3686        update_thresh_freq_fact(cpi, x, bsize, INTRA_FRAME, best_mode_idx,
   3687                                intra_mode_list[mode_index]);
   3688      }
   3689    } else {
   3690      PREDICTION_MODE this_mode;
   3691      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
   3692        update_thresh_freq_fact(cpi, x, bsize, best_pickmode->best_ref_frame,
   3693                                best_mode_idx, this_mode);
   3694      }
   3695    }
   3696  }
   3697 
   3698 #if CONFIG_INTERNAL_STATS
   3699  store_coding_context_nonrd(x, ctx, mi->mode);
   3700 #else
   3701  store_coding_context_nonrd(x, ctx);
   3702 #endif  // CONFIG_INTERNAL_STATS
   3703 
   3704 #if COLLECT_NONRD_PICK_MODE_STAT
   3705  aom_usec_timer_mark(&x->ms_stat_nonrd.bsize_timer);
   3706  x->ms_stat_nonrd.total_block_times[bsize] +=
   3707      aom_usec_timer_elapsed(&x->ms_stat_nonrd.bsize_timer);
   3708  print_time(&x->ms_stat_nonrd, bsize, cm->mi_params.mi_rows,
   3709             cm->mi_params.mi_cols, mi_row, mi_col);
   3710 #endif  // COLLECT_NONRD_PICK_MODE_STAT
   3711 
   3712  *rd_cost = search_state.best_rdc;
   3713 
   3714  // Reset the xd->block_ref_scale_factors[i], as they may have
   3715  // been set to pointer &sf_no_scale, which becomes invalid afer
   3716  // this function.
   3717  set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
   3718 }