vp9_rdopt.c

/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
#include <math.h>

#include "./vp9_rtcd.h"

#include "vpx_mem/vpx_mem.h"

#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_systemdependent.h"

#include "vp9/encoder/vp9_cost.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_encoder.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_variance.h"

#define RD_THRESH_MAX_FACT 64
#define RD_THRESH_INC      1

#define LAST_FRAME_MODE_MASK    0xFFEDCD60
#define GOLDEN_FRAME_MODE_MASK  0xFFDA3BB0
#define ALT_REF_MODE_MASK       0xFFC648D0

#define MIN_EARLY_TERM_INDEX    3

typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

typedef struct {
  MV_REFERENCE_FRAME ref_frame[2];
} REF_DEFINITION;

struct rdcost_block_args {
  MACROBLOCK *x;
  ENTROPY_CONTEXT t_above[16];
  ENTROPY_CONTEXT t_left[16];
  int rate;
  int64_t dist;
  int64_t sse;
  int this_rate;
  int64_t this_dist;
  int64_t this_sse;
  int64_t this_rd;
  int64_t best_rd;
  int skip;
  int use_fast_coef_costing;
  const scan_order *so;
};

static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
  {NEARESTMV, {LAST_FRAME,   NONE}},
  {NEARESTMV, {ALTREF_FRAME, NONE}},
  {NEARESTMV, {GOLDEN_FRAME, NONE}},

  {DC_PRED,   {INTRA_FRAME,  NONE}},

  {NEWMV,     {LAST_FRAME,   NONE}},
  {NEWMV,     {ALTREF_FRAME, NONE}},
  {NEWMV,     {GOLDEN_FRAME, NONE}},

  {NEARMV,    {LAST_FRAME,   NONE}},
  {NEARMV,    {ALTREF_FRAME, NONE}},
  {NEARESTMV, {LAST_FRAME,   ALTREF_FRAME}},
  {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},

  {TM_PRED,   {INTRA_FRAME,  NONE}},

  {NEARMV,    {LAST_FRAME,   ALTREF_FRAME}},
  {NEWMV,     {LAST_FRAME,   ALTREF_FRAME}},
  {NEARMV,    {GOLDEN_FRAME, NONE}},
  {NEARMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
  {NEWMV,     {GOLDEN_FRAME, ALTREF_FRAME}},

  {ZEROMV,    {LAST_FRAME,   NONE}},
  {ZEROMV,    {GOLDEN_FRAME, NONE}},
  {ZEROMV,    {ALTREF_FRAME, NONE}},
  {ZEROMV,    {LAST_FRAME,   ALTREF_FRAME}},
  {ZEROMV,    {GOLDEN_FRAME, ALTREF_FRAME}},

  {H_PRED,    {INTRA_FRAME,  NONE}},
  {V_PRED,    {INTRA_FRAME,  NONE}},
  {D135_PRED, {INTRA_FRAME,  NONE}},
  {D207_PRED, {INTRA_FRAME,  NONE}},
  {D153_PRED, {INTRA_FRAME,  NONE}},
  {D63_PRED,  {INTRA_FRAME,  NONE}},
  {D117_PRED, {INTRA_FRAME,  NONE}},
  {D45_PRED,  {INTRA_FRAME,  NONE}},
};

static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
  {{LAST_FRAME,   NONE}},
  {{GOLDEN_FRAME, NONE}},
  {{ALTREF_FRAME, NONE}},
  {{LAST_FRAME,   ALTREF_FRAME}},
  {{GOLDEN_FRAME, ALTREF_FRAME}},
  {{INTRA_FRAME,  NONE}},
};

static int raster_block_offset(BLOCK_SIZE plane_bsize,
                               int raster_block, int stride) {
  const int bw = b_width_log2(plane_bsize);
  const int y = 4 * (raster_block >> bw);
  const int x = 4 * (raster_block & ((1 << bw) - 1));
  return y * stride + x;
}
static int16_t* raster_block_offset_int16(BLOCK_SIZE plane_bsize,
                                          int raster_block, int16_t *base) {
  const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  return base + raster_block_offset(plane_bsize, raster_block, stride);
}

static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
                           int m, int n, int min_plane, int max_plane) {
  int i;

  for (i = min_plane; i < max_plane; ++i) {
    struct macroblock_plane *const p = &x->plane[i];
    struct macroblockd_plane *const pd = &x->e_mbd.plane[i];

    p->coeff    = ctx->coeff_pbuf[i][m];
    p->qcoeff   = ctx->qcoeff_pbuf[i][m];
    pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
    p->eobs     = ctx->eobs_pbuf[i][m];

    ctx->coeff_pbuf[i][m]   = ctx->coeff_pbuf[i][n];
    ctx->qcoeff_pbuf[i][m]  = ctx->qcoeff_pbuf[i][n];
    ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
    ctx->eobs_pbuf[i][m]    = ctx->eobs_pbuf[i][n];

    ctx->coeff_pbuf[i][n]   = p->coeff;
    ctx->qcoeff_pbuf[i][n]  = p->qcoeff;
    ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
    ctx->eobs_pbuf[i][n]    = p->eobs;
  }
}

static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
                            MACROBLOCK *x, MACROBLOCKD *xd,
                            int *out_rate_sum, int64_t *out_dist_sum) {
  // Note our transform coeffs are 8 times an orthogonal transform.
  // Hence quantizer step is also 8 times. To get effective quantizer
  // we need to divide by 8 before sending to modeling function.
  int i;
  int64_t rate_sum = 0;
  int64_t dist_sum = 0;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];
  unsigned int sse;
  unsigned int var = 0;
  unsigned int sum_sse = 0;
  const int shift = 8;
  int rate;
  int64_t dist;

  x->pred_sse[ref] = 0;

  for (i = 0; i < MAX_MB_PLANE; ++i) {
    struct macroblock_plane *const p = &x->plane[i];
    struct macroblockd_plane *const pd = &xd->plane[i];
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
    const TX_SIZE max_tx_size = max_txsize_lookup[bs];
    const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
    int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
    int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
    int idx, idy;
    int lw = b_width_log2_lookup[unit_size] + 2;
    int lh = b_height_log2_lookup[unit_size] + 2;

    sum_sse = 0;

    for (idy = 0; idy < bh; ++idy) {
      for (idx = 0; idx < bw; ++idx) {
        uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
        uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
        int block_idx = (idy << 1) + idx;

        var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
                                        dst, pd->dst.stride, &sse);
        x->bsse[(i << 2) + block_idx] = sse;
        sum_sse += sse;

        if (!x->select_tx_size) {
          if (x->bsse[(i << 2) + block_idx] < p->quant_thred[0] >> shift)
            x->skip_txfm[(i << 2) + block_idx] = 1;
          else if (var < p->quant_thred[1] >> shift)
            x->skip_txfm[(i << 2) + block_idx] = 2;
          else
            x->skip_txfm[(i << 2) + block_idx] = 0;
        }

        if (i == 0)
          x->pred_sse[ref] += sse;
      }
    }

    // Fast approximate the modelling function.
    if (cpi->oxcf.speed > 4) {
      int64_t rate;
      int64_t dist;
      int64_t square_error = sse;
      int quantizer = (pd->dequant[1] >> 3);

      if (quantizer < 120)
        rate = (square_error * (280 - quantizer)) >> 8;
      else
        rate = 0;
      dist = (square_error * quantizer) >> 8;
      rate_sum += rate;
      dist_sum += dist;
    } else {
      vp9_model_rd_from_var_lapndz(sum_sse, 1 << num_pels_log2_lookup[bs],
                                   pd->dequant[1] >> 3, &rate, &dist);
      rate_sum += rate;
      dist_sum += dist;
    }
  }

  *out_rate_sum = (int)rate_sum;
  *out_dist_sum = dist_sum << 4;
}

int64_t vp9_block_error_c(const int16_t *coeff, const int16_t *dqcoeff,
                          intptr_t block_size, int64_t *ssz) {
  int i;
  int64_t error = 0, sqcoeff = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
    error +=  diff * diff;
    sqcoeff += coeff[i] * coeff[i];
  }

  *ssz = sqcoeff;
  return error;
}

/* The trailing '0' is a terminator which is used inside cost_coeffs() to
 * decide whether to include cost of a trailing EOB node or not (i.e. we
 * can skip this if the last coefficient in this transform block, e.g. the
 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
 * were non-zero). */
static const int16_t band_counts[TX_SIZES][8] = {
  { 1, 2, 3, 4,  3,   16 - 13, 0 },
  { 1, 2, 3, 4, 11,   64 - 21, 0 },
  { 1, 2, 3, 4, 11,  256 - 21, 0 },
  { 1, 2, 3, 4, 11, 1024 - 21, 0 },
};
static INLINE int cost_coeffs(MACROBLOCK *x,
                              int plane, int block,
                              ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
                              TX_SIZE tx_size,
                              const int16_t *scan, const int16_t *nb,
                              int use_fast_coef_costing) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblock_plane *p = &x->plane[plane];
  const struct macroblockd_plane *pd = &xd->plane[plane];
  const PLANE_TYPE type = pd->plane_type;
  const int16_t *band_count = &band_counts[tx_size][1];
  const int eob = p->eobs[block];
  const int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
                   x->token_costs[tx_size][type][is_inter_block(mbmi)];
  uint8_t token_cache[32 * 32];
  int pt = combine_entropy_contexts(*A, *L);
  int c, cost;
  // Check for consistency of tx_size with mode info
  assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
                              : get_uv_tx_size(mbmi, pd) == tx_size);

  if (eob == 0) {
    // single eob token
    cost = token_costs[0][0][pt][EOB_TOKEN];
    c = 0;
  } else {
    int band_left = *band_count++;

    // dc token
    int v = qcoeff[0];
    int prev_t = vp9_dct_value_tokens_ptr[v].token;
    cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
    token_cache[0] = vp9_pt_energy_class[prev_t];
    ++token_costs;

    // ac tokens
    for (c = 1; c < eob; c++) {
      const int rc = scan[c];
      int t;

      v = qcoeff[rc];
      t = vp9_dct_value_tokens_ptr[v].token;
      if (use_fast_coef_costing) {
        cost += (*token_costs)[!prev_t][!prev_t][t] + vp9_dct_value_cost_ptr[v];
      } else {
        pt = get_coef_context(nb, token_cache, c);
        cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
        token_cache[rc] = vp9_pt_energy_class[t];
      }
      prev_t = t;
      if (!--band_left) {
        band_left = *band_count++;
        ++token_costs;
      }
    }

    // eob token
    if (band_left) {
      if (use_fast_coef_costing) {
        cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
      } else {
        pt = get_coef_context(nb, token_cache, c);
        cost += (*token_costs)[0][pt][EOB_TOKEN];
      }
    }
  }

  // is eob first coefficient;
  *A = *L = (c > 0);

  return cost;
}
static void dist_block(int plane, int block, TX_SIZE tx_size,
                       struct rdcost_block_args* args) {
  const int ss_txfrm_size = tx_size << 1;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  int64_t this_sse;
  int shift = tx_size == TX_32X32 ? 0 : 2;
  int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
                               &this_sse) >> shift;
  args->sse  = this_sse >> shift;

  if (x->skip_encode && !is_inter_block(&xd->mi[0]->mbmi)) {
    // TODO(jingning): tune the model to better capture the distortion.
    int64_t p = (pd->dequant[1] * pd->dequant[1] *
                    (1 << ss_txfrm_size)) >> (shift + 2);
    args->dist += (p >> 4);
    args->sse  += p;
  }
}

static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
                       TX_SIZE tx_size, struct rdcost_block_args* args) {
  int x_idx, y_idx;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);

  args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
                           args->t_left + y_idx, tx_size,
                           args->so->scan, args->so->neighbors,
                           args->use_fast_coef_costing);
}

static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
                          TX_SIZE tx_size, void *arg) {
  struct rdcost_block_args *args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  int64_t rd1, rd2, rd;

  if (args->skip)
    return;

  if (!is_inter_block(mbmi)) {
    vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
    dist_block(plane, block, tx_size, args);
  } else if (max_txsize_lookup[plane_bsize] == tx_size) {
    if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) {
      // full forward transform and quantization
      vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
      dist_block(plane, block, tx_size, args);
    } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) {
      // compute DC coefficient
      int16_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
      int16_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
      vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
      args->sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
      args->dist = args->sse;
      if (!x->plane[plane].eobs[block])
        args->dist = args->sse - ((coeff[0] * coeff[0] -
            (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2);
    } else {
      // skip forward transform
      x->plane[plane].eobs[block] = 0;
      args->sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
      args->dist = args->sse;
    }
  } else {
    // full forward transform and quantization
    vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
    dist_block(plane, block, tx_size, args);
  }

  rate_block(plane, block, plane_bsize, tx_size, args);
  rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
  rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);

  // TODO(jingning): temporarily enabled only for luma component
  rd = MIN(rd1, rd2);
  if (plane == 0)
    x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
                                    (rd1 > rd2 && !xd->lossless);

  args->this_rate += args->rate;
  args->this_dist += args->dist;
  args->this_sse  += args->sse;
  args->this_rd += rd;

  if (args->this_rd > args->best_rd) {
    args->skip = 1;
    return;
  }
}

static void txfm_rd_in_plane(MACROBLOCK *x,
                             int *rate, int64_t *distortion,
                             int *skippable, int64_t *sse,
                             int64_t ref_best_rd, int plane,
                             BLOCK_SIZE bsize, TX_SIZE tx_size,
                             int use_fast_coef_casting) {
  MACROBLOCKD *const xd = &x->e_mbd;
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  struct rdcost_block_args args;
  vp9_zero(args);
  args.x = x;
  args.best_rd = ref_best_rd;
  args.use_fast_coef_costing = use_fast_coef_casting;

  if (plane == 0)
    xd->mi[0]->mbmi.tx_size = tx_size;

  vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);

  args.so = get_scan(xd, tx_size, pd->plane_type, 0);

  vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
                                         block_rd_txfm, &args);
  if (args.skip) {
    *rate       = INT_MAX;
    *distortion = INT64_MAX;
    *sse        = INT64_MAX;
    *skippable  = 0;
  } else {
    *distortion = args.this_dist;
    *rate       = args.this_rate;
    *sse        = args.this_sse;
    *skippable  = vp9_is_skippable_in_plane(x, bsize, plane);
  }
}

static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
                                   int *rate, int64_t *distortion,
                                   int *skip, int64_t *sse,
                                   int64_t ref_best_rd,
                                   BLOCK_SIZE bs) {
  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
  VP9_COMMON *const cm = &cpi->common;
  const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;

  mbmi->tx_size = MIN(max_tx_size, largest_tx_size);

  txfm_rd_in_plane(x, rate, distortion, skip,
                   sse, ref_best_rd, 0, bs,
                   mbmi->tx_size, cpi->sf.use_fast_coef_costing);
}

static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
                                   int *rate,
                                   int64_t *distortion,
                                   int *skip,
                                   int64_t *psse,
                                   int64_t tx_cache[TX_MODES],
                                   int64_t ref_best_rd,
                                   BLOCK_SIZE bs) {
  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
  int r[TX_SIZES][2], s[TX_SIZES];
  int64_t d[TX_SIZES], sse[TX_SIZES];
  int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
                             {INT64_MAX, INT64_MAX},
                             {INT64_MAX, INT64_MAX},
                             {INT64_MAX, INT64_MAX}};
  int n, m;
  int s0, s1;
  const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
  int64_t best_rd = INT64_MAX;
  TX_SIZE best_tx = max_tx_size;

  const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
  assert(skip_prob > 0);
  s0 = vp9_cost_bit(skip_prob, 0);
  s1 = vp9_cost_bit(skip_prob, 1);

  for (n = max_tx_size; n >= 0;  n--) {
    txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
                     &sse[n], ref_best_rd, 0, bs, n,
                     cpi->sf.use_fast_coef_costing);
    r[n][1] = r[n][0];
    if (r[n][0] < INT_MAX) {
      for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
        if (m == n)
          r[n][1] += vp9_cost_zero(tx_probs[m]);
        else
          r[n][1] += vp9_cost_one(tx_probs[m]);
      }
    }
    if (d[n] == INT64_MAX) {
      rd[n][0] = rd[n][1] = INT64_MAX;
    } else if (s[n]) {
      rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
    } else {
      rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
      rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
    }

    // Early termination in transform size search.
    if (cpi->sf.tx_size_search_breakout &&
        (rd[n][1] == INT64_MAX ||
        (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
        s[n] == 1))
      break;

    if (rd[n][1] < best_rd) {
      best_tx = n;
      best_rd = rd[n][1];
    }
  }
  mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
                      best_tx : MIN(max_tx_size, max_mode_tx_size);


  *distortion = d[mbmi->tx_size];
  *rate       = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
  *skip       = s[mbmi->tx_size];
  *psse       = sse[mbmi->tx_size];

  tx_cache[ONLY_4X4] = rd[TX_4X4][0];
  tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
  tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
  tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];

  if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {
    tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
  } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) {
    tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
  } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
    tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1];
  } else {
    tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1];
  }
}

static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
                            int64_t *distortion, int *skip,
                            int64_t *psse, BLOCK_SIZE bs,
                            int64_t txfm_cache[TX_MODES],
                            int64_t ref_best_rd) {
  MACROBLOCKD *xd = &x->e_mbd;
  int64_t sse;
  int64_t *ret_sse = psse ? psse : &sse;

  assert(bs == xd->mi[0]->mbmi.sb_type);

  if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
    vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
    choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
                           bs);
  } else {
    choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
                           txfm_cache, ref_best_rd, bs);
  }
}

static int conditional_skipintra(PREDICTION_MODE mode,
                                 PREDICTION_MODE best_intra_mode) {
  if (mode == D117_PRED &&
      best_intra_mode != V_PRED &&
      best_intra_mode != D135_PRED)
    return 1;
  if (mode == D63_PRED &&
      best_intra_mode != V_PRED &&
      best_intra_mode != D45_PRED)
    return 1;
  if (mode == D207_PRED &&
      best_intra_mode != H_PRED &&
      best_intra_mode != D45_PRED)
    return 1;
  if (mode == D153_PRED &&
      best_intra_mode != H_PRED &&
      best_intra_mode != D135_PRED)
    return 1;
  return 0;
}

static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
                                     PREDICTION_MODE *best_mode,
                                     const int *bmode_costs,
                                     ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
                                     int *bestrate, int *bestratey,
                                     int64_t *bestdistortion,
                                     BLOCK_SIZE bsize, int64_t rd_thresh) {
  PREDICTION_MODE mode;
  MACROBLOCKD *const xd = &x->e_mbd;
  int64_t best_rd = rd_thresh;

  struct macroblock_plane *p = &x->plane[0];
  struct macroblockd_plane *pd = &xd->plane[0];
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  const uint8_t *src_init = &p->src.buf[raster_block_offset(BLOCK_8X8, ib,
                                                            src_stride)];
  uint8_t *dst_init = &pd->dst.buf[raster_block_offset(BLOCK_8X8, ib,
                                                       dst_stride)];
  ENTROPY_CONTEXT ta[2], tempa[2];
  ENTROPY_CONTEXT tl[2], templ[2];

  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  int idx, idy;
  uint8_t best_dst[8 * 8];

  assert(ib < 4);

  vpx_memcpy(ta, a, sizeof(ta));
  vpx_memcpy(tl, l, sizeof(tl));
  xd->mi[0]->mbmi.tx_size = TX_4X4;

  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
    int64_t this_rd;
    int ratey = 0;
    int64_t distortion = 0;
    int rate = bmode_costs[mode];

    if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
      continue;

    // Only do the oblique modes if the best so far is
    // one of the neighboring directional modes
    if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
      if (conditional_skipintra(mode, *best_mode))
          continue;
    }

    vpx_memcpy(tempa, ta, sizeof(ta));
    vpx_memcpy(templ, tl, sizeof(tl));

    for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
      for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
        const int block = ib + idy * 2 + idx;
        const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
        uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
        int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block,
                                                            p->src_diff);
        int16_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
        xd->mi[0]->bmi[block].as_mode = mode;
        vp9_predict_intra_block(xd, block, 1,
                                TX_4X4, mode,
                                x->skip_encode ? src : dst,
                                x->skip_encode ? src_stride : dst_stride,
                                dst, dst_stride, idx, idy, 0);
        vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);

        if (xd->lossless) {
          const scan_order *so = &vp9_default_scan_orders[TX_4X4];
          vp9_fwht4x4(src_diff, coeff, 8);
          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
                               so->scan, so->neighbors,
                               cpi->sf.use_fast_coef_costing);
          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
            goto next;
          vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
                          p->eobs[block]);
        } else {
          int64_t unused;
          const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
          const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
          vp9_fht4x4(src_diff, coeff, 8, tx_type);
          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
                             so->scan, so->neighbors,
                             cpi->sf.use_fast_coef_costing);
          distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
                                        16, &unused) >> 2;
          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
            goto next;
          vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
                         dst, dst_stride, p->eobs[block]);
        }
      }
    }

    rate += ratey;
    this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);

    if (this_rd < best_rd) {
      *bestrate = rate;
      *bestratey = ratey;
      *bestdistortion = distortion;
      best_rd = this_rd;
      *best_mode = mode;
      vpx_memcpy(a, tempa, sizeof(tempa));
      vpx_memcpy(l, templ, sizeof(templ));
      for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
        vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
                   num_4x4_blocks_wide * 4);
    }
  next:
    {}
  }

  if (best_rd >= rd_thresh || x->skip_encode)
    return best_rd;

  for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
    vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
               num_4x4_blocks_wide * 4);

  return best_rd;
}

static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
                                            int *rate, int *rate_y,
                                            int64_t *distortion,
                                            int64_t best_rd) {
  int i, j;
  const MACROBLOCKD *const xd = &mb->e_mbd;
  MODE_INFO *const mic = xd->mi[0];
  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  int idx, idy;
  int cost = 0;
  int64_t total_distortion = 0;
  int tot_rate_y = 0;
  int64_t total_rd = 0;
  ENTROPY_CONTEXT t_above[4], t_left[4];
  const int *bmode_costs = cpi->mbmode_cost;

  vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
  vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));

  // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
      PREDICTION_MODE best_mode = DC_PRED;
      int r = INT_MAX, ry = INT_MAX;
      int64_t d = INT64_MAX, this_rd = INT64_MAX;
      i = idy * 2 + idx;
      if (cpi->common.frame_type == KEY_FRAME) {
        const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
        const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);

        bmode_costs  = cpi->y_mode_costs[A][L];
      }

      this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
                                      t_above + idx, t_left + idy, &r, &ry, &d,
                                      bsize, best_rd - total_rd);
      if (this_rd >= best_rd - total_rd)
        return INT64_MAX;

      total_rd += this_rd;
      cost += r;
      total_distortion += d;
      tot_rate_y += ry;

      mic->bmi[i].as_mode = best_mode;
      for (j = 1; j < num_4x4_blocks_high; ++j)
        mic->bmi[i + j * 2].as_mode = best_mode;
      for (j = 1; j < num_4x4_blocks_wide; ++j)
        mic->bmi[i + j].as_mode = best_mode;

      if (total_rd >= best_rd)
        return INT64_MAX;
    }
  }

  *rate = cost;
  *rate_y = tot_rate_y;
  *distortion = total_distortion;
  mic->mbmi.mode = mic->bmi[3].as_mode;

  return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
}

static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                      int *rate, int *rate_tokenonly,
                                      int64_t *distortion, int *skippable,
                                      BLOCK_SIZE bsize,
                                      int64_t tx_cache[TX_MODES],
                                      int64_t best_rd) {
  PREDICTION_MODE mode;
  PREDICTION_MODE mode_selected = DC_PRED;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *const mic = xd->mi[0];
  int this_rate, this_rate_tokenonly, s;
  int64_t this_distortion, this_rd;
  TX_SIZE best_tx = TX_4X4;