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

#include "vp9/common/vp9_pragmas.h"
#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/encoder/vp9_treewriter.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_modecosts.h"
#include "vp9/encoder/vp9_encodeintra.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_variance.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vpx_mem/vpx_mem.h"
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9_rtcd.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_common.h"

#define INVALID_MV 0x80008000

/* Factor to weigh the rate for switchable interp filters */
#define SWITCHABLE_INTERP_RATE_FACTOR 1

DECLARE_ALIGNED(16, extern const uint8_t,
                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);

#define I4X4_PRED 0x8000
#define SPLITMV 0x10000

const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
  {ZEROMV,    LAST_FRAME,   NONE},
  {DC_PRED,   INTRA_FRAME,  NONE},

  {NEARESTMV, LAST_FRAME,   NONE},
  {NEARMV,    LAST_FRAME,   NONE},

  {ZEROMV,    GOLDEN_FRAME, NONE},
  {NEARESTMV, GOLDEN_FRAME, NONE},

  {ZEROMV,    ALTREF_FRAME, NONE},
  {NEARESTMV, ALTREF_FRAME, NONE},

  {NEARMV,    GOLDEN_FRAME, NONE},
  {NEARMV,    ALTREF_FRAME, NONE},

  {V_PRED,    INTRA_FRAME,  NONE},
  {H_PRED,    INTRA_FRAME,  NONE},
  {D45_PRED,  INTRA_FRAME,  NONE},
  {D135_PRED, INTRA_FRAME,  NONE},
  {D117_PRED, INTRA_FRAME,  NONE},
  {D153_PRED, INTRA_FRAME,  NONE},
  {D27_PRED,  INTRA_FRAME,  NONE},
  {D63_PRED,  INTRA_FRAME,  NONE},

  {TM_PRED,   INTRA_FRAME,  NONE},

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

  {SPLITMV,   LAST_FRAME,   NONE},
  {SPLITMV,   GOLDEN_FRAME, NONE},
  {SPLITMV,   ALTREF_FRAME, NONE},

  {I4X4_PRED, INTRA_FRAME,  NONE},

  /* compound prediction modes */
  {ZEROMV,    LAST_FRAME,   ALTREF_FRAME},
  {NEARESTMV, LAST_FRAME,   ALTREF_FRAME},
  {NEARMV,    LAST_FRAME,   ALTREF_FRAME},

  {ZEROMV,    GOLDEN_FRAME, ALTREF_FRAME},
  {NEARESTMV, GOLDEN_FRAME, ALTREF_FRAME},
  {NEARMV,    GOLDEN_FRAME, ALTREF_FRAME},

  {NEWMV,     LAST_FRAME,   ALTREF_FRAME},
  {NEWMV,     GOLDEN_FRAME, ALTREF_FRAME},

  {SPLITMV,   LAST_FRAME,   ALTREF_FRAME},
  {SPLITMV,   GOLDEN_FRAME, ALTREF_FRAME},
};

// The baseline rd thresholds for breaking out of the rd loop for
// certain modes are assumed to be based on 8x8 blocks.
// This table is used to correct for blocks size.
// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
static int rd_thresh_block_size_factor[BLOCK_SIZE_TYPES] =
  {2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};

#define BASE_RD_THRESH_FREQ_FACT 16
#define MAX_RD_THRESH_FREQ_FACT 32
#define MAX_RD_THRESH_FREQ_INC 1

static void fill_token_costs(vp9_coeff_count (*c)[BLOCK_TYPES],
                             vp9_coeff_count (*cnoskip)[BLOCK_TYPES],
                             vp9_coeff_probs_model (*p)[BLOCK_TYPES]) {
  int i, j, k, l;
  TX_SIZE t;
  for (t = TX_4X4; t <= TX_32X32; t++)
    for (i = 0; i < BLOCK_TYPES; i++)
      for (j = 0; j < REF_TYPES; j++)
        for (k = 0; k < COEF_BANDS; k++)
          for (l = 0; l < PREV_COEF_CONTEXTS; l++) {
            vp9_prob probs[ENTROPY_NODES];
            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
            vp9_cost_tokens((int *)cnoskip[t][i][j][k][l], probs,
                            vp9_coef_tree);
#if CONFIG_BALANCED_COEFTREE
            // Replace the eob node prob with a very small value so that the
            // cost approximately equals the cost without the eob node
            probs[1] = 1;
            vp9_cost_tokens((int *)c[t][i][j][k][l], probs, vp9_coef_tree);
#else
            vp9_cost_tokens_skip((int *)c[t][i][j][k][l], probs,
                                 vp9_coef_tree);
            assert(c[t][i][j][k][l][DCT_EOB_TOKEN] ==
                   cnoskip[t][i][j][k][l][DCT_EOB_TOKEN]);
#endif
          }
}

static int rd_iifactor[32] =  { 4, 4, 3, 2, 1, 0, 0, 0,
                                0, 0, 0, 0, 0, 0, 0, 0,
                                0, 0, 0, 0, 0, 0, 0, 0,
                                0, 0, 0, 0, 0, 0, 0, 0, };

// 3* dc_qlookup[Q]*dc_qlookup[Q];

/* values are now correlated to quantizer */
static int sad_per_bit16lut[QINDEX_RANGE];
static int sad_per_bit4lut[QINDEX_RANGE];

void vp9_init_me_luts() {
  int i;

  // Initialize the sad lut tables using a formulaic calculation for now
  // This is to make it easier to resolve the impact of experimental changes
  // to the quantizer tables.
  for (i = 0; i < QINDEX_RANGE; i++) {
    sad_per_bit16lut[i] =
      (int)((0.0418 * vp9_convert_qindex_to_q(i)) + 2.4107);
    sad_per_bit4lut[i] = (int)(0.063 * vp9_convert_qindex_to_q(i) + 2.742);
  }
}

static int compute_rd_mult(int qindex) {
  const int q = vp9_dc_quant(qindex, 0);
  return (11 * q * q) >> 2;
}

void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) {
  cpi->mb.sadperbit16 = sad_per_bit16lut[qindex];
  cpi->mb.sadperbit4 = sad_per_bit4lut[qindex];
}


void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) {
  int q, i, bsize;

  vp9_clear_system_state();  // __asm emms;

  // Further tests required to see if optimum is different
  // for key frames, golden frames and arf frames.
  // if (cpi->common.refresh_golden_frame ||
  //     cpi->common.refresh_alt_ref_frame)
  qindex = clamp(qindex, 0, MAXQ);

  cpi->RDMULT = compute_rd_mult(qindex);
  if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
    if (cpi->twopass.next_iiratio > 31)
      cpi->RDMULT += (cpi->RDMULT * rd_iifactor[31]) >> 4;
    else
      cpi->RDMULT +=
          (cpi->RDMULT * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
  }
  cpi->mb.errorperbit = cpi->RDMULT >> 6;
  cpi->mb.errorperbit += (cpi->mb.errorperbit == 0);

  vp9_set_speed_features(cpi);

  q = (int)pow(vp9_dc_quant(qindex, 0) >> 2, 1.25);
  q <<= 2;
  if (q < 8)
    q = 8;

  if (cpi->RDMULT > 1000) {
    cpi->RDDIV = 1;
    cpi->RDMULT /= 100;

    for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
      for (i = 0; i < MAX_MODES; ++i) {
        // Threshold here seem unecessarily harsh but fine given actual
        // range of values used for cpi->sf.thresh_mult[]
        int thresh_max = INT_MAX / (q * rd_thresh_block_size_factor[bsize]);

        // *4 relates to the scaling of rd_thresh_block_size_factor[]
        if ((int64_t)cpi->sf.thresh_mult[i] < thresh_max) {
          cpi->rd_threshes[bsize][i] =
            cpi->sf.thresh_mult[i] * q *
            rd_thresh_block_size_factor[bsize] / (4 * 100);
        } else {
          cpi->rd_threshes[bsize][i] = INT_MAX;
        }
        cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
        cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
      }
    }
  } else {
    cpi->RDDIV = 100;

    for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
      for (i = 0; i < MAX_MODES; i++) {
        // Threshold here seem unecessarily harsh but fine given actual
        // range of values used for cpi->sf.thresh_mult[]
        int thresh_max = INT_MAX / (q * rd_thresh_block_size_factor[bsize]);

        if (cpi->sf.thresh_mult[i] < thresh_max) {
          cpi->rd_threshes[bsize][i] =
            cpi->sf.thresh_mult[i] * q *
            rd_thresh_block_size_factor[bsize] / 4;
        } else {
          cpi->rd_threshes[bsize][i] = INT_MAX;
        }
        cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
        cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
      }
    }
  }

  fill_token_costs(cpi->mb.token_costs,
                   cpi->mb.token_costs_noskip,
                   cpi->common.fc.coef_probs);

  for (i = 0; i < NUM_PARTITION_CONTEXTS; i++)
    vp9_cost_tokens(cpi->mb.partition_cost[i],
                    cpi->common.fc.partition_prob[cpi->common.frame_type][i],
                    vp9_partition_tree);

  /*rough estimate for costing*/
  vp9_init_mode_costs(cpi);

  if (cpi->common.frame_type != KEY_FRAME) {
    vp9_build_nmv_cost_table(
        cpi->mb.nmvjointcost,
        cpi->mb.e_mbd.allow_high_precision_mv ?
        cpi->mb.nmvcost_hp : cpi->mb.nmvcost,
        &cpi->common.fc.nmvc,
        cpi->mb.e_mbd.allow_high_precision_mv, 1, 1);
  }
}

int vp9_block_error_c(int16_t *coeff, int16_t *dqcoeff, int block_size) {
  int i, error = 0;

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

  return error;
}

static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb,
                              int plane, int block, PLANE_TYPE type,
                              ENTROPY_CONTEXT *A,
                              ENTROPY_CONTEXT *L,
                              TX_SIZE tx_size,
                              int y_blocks) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
  int pt;
  int c = 0;
  int cost = 0, pad;
  const int *scan, *nb;
  const int eob = xd->plane[plane].eobs[block];
  const int16_t *qcoeff_ptr = BLOCK_OFFSET(xd->plane[plane].qcoeff,
                                           block, 16);
  const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
  unsigned int (*token_costs)[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] =
      mb->token_costs[tx_size][type][ref];
  ENTROPY_CONTEXT above_ec, left_ec;
  TX_TYPE tx_type = DCT_DCT;

  const int segment_id = xd->mode_info_context->mbmi.segment_id;
  unsigned int (*token_costs_noskip)[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] =
      mb->token_costs_noskip[tx_size][type][ref];

  int seg_eob, default_eob;
  uint8_t token_cache[1024];
  const uint8_t * band_translate;

  // Check for consistency of tx_size with mode info
  assert((!type && !plane) || (type && plane));
  if (type == PLANE_TYPE_Y_WITH_DC) {
    assert(xd->mode_info_context->mbmi.txfm_size == tx_size);
  } else {
    TX_SIZE tx_size_uv = get_uv_tx_size(mbmi);
    assert(tx_size == tx_size_uv);
  }

  switch (tx_size) {
    case TX_4X4: {
      tx_type = (type == PLANE_TYPE_Y_WITH_DC) ?
          get_tx_type_4x4(xd, block) : DCT_DCT;
      above_ec = A[0] != 0;
      left_ec = L[0] != 0;
      seg_eob = 16;
      scan = get_scan_4x4(tx_type);
      band_translate = vp9_coefband_trans_4x4;
      break;
    }
    case TX_8X8: {
      const BLOCK_SIZE_TYPE sb_type = xd->mode_info_context->mbmi.sb_type;
      const int sz = 1 + b_width_log2(sb_type);
      const int x = block & ((1 << sz) - 1), y = block - x;
      TX_TYPE tx_type = (type == PLANE_TYPE_Y_WITH_DC) ?
          get_tx_type_8x8(xd, y + (x >> 1)) : DCT_DCT;
      above_ec = (A[0] + A[1]) != 0;
      left_ec = (L[0] + L[1]) != 0;
      scan = get_scan_8x8(tx_type);
      seg_eob = 64;
      band_translate = vp9_coefband_trans_8x8plus;
      break;
    }
    case TX_16X16: {
      const BLOCK_SIZE_TYPE sb_type = xd->mode_info_context->mbmi.sb_type;
      const int sz = 2 + b_width_log2(sb_type);
      const int x = block & ((1 << sz) - 1), y = block - x;
      TX_TYPE tx_type = (type == PLANE_TYPE_Y_WITH_DC) ?
          get_tx_type_16x16(xd, y + (x >> 2)) : DCT_DCT;
      scan = get_scan_16x16(tx_type);
      seg_eob = 256;
      above_ec = (A[0] + A[1] + A[2] + A[3]) != 0;
      left_ec = (L[0] + L[1] + L[2] + L[3]) != 0;
      band_translate = vp9_coefband_trans_8x8plus;
      break;
    }
    case TX_32X32:
      scan = vp9_default_scan_32x32;
      seg_eob = 1024;
      above_ec = (A[0] + A[1] + A[2] + A[3] + A[4] + A[5] + A[6] + A[7]) != 0;
      left_ec = (L[0] + L[1] + L[2] + L[3] + L[4] + L[5] + L[6] + L[7]) != 0;
      band_translate = vp9_coefband_trans_8x8plus;
      break;
    default:
      abort();
      break;
  }
  assert(eob <= seg_eob);

  pt = combine_entropy_contexts(above_ec, left_ec);
  nb = vp9_get_coef_neighbors_handle(scan, &pad);
  default_eob = seg_eob;

  if (vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))
    seg_eob = 0;

  /* sanity check to ensure that we do not have spurious non-zero q values */
  if (eob < seg_eob)
    assert(qcoeff_ptr[scan[eob]] == 0);

  {
    for (c = 0; c < eob; c++) {
      int v = qcoeff_ptr[scan[c]];
      int t = vp9_dct_value_tokens_ptr[v].token;
      int band = get_coef_band(band_translate, c);
      if (c)
        pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob);

      if (!c || token_cache[scan[c - 1]])  // do not skip eob
        cost += token_costs_noskip[band][pt][t] + vp9_dct_value_cost_ptr[v];
      else
        cost += token_costs[band][pt][t] + vp9_dct_value_cost_ptr[v];
      token_cache[scan[c]] = vp9_pt_energy_class[t];
    }
    if (c < seg_eob) {
      if (c)
        pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob);
      cost += mb->token_costs_noskip[tx_size][type][ref]
          [get_coef_band(band_translate, c)]
          [pt][DCT_EOB_TOKEN];
    }
  }

  // is eob first coefficient;
  for (pt = 0; pt < (1 << tx_size); pt++) {
    A[pt] = L[pt] = c > 0;
  }

  return cost;
}

static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
                                     int (*r)[2], int *rate,
                                     int *d, int *distortion,
                                     int *s, int *skip,
                                     int64_t txfm_cache[NB_TXFM_MODES],
                                     TX_SIZE max_txfm_size) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
  vp9_prob skip_prob = vp9_get_pred_prob(cm, xd, PRED_MBSKIP);
  int64_t rd[TX_SIZE_MAX_SB][2];
  int n, m;
  int s0, s1;

  const vp9_prob *tx_probs = vp9_get_pred_probs(cm, xd, PRED_TX_SIZE);

  for (n = TX_4X4; n <= max_txfm_size; n++) {
    r[n][1] = r[n][0];
    for (m = 0; m <= n - (n == max_txfm_size); m++) {
      if (m == n)
        r[n][1] += vp9_cost_zero(tx_probs[m]);
      else
        r[n][1] += vp9_cost_one(tx_probs[m]);
    }
  }

  assert(skip_prob > 0);
  s0 = vp9_cost_bit(skip_prob, 0);
  s1 = vp9_cost_bit(skip_prob, 1);

  for (n = TX_4X4; n <= max_txfm_size; n++) {
    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]);
    }
  }

  if (max_txfm_size == TX_32X32 &&
      (cm->txfm_mode == ALLOW_32X32 ||
       (cm->txfm_mode == TX_MODE_SELECT &&
        rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
        rd[TX_32X32][1] < rd[TX_4X4][1]))) {
    mbmi->txfm_size = TX_32X32;
  } else if (max_txfm_size >= TX_16X16 &&
             (cm->txfm_mode == ALLOW_16X16 ||
              cm->txfm_mode == ALLOW_32X32 ||
              (cm->txfm_mode == TX_MODE_SELECT &&
               rd[TX_16X16][1] < rd[TX_8X8][1] &&
               rd[TX_16X16][1] < rd[TX_4X4][1]))) {
    mbmi->txfm_size = TX_16X16;
  } else if (cm->txfm_mode == ALLOW_8X8 ||
             cm->txfm_mode == ALLOW_16X16 ||
             cm->txfm_mode == ALLOW_32X32 ||
           (cm->txfm_mode == TX_MODE_SELECT && rd[TX_8X8][1] < rd[TX_4X4][1])) {
    mbmi->txfm_size = TX_8X8;
  } else {
    mbmi->txfm_size = TX_4X4;
  }

  *distortion = d[mbmi->txfm_size];
  *rate       = r[mbmi->txfm_size][cm->txfm_mode == TX_MODE_SELECT];
  *skip       = s[mbmi->txfm_size];

  txfm_cache[ONLY_4X4] = rd[TX_4X4][0];
  txfm_cache[ALLOW_8X8] = rd[TX_8X8][0];
  txfm_cache[ALLOW_16X16] = rd[MIN(max_txfm_size, TX_16X16)][0];
  txfm_cache[ALLOW_32X32] = rd[MIN(max_txfm_size, TX_32X32)][0];
  if (max_txfm_size == TX_32X32 &&
      rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
      rd[TX_32X32][1] < rd[TX_4X4][1])
    txfm_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
  else if (max_txfm_size >= TX_16X16 &&
           rd[TX_16X16][1] < rd[TX_8X8][1] && rd[TX_16X16][1] < rd[TX_4X4][1])
    txfm_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
  else
    txfm_cache[TX_MODE_SELECT] = rd[TX_4X4][1] < rd[TX_8X8][1] ?
                                 rd[TX_4X4][1] : rd[TX_8X8][1];
}

static int block_error(int16_t *coeff, int16_t *dqcoeff,
                       int block_size, int shift) {
  int i;
  int64_t error = 0;

  for (i = 0; i < block_size; i++) {
    int this_diff = coeff[i] - dqcoeff[i];
    error += (unsigned)this_diff * this_diff;
  }
  error >>= shift;

  return error > INT_MAX ? INT_MAX : (int)error;
}

static int block_error_sby(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize, int shift) {
  const int bwl = b_width_log2(bsize), bhl = b_height_log2(bsize);
  return block_error(x->plane[0].coeff, x->e_mbd.plane[0].dqcoeff,
                     16 << (bwl + bhl), shift);
}

static int block_error_sbuv(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize, int shift) {
  const int bwl = b_width_log2(bsize), bhl = b_height_log2(bsize);
  int64_t sum = 0;
  int plane;

  for (plane = 1; plane < MAX_MB_PLANE; plane++) {
    const int subsampling = x->e_mbd.plane[plane].subsampling_x +
                            x->e_mbd.plane[plane].subsampling_y;
    sum += block_error(x->plane[plane].coeff, x->e_mbd.plane[plane].dqcoeff,
                       16 << (bwl + bhl - subsampling), 0);
  }
  sum >>= shift;
  return sum > INT_MAX ? INT_MAX : (int)sum;
}

struct rdcost_block_args {
  VP9_COMMON *cm;
  MACROBLOCK *x;
  ENTROPY_CONTEXT t_above[16];
  ENTROPY_CONTEXT t_left[16];
  TX_SIZE tx_size;
  int bw;
  int bh;
  int cost;
};

static void rdcost_block(int plane, int block, BLOCK_SIZE_TYPE bsize,
                         int ss_txfrm_size, void *arg) {
  struct rdcost_block_args* args = arg;
  int x_idx, y_idx;
  MACROBLOCKD * const xd = &args->x->e_mbd;

  txfrm_block_to_raster_xy(xd, bsize, plane, block, args->tx_size * 2, &x_idx,
                           &y_idx);

  args->cost += cost_coeffs(args->cm, args->x, plane, block,
                            xd->plane[plane].plane_type, args->t_above + x_idx,
                            args->t_left + y_idx, args->tx_size,
                            args->bw * args->bh);
}

static int rdcost_plane(VP9_COMMON * const cm, MACROBLOCK *x, int plane,
                        BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
  MACROBLOCKD * const xd = &x->e_mbd;
  const int bwl = b_width_log2(bsize) - xd->plane[plane].subsampling_x;
  const int bhl = b_height_log2(bsize) - xd->plane[plane].subsampling_y;
  const int bw = 1 << bwl, bh = 1 << bhl;
  struct rdcost_block_args args = { cm, x, { 0 }, { 0 }, tx_size, bw, bh, 0 };

  vpx_memcpy(&args.t_above, xd->plane[plane].above_context,
             sizeof(ENTROPY_CONTEXT) * bw);
  vpx_memcpy(&args.t_left, xd->plane[plane].left_context,
             sizeof(ENTROPY_CONTEXT) * bh);

  foreach_transformed_block_in_plane(xd, bsize, plane, rdcost_block, &args);

  return args.cost;
}

static int rdcost_uv(VP9_COMMON *const cm, MACROBLOCK *x,
                     BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
  int cost = 0, plane;

  for (plane = 1; plane < MAX_MB_PLANE; plane++) {
    cost += rdcost_plane(cm, x, plane, bsize, tx_size);
  }
  return cost;
}

static void super_block_yrd_for_txfm(VP9_COMMON *const cm, MACROBLOCK *x,
                                     int *rate, int *distortion, int *skippable,
                                     BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  xd->mode_info_context->mbmi.txfm_size = tx_size;

  if (xd->mode_info_context->mbmi.ref_frame[0] == INTRA_FRAME)
    vp9_encode_intra_block_y(cm, x, bsize);
  else
    vp9_xform_quant_sby(cm, x, bsize);

  *distortion = block_error_sby(x, bsize, tx_size == TX_32X32 ? 0 : 2);
  *rate       = rdcost_plane(cm, x, 0, bsize, tx_size);
  *skippable  = vp9_sby_is_skippable(xd, bsize);
}

static void super_block_yrd(VP9_COMP *cpi,
                            MACROBLOCK *x, int *rate, int *distortion,
                            int *skip, BLOCK_SIZE_TYPE bs,
                            int64_t txfm_cache[NB_TXFM_MODES]) {
  VP9_COMMON *const cm = &cpi->common;
  int r[TX_SIZE_MAX_SB][2], d[TX_SIZE_MAX_SB], s[TX_SIZE_MAX_SB];
  MACROBLOCKD *xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;

  assert(bs == mbmi->sb_type);
  if (mbmi->ref_frame[0] > INTRA_FRAME)
    vp9_subtract_sby(x, bs);

  if (cpi->speed > 4) {
    if (bs >= BLOCK_SIZE_SB32X32) {
      mbmi->txfm_size = TX_32X32;
    } else if (bs >= BLOCK_SIZE_MB16X16) {
      mbmi->txfm_size = TX_16X16;
    } else if (bs >= BLOCK_SIZE_SB8X8) {
      mbmi->txfm_size = TX_8X8;
    } else {
      mbmi->txfm_size = TX_4X4;
    }
    vpx_memset(txfm_cache, 0, NB_TXFM_MODES * sizeof(int64_t));
    super_block_yrd_for_txfm(cm, x, rate, distortion, skip, bs,
                             mbmi->txfm_size);
    return;
  }
  if (bs >= BLOCK_SIZE_SB32X32)
    super_block_yrd_for_txfm(cm, x, &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32],
                             bs, TX_32X32);
  if (bs >= BLOCK_SIZE_MB16X16)
    super_block_yrd_for_txfm(cm, x, &r[TX_16X16][0], &d[TX_16X16], &s[TX_16X16],
                             bs, TX_16X16);
  super_block_yrd_for_txfm(cm, x, &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8], bs,
                           TX_8X8);
  super_block_yrd_for_txfm(cm, x, &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4], bs,
                           TX_4X4);

  choose_txfm_size_from_rd(cpi, x, r, rate, d, distortion, s,
                           skip, txfm_cache,
                           TX_32X32 - (bs < BLOCK_SIZE_SB32X32)
                           - (bs < BLOCK_SIZE_MB16X16));
}

static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
                                     MB_PREDICTION_MODE *best_mode,
                                     int *bmode_costs,
                                     ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
                                     int *bestrate, int *bestratey,
                                     int *bestdistortion,
                                     BLOCK_SIZE_TYPE bsize) {
  MB_PREDICTION_MODE mode;
  MACROBLOCKD *xd = &x->e_mbd;
  int64_t best_rd = INT64_MAX;
  int rate = 0;
  int distortion;
  VP9_COMMON *const cm = &cpi->common;
  const int src_stride = x->plane[0].src.stride;
  uint8_t *src, *dst;
  int16_t *src_diff, *coeff;

  ENTROPY_CONTEXT ta[2], tempa[2];
  ENTROPY_CONTEXT tl[2], templ[2];
  TX_TYPE tx_type = DCT_DCT;
  TX_TYPE best_tx_type = DCT_DCT;
  int bw = 1 << b_width_log2(bsize);
  int bh = 1 << b_height_log2(bsize);
  int idx, idy, block;
  DECLARE_ALIGNED(16, int16_t, best_dqcoeff[4][16]);

  assert(ib < 4);

  vpx_memcpy(ta, a, sizeof(ta));
  vpx_memcpy(tl, l, sizeof(tl));
  xd->mode_info_context->mbmi.txfm_size = TX_4X4;

  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
    int64_t this_rd;
    int ratey = 0;

    rate = bmode_costs[mode];
    distortion = 0;

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

    for (idy = 0; idy < bh; ++idy) {
      for (idx = 0; idx < bw; ++idx) {
        block = ib + idy * 2 + idx;
        xd->mode_info_context->bmi[block].as_mode.first = mode;
        src = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
                                        x->plane[0].src.buf, src_stride);
        src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, block,
                                             x->plane[0].src_diff);
        coeff = BLOCK_OFFSET(x->plane[0].coeff, block, 16);
        dst = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
                                        xd->plane[0].dst.buf,
                                        xd->plane[0].dst.stride);
        vp9_intra4x4_predict(xd, block, BLOCK_SIZE_SB8X8, mode,
                             dst, xd->plane[0].dst.stride);
        vp9_subtract_block(4, 4, src_diff, 8,
                           src, src_stride,
                           dst, xd->plane[0].dst.stride);

        tx_type = get_tx_type_4x4(xd, block);
        if (tx_type != DCT_DCT) {
          vp9_short_fht4x4(src_diff, coeff, 8, tx_type);
          x->quantize_b_4x4(x, block, tx_type, 16);
        } else {
          x->fwd_txm4x4(src_diff, coeff, 16);
          x->quantize_b_4x4(x, block, tx_type, 16);
        }

        ratey += cost_coeffs(cm, x, 0, block, PLANE_TYPE_Y_WITH_DC,
                             tempa + idx, templ + idy, TX_4X4, 16);
        distortion += vp9_block_error(coeff, BLOCK_OFFSET(xd->plane[0].dqcoeff,
                                                         block, 16), 16) >> 2;

        if (best_tx_type != DCT_DCT)
          vp9_short_iht4x4_add(BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16),
                               dst, xd->plane[0].dst.stride, best_tx_type);
        else
          xd->inv_txm4x4_add(BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16),
                             dst, xd->plane[0].dst.stride);
      }
    }

    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;
      best_tx_type = tx_type;
      vpx_memcpy(a, tempa, sizeof(tempa));
      vpx_memcpy(l, templ, sizeof(templ));
      for (idy = 0; idy < bh; ++idy) {
        for (idx = 0; idx < bw; ++idx) {
          block = ib + idy * 2 + idx;
          vpx_memcpy(best_dqcoeff[idy * 2 + idx],
                     BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16),
                     sizeof(best_dqcoeff[0]));
        }
      }
    }
  }

  for (idy = 0; idy < bh; ++idy) {
    for (idx = 0; idx < bw; ++idx) {
      block = ib + idy * 2 + idx;
      xd->mode_info_context->bmi[block].as_mode.first = *best_mode;
      dst = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
                                      xd->plane[0].dst.buf,
                                      xd->plane[0].dst.stride);

      vp9_intra4x4_predict(xd, block, BLOCK_SIZE_SB8X8, *best_mode,
                           dst, xd->plane[0].dst.stride);
      // inverse transform
      if (best_tx_type != DCT_DCT)
        vp9_short_iht4x4_add(best_dqcoeff[idy * 2 + idx], dst,
                             xd->plane[0].dst.stride, best_tx_type);
      else
        xd->inv_txm4x4_add(best_dqcoeff[idy * 2 + idx], dst,
                           xd->plane[0].dst.stride);
    }
  }

  return best_rd;
}

static int64_t rd_pick_intra4x4mby_modes(VP9_COMP *cpi, MACROBLOCK *mb,
                                         int *Rate, int *rate_y,
                                         int *Distortion, int64_t best_rd) {
  int i, j;
  MACROBLOCKD *const xd = &mb->e_mbd;
  BLOCK_SIZE_TYPE bsize = xd->mode_info_context->mbmi.sb_type;
  int bw = 1 << b_width_log2(bsize);
  int bh = 1 << b_height_log2(bsize);
  int idx, idy;
  int cost = 0;
  int distortion = 0;
  int tot_rate_y = 0;
  int64_t total_rd = 0;
  ENTROPY_CONTEXT t_above[4], t_left[4];
  int *bmode_costs;
  MODE_INFO *const mic = xd->mode_info_context;

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

  bmode_costs = mb->mbmode_cost;

  for (idy = 0; idy < 2; idy += bh) {
    for (idx = 0; idx < 2; idx += bw) {
      const int mis = xd->mode_info_stride;
      MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
      int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry);
      int UNINITIALIZED_IS_SAFE(d);
      i = idy * 2 + idx;

      if (xd->frame_type == KEY_FRAME) {
        const MB_PREDICTION_MODE A = above_block_mode(mic, i, mis);
        const MB_PREDICTION_MODE L = (xd->left_available || idx) ?
                                     left_block_mode(mic, i) : DC_PRED;

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

      total_rd += rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
                                        t_above + idx, t_left + idy,
                                        &r, &ry, &d, bsize);
      cost += r;
      distortion += d;
      tot_rate_y += ry;

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

      if (total_rd >= best_rd)
        break;
    }
  }

  if (total_rd >= best_rd)
    return INT64_MAX;

  *Rate = cost;
  *rate_y = tot_rate_y;
  *Distortion = distortion;
  xd->mode_info_context->mbmi.mode = mic->bmi[3].as_mode.first;

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

static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                      int *rate, int *rate_tokenonly,
                                      int *distortion, int *skippable,
                                      BLOCK_SIZE_TYPE bsize,
                                      int64_t txfm_cache[NB_TXFM_MODES]) {
  MB_PREDICTION_MODE mode;
  MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
  MACROBLOCKD *const xd = &x->e_mbd;
  int this_rate, this_rate_tokenonly;
  int this_distortion, s;
  int64_t best_rd = INT64_MAX, this_rd;
  TX_SIZE UNINITIALIZED_IS_SAFE(best_tx);
  int i;
  int *bmode_costs = x->mbmode_cost;

  if (bsize < BLOCK_SIZE_SB8X8) {
    x->e_mbd.mode_info_context->mbmi.txfm_size = TX_4X4;
    return best_rd;
  }

  for (i = 0; i < NB_TXFM_MODES; i++)
    txfm_cache[i] = INT64_MAX;

  /* Y Search for 32x32 intra prediction mode */
  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
    int64_t local_txfm_cache[NB_TXFM_MODES];
    MODE_INFO *const mic = xd->mode_info_context;
    const int mis = xd->mode_info_stride;

    if (cpi->common.frame_type == KEY_FRAME) {
      const MB_PREDICTION_MODE A = above_block_mode(mic, 0, mis);
      const MB_PREDICTION_MODE L = xd->left_available ?
                                   left_block_mode(mic, 0) : DC_PRED;

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