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 <limits.h>
#include <math.h>
#include <stdio.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_pragmas.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_mcomp.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/encoder/vp9_variance.h"

#define RD_THRESH_MAX_FACT 64
#define RD_THRESH_INC      1
#define RD_THRESH_POW      1.25
#define RD_MULT_EPB_RATIO  64

/* Factor to weigh the rate for switchable interp filters */
#define SWITCHABLE_INTERP_RATE_FACTOR 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 {
  MB_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}},
};

// 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 const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
};

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 fill_mode_costs(VP9_COMP *cpi) {
  MACROBLOCK *const x = &cpi->mb;
  const FRAME_CONTEXT *const fc = &cpi->common.fc;
  int i, j;

  for (i = 0; i < INTRA_MODES; i++)
    for (j = 0; j < INTRA_MODES; j++)
      vp9_cost_tokens((int *)x->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
                      vp9_intra_mode_tree);

  // TODO(rbultje) separate tables for superblock costing?
  vp9_cost_tokens(x->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
  vp9_cost_tokens(x->intra_uv_mode_cost[KEY_FRAME],
                  vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
  vp9_cost_tokens(x->intra_uv_mode_cost[INTER_FRAME],
                  fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree);

  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
    vp9_cost_tokens((int *)x->switchable_interp_costs[i],
                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
}

static void fill_token_costs(vp9_coeff_cost *c,
                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
  int i, j, k, l;
  TX_SIZE t;
  for (t = TX_4X4; t <= TX_32X32; ++t)
    for (i = 0; i < PLANE_TYPES; ++i)
      for (j = 0; j < REF_TYPES; ++j)
        for (k = 0; k < COEF_BANDS; ++k)
          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
            vp9_prob probs[ENTROPY_NODES];
            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
                            vp9_coef_tree);
            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
                                 vp9_coef_tree);
            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
                   c[t][i][j][k][1][l][EOB_TOKEN]);
          }
}

static const uint8_t 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++) {
    const double q = vp9_convert_qindex_to_q(i);
    sad_per_bit16lut[i] = (int)(0.0418 * q + 2.4107);
    sad_per_bit4lut[i] = (int)(0.063 * q + 2.742);
  }
}

int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
  const int q = vp9_dc_quant(qindex, 0);
  // TODO(debargha): Adjust the function below
  int rdmult = 88 * q * q / 25;
  if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
    if (cpi->twopass.next_iiratio > 31)
      rdmult += (rdmult * rd_iifactor[31]) >> 4;
    else
      rdmult += (rdmult * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
  }
  return rdmult;
}

static int compute_rd_thresh_factor(int qindex) {
  // TODO(debargha): Adjust the function below
  const int q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12);
  return MAX(q, 8);
}

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

static void set_block_thresholds(VP9_COMP *cpi) {
  const VP9_COMMON *const cm = &cpi->common;
  RD_OPT *const rd = &cpi->rd;
  int i, bsize, segment_id;

  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
    const int qindex = clamp(vp9_get_qindex(&cm->seg, segment_id,
                                            cm->base_qindex) + cm->y_dc_delta_q,
                             0, MAXQ);
    const int q = compute_rd_thresh_factor(qindex);

    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
      // Threshold here seems unnecessarily harsh but fine given actual
      // range of values used for cpi->sf.thresh_mult[].
      const int t = q * rd_thresh_block_size_factor[bsize];
      const int thresh_max = INT_MAX / t;

      for (i = 0; i < MAX_MODES; ++i)
        rd->threshes[segment_id][bsize][i] =
            rd->thresh_mult[i] < thresh_max ? rd->thresh_mult[i] * t / 4
                                            : INT_MAX;

      for (i = 0; i < MAX_REFS; ++i) {
        rd->thresh_sub8x8[segment_id][bsize][i] =
            rd->thresh_mult_sub8x8[i] < thresh_max
                ? rd->thresh_mult_sub8x8[i] * t / 4
                : INT_MAX;
      }
    }
  }
}

void vp9_initialize_rd_consts(VP9_COMP *cpi) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  int i;

  vp9_clear_system_state();

  cpi->rd.RDDIV = RDDIV_BITS;  // in bits (to multiply D by 128)
  cpi->rd.RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);

  x->errorperbit = cpi->rd.RDMULT / RD_MULT_EPB_RATIO;
  x->errorperbit += (x->errorperbit == 0);

  x->select_txfm_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
                         cm->frame_type != KEY_FRAME) ? 0 : 1;

  set_block_thresholds(cpi);

  if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME) {
    fill_token_costs(x->token_costs, cm->fc.coef_probs);

    for (i = 0; i < PARTITION_CONTEXTS; i++)
      vp9_cost_tokens(x->partition_cost[i], get_partition_probs(cm, i),
                      vp9_partition_tree);
  }

  if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
      cm->frame_type == KEY_FRAME) {
    fill_mode_costs(cpi);

    if (!frame_is_intra_only(cm)) {
      vp9_build_nmv_cost_table(x->nmvjointcost,
                               cm->allow_high_precision_mv ? x->nmvcost_hp
                                                           : x->nmvcost,
                               &cm->fc.nmvc, cm->allow_high_precision_mv);

      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
        vp9_cost_tokens((int *)x->inter_mode_cost[i],
                        cm->fc.inter_mode_probs[i], vp9_inter_mode_tree);
    }
  }
}

static const int MAX_XSQ_Q10 = 245727;

static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
  // NOTE: The tables below must be of the same size

  // The functions described below are sampled at the four most significant
  // bits of x^2 + 8 / 256

  // Normalized rate
  // This table models the rate for a Laplacian source
  // source with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expression is:
  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
  // and H(x) is the binary entropy function.
  static const int rate_tab_q10[] = {
    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
     1159,  1086,  1021,   963,   911,   864,   821,   781,
      745,   680,   623,   574,   530,   490,   455,   424,
      395,   345,   304,   269,   239,   213,   190,   171,
      154,   126,   104,    87,    73,    61,    52,    44,
       38,    28,    21,    16,    12,    10,     8,     6,
        5,     3,     2,     1,     1,     1,     0,     0,
  };
  // Normalized distortion
  // This table models the normalized distortion for a Laplacian source
  // source with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expression is:
  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
  // where x = qpstep / sqrt(variance)
  // Note the actual distortion is Dn * variance.
  static const int dist_tab_q10[] = {
       0,     0,     1,     1,     1,     2,     2,     2,
       3,     3,     4,     5,     5,     6,     7,     7,
       8,     9,    11,    12,    13,    15,    16,    17,
      18,    21,    24,    26,    29,    31,    34,    36,
      39,    44,    49,    54,    59,    64,    69,    73,
      78,    88,    97,   106,   115,   124,   133,   142,
     151,   167,   184,   200,   215,   231,   245,   260,
     274,   301,   327,   351,   375,   397,   418,   439,
     458,   495,   528,   559,   587,   613,   637,   659,
     680,   717,   749,   777,   801,   823,   842,   859,
     874,   899,   919,   936,   949,   960,   969,   977,
     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
  };
  static const int xsq_iq_q10[] = {
         0,      4,      8,     12,     16,     20,     24,     28,
        32,     40,     48,     56,     64,     72,     80,     88,
        96,    112,    128,    144,    160,    176,    192,    208,
       224,    256,    288,    320,    352,    384,    416,    448,
       480,    544,    608,    672,    736,    800,    864,    928,
       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
  };
  /*
  static const int tab_size = sizeof(rate_tab_q10) / sizeof(rate_tab_q10[0]);
  assert(sizeof(dist_tab_q10) / sizeof(dist_tab_q10[0]) == tab_size);
  assert(sizeof(xsq_iq_q10) / sizeof(xsq_iq_q10[0]) == tab_size);
  assert(MAX_XSQ_Q10 + 1 == xsq_iq_q10[tab_size - 1]);
  */
  int tmp = (xsq_q10 >> 2) + 8;
  int k = get_msb(tmp) - 3;
  int xq = (k << 3) + ((tmp >> k) & 0x7);
  const int one_q10 = 1 << 10;
  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
  const int b_q10 = one_q10 - a_q10;
  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
}

void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
                                  unsigned int qstep, int *rate,
                                  int64_t *dist) {
  // This function models the rate and distortion for a Laplacian
  // source with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expressions are in:
  // Hang and Chen, "Source Model for transform video coder and its
  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
  // Sys. for Video Tech., April 1997.
  if (var == 0) {
    *rate = 0;
    *dist = 0;
  } else {
    int d_q10, r_q10;
    const uint64_t xsq_q10_64 =
        ((((uint64_t)qstep * qstep * n) << 10) + (var >> 1)) / var;
    const int xsq_q10 = xsq_q10_64 > MAX_XSQ_Q10 ?
                        MAX_XSQ_Q10 : (int)xsq_q10_64;
    model_rd_norm(xsq_q10, &r_q10, &d_q10);
    *rate = (n * r_q10 + 2) >> 2;
    *dist = (var * (int64_t)d_q10 + 512) >> 10;
  }
}

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;

  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);

    (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
                              pd->dst.buf, pd->dst.stride, &sse);

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

    // Fast approximate the modelling function.
    if (cpi->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 {
      int rate;
      int64_t dist;
      vp9_model_rd_from_var_lapndz(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;
}

static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE bsize,
                                 TX_SIZE tx_size,
                                 MACROBLOCK *x, MACROBLOCKD *xd,
                                 int *out_rate_sum, int64_t *out_dist_sum,
                                 int *out_skip) {
  int j, k;
  BLOCK_SIZE bs;
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
  const int width = 4 * num_4x4_blocks_wide_lookup[bsize];
  const int height = 4 * num_4x4_blocks_high_lookup[bsize];
  int rate_sum = 0;
  int64_t dist_sum = 0;
  const int t = 4 << tx_size;

  if (tx_size == TX_4X4) {
    bs = BLOCK_4X4;
  } else if (tx_size == TX_8X8) {
    bs = BLOCK_8X8;
  } else if (tx_size == TX_16X16) {
    bs = BLOCK_16X16;
  } else if (tx_size == TX_32X32) {
    bs = BLOCK_32X32;
  } else {
    assert(0);
  }

  *out_skip = 1;
  for (j = 0; j < height; j += t) {
    for (k = 0; k < width; k += t) {
      int rate;
      int64_t dist;
      unsigned int sse;
      cpi->fn_ptr[bs].vf(&p->src.buf[j * p->src.stride + k], p->src.stride,
                         &pd->dst.buf[j * pd->dst.stride + k], pd->dst.stride,
                         &sse);
      // sse works better than var, since there is no dc prediction used
      vp9_model_rd_from_var_lapndz(sse, t * t, pd->dequant[1] >> 3,
                                   &rate, &dist);
      rate_sum += rate;
      dist_sum += dist;
      *out_skip &= (rate < 1024);
    }
  }

  *out_rate_sum = 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) == 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);
  else
    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;
  }
}

void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
                              const struct macroblockd_plane *pd,
                              ENTROPY_CONTEXT t_above[16],
                              ENTROPY_CONTEXT t_left[16]) {
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
  const ENTROPY_CONTEXT *const above = pd->above_context;
  const ENTROPY_CONTEXT *const left = pd->left_context;

  int i;
  switch (tx_size) {
    case TX_4X4:
      vpx_memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      vpx_memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
      break;
    case TX_8X8:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_16X16:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_32X32:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    default:
      assert(0 && "Invalid transform size.");
  }
}

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 = { 0 };
  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_txfm_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[mbmi->tx_size], ref_best_rd, 0, bs,
                   mbmi->tx_size, cpi->sf.use_fast_coef_costing);
  cpi->tx_stepdown_count[0]++;
}

static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
                                     int (*r)[2], int *rate,
                                     int64_t *d, int64_t *distortion,
                                     int *s, int *skip,
                                     int64_t tx_cache[TX_MODES],
                                     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);
  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 = TX_4X4;

  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 = TX_4X4; n <= max_tx_size; n++) {
    r[n][1] = r[n][0];
    if (r[n][0] < INT_MAX) {
      for (m = 0; m <= n - (n == 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]);
    }

    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];

  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];
    cpi->tx_stepdown_count[0]++;
  } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) {
    tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
    cpi->tx_stepdown_count[max_tx_size - TX_16X16]++;
  } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
    tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1];
    cpi->tx_stepdown_count[max_tx_size - TX_8X8]++;
  } else {
    tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1];
    cpi->tx_stepdown_count[max_tx_size - TX_4X4]++;
  }
}

static int64_t scaled_rd_cost(int rdmult, int rddiv,
                              int rate, int64_t dist, double scale) {
  return (int64_t) (RDCOST(rdmult, rddiv, rate, dist) * scale);
}

static void choose_txfm_size_from_modelrd(VP9_COMP *cpi, MACROBLOCK *x,
                                          int (*r)[2], int *rate,
                                          int64_t *d, int64_t *distortion,
                                          int *s, 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;
  MACROBLOCKD *const xd = &x->e_mbd;