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/common/vp9_entropymode.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.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

#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;

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}},
};

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

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

#define MV_COST_WEIGHT      108
#define MV_COST_WEIGHT_SUB  120

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 *c) {
  VP9_COMMON *const cm = &c->common;
  int i, j;

  for (i = 0; i < INTRA_MODES; i++)
    for (j = 0; j < INTRA_MODES; j++)
      vp9_cost_tokens((int *)c->mb.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(c->mb.mbmode_cost, cm->fc.y_mode_prob[1],
                  vp9_intra_mode_tree);
  vp9_cost_tokens(c->mb.intra_uv_mode_cost[1],
                  cm->fc.uv_mode_prob[INTRA_MODES - 1], vp9_intra_mode_tree);
  vp9_cost_tokens(c->mb.intra_uv_mode_cost[0],
                  vp9_kf_uv_mode_prob[INTRA_MODES - 1],
                  vp9_intra_mode_tree);

  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
    vp9_cost_tokens((int *)c->mb.switchable_interp_costs[i],
                    cm->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 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);
  }
}

int vp9_compute_rd_mult(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) {
  int q;
  // TODO(debargha): Adjust the function below
  q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12);
  if (q < 8)
    q = 8;
  return q;
}

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) {
  int i, bsize, segment_id;
  VP9_COMMON *cm = &cpi->common;

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

    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
      // 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]);

      for (i = 0; i < MAX_MODES; ++i) {
        if (cpi->sf.thresh_mult[i] < thresh_max) {
          cpi->rd_threshes[segment_id][bsize][i] =
              cpi->sf.thresh_mult[i] * q *
              rd_thresh_block_size_factor[bsize] / 4;
        } else {
          cpi->rd_threshes[segment_id][bsize][i] = INT_MAX;
        }
      }

      for (i = 0; i < MAX_REFS; ++i) {
        if (cpi->sf.thresh_mult_sub8x8[i] < thresh_max) {
          cpi->rd_thresh_sub8x8[segment_id][bsize][i] =
              cpi->sf.thresh_mult_sub8x8[i] * q *
              rd_thresh_block_size_factor[bsize] / 4;
        } else {
          cpi->rd_thresh_sub8x8[segment_id][bsize][i] = INT_MAX;
        }
      }
    }
  }
}

void vp9_initialize_rd_consts(VP9_COMP *cpi) {
  VP9_COMMON *cm = &cpi->common;
  int qindex, i;

  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(cm->base_qindex + cm->y_dc_delta_q, 0, MAXQ);

  cpi->RDDIV = RDDIV_BITS;  // in bits (to multiply D by 128)
  cpi->RDMULT = vp9_compute_rd_mult(cpi, qindex);

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

  vp9_set_speed_features(cpi);

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

  set_block_thresholds(cpi);

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

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

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

  if (!frame_is_intra_only(cm)) {
    vp9_build_nmv_cost_table(
        cpi->mb.nmvjointcost,
        cm->allow_high_precision_mv ? cpi->mb.nmvcost_hp : cpi->mb.nmvcost,
        &cm->fc.nmvc,
        cm->allow_high_precision_mv, 1, 1);

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

static INLINE void linear_interpolate2(double x, int ntab, int inv_step,
                                       const double *tab1, const double *tab2,
                                       double *v1, double *v2) {
  double y = x * inv_step;
  int d = (int) y;
  if (d >= ntab - 1) {
    *v1 = tab1[ntab - 1];
    *v2 = tab2[ntab - 1];
  } else {
    double a = y - d;
    *v1 = tab1[d] * (1 - a) + tab1[d + 1] * a;
    *v2 = tab2[d] * (1 - a) + tab2[d + 1] * a;
  }
}

static void model_rd_norm(double x, double *R, double *D) {
  static const int inv_tab_step = 8;
  static const int tab_size = 120;
  // NOTE: The tables below must be of the same size
  //
  // 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 double rate_tab[] = {
    64.00, 4.944, 3.949, 3.372, 2.966, 2.655, 2.403, 2.194,
    2.014, 1.858, 1.720, 1.596, 1.485, 1.384, 1.291, 1.206,
    1.127, 1.054, 0.986, 0.923, 0.863, 0.808, 0.756, 0.708,
    0.662, 0.619, 0.579, 0.541, 0.506, 0.473, 0.442, 0.412,
    0.385, 0.359, 0.335, 0.313, 0.291, 0.272, 0.253, 0.236,
    0.220, 0.204, 0.190, 0.177, 0.165, 0.153, 0.142, 0.132,
    0.123, 0.114, 0.106, 0.099, 0.091, 0.085, 0.079, 0.073,
    0.068, 0.063, 0.058, 0.054, 0.050, 0.047, 0.043, 0.040,
    0.037, 0.034, 0.032, 0.029, 0.027, 0.025, 0.023, 0.022,
    0.020, 0.019, 0.017, 0.016, 0.015, 0.014, 0.013, 0.012,
    0.011, 0.010, 0.009, 0.008, 0.008, 0.007, 0.007, 0.006,
    0.006, 0.005, 0.005, 0.005, 0.004, 0.004, 0.004, 0.003,
    0.003, 0.003, 0.003, 0.002, 0.002, 0.002, 0.002, 0.002,
    0.002, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001,
    0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.000,
  };
  // 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 double dist_tab[] = {
    0.000, 0.001, 0.005, 0.012, 0.021, 0.032, 0.045, 0.061,
    0.079, 0.098, 0.119, 0.142, 0.166, 0.190, 0.216, 0.242,
    0.269, 0.296, 0.324, 0.351, 0.378, 0.405, 0.432, 0.458,
    0.484, 0.509, 0.534, 0.557, 0.580, 0.603, 0.624, 0.645,
    0.664, 0.683, 0.702, 0.719, 0.735, 0.751, 0.766, 0.780,
    0.794, 0.807, 0.819, 0.830, 0.841, 0.851, 0.861, 0.870,
    0.878, 0.886, 0.894, 0.901, 0.907, 0.913, 0.919, 0.925,
    0.930, 0.935, 0.939, 0.943, 0.947, 0.951, 0.954, 0.957,
    0.960, 0.963, 0.966, 0.968, 0.971, 0.973, 0.975, 0.976,
    0.978, 0.980, 0.981, 0.982, 0.984, 0.985, 0.986, 0.987,
    0.988, 0.989, 0.990, 0.990, 0.991, 0.992, 0.992, 0.993,
    0.993, 0.994, 0.994, 0.995, 0.995, 0.996, 0.996, 0.996,
    0.996, 0.997, 0.997, 0.997, 0.997, 0.998, 0.998, 0.998,
    0.998, 0.998, 0.998, 0.999, 0.999, 0.999, 0.999, 0.999,
    0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 1.000,
  };
  /*
  assert(sizeof(rate_tab) == tab_size * sizeof(rate_tab[0]);
  assert(sizeof(dist_tab) == tab_size * sizeof(dist_tab[0]);
  assert(sizeof(rate_tab) == sizeof(dist_tab));
  */
  assert(x >= 0.0);
  linear_interpolate2(x, tab_size, inv_tab_step,
                      rate_tab, dist_tab, R, D);
}

static void model_rd_from_var_lapndz(int var, int n, 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.
  vp9_clear_system_state();
  if (var == 0 || n == 0) {
    *rate = 0;
    *dist = 0;
  } else {
    double D, R;
    double s2 = (double) var / n;
    double x = qstep / sqrt(s2);
    model_rd_norm(x, &R, &D);
    *rate = (int)((n << 8) * R + 0.5);
    *dist = (int)(var * D + 0.5);
  }
  vp9_clear_system_state();
}

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, rate_sum = 0, dist_sum = 0;
  int ref = xd->mi_8x8[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);
    int rate;
    int64_t dist;
    (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;
    // sse works better than var, since there is no dc prediction used
    model_rd_from_var_lapndz(sse, 1 << num_pels_log2_lookup[bs],
                             pd->dequant[1] >> 3, &rate, &dist);

    rate_sum += rate;
    dist_sum += (int)dist;
  }

  *out_rate_sum = 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;
  struct macroblock_plane *const p = &x->plane[0];
  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
      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(int16_t *coeff, 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++) {
    int this_diff = coeff[i] - dqcoeff[i];
    error += (unsigned)this_diff * this_diff;
    sqcoeff += (unsigned) 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) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct macroblock_plane *p = &x->plane[plane];
  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_ptr = BLOCK_OFFSET(p->qcoeff, block);
  const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
  unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
                   x->token_costs[tx_size][type][ref];
  const ENTROPY_CONTEXT above_ec = !!*A, left_ec = !!*L;
  uint8_t *p_tok = x->token_cache;
  int pt = combine_entropy_contexts(above_ec, left_ec);
  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_ptr[0];
    int prev_t = vp9_dct_value_tokens_ptr[v].token;
    cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
    p_tok[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_ptr[rc];
      t = vp9_dct_value_tokens_ptr[v].token;
      pt = get_coef_context(nb, p_tok, c);
      cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
      p_tok[rc] = vp9_pt_energy_class[t];
      prev_t = t;
      if (!--band_left) {
        band_left = *band_count++;
        ++token_costs;
      }
    }

    // eob token
    if (band_left) {
      pt = get_coef_context(nb, p_tok, 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, void *arg) {
  const int ss_txfrm_size = tx_size << 1;
  struct rdcost_block_args* args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int64_t this_sse;
  int shift = args->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_8x8[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, void *arg) {
  struct rdcost_block_args* args = arg;

  int x_idx, y_idx;
  txfrm_block_to_raster_xy(plane_bsize, args->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, args->tx_size,
                           args->scan, args->nb);
}

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;
  struct encode_b_args encode_args = {x, NULL};
  int64_t rd1, rd2, rd;

  if (args->skip)
    return;

  if (!is_inter_block(&xd->mi_8x8[0]->mbmi))
    vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &encode_args);
  else
    vp9_xform_quant(plane, block, plane_bsize, tx_size, &encode_args);

  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(TX_SIZE tx_size,
    ENTROPY_CONTEXT t_above[16], ENTROPY_CONTEXT t_left[16],
    const ENTROPY_CONTEXT *above, const ENTROPY_CONTEXT *left,
    int num_4x4_w, int num_4x4_h) {
  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 init_rdcost_stack(MACROBLOCK *x, TX_SIZE tx_size,
                              const int num_4x4_w, const int num_4x4_h,
                              const int64_t ref_rdcost,
                              struct rdcost_block_args *arg) {
  vpx_memset(arg, 0, sizeof(struct rdcost_block_args));
  arg->x = x;
  arg->tx_size = tx_size;
  arg->bw = num_4x4_w;
  arg->bh = num_4x4_h;
  arg->best_rd = ref_rdcost;
}

static void txfm_rd_in_plane(MACROBLOCK *x,
                             struct rdcost_block_args *rd_stack,
                             int *rate, int64_t *distortion,
                             int *skippable, int64_t *sse,
                             int64_t ref_best_rd, int plane,
                             BLOCK_SIZE bsize, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
  const int num_4x4_w = num_4x4_blocks_wide_lookup[bs];
  const int num_4x4_h = num_4x4_blocks_high_lookup[bs];
  const scan_order *so;

  init_rdcost_stack(x, tx_size, num_4x4_w, num_4x4_h,
                    ref_best_rd, rd_stack);
  if (plane == 0)
    xd->mi_8x8[0]->mbmi.tx_size = tx_size;

  vp9_get_entropy_contexts(tx_size, rd_stack->t_above, rd_stack->t_left,
                           pd->above_context, pd->left_context,
                           num_4x4_w, num_4x4_h);

  so = get_scan(xd, tx_size, pd->plane_type, 0);
  rd_stack->scan = so->scan;
  rd_stack->nb = so->neighbors;

  foreach_transformed_block_in_plane(xd, bsize, plane,
                                     block_rd_txfm, rd_stack);
  if (rd_stack->skip) {
    *rate       = INT_MAX;
    *distortion = INT64_MAX;
    *sse        = INT64_MAX;
    *skippable  = 0;
  } else {
    *distortion = rd_stack->this_dist;
    *rate       = rd_stack->this_rate;
    *sse        = rd_stack->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_8x8[0]->mbmi;

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

  txfm_rd_in_plane(x, &cpi->rdcost_stack, rate, distortion, skip,
                   &sse[mbmi->tx_size], ref_best_rd, 0, bs,
                   mbmi->tx_size);
  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_8x8[0]->mbmi;
  vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
  int64_t rd[TX_SIZES][2];
  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 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;
  MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
  vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
  int64_t rd[TX_SIZES][2];
  int n, m;
  int s0, s1;
  double scale_rd[TX_SIZES] = {1.73, 1.44, 1.20, 1.00};
  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++) {
    double scale = scale_rd[n];
    r[n][1] = r[n][0];
    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 (s[n]) {
      rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]) * scale;
    } else {
      rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]) * scale;
      rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]) * scale;
    }
    if (rd[n][1] < best_rd) {
      best_rd = rd[n][1];
      best_tx = n;
    }
  }

  mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
                      best_tx : MIN(max_tx_size, max_mode_tx_size);

  // Actually encode using the chosen mode if a model was used, but do not
  // update the r, d costs
  txfm_rd_in_plane(x, &cpi->rdcost_stack, rate, distortion, skip,
                   &sse[mbmi->tx_size], ref_best_rd, 0, bs, mbmi->tx_size);

  if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {