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][2],
                             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 *)c[t][i][j][0][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][1][k][l], probs, vp9_coef_tree);
#else
            vp9_cost_tokens_skip((int *)c[t][i][j][1][k][l], probs,
                                 vp9_coef_tree);
            assert(c[t][i][j][0][k][l][DCT_EOB_TOKEN] ==
                   c[t][i][j][1][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];

        if (cpi->sf.adaptive_rd_thresh)
          cpi->rd_thresh_freq_fact[bsize][i] = MAX_RD_THRESH_FREQ_FACT;
        else
          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];

        if (cpi->sf.adaptive_rd_thresh)
          cpi->rd_thresh_freq_fact[bsize][i] = MAX_RD_THRESH_FREQ_FACT;
        else
          cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
      }
    }
  }

  fill_token_costs(cpi->mb.token_costs, 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);

    for (i = 0; i < INTER_MODE_CONTEXTS; i++) {
      MB_PREDICTION_MODE m;

      for (m = NEARESTMV; m < MB_MODE_COUNT; m++)
        cpi->mb.inter_mode_cost[i][m - NEARESTMV] =
            cost_token(vp9_sb_mv_ref_tree,
                       cpi->common.fc.inter_mode_probs[i],
                       vp9_sb_mv_ref_encoding_array - NEARESTMV + m);
    }
  }
}

static INLINE BLOCK_SIZE_TYPE get_block_size(int bwl, int bhl) {
  return bsize_from_dim_lookup[bwl][bhl];
}

static BLOCK_SIZE_TYPE get_plane_block_size(BLOCK_SIZE_TYPE bsize,
                                            struct macroblockd_plane *pd) {
  return get_block_size(plane_block_width_log2by4(bsize, pd),
                        plane_block_height_log2by4(bsize, pd));
}

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 = ((n << 8) * R + 0.5);
    *dist = (var * D + 0.5);
  }
  vp9_clear_system_state();
}

static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE_TYPE 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;

  for (i = 0; i < MAX_MB_PLANE; ++i) {
    struct macroblock_plane *const p = &x->plane[i];
    struct macroblockd_plane *const pd = &xd->plane[i];

    // TODO(dkovalev) the same code in get_plane_block_size
    const int bwl = plane_block_width_log2by4(bsize, pd);
    const int bhl = plane_block_height_log2by4(bsize, pd);
    const BLOCK_SIZE_TYPE bs = get_block_size(bwl, bhl);
    unsigned int sse;
    int rate;
    int64_t dist;
    (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
                              pd->dst.buf, pd->dst.stride, &sse);
    // sse works better than var, since there is no dc prediction used
    model_rd_from_var_lapndz(sse, 16 << (bwl + bhl),
                             pd->dequant[1] >> 3, &rate, &dist);

    rate_sum += rate;
    dist_sum += dist;
  }

  *out_rate_sum = rate_sum;
  *out_dist_sum = dist_sum << 4;
}

static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE_TYPE 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.
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];

  // TODO(dkovalev) the same code in get_plane_block_size
  const int bwl = plane_block_width_log2by4(bsize, pd);
  const int bhl = plane_block_height_log2by4(bsize, pd);
  const BLOCK_SIZE_TYPE bs = get_block_size(bwl, bhl);
  unsigned int sse;
  int rate;
  int64_t dist;
  (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
                            pd->dst.buf, pd->dst.stride, &sse);
  // sse works better than var, since there is no dc prediction used
  model_rd_from_var_lapndz(sse, 16 << (bwl + bhl),
                           pd->dequant[1] >> 3, &rate, &dist);

  *out_rate_sum = rate;
  *out_dist_sum = dist << 4;
}

static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE_TYPE bsize,
                                 TX_SIZE tx_size,
                                 MACROBLOCK *x, MACROBLOCKD *xd,
                                 int *out_rate_sum, int64_t *out_dist_sum,
                                 int *out_skip) {
  int t = 4, j, k;
  BLOCK_SIZE_TYPE bs = BLOCK_SIZE_AB4X4;
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
  const int bwl = plane_block_width_log2by4(bsize, pd);
  const int bhl = plane_block_height_log2by4(bsize, pd);
  const int bw = 4 << bwl;
  const int bh = 4 << bhl;
  int rate_sum = 0;
  int64_t dist_sum = 0;

  if (tx_size == TX_4X4) {
    bs = BLOCK_4X4;
    t = 4;
  } else if (tx_size == TX_8X8) {
    bs = BLOCK_8X8;
    t = 8;
  } else if (tx_size == TX_16X16) {
    bs = BLOCK_16X16;
    t = 16;
  } else if (tx_size == TX_32X32) {
    bs = BLOCK_32X32;
    t = 32;
  } else {
    assert(0);
  }
  assert(bs <= get_block_size(bwl, bhl));
  *out_skip = 1;
  for (j = 0; j < bh; j+=t) {
    for (k = 0; k < bw; k+=t) {
      int rate;
      int64_t dist;
      unsigned int sse;
      (void) 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;
}

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;
  const int16_t *scan = NULL, *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)[COEF_BANDS][PREV_COEF_CONTEXTS]
                    [MAX_ENTROPY_TOKENS] = mb->token_costs[tx_size][type][ref];
  ENTROPY_CONTEXT above_ec = 0, left_ec = 0;
  TX_TYPE tx_type = DCT_DCT;
  const int segment_id = xd->mode_info_context->mbmi.segment_id;
  int seg_eob = 0;
  uint8_t token_cache[1024];
  const uint8_t *band_translate = NULL;

  // 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 {
    assert(tx_size == get_uv_tx_size(mbmi));
  }

  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 TX_TYPE tx_type = type == PLANE_TYPE_Y_WITH_DC ?
                                  get_tx_type_8x8(xd) : 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 TX_TYPE tx_type = type == PLANE_TYPE_Y_WITH_DC ?
                                  get_tx_type_16x16(xd) : 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:
      assert(0);
      break;
  }
  assert(eob <= seg_eob);

  pt = combine_entropy_contexts(above_ec, left_ec);
  nb = vp9_get_coef_neighbors_handle(scan);

  if (vp9_segfeature_active(&xd->seg, 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);

  if (eob == 0) {
    // single eob token
    cost += token_costs[0][0][pt][DCT_EOB_TOKEN];
  } else {
    int v, prev_t;

    // dc token
    v = qcoeff_ptr[0];
    prev_t = vp9_dct_value_tokens_ptr[v].token;
    cost += token_costs[0][0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
    token_cache[0] = vp9_pt_energy_class[prev_t];

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

      v = qcoeff_ptr[rc];
      t = vp9_dct_value_tokens_ptr[v].token;
      pt = get_coef_context(nb, token_cache, c);
      cost += token_costs[!prev_t][band][pt][t] + vp9_dct_value_cost_ptr[v];
      token_cache[rc] = vp9_pt_energy_class[t];
      prev_t = t;
    }

    // eob token
    if (c < seg_eob) {
      pt = get_coef_context(nb, token_cache, c);
      cost += token_costs[0][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;
}

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 rate;
  int64_t dist;
  int64_t sse;
  int64_t best_rd;
  int skip;
};

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

static void rate_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->rate += 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);
}

// FIXME(jingning): need to make the rd test of chroma components consistent
// with that of luma component. this function should be deprecated afterwards.
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, 0, 0, INT64_MAX, 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, rate_block, &args);
  return args.rate;
}

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 int block_error_sby(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize,
                           int shift, int64_t *sse) {
  struct macroblockd_plane *p = &x->e_mbd.plane[0];
  const int bwl = plane_block_width_log2by4(bsize, p);
  const int bhl = plane_block_height_log2by4(bsize, p);
  int64_t e = vp9_block_error(x->plane[0].coeff, x->e_mbd.plane[0].dqcoeff,
                              16 << (bwl + bhl), sse) >> shift;
  *sse >>= shift;
  return e;
}

static int64_t block_error_sbuv(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize,
                                int shift, int64_t *sse) {
  int64_t sum = 0, this_sse;
  int plane;

  *sse = 0;
  for (plane = 1; plane < MAX_MB_PLANE; plane++) {
    struct macroblockd_plane *p = &x->e_mbd.plane[plane];
    const int bwl = plane_block_width_log2by4(bsize, p);
    const int bhl = plane_block_height_log2by4(bsize, p);
    sum += vp9_block_error(x->plane[plane].coeff, x->e_mbd.plane[plane].dqcoeff,
                           16 << (bwl + bhl), &this_sse);
    *sse += this_sse;
  }
  *sse >>= shift;
  return sum >> shift;
}

static void block_yrd_txfm(int plane, int block, BLOCK_SIZE_TYPE bsize,
                           int ss_txfrm_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 = {args->cm, x, NULL};

  if (args->skip)
    return;
  if (RDCOST(x->rdmult, x->rddiv, args->rate, args->dist) > args->best_rd) {
    args->skip = 1;
    args->rate = INT_MAX;
    args->dist = INT64_MAX;
    args->sse  = INT64_MAX;
    return;
  }

  if (xd->mode_info_context->mbmi.ref_frame[0] == INTRA_FRAME)
    encode_block_intra(plane, block, bsize, ss_txfrm_size, &encode_args);
  else
    xform_quant(plane, block, bsize, ss_txfrm_size, &encode_args);

  dist_block(plane, block, bsize, ss_txfrm_size, args);
  rate_block(plane, block, bsize, ss_txfrm_size, args);
}

static void super_block_yrd_for_txfm(VP9_COMMON *const cm, MACROBLOCK *x,
                                     int *rate, int64_t *distortion,
                                     int *skippable, int64_t *sse,
                                     int64_t ref_best_rd,
                                     BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblockd_plane *const pd = &xd->plane[0];
  const int bwl = b_width_log2(bsize) - xd->plane[0].subsampling_x;
  const int bhl = b_height_log2(bsize) - xd->plane[0].subsampling_y;
  const int bw = 1 << bwl, bh = 1 << bhl;
  struct rdcost_block_args args = { cm, x, { 0 }, { 0 }, tx_size, bw, bh,
                                    0, 0, 0, ref_best_rd, 0 };
  xd->mode_info_context->mbmi.txfm_size = tx_size;
  vpx_memcpy(&args.t_above, pd->above_context, sizeof(ENTROPY_CONTEXT) * bw);
  vpx_memcpy(&args.t_left, pd->left_context, sizeof(ENTROPY_CONTEXT) * bh);

  foreach_transformed_block_in_plane(xd, bsize, 0, block_yrd_txfm, &args);
  *distortion = args.dist;
  *rate       = args.rate;
  *sse        = args.sse;
  *skippable  = vp9_sby_is_skippable(xd, bsize) && (!args.skip);
}

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_TYPE bs) {
  const TX_SIZE max_txfm_size = TX_32X32
      - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
  if (max_txfm_size == TX_32X32 &&
      (cm->txfm_mode == ALLOW_32X32 ||
       cm->txfm_mode == TX_MODE_SELECT)) {
    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)) {
    mbmi->txfm_size = TX_16X16;
  } else if (cm->txfm_mode != ONLY_4X4) {
    mbmi->txfm_size = TX_8X8;
  } else {
    mbmi->txfm_size = TX_4X4;
  }
  super_block_yrd_for_txfm(cm, x, rate, distortion, skip,
                           &sse[mbmi->txfm_size], ref_best_rd, bs,
                           mbmi->txfm_size);
  cpi->txfm_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 txfm_cache[NB_TXFM_MODES],
                                     BLOCK_SIZE_TYPE bs) {
  const TX_SIZE max_txfm_size = TX_32X32
      - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
  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_mbskip(cm, xd);
  int64_t rd[TX_SIZE_MAX_SB][2];
  int n, m;
  int s0, s1;

  const vp9_prob *tx_probs = vp9_get_pred_probs_tx_size(cm, xd);

  for (n = TX_4X4; n <= max_txfm_size; n++) {
    r[n][1] = r[n][0];
    if (r[n][0] == INT_MAX)
      continue;
    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 (d[n] == INT64_MAX) {
      rd[n][0] = rd[n][1] = INT64_MAX;
      continue;
    }
    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];

  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]) {
    cpi->txfm_stepdown_count[0]++;
  } else if (max_txfm_size >= TX_16X16 &&
             rd[TX_16X16][1] < rd[TX_8X8][1] &&
             rd[TX_16X16][1] < rd[TX_4X4][1]) {
    cpi->txfm_stepdown_count[max_txfm_size - TX_16X16]++;
  } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
    cpi->txfm_stepdown_count[max_txfm_size - TX_8X8]++;
  } else {
    cpi->txfm_stepdown_count[max_txfm_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_TYPE bs,
                                          int *model_used) {
  const TX_SIZE max_txfm_size = TX_32X32
      - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
  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_mbskip(cm, xd);
  int64_t rd[TX_SIZE_MAX_SB][2];
  int n, m;
  int s0, s1;
  double scale_rd[TX_SIZE_MAX_SB] = {1.73, 1.44, 1.20, 1.00};
  // double scale_r[TX_SIZE_MAX_SB] = {2.82, 2.00, 1.41, 1.00};

  const vp9_prob *tx_probs = vp9_get_pred_probs_tx_size(cm, xd);

  // for (n = TX_4X4; n <= max_txfm_size; n++)
  //   r[n][0] = (r[n][0] * scale_r[n]);

  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]);
    }
  }
  for (n = TX_4X4; n <= max_txfm_size; n++) {
    rd[n][0] = (scale_rd[n] * rd[n][0]);
    rd[n][1] = (scale_rd[n] * rd[n][1]);
  }

  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;
  }

  if (model_used[mbmi->txfm_size]) {
    // Actually encode using the chosen mode if a model was used, but do not
    // update the r, d costs
    super_block_yrd_for_txfm(cm, x, rate, distortion, skip,
                             &sse[mbmi->txfm_size], ref_best_rd,
                             bs, mbmi->txfm_size);