vp9/encoder/vp9_encodemv.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 "vp9/common/vp9_common.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_systemdependent.h"

#include <math.h>

#ifdef ENTROPY_STATS
extern unsigned int active_section;
#endif

#ifdef NMV_STATS
nmv_context_counts tnmvcounts;
#endif

static void encode_mv_component(vp9_writer* w, int comp,
                                const nmv_component* mvcomp, int usehp) {
  int offset;
  const int sign = comp < 0;
  const int mag = sign ? -comp : comp;
  const int mv_class = vp9_get_mv_class(mag - 1, &offset);
  const int d = offset >> 3;                // int mv data
  const int fr = (offset >> 1) & 3;         // fractional mv data
  const int hp = offset & 1;                // high precision mv data

  assert(comp != 0);

  // Sign
  vp9_write(w, sign, mvcomp->sign);

  // Class
  write_token(w, vp9_mv_class_tree, mvcomp->classes,
              &vp9_mv_class_encodings[mv_class]);

  // Integer bits
  if (mv_class == MV_CLASS_0) {
    write_token(w, vp9_mv_class0_tree, mvcomp->class0,
                &vp9_mv_class0_encodings[d]);
  } else {
    int i;
    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
    for (i = 0; i < n; ++i)
      vp9_write(w, (d >> i) & 1, mvcomp->bits[i]);
  }

  // Fractional bits
  write_token(w, vp9_mv_fp_tree,
              mv_class == MV_CLASS_0 ?  mvcomp->class0_fp[d] : mvcomp->fp,
              &vp9_mv_fp_encodings[fr]);

  // High precision bit
  if (usehp)
    vp9_write(w, hp,
              mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
}


static void build_nmv_component_cost_table(int *mvcost,
                                           const nmv_component* const mvcomp,
                                           int usehp) {
  int i, v;
  int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
  int bits_cost[MV_OFFSET_BITS][2];
  int class0_fp_cost[CLASS0_SIZE][4], fp_cost[4];
  int class0_hp_cost[2], hp_cost[2];

  sign_cost[0] = vp9_cost_zero(mvcomp->sign);
  sign_cost[1] = vp9_cost_one(mvcomp->sign);
  vp9_cost_tokens(class_cost, mvcomp->classes, vp9_mv_class_tree);
  vp9_cost_tokens(class0_cost, mvcomp->class0, vp9_mv_class0_tree);
  for (i = 0; i < MV_OFFSET_BITS; ++i) {
    bits_cost[i][0] = vp9_cost_zero(mvcomp->bits[i]);
    bits_cost[i][1] = vp9_cost_one(mvcomp->bits[i]);
  }

  for (i = 0; i < CLASS0_SIZE; ++i)
    vp9_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], vp9_mv_fp_tree);
  vp9_cost_tokens(fp_cost, mvcomp->fp, vp9_mv_fp_tree);

  if (usehp) {
    class0_hp_cost[0] = vp9_cost_zero(mvcomp->class0_hp);
    class0_hp_cost[1] = vp9_cost_one(mvcomp->class0_hp);
    hp_cost[0] = vp9_cost_zero(mvcomp->hp);
    hp_cost[1] = vp9_cost_one(mvcomp->hp);
  }
  mvcost[0] = 0;
  for (v = 1; v <= MV_MAX; ++v) {
    int z, c, o, d, e, f, cost = 0;
    z = v - 1;
    c = vp9_get_mv_class(z, &o);
    cost += class_cost[c];
    d = (o >> 3);               /* int mv data */
    f = (o >> 1) & 3;           /* fractional pel mv data */
    e = (o & 1);                /* high precision mv data */
    if (c == MV_CLASS_0) {
      cost += class0_cost[d];
    } else {
      int i, b;
      b = c + CLASS0_BITS - 1;  /* number of bits */
      for (i = 0; i < b; ++i)
        cost += bits_cost[i][((d >> i) & 1)];
    }
    if (c == MV_CLASS_0) {
      cost += class0_fp_cost[d][f];
    } else {
      cost += fp_cost[f];
    }
    if (usehp) {
      if (c == MV_CLASS_0) {
        cost += class0_hp_cost[e];
      } else {
        cost += hp_cost[e];
      }
    }
    mvcost[v] = cost + sign_cost[0];
    mvcost[-v] = cost + sign_cost[1];
  }
}

static int update_nmv_savings(const unsigned int ct[2],
                              const vp9_prob cur_p,
                              const vp9_prob new_p,
                              const vp9_prob upd_p) {

#ifdef LOW_PRECISION_MV_UPDATE
  vp9_prob mod_p = new_p | 1;
#else
  vp9_prob mod_p = new_p;
#endif
  const int cur_b = cost_branch256(ct, cur_p);
  const int mod_b = cost_branch256(ct, mod_p);
  const int cost = 7 * 256 +
#ifndef LOW_PRECISION_MV_UPDATE
      256 +
#endif
      (vp9_cost_one(upd_p) - vp9_cost_zero(upd_p));
  if (cur_b - mod_b - cost > 0) {
    return cur_b - mod_b - cost;
  } else {
    return 0 - vp9_cost_zero(upd_p);
  }
}

static int update_mv(vp9_writer *bc, const unsigned int ct[2],
                     vp9_prob *cur_p, vp9_prob new_p, vp9_prob upd_p) {

#ifdef LOW_PRECISION_MV_UPDATE
  vp9_prob mod_p = new_p | 1;
#else
  vp9_prob mod_p = new_p;
#endif

  const int cur_b = cost_branch256(ct, *cur_p);
  const int mod_b = cost_branch256(ct, mod_p);
  const int cost = 7 * 256 +
#ifndef LOW_PRECISION_MV_UPDATE
      256 +
#endif
      (vp9_cost_one(upd_p) - vp9_cost_zero(upd_p));

  if (cur_b - mod_b > cost) {
    *cur_p = mod_p;
    vp9_write(bc, 1, upd_p);
#ifdef LOW_PRECISION_MV_UPDATE
    vp9_write_literal(bc, mod_p >> 1, 7);
#else
    vp9_write_literal(bc, mod_p, 8);
#endif
    return 1;
  } else {
    vp9_write(bc, 0, upd_p);
    return 0;
  }
}

static void counts_to_nmv_context(
    nmv_context_counts *nmv_count,
    nmv_context *prob,
    int usehp,
    unsigned int (*branch_ct_joint)[2],
    unsigned int (*branch_ct_sign)[2],
    unsigned int (*branch_ct_classes)[MV_CLASSES - 1][2],
    unsigned int (*branch_ct_class0)[CLASS0_SIZE - 1][2],
    unsigned int (*branch_ct_bits)[MV_OFFSET_BITS][2],
    unsigned int (*branch_ct_class0_fp)[CLASS0_SIZE][4 - 1][2],
    unsigned int (*branch_ct_fp)[4 - 1][2],
    unsigned int (*branch_ct_class0_hp)[2],
    unsigned int (*branch_ct_hp)[2]) {
  int i, j, k;
  vp9_counts_process(nmv_count, usehp);
  vp9_tree_probs_from_distribution(vp9_mv_joint_tree,
                                   prob->joints,
                                   branch_ct_joint,
                                   nmv_count->joints, 0);
  for (i = 0; i < 2; ++i) {
    const uint32_t s0 = nmv_count->comps[i].sign[0];
    const uint32_t s1 = nmv_count->comps[i].sign[1];

    prob->comps[i].sign = get_binary_prob(s0, s1);
    branch_ct_sign[i][0] = s0;
    branch_ct_sign[i][1] = s1;
    vp9_tree_probs_from_distribution(vp9_mv_class_tree,
                                     prob->comps[i].classes,
                                     branch_ct_classes[i],
                                     nmv_count->comps[i].classes, 0);
    vp9_tree_probs_from_distribution(vp9_mv_class0_tree,
                                     prob->comps[i].class0,
                                     branch_ct_class0[i],
                                     nmv_count->comps[i].class0, 0);
    for (j = 0; j < MV_OFFSET_BITS; ++j) {
      const uint32_t b0 = nmv_count->comps[i].bits[j][0];
      const uint32_t b1 = nmv_count->comps[i].bits[j][1];

      prob->comps[i].bits[j] = get_binary_prob(b0, b1);
      branch_ct_bits[i][j][0] = b0;
      branch_ct_bits[i][j][1] = b1;
    }
  }
  for (i = 0; i < 2; ++i) {
    for (k = 0; k < CLASS0_SIZE; ++k) {
      vp9_tree_probs_from_distribution(vp9_mv_fp_tree,
                                       prob->comps[i].class0_fp[k],
                                       branch_ct_class0_fp[i][k],
                                       nmv_count->comps[i].class0_fp[k], 0);
    }
    vp9_tree_probs_from_distribution(vp9_mv_fp_tree,
                                     prob->comps[i].fp,
                                     branch_ct_fp[i],
                                     nmv_count->comps[i].fp, 0);
  }
  if (usehp) {
    for (i = 0; i < 2; ++i) {
      const uint32_t c0_hp0 = nmv_count->comps[i].class0_hp[0];
      const uint32_t c0_hp1 = nmv_count->comps[i].class0_hp[1];
      const uint32_t hp0 = nmv_count->comps[i].hp[0];
      const uint32_t hp1 = nmv_count->comps[i].hp[1];

      prob->comps[i].class0_hp = get_binary_prob(c0_hp0, c0_hp1);
      branch_ct_class0_hp[i][0] = c0_hp0;
      branch_ct_class0_hp[i][1] = c0_hp1;

      prob->comps[i].hp = get_binary_prob(hp0, hp1);
      branch_ct_hp[i][0] = hp0;
      branch_ct_hp[i][1] = hp1;
    }
  }
}

#ifdef NMV_STATS
void init_nmvstats() {
  vp9_zero(tnmvcounts);
}

void print_nmvstats() {
  nmv_context prob;
  unsigned int branch_ct_joint[MV_JOINTS - 1][2];
  unsigned int branch_ct_sign[2][2];
  unsigned int branch_ct_classes[2][MV_CLASSES - 1][2];
  unsigned int branch_ct_class0[2][CLASS0_SIZE - 1][2];
  unsigned int branch_ct_bits[2][MV_OFFSET_BITS][2];
  unsigned int branch_ct_class0_fp[2][CLASS0_SIZE][4 - 1][2];
  unsigned int branch_ct_fp[2][4 - 1][2];
  unsigned int branch_ct_class0_hp[2][2];
  unsigned int branch_ct_hp[2][2];
  int i, j, k;
  counts_to_nmv_context(&tnmvcounts, &prob, 1,
                        branch_ct_joint, branch_ct_sign, branch_ct_classes,
                        branch_ct_class0, branch_ct_bits,
                        branch_ct_class0_fp, branch_ct_fp,
                        branch_ct_class0_hp, branch_ct_hp);

  printf("\nCounts =\n  { ");
  for (j = 0; j < MV_JOINTS; ++j)
    printf("%d, ", tnmvcounts.joints[j]);
  printf("},\n");
  for (i = 0; i < 2; ++i) {
    printf("  {\n");
    printf("    %d/%d,\n", tnmvcounts.comps[i].sign[0],
                           tnmvcounts.comps[i].sign[1]);
    printf("    { ");
    for (j = 0; j < MV_CLASSES; ++j)
      printf("%d, ", tnmvcounts.comps[i].classes[j]);
    printf("},\n");
    printf("    { ");
    for (j = 0; j < CLASS0_SIZE; ++j)
      printf("%d, ", tnmvcounts.comps[i].class0[j]);
    printf("},\n");
    printf("    { ");
    for (j = 0; j < MV_OFFSET_BITS; ++j)
      printf("%d/%d, ", tnmvcounts.comps[i].bits[j][0],
                        tnmvcounts.comps[i].bits[j][1]);
    printf("},\n");

    printf("    {");
    for (j = 0; j < CLASS0_SIZE; ++j) {
      printf("{");
      for (k = 0; k < 4; ++k)
        printf("%d, ", tnmvcounts.comps[i].class0_fp[j][k]);
      printf("}, ");
    }
    printf("},\n");

    printf("    { ");
    for (j = 0; j < 4; ++j)
      printf("%d, ", tnmvcounts.comps[i].fp[j]);
    printf("},\n");

    printf("    %d/%d,\n",
           tnmvcounts.comps[i].class0_hp[0],
           tnmvcounts.comps[i].class0_hp[1]);
    printf("    %d/%d,\n",
           tnmvcounts.comps[i].hp[0],
           tnmvcounts.comps[i].hp[1]);
    printf("  },\n");
  }

  printf("\nProbs =\n  { ");
  for (j = 0; j < MV_JOINTS - 1; ++j)
    printf("%d, ", prob.joints[j]);
  printf("},\n");
  for (i=0; i< 2; ++i) {
    printf("  {\n");
    printf("    %d,\n", prob.comps[i].sign);
    printf("    { ");
    for (j = 0; j < MV_CLASSES - 1; ++j)
      printf("%d, ", prob.comps[i].classes[j]);
    printf("},\n");
    printf("    { ");
    for (j = 0; j < CLASS0_SIZE - 1; ++j)
      printf("%d, ", prob.comps[i].class0[j]);
    printf("},\n");
    printf("    { ");
    for (j = 0; j < MV_OFFSET_BITS; ++j)
      printf("%d, ", prob.comps[i].bits[j]);
    printf("},\n");
    printf("    { ");
    for (j = 0; j < CLASS0_SIZE; ++j) {
      printf("{");
      for (k = 0; k < 3; ++k)
        printf("%d, ", prob.comps[i].class0_fp[j][k]);
      printf("}, ");
    }
    printf("},\n");
    printf("    { ");
    for (j = 0; j < 3; ++j)
      printf("%d, ", prob.comps[i].fp[j]);
    printf("},\n");

    printf("    %d,\n", prob.comps[i].class0_hp);
    printf("    %d,\n", prob.comps[i].hp);
    printf("  },\n");
  }
}

static void add_nmvcount(nmv_context_counts* const dst,
                         const nmv_context_counts* const src) {
  int i, j, k;
  for (j = 0; j < MV_JOINTS; ++j) {
    dst->joints[j] += src->joints[j];
  }
  for (i = 0; i < 2; ++i) {
    for (j = 0; j < MV_VALS; ++j) {
      dst->comps[i].mvcount[j] += src->comps[i].mvcount[j];
    }
    dst->comps[i].sign[0] += src->comps[i].sign[0];
    dst->comps[i].sign[1] += src->comps[i].sign[1];
    for (j = 0; j < MV_CLASSES; ++j) {
      dst->comps[i].classes[j] += src->comps[i].classes[j];
    }
    for (j = 0; j < CLASS0_SIZE; ++j) {
      dst->comps[i].class0[j] += src->comps[i].class0[j];
    }
    for (j = 0; j < MV_OFFSET_BITS; ++j) {
      dst->comps[i].bits[j][0] += src->comps[i].bits[j][0];
      dst->comps[i].bits[j][1] += src->comps[i].bits[j][1];
    }
  }
  for (i = 0; i < 2; ++i) {
    for (j = 0; j < CLASS0_SIZE; ++j) {
      for (k = 0; k < 4; ++k) {
        dst->comps[i].class0_fp[j][k] += src->comps[i].class0_fp[j][k];
      }
    }
    for (j = 0; j < 4; ++j) {
      dst->comps[i].fp[j] += src->comps[i].fp[j];
    }
    dst->comps[i].class0_hp[0] += src->comps[i].class0_hp[0];
    dst->comps[i].class0_hp[1] += src->comps[i].class0_hp[1];
    dst->comps[i].hp[0] += src->comps[i].hp[0];
    dst->comps[i].hp[1] += src->comps[i].hp[1];
  }
}
#endif

void vp9_write_nmv_probs(VP9_COMP* const cpi, int usehp, vp9_writer* const bc) {
  int i, j;
  nmv_context prob;
  unsigned int branch_ct_joint[MV_JOINTS - 1][2];
  unsigned int branch_ct_sign[2][2];
  unsigned int branch_ct_classes[2][MV_CLASSES - 1][2];
  unsigned int branch_ct_class0[2][CLASS0_SIZE - 1][2];
  unsigned int branch_ct_bits[2][MV_OFFSET_BITS][2];
  unsigned int branch_ct_class0_fp[2][CLASS0_SIZE][4 - 1][2];
  unsigned int branch_ct_fp[2][4 - 1][2];
  unsigned int branch_ct_class0_hp[2][2];
  unsigned int branch_ct_hp[2][2];
  nmv_context *mvc = &cpi->common.fc.nmvc;

#ifdef MV_GROUP_UPDATE
  int savings = 0;
#endif

#ifdef NMV_STATS
  if (!cpi->dummy_packing)
    add_nmvcount(&tnmvcounts, &cpi->NMVcount);
#endif
  counts_to_nmv_context(&cpi->NMVcount, &prob, usehp,
                        branch_ct_joint, branch_ct_sign, branch_ct_classes,
                        branch_ct_class0, branch_ct_bits,
                        branch_ct_class0_fp, branch_ct_fp,
                        branch_ct_class0_hp, branch_ct_hp);
  /* write updates if they help */
#ifdef MV_GROUP_UPDATE
  for (j = 0; j < MV_JOINTS - 1; ++j) {
    savings += update_nmv_savings(branch_ct_joint[j],
                                  cpi->common.fc.nmvc.joints[j],
                                  prob.joints[j],
                                  VP9_NMV_UPDATE_PROB);
  }
  for (i = 0; i < 2; ++i) {
    savings += update_nmv_savings(branch_ct_sign[i],
                                  cpi->common.fc.nmvc.comps[i].sign,
                                  prob.comps[i].sign,
                                  VP9_NMV_UPDATE_PROB);
    for (j = 0; j < MV_CLASSES - 1; ++j) {
      savings += update_nmv_savings(branch_ct_classes[i][j],
                                    cpi->common.fc.nmvc.comps[i].classes[j],
                                    prob.comps[i].classes[j],
                                    VP9_NMV_UPDATE_PROB);
    }
    for (j = 0; j < CLASS0_SIZE - 1; ++j) {
      savings += update_nmv_savings(branch_ct_class0[i][j],
                                    cpi->common.fc.nmvc.comps[i].class0[j],
                                    prob.comps[i].class0[j],
                                    VP9_NMV_UPDATE_PROB);
    }
    for (j = 0; j < MV_OFFSET_BITS; ++j) {
      savings += update_nmv_savings(branch_ct_bits[i][j],
                                    cpi->common.fc.nmvc.comps[i].bits[j],
                                    prob.comps[i].bits[j],
                                    VP9_NMV_UPDATE_PROB);
    }
  }
  for (i = 0; i < 2; ++i) {
    for (j = 0; j < CLASS0_SIZE; ++j) {
      int k;
      for (k = 0; k < 3; ++k) {
        savings += update_nmv_savings(branch_ct_class0_fp[i][j][k],
                                      cpi->common.fc.nmvc.comps[i].class0_fp[j][k],
                                      prob.comps[i].class0_fp[j][k],
                                      VP9_NMV_UPDATE_PROB);
      }
    }
    for (j = 0; j < 3; ++j) {
      savings += update_nmv_savings(branch_ct_fp[i][j],
                                    cpi->common.fc.nmvc.comps[i].fp[j],
                                    prob.comps[i].fp[j],
                                    VP9_NMV_UPDATE_PROB);
    }
  }
  if (usehp) {
    for (i = 0; i < 2; ++i) {
      savings += update_nmv_savings(branch_ct_class0_hp[i],
                                    cpi->common.fc.nmvc.comps[i].class0_hp,
                                    prob.comps[i].class0_hp,
                                    VP9_NMV_UPDATE_PROB);
      savings += update_nmv_savings(branch_ct_hp[i],
                                    cpi->common.fc.nmvc.comps[i].hp,
                                    prob.comps[i].hp,
                                    VP9_NMV_UPDATE_PROB);
    }
  }
  if (savings <= 0) {
    vp9_write_bit(bc, 0);
    return;
  }
  vp9_write_bit(bc, 1);
#endif

  for (j = 0; j < MV_JOINTS - 1; ++j)
    update_mv(bc, branch_ct_joint[j], &mvc->joints[j], prob.joints[j],
              VP9_NMV_UPDATE_PROB);

  for (i = 0; i < 2; ++i) {
    update_mv(bc, branch_ct_sign[i], &mvc->comps[i].sign,
              prob.comps[i].sign, VP9_NMV_UPDATE_PROB);
    for (j = 0; j < MV_CLASSES - 1; ++j)
      update_mv(bc, branch_ct_classes[i][j], &mvc->comps[i].classes[j],
                prob.comps[i].classes[j], VP9_NMV_UPDATE_PROB);

    for (j = 0; j < CLASS0_SIZE - 1; ++j)
      update_mv(bc, branch_ct_class0[i][j], &mvc->comps[i].class0[j],
                prob.comps[i].class0[j], VP9_NMV_UPDATE_PROB);

    for (j = 0; j < MV_OFFSET_BITS; ++j)
      update_mv(bc, branch_ct_bits[i][j], &mvc->comps[i].bits[j],
                prob.comps[i].bits[j], VP9_NMV_UPDATE_PROB);
  }

  for (i = 0; i < 2; ++i) {
    for (j = 0; j < CLASS0_SIZE; ++j) {
      int k;
      for (k = 0; k < 3; ++k)
        update_mv(bc, branch_ct_class0_fp[i][j][k],
                  &mvc->comps[i].class0_fp[j][k],
                  prob.comps[i].class0_fp[j][k], VP9_NMV_UPDATE_PROB);
    }

    for (j = 0; j < 3; ++j)
      update_mv(bc, branch_ct_fp[i][j], &mvc->comps[i].fp[j],
                prob.comps[i].fp[j], VP9_NMV_UPDATE_PROB);
  }

  if (usehp) {
    for (i = 0; i < 2; ++i) {
      update_mv(bc, branch_ct_class0_hp[i], &mvc->comps[i].class0_hp,
                prob.comps[i].class0_hp, VP9_NMV_UPDATE_PROB);
      update_mv(bc, branch_ct_hp[i], &mvc->comps[i].hp,
                prob.comps[i].hp, VP9_NMV_UPDATE_PROB);
    }
  }
}

void vp9_encode_mv(VP9_COMP* cpi, vp9_writer* w,
                   const MV* mv, const MV* ref,
                   const nmv_context* mvctx, int usehp) {
  const MV diff = {mv->row - ref->row,
                   mv->col - ref->col};
  const MV_JOINT_TYPE j = vp9_get_mv_joint(&diff);
  usehp = usehp && vp9_use_mv_hp(ref);

  write_token(w, vp9_mv_joint_tree, mvctx->joints, &vp9_mv_joint_encodings[j]);
  if (mv_joint_vertical(j))
    encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);

  if (mv_joint_horizontal(j))
    encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);

  // If auto_mv_step_size is enabled and it is an arf/non shown frame
  // then keep track of the largest motion vector component used.
  if (cpi->sf.auto_mv_step_size && !cpi->common.show_frame) {
    cpi->max_mv_magnitude = MAX((MAX(abs(mv->row), abs(mv->col)) >> 3),
                                cpi->max_mv_magnitude);
  }
}

void vp9_build_nmv_cost_table(int *mvjoint,
                              int *mvcost[2],
                              const nmv_context* const mvctx,
                              int usehp,
                              int mvc_flag_v,
                              int mvc_flag_h) {
  vp9_clear_system_state();
  vp9_cost_tokens(mvjoint, mvctx->joints, vp9_mv_joint_tree);
  if (mvc_flag_v)
    build_nmv_component_cost_table(mvcost[0], &mvctx->comps[0], usehp);
  if (mvc_flag_h)
    build_nmv_component_cost_table(mvcost[1], &mvctx->comps[1], usehp);
}

void vp9_update_nmv_count(VP9_COMP *cpi, MACROBLOCK *x,
                         int_mv *best_ref_mv, int_mv *second_best_ref_mv) {
  MB_MODE_INFO *mbmi = &x->e_mbd.mode_info_context->mbmi;
  MV diff;
  const int bw = 1 << b_width_log2(mbmi->sb_type);
  const int bh = 1 << b_height_log2(mbmi->sb_type);
  int idx, idy;

  if (mbmi->sb_type < BLOCK_SIZE_SB8X8) {
    PARTITION_INFO *pi = x->partition_info;
    for (idy = 0; idy < 2; idy += bh) {
      for (idx = 0; idx < 2; idx += bw) {
        const int i = idy * 2 + idx;
        if (pi->bmi[i].mode == NEWMV) {
          diff.row = pi->bmi[i].mv.as_mv.row - best_ref_mv->as_mv.row;
          diff.col = pi->bmi[i].mv.as_mv.col - best_ref_mv->as_mv.col;
          vp9_inc_mv(&diff, &cpi->NMVcount);

          if (x->e_mbd.mode_info_context->mbmi.ref_frame[1] > INTRA_FRAME) {
            diff.row = pi->bmi[i].second_mv.as_mv.row -
                         second_best_ref_mv->as_mv.row;
            diff.col = pi->bmi[i].second_mv.as_mv.col -
                         second_best_ref_mv->as_mv.col;
            vp9_inc_mv(&diff, &cpi->NMVcount);
          }
        }
      }
    }
  } else if (mbmi->mode == NEWMV) {
    diff.row = mbmi->mv[0].as_mv.row - best_ref_mv->as_mv.row;
    diff.col = mbmi->mv[0].as_mv.col - best_ref_mv->as_mv.col;
    vp9_inc_mv(&diff, &cpi->NMVcount);

    if (mbmi->ref_frame[1] > INTRA_FRAME) {
      diff.row = mbmi->mv[1].as_mv.row - second_best_ref_mv->as_mv.row;
      diff.col = mbmi->mv[1].as_mv.col - second_best_ref_mv->as_mv.col;
      vp9_inc_mv(&diff, &cpi->NMVcount);
    }
  }
}