1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/vp9_encodemv.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,332 @@
1.4 +/*
1.5 + * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include <math.h>
1.15 +
1.16 +#include "vp9/common/vp9_common.h"
1.17 +#include "vp9/common/vp9_entropymode.h"
1.18 +#include "vp9/common/vp9_systemdependent.h"
1.19 +#include "vp9/encoder/vp9_encodemv.h"
1.20 +
1.21 +
1.22 +#ifdef ENTROPY_STATS
1.23 +extern unsigned int active_section;
1.24 +#endif
1.25 +
1.26 +static void encode_mv_component(vp9_writer* w, int comp,
1.27 + const nmv_component* mvcomp, int usehp) {
1.28 + int offset;
1.29 + const int sign = comp < 0;
1.30 + const int mag = sign ? -comp : comp;
1.31 + const int mv_class = vp9_get_mv_class(mag - 1, &offset);
1.32 + const int d = offset >> 3; // int mv data
1.33 + const int fr = (offset >> 1) & 3; // fractional mv data
1.34 + const int hp = offset & 1; // high precision mv data
1.35 +
1.36 + assert(comp != 0);
1.37 +
1.38 + // Sign
1.39 + vp9_write(w, sign, mvcomp->sign);
1.40 +
1.41 + // Class
1.42 + write_token(w, vp9_mv_class_tree, mvcomp->classes,
1.43 + &vp9_mv_class_encodings[mv_class]);
1.44 +
1.45 + // Integer bits
1.46 + if (mv_class == MV_CLASS_0) {
1.47 + write_token(w, vp9_mv_class0_tree, mvcomp->class0,
1.48 + &vp9_mv_class0_encodings[d]);
1.49 + } else {
1.50 + int i;
1.51 + const int n = mv_class + CLASS0_BITS - 1; // number of bits
1.52 + for (i = 0; i < n; ++i)
1.53 + vp9_write(w, (d >> i) & 1, mvcomp->bits[i]);
1.54 + }
1.55 +
1.56 + // Fractional bits
1.57 + write_token(w, vp9_mv_fp_tree,
1.58 + mv_class == MV_CLASS_0 ? mvcomp->class0_fp[d] : mvcomp->fp,
1.59 + &vp9_mv_fp_encodings[fr]);
1.60 +
1.61 + // High precision bit
1.62 + if (usehp)
1.63 + vp9_write(w, hp,
1.64 + mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
1.65 +}
1.66 +
1.67 +
1.68 +static void build_nmv_component_cost_table(int *mvcost,
1.69 + const nmv_component* const mvcomp,
1.70 + int usehp) {
1.71 + int i, v;
1.72 + int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
1.73 + int bits_cost[MV_OFFSET_BITS][2];
1.74 + int class0_fp_cost[CLASS0_SIZE][4], fp_cost[4];
1.75 + int class0_hp_cost[2], hp_cost[2];
1.76 +
1.77 + sign_cost[0] = vp9_cost_zero(mvcomp->sign);
1.78 + sign_cost[1] = vp9_cost_one(mvcomp->sign);
1.79 + vp9_cost_tokens(class_cost, mvcomp->classes, vp9_mv_class_tree);
1.80 + vp9_cost_tokens(class0_cost, mvcomp->class0, vp9_mv_class0_tree);
1.81 + for (i = 0; i < MV_OFFSET_BITS; ++i) {
1.82 + bits_cost[i][0] = vp9_cost_zero(mvcomp->bits[i]);
1.83 + bits_cost[i][1] = vp9_cost_one(mvcomp->bits[i]);
1.84 + }
1.85 +
1.86 + for (i = 0; i < CLASS0_SIZE; ++i)
1.87 + vp9_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], vp9_mv_fp_tree);
1.88 + vp9_cost_tokens(fp_cost, mvcomp->fp, vp9_mv_fp_tree);
1.89 +
1.90 + if (usehp) {
1.91 + class0_hp_cost[0] = vp9_cost_zero(mvcomp->class0_hp);
1.92 + class0_hp_cost[1] = vp9_cost_one(mvcomp->class0_hp);
1.93 + hp_cost[0] = vp9_cost_zero(mvcomp->hp);
1.94 + hp_cost[1] = vp9_cost_one(mvcomp->hp);
1.95 + }
1.96 + mvcost[0] = 0;
1.97 + for (v = 1; v <= MV_MAX; ++v) {
1.98 + int z, c, o, d, e, f, cost = 0;
1.99 + z = v - 1;
1.100 + c = vp9_get_mv_class(z, &o);
1.101 + cost += class_cost[c];
1.102 + d = (o >> 3); /* int mv data */
1.103 + f = (o >> 1) & 3; /* fractional pel mv data */
1.104 + e = (o & 1); /* high precision mv data */
1.105 + if (c == MV_CLASS_0) {
1.106 + cost += class0_cost[d];
1.107 + } else {
1.108 + int i, b;
1.109 + b = c + CLASS0_BITS - 1; /* number of bits */
1.110 + for (i = 0; i < b; ++i)
1.111 + cost += bits_cost[i][((d >> i) & 1)];
1.112 + }
1.113 + if (c == MV_CLASS_0) {
1.114 + cost += class0_fp_cost[d][f];
1.115 + } else {
1.116 + cost += fp_cost[f];
1.117 + }
1.118 + if (usehp) {
1.119 + if (c == MV_CLASS_0) {
1.120 + cost += class0_hp_cost[e];
1.121 + } else {
1.122 + cost += hp_cost[e];
1.123 + }
1.124 + }
1.125 + mvcost[v] = cost + sign_cost[0];
1.126 + mvcost[-v] = cost + sign_cost[1];
1.127 + }
1.128 +}
1.129 +
1.130 +static int update_mv(vp9_writer *w, const unsigned int ct[2], vp9_prob *cur_p,
1.131 + vp9_prob upd_p) {
1.132 + const vp9_prob new_p = get_binary_prob(ct[0], ct[1]);
1.133 + vp9_prob mod_p = new_p | 1;
1.134 + const int cur_b = cost_branch256(ct, *cur_p);
1.135 + const int mod_b = cost_branch256(ct, mod_p);
1.136 + const int cost = 7 * 256 + (vp9_cost_one(upd_p) - vp9_cost_zero(upd_p));
1.137 + if (cur_b - mod_b > cost) {
1.138 + *cur_p = mod_p;
1.139 + vp9_write(w, 1, upd_p);
1.140 + vp9_write_literal(w, mod_p >> 1, 7);
1.141 + return 1;
1.142 + } else {
1.143 + vp9_write(w, 0, upd_p);
1.144 + return 0;
1.145 + }
1.146 +}
1.147 +
1.148 +static void counts_to_nmv_context(
1.149 + nmv_context_counts *nmv_count,
1.150 + int usehp,
1.151 + unsigned int (*branch_ct_joint)[2],
1.152 + unsigned int (*branch_ct_sign)[2],
1.153 + unsigned int (*branch_ct_classes)[MV_CLASSES - 1][2],
1.154 + unsigned int (*branch_ct_class0)[CLASS0_SIZE - 1][2],
1.155 + unsigned int (*branch_ct_bits)[MV_OFFSET_BITS][2],
1.156 + unsigned int (*branch_ct_class0_fp)[CLASS0_SIZE][4 - 1][2],
1.157 + unsigned int (*branch_ct_fp)[4 - 1][2],
1.158 + unsigned int (*branch_ct_class0_hp)[2],
1.159 + unsigned int (*branch_ct_hp)[2]) {
1.160 + int i, j, k;
1.161 + vp9_tree_probs_from_distribution(vp9_mv_joint_tree, branch_ct_joint,
1.162 + nmv_count->joints);
1.163 + for (i = 0; i < 2; ++i) {
1.164 + const uint32_t s0 = nmv_count->comps[i].sign[0];
1.165 + const uint32_t s1 = nmv_count->comps[i].sign[1];
1.166 +
1.167 + branch_ct_sign[i][0] = s0;
1.168 + branch_ct_sign[i][1] = s1;
1.169 + vp9_tree_probs_from_distribution(vp9_mv_class_tree,
1.170 + branch_ct_classes[i],
1.171 + nmv_count->comps[i].classes);
1.172 + vp9_tree_probs_from_distribution(vp9_mv_class0_tree,
1.173 + branch_ct_class0[i],
1.174 + nmv_count->comps[i].class0);
1.175 + for (j = 0; j < MV_OFFSET_BITS; ++j) {
1.176 + const uint32_t b0 = nmv_count->comps[i].bits[j][0];
1.177 + const uint32_t b1 = nmv_count->comps[i].bits[j][1];
1.178 +
1.179 + branch_ct_bits[i][j][0] = b0;
1.180 + branch_ct_bits[i][j][1] = b1;
1.181 + }
1.182 + }
1.183 + for (i = 0; i < 2; ++i) {
1.184 + for (k = 0; k < CLASS0_SIZE; ++k) {
1.185 + vp9_tree_probs_from_distribution(vp9_mv_fp_tree,
1.186 + branch_ct_class0_fp[i][k],
1.187 + nmv_count->comps[i].class0_fp[k]);
1.188 + }
1.189 + vp9_tree_probs_from_distribution(vp9_mv_fp_tree,
1.190 + branch_ct_fp[i],
1.191 + nmv_count->comps[i].fp);
1.192 + }
1.193 + if (usehp) {
1.194 + for (i = 0; i < 2; ++i) {
1.195 + const uint32_t c0_hp0 = nmv_count->comps[i].class0_hp[0];
1.196 + const uint32_t c0_hp1 = nmv_count->comps[i].class0_hp[1];
1.197 + const uint32_t hp0 = nmv_count->comps[i].hp[0];
1.198 + const uint32_t hp1 = nmv_count->comps[i].hp[1];
1.199 +
1.200 + branch_ct_class0_hp[i][0] = c0_hp0;
1.201 + branch_ct_class0_hp[i][1] = c0_hp1;
1.202 +
1.203 + branch_ct_hp[i][0] = hp0;
1.204 + branch_ct_hp[i][1] = hp1;
1.205 + }
1.206 + }
1.207 +}
1.208 +
1.209 +void vp9_write_nmv_probs(VP9_COMP* const cpi, int usehp, vp9_writer* const bc) {
1.210 + int i, j;
1.211 + unsigned int branch_ct_joint[MV_JOINTS - 1][2];
1.212 + unsigned int branch_ct_sign[2][2];
1.213 + unsigned int branch_ct_classes[2][MV_CLASSES - 1][2];
1.214 + unsigned int branch_ct_class0[2][CLASS0_SIZE - 1][2];
1.215 + unsigned int branch_ct_bits[2][MV_OFFSET_BITS][2];
1.216 + unsigned int branch_ct_class0_fp[2][CLASS0_SIZE][4 - 1][2];
1.217 + unsigned int branch_ct_fp[2][4 - 1][2];
1.218 + unsigned int branch_ct_class0_hp[2][2];
1.219 + unsigned int branch_ct_hp[2][2];
1.220 + nmv_context *mvc = &cpi->common.fc.nmvc;
1.221 +
1.222 + counts_to_nmv_context(&cpi->NMVcount, usehp,
1.223 + branch_ct_joint, branch_ct_sign, branch_ct_classes,
1.224 + branch_ct_class0, branch_ct_bits,
1.225 + branch_ct_class0_fp, branch_ct_fp,
1.226 + branch_ct_class0_hp, branch_ct_hp);
1.227 +
1.228 + for (j = 0; j < MV_JOINTS - 1; ++j)
1.229 + update_mv(bc, branch_ct_joint[j], &mvc->joints[j], NMV_UPDATE_PROB);
1.230 +
1.231 + for (i = 0; i < 2; ++i) {
1.232 + update_mv(bc, branch_ct_sign[i], &mvc->comps[i].sign, NMV_UPDATE_PROB);
1.233 + for (j = 0; j < MV_CLASSES - 1; ++j)
1.234 + update_mv(bc, branch_ct_classes[i][j], &mvc->comps[i].classes[j],
1.235 + NMV_UPDATE_PROB);
1.236 +
1.237 + for (j = 0; j < CLASS0_SIZE - 1; ++j)
1.238 + update_mv(bc, branch_ct_class0[i][j], &mvc->comps[i].class0[j],
1.239 + NMV_UPDATE_PROB);
1.240 +
1.241 + for (j = 0; j < MV_OFFSET_BITS; ++j)
1.242 + update_mv(bc, branch_ct_bits[i][j], &mvc->comps[i].bits[j],
1.243 + NMV_UPDATE_PROB);
1.244 + }
1.245 +
1.246 + for (i = 0; i < 2; ++i) {
1.247 + for (j = 0; j < CLASS0_SIZE; ++j) {
1.248 + int k;
1.249 + for (k = 0; k < 3; ++k)
1.250 + update_mv(bc, branch_ct_class0_fp[i][j][k],
1.251 + &mvc->comps[i].class0_fp[j][k], NMV_UPDATE_PROB);
1.252 + }
1.253 +
1.254 + for (j = 0; j < 3; ++j)
1.255 + update_mv(bc, branch_ct_fp[i][j], &mvc->comps[i].fp[j], NMV_UPDATE_PROB);
1.256 + }
1.257 +
1.258 + if (usehp) {
1.259 + for (i = 0; i < 2; ++i) {
1.260 + update_mv(bc, branch_ct_class0_hp[i], &mvc->comps[i].class0_hp,
1.261 + NMV_UPDATE_PROB);
1.262 + update_mv(bc, branch_ct_hp[i], &mvc->comps[i].hp,
1.263 + NMV_UPDATE_PROB);
1.264 + }
1.265 + }
1.266 +}
1.267 +
1.268 +void vp9_encode_mv(VP9_COMP* cpi, vp9_writer* w,
1.269 + const MV* mv, const MV* ref,
1.270 + const nmv_context* mvctx, int usehp) {
1.271 + const MV diff = {mv->row - ref->row,
1.272 + mv->col - ref->col};
1.273 + const MV_JOINT_TYPE j = vp9_get_mv_joint(&diff);
1.274 + usehp = usehp && vp9_use_mv_hp(ref);
1.275 +
1.276 + write_token(w, vp9_mv_joint_tree, mvctx->joints, &vp9_mv_joint_encodings[j]);
1.277 + if (mv_joint_vertical(j))
1.278 + encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);
1.279 +
1.280 + if (mv_joint_horizontal(j))
1.281 + encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);
1.282 +
1.283 + // If auto_mv_step_size is enabled then keep track of the largest
1.284 + // motion vector component used.
1.285 + if (!cpi->dummy_packing && cpi->sf.auto_mv_step_size) {
1.286 + unsigned int maxv = MAX(abs(mv->row), abs(mv->col)) >> 3;
1.287 + cpi->max_mv_magnitude = MAX(maxv, cpi->max_mv_magnitude);
1.288 + }
1.289 +}
1.290 +
1.291 +void vp9_build_nmv_cost_table(int *mvjoint,
1.292 + int *mvcost[2],
1.293 + const nmv_context* const mvctx,
1.294 + int usehp,
1.295 + int mvc_flag_v,
1.296 + int mvc_flag_h) {
1.297 + vp9_clear_system_state();
1.298 + vp9_cost_tokens(mvjoint, mvctx->joints, vp9_mv_joint_tree);
1.299 + if (mvc_flag_v)
1.300 + build_nmv_component_cost_table(mvcost[0], &mvctx->comps[0], usehp);
1.301 + if (mvc_flag_h)
1.302 + build_nmv_component_cost_table(mvcost[1], &mvctx->comps[1], usehp);
1.303 +}
1.304 +
1.305 +static void inc_mvs(int_mv mv[2], int_mv ref[2], int is_compound,
1.306 + nmv_context_counts *counts) {
1.307 + int i;
1.308 + for (i = 0; i < 1 + is_compound; ++i) {
1.309 + const MV diff = { mv[i].as_mv.row - ref[i].as_mv.row,
1.310 + mv[i].as_mv.col - ref[i].as_mv.col };
1.311 + vp9_inc_mv(&diff, counts);
1.312 + }
1.313 +}
1.314 +
1.315 +void vp9_update_mv_count(VP9_COMP *cpi, MACROBLOCK *x, int_mv best_ref_mv[2]) {
1.316 + MODE_INFO *mi = x->e_mbd.mi_8x8[0];
1.317 + MB_MODE_INFO *const mbmi = &mi->mbmi;
1.318 + const int is_compound = has_second_ref(mbmi);
1.319 +
1.320 + if (mbmi->sb_type < BLOCK_8X8) {
1.321 + const int num_4x4_w = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1.322 + const int num_4x4_h = num_4x4_blocks_high_lookup[mbmi->sb_type];
1.323 + int idx, idy;
1.324 +
1.325 + for (idy = 0; idy < 2; idy += num_4x4_h) {
1.326 + for (idx = 0; idx < 2; idx += num_4x4_w) {
1.327 + const int i = idy * 2 + idx;
1.328 + if (mi->bmi[i].as_mode == NEWMV)
1.329 + inc_mvs(mi->bmi[i].as_mv, best_ref_mv, is_compound, &cpi->NMVcount);
1.330 + }
1.331 + }
1.332 + } else if (mbmi->mode == NEWMV) {
1.333 + inc_mvs(mbmi->mv, best_ref_mv, is_compound, &cpi->NMVcount);
1.334 + }
1.335 +}
