1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/vp9_rdopt.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,4582 @@
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 <stdio.h>
1.15 +#include <math.h>
1.16 +#include <limits.h>
1.17 +#include <assert.h>
1.18 +
1.19 +#include "vp9/common/vp9_pragmas.h"
1.20 +#include "vp9/encoder/vp9_tokenize.h"
1.21 +#include "vp9/encoder/vp9_treewriter.h"
1.22 +#include "vp9/encoder/vp9_onyx_int.h"
1.23 +#include "vp9/encoder/vp9_modecosts.h"
1.24 +#include "vp9/encoder/vp9_encodeintra.h"
1.25 +#include "vp9/common/vp9_entropymode.h"
1.26 +#include "vp9/common/vp9_reconinter.h"
1.27 +#include "vp9/common/vp9_reconintra.h"
1.28 +#include "vp9/common/vp9_findnearmv.h"
1.29 +#include "vp9/common/vp9_quant_common.h"
1.30 +#include "vp9/encoder/vp9_encodemb.h"
1.31 +#include "vp9/encoder/vp9_quantize.h"
1.32 +#include "vp9/encoder/vp9_variance.h"
1.33 +#include "vp9/encoder/vp9_mcomp.h"
1.34 +#include "vp9/encoder/vp9_rdopt.h"
1.35 +#include "vp9/encoder/vp9_ratectrl.h"
1.36 +#include "vpx_mem/vpx_mem.h"
1.37 +#include "vp9/common/vp9_systemdependent.h"
1.38 +#include "vp9/encoder/vp9_encodemv.h"
1.39 +#include "vp9/common/vp9_seg_common.h"
1.40 +#include "vp9/common/vp9_pred_common.h"
1.41 +#include "vp9/common/vp9_entropy.h"
1.42 +#include "./vp9_rtcd.h"
1.43 +#include "vp9/common/vp9_mvref_common.h"
1.44 +#include "vp9/common/vp9_common.h"
1.45 +
1.46 +#define INVALID_MV 0x80008000
1.47 +
1.48 +/* Factor to weigh the rate for switchable interp filters */
1.49 +#define SWITCHABLE_INTERP_RATE_FACTOR 1
1.50 +
1.51 +#define LAST_FRAME_MODE_MASK 0xFFEDCD60
1.52 +#define GOLDEN_FRAME_MODE_MASK 0xFFDA3BB0
1.53 +#define ALT_REF_MODE_MASK 0xFFC648D0
1.54 +
1.55 +#define MIN_EARLY_TERM_INDEX 3
1.56 +
1.57 +const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
1.58 + {NEARESTMV, LAST_FRAME, NONE},
1.59 + {NEARESTMV, ALTREF_FRAME, NONE},
1.60 + {NEARESTMV, GOLDEN_FRAME, NONE},
1.61 +
1.62 + {DC_PRED, INTRA_FRAME, NONE},
1.63 +
1.64 + {NEWMV, LAST_FRAME, NONE},
1.65 + {NEWMV, ALTREF_FRAME, NONE},
1.66 + {NEWMV, GOLDEN_FRAME, NONE},
1.67 +
1.68 + {NEARMV, LAST_FRAME, NONE},
1.69 + {NEARMV, ALTREF_FRAME, NONE},
1.70 + {NEARESTMV, LAST_FRAME, ALTREF_FRAME},
1.71 + {NEARESTMV, GOLDEN_FRAME, ALTREF_FRAME},
1.72 +
1.73 + {TM_PRED, INTRA_FRAME, NONE},
1.74 +
1.75 + {NEARMV, LAST_FRAME, ALTREF_FRAME},
1.76 + {NEWMV, LAST_FRAME, ALTREF_FRAME},
1.77 + {NEARMV, GOLDEN_FRAME, NONE},
1.78 + {NEARMV, GOLDEN_FRAME, ALTREF_FRAME},
1.79 + {NEWMV, GOLDEN_FRAME, ALTREF_FRAME},
1.80 +
1.81 + {ZEROMV, LAST_FRAME, NONE},
1.82 + {ZEROMV, GOLDEN_FRAME, NONE},
1.83 + {ZEROMV, ALTREF_FRAME, NONE},
1.84 + {ZEROMV, LAST_FRAME, ALTREF_FRAME},
1.85 + {ZEROMV, GOLDEN_FRAME, ALTREF_FRAME},
1.86 +
1.87 + {H_PRED, INTRA_FRAME, NONE},
1.88 + {V_PRED, INTRA_FRAME, NONE},
1.89 + {D135_PRED, INTRA_FRAME, NONE},
1.90 + {D207_PRED, INTRA_FRAME, NONE},
1.91 + {D153_PRED, INTRA_FRAME, NONE},
1.92 + {D63_PRED, INTRA_FRAME, NONE},
1.93 + {D117_PRED, INTRA_FRAME, NONE},
1.94 + {D45_PRED, INTRA_FRAME, NONE},
1.95 +};
1.96 +
1.97 +const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
1.98 + {LAST_FRAME, NONE},
1.99 + {GOLDEN_FRAME, NONE},
1.100 + {ALTREF_FRAME, NONE},
1.101 + {LAST_FRAME, ALTREF_FRAME},
1.102 + {GOLDEN_FRAME, ALTREF_FRAME},
1.103 + {INTRA_FRAME, NONE},
1.104 +};
1.105 +
1.106 +// The baseline rd thresholds for breaking out of the rd loop for
1.107 +// certain modes are assumed to be based on 8x8 blocks.
1.108 +// This table is used to correct for blocks size.
1.109 +// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
1.110 +static int rd_thresh_block_size_factor[BLOCK_SIZES] =
1.111 + {2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};
1.112 +
1.113 +#define RD_THRESH_MAX_FACT 64
1.114 +#define RD_THRESH_INC 1
1.115 +#define RD_THRESH_POW 1.25
1.116 +#define RD_MULT_EPB_RATIO 64
1.117 +
1.118 +#define MV_COST_WEIGHT 108
1.119 +#define MV_COST_WEIGHT_SUB 120
1.120 +
1.121 +static void fill_token_costs(vp9_coeff_cost *c,
1.122 + vp9_coeff_probs_model (*p)[BLOCK_TYPES]) {
1.123 + int i, j, k, l;
1.124 + TX_SIZE t;
1.125 + for (t = TX_4X4; t <= TX_32X32; t++)
1.126 + for (i = 0; i < BLOCK_TYPES; i++)
1.127 + for (j = 0; j < REF_TYPES; j++)
1.128 + for (k = 0; k < COEF_BANDS; k++)
1.129 + for (l = 0; l < PREV_COEF_CONTEXTS; l++) {
1.130 + vp9_prob probs[ENTROPY_NODES];
1.131 + vp9_model_to_full_probs(p[t][i][j][k][l], probs);
1.132 + vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
1.133 + vp9_coef_tree);
1.134 + vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
1.135 + vp9_coef_tree);
1.136 + assert(c[t][i][j][k][0][l][DCT_EOB_TOKEN] ==
1.137 + c[t][i][j][k][1][l][DCT_EOB_TOKEN]);
1.138 + }
1.139 +}
1.140 +
1.141 +static const int rd_iifactor[32] = {
1.142 + 4, 4, 3, 2, 1, 0, 0, 0,
1.143 + 0, 0, 0, 0, 0, 0, 0, 0,
1.144 + 0, 0, 0, 0, 0, 0, 0, 0,
1.145 + 0, 0, 0, 0, 0, 0, 0, 0,
1.146 +};
1.147 +
1.148 +// 3* dc_qlookup[Q]*dc_qlookup[Q];
1.149 +
1.150 +/* values are now correlated to quantizer */
1.151 +static int sad_per_bit16lut[QINDEX_RANGE];
1.152 +static int sad_per_bit4lut[QINDEX_RANGE];
1.153 +
1.154 +void vp9_init_me_luts() {
1.155 + int i;
1.156 +
1.157 + // Initialize the sad lut tables using a formulaic calculation for now
1.158 + // This is to make it easier to resolve the impact of experimental changes
1.159 + // to the quantizer tables.
1.160 + for (i = 0; i < QINDEX_RANGE; i++) {
1.161 + sad_per_bit16lut[i] =
1.162 + (int)((0.0418 * vp9_convert_qindex_to_q(i)) + 2.4107);
1.163 + sad_per_bit4lut[i] = (int)(0.063 * vp9_convert_qindex_to_q(i) + 2.742);
1.164 + }
1.165 +}
1.166 +
1.167 +int vp9_compute_rd_mult(VP9_COMP *cpi, int qindex) {
1.168 + const int q = vp9_dc_quant(qindex, 0);
1.169 + // TODO(debargha): Adjust the function below
1.170 + int rdmult = 88 * q * q / 25;
1.171 + if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
1.172 + if (cpi->twopass.next_iiratio > 31)
1.173 + rdmult += (rdmult * rd_iifactor[31]) >> 4;
1.174 + else
1.175 + rdmult += (rdmult * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
1.176 + }
1.177 + return rdmult;
1.178 +}
1.179 +
1.180 +static int compute_rd_thresh_factor(int qindex) {
1.181 + int q;
1.182 + // TODO(debargha): Adjust the function below
1.183 + q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12);
1.184 + if (q < 8)
1.185 + q = 8;
1.186 + return q;
1.187 +}
1.188 +
1.189 +void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) {
1.190 + cpi->mb.sadperbit16 = sad_per_bit16lut[qindex];
1.191 + cpi->mb.sadperbit4 = sad_per_bit4lut[qindex];
1.192 +}
1.193 +
1.194 +static void set_block_thresholds(VP9_COMP *cpi) {
1.195 + int i, bsize, segment_id;
1.196 + VP9_COMMON *cm = &cpi->common;
1.197 +
1.198 + for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
1.199 + int q;
1.200 + int segment_qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1.201 + segment_qindex = clamp(segment_qindex + cm->y_dc_delta_q, 0, MAXQ);
1.202 + q = compute_rd_thresh_factor(segment_qindex);
1.203 +
1.204 + for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
1.205 + // Threshold here seem unecessarily harsh but fine given actual
1.206 + // range of values used for cpi->sf.thresh_mult[]
1.207 + int thresh_max = INT_MAX / (q * rd_thresh_block_size_factor[bsize]);
1.208 +
1.209 + for (i = 0; i < MAX_MODES; ++i) {
1.210 + if (cpi->sf.thresh_mult[i] < thresh_max) {
1.211 + cpi->rd_threshes[segment_id][bsize][i] =
1.212 + cpi->sf.thresh_mult[i] * q *
1.213 + rd_thresh_block_size_factor[bsize] / 4;
1.214 + } else {
1.215 + cpi->rd_threshes[segment_id][bsize][i] = INT_MAX;
1.216 + }
1.217 + }
1.218 +
1.219 + for (i = 0; i < MAX_REFS; ++i) {
1.220 + if (cpi->sf.thresh_mult_sub8x8[i] < thresh_max) {
1.221 + cpi->rd_thresh_sub8x8[segment_id][bsize][i] =
1.222 + cpi->sf.thresh_mult_sub8x8[i] * q *
1.223 + rd_thresh_block_size_factor[bsize] / 4;
1.224 + } else {
1.225 + cpi->rd_thresh_sub8x8[segment_id][bsize][i] = INT_MAX;
1.226 + }
1.227 + }
1.228 + }
1.229 + }
1.230 +}
1.231 +
1.232 +void vp9_initialize_rd_consts(VP9_COMP *cpi) {
1.233 + VP9_COMMON *cm = &cpi->common;
1.234 + int qindex, i;
1.235 +
1.236 + vp9_clear_system_state(); // __asm emms;
1.237 +
1.238 + // Further tests required to see if optimum is different
1.239 + // for key frames, golden frames and arf frames.
1.240 + // if (cpi->common.refresh_golden_frame ||
1.241 + // cpi->common.refresh_alt_ref_frame)
1.242 + qindex = clamp(cm->base_qindex + cm->y_dc_delta_q, 0, MAXQ);
1.243 +
1.244 + cpi->RDDIV = RDDIV_BITS; // in bits (to multiply D by 128)
1.245 + cpi->RDMULT = vp9_compute_rd_mult(cpi, qindex);
1.246 +
1.247 + cpi->mb.errorperbit = cpi->RDMULT / RD_MULT_EPB_RATIO;
1.248 + cpi->mb.errorperbit += (cpi->mb.errorperbit == 0);
1.249 +
1.250 + vp9_set_speed_features(cpi);
1.251 +
1.252 + cpi->mb.select_txfm_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
1.253 + cm->frame_type != KEY_FRAME) ?
1.254 + 0 : 1;
1.255 +
1.256 + set_block_thresholds(cpi);
1.257 +
1.258 + fill_token_costs(cpi->mb.token_costs, cm->fc.coef_probs);
1.259 +
1.260 + for (i = 0; i < PARTITION_CONTEXTS; i++)
1.261 + vp9_cost_tokens(cpi->mb.partition_cost[i], get_partition_probs(cm, i),
1.262 + vp9_partition_tree);
1.263 +
1.264 + /*rough estimate for costing*/
1.265 + vp9_init_mode_costs(cpi);
1.266 +
1.267 + if (!frame_is_intra_only(cm)) {
1.268 + vp9_build_nmv_cost_table(
1.269 + cpi->mb.nmvjointcost,
1.270 + cm->allow_high_precision_mv ? cpi->mb.nmvcost_hp : cpi->mb.nmvcost,
1.271 + &cm->fc.nmvc,
1.272 + cm->allow_high_precision_mv, 1, 1);
1.273 +
1.274 + for (i = 0; i < INTER_MODE_CONTEXTS; i++) {
1.275 + MB_PREDICTION_MODE m;
1.276 +
1.277 + for (m = NEARESTMV; m < MB_MODE_COUNT; m++)
1.278 + cpi->mb.inter_mode_cost[i][INTER_OFFSET(m)] =
1.279 + cost_token(vp9_inter_mode_tree,
1.280 + cm->fc.inter_mode_probs[i],
1.281 + &vp9_inter_mode_encodings[INTER_OFFSET(m)]);
1.282 + }
1.283 + }
1.284 +}
1.285 +
1.286 +static INLINE void linear_interpolate2(double x, int ntab, int inv_step,
1.287 + const double *tab1, const double *tab2,
1.288 + double *v1, double *v2) {
1.289 + double y = x * inv_step;
1.290 + int d = (int) y;
1.291 + if (d >= ntab - 1) {
1.292 + *v1 = tab1[ntab - 1];
1.293 + *v2 = tab2[ntab - 1];
1.294 + } else {
1.295 + double a = y - d;
1.296 + *v1 = tab1[d] * (1 - a) + tab1[d + 1] * a;
1.297 + *v2 = tab2[d] * (1 - a) + tab2[d + 1] * a;
1.298 + }
1.299 +}
1.300 +
1.301 +static void model_rd_norm(double x, double *R, double *D) {
1.302 + static const int inv_tab_step = 8;
1.303 + static const int tab_size = 120;
1.304 + // NOTE: The tables below must be of the same size
1.305 + //
1.306 + // Normalized rate
1.307 + // This table models the rate for a Laplacian source
1.308 + // source with given variance when quantized with a uniform quantizer
1.309 + // with given stepsize. The closed form expression is:
1.310 + // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
1.311 + // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
1.312 + // and H(x) is the binary entropy function.
1.313 + static const double rate_tab[] = {
1.314 + 64.00, 4.944, 3.949, 3.372, 2.966, 2.655, 2.403, 2.194,
1.315 + 2.014, 1.858, 1.720, 1.596, 1.485, 1.384, 1.291, 1.206,
1.316 + 1.127, 1.054, 0.986, 0.923, 0.863, 0.808, 0.756, 0.708,
1.317 + 0.662, 0.619, 0.579, 0.541, 0.506, 0.473, 0.442, 0.412,
1.318 + 0.385, 0.359, 0.335, 0.313, 0.291, 0.272, 0.253, 0.236,
1.319 + 0.220, 0.204, 0.190, 0.177, 0.165, 0.153, 0.142, 0.132,
1.320 + 0.123, 0.114, 0.106, 0.099, 0.091, 0.085, 0.079, 0.073,
1.321 + 0.068, 0.063, 0.058, 0.054, 0.050, 0.047, 0.043, 0.040,
1.322 + 0.037, 0.034, 0.032, 0.029, 0.027, 0.025, 0.023, 0.022,
1.323 + 0.020, 0.019, 0.017, 0.016, 0.015, 0.014, 0.013, 0.012,
1.324 + 0.011, 0.010, 0.009, 0.008, 0.008, 0.007, 0.007, 0.006,
1.325 + 0.006, 0.005, 0.005, 0.005, 0.004, 0.004, 0.004, 0.003,
1.326 + 0.003, 0.003, 0.003, 0.002, 0.002, 0.002, 0.002, 0.002,
1.327 + 0.002, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001,
1.328 + 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.000,
1.329 + };
1.330 + // Normalized distortion
1.331 + // This table models the normalized distortion for a Laplacian source
1.332 + // source with given variance when quantized with a uniform quantizer
1.333 + // with given stepsize. The closed form expression is:
1.334 + // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
1.335 + // where x = qpstep / sqrt(variance)
1.336 + // Note the actual distortion is Dn * variance.
1.337 + static const double dist_tab[] = {
1.338 + 0.000, 0.001, 0.005, 0.012, 0.021, 0.032, 0.045, 0.061,
1.339 + 0.079, 0.098, 0.119, 0.142, 0.166, 0.190, 0.216, 0.242,
1.340 + 0.269, 0.296, 0.324, 0.351, 0.378, 0.405, 0.432, 0.458,
1.341 + 0.484, 0.509, 0.534, 0.557, 0.580, 0.603, 0.624, 0.645,
1.342 + 0.664, 0.683, 0.702, 0.719, 0.735, 0.751, 0.766, 0.780,
1.343 + 0.794, 0.807, 0.819, 0.830, 0.841, 0.851, 0.861, 0.870,
1.344 + 0.878, 0.886, 0.894, 0.901, 0.907, 0.913, 0.919, 0.925,
1.345 + 0.930, 0.935, 0.939, 0.943, 0.947, 0.951, 0.954, 0.957,
1.346 + 0.960, 0.963, 0.966, 0.968, 0.971, 0.973, 0.975, 0.976,
1.347 + 0.978, 0.980, 0.981, 0.982, 0.984, 0.985, 0.986, 0.987,
1.348 + 0.988, 0.989, 0.990, 0.990, 0.991, 0.992, 0.992, 0.993,
1.349 + 0.993, 0.994, 0.994, 0.995, 0.995, 0.996, 0.996, 0.996,
1.350 + 0.996, 0.997, 0.997, 0.997, 0.997, 0.998, 0.998, 0.998,
1.351 + 0.998, 0.998, 0.998, 0.999, 0.999, 0.999, 0.999, 0.999,
1.352 + 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 0.999, 1.000,
1.353 + };
1.354 + /*
1.355 + assert(sizeof(rate_tab) == tab_size * sizeof(rate_tab[0]);
1.356 + assert(sizeof(dist_tab) == tab_size * sizeof(dist_tab[0]);
1.357 + assert(sizeof(rate_tab) == sizeof(dist_tab));
1.358 + */
1.359 + assert(x >= 0.0);
1.360 + linear_interpolate2(x, tab_size, inv_tab_step,
1.361 + rate_tab, dist_tab, R, D);
1.362 +}
1.363 +
1.364 +static void model_rd_from_var_lapndz(int var, int n, int qstep,
1.365 + int *rate, int64_t *dist) {
1.366 + // This function models the rate and distortion for a Laplacian
1.367 + // source with given variance when quantized with a uniform quantizer
1.368 + // with given stepsize. The closed form expressions are in:
1.369 + // Hang and Chen, "Source Model for transform video coder and its
1.370 + // application - Part I: Fundamental Theory", IEEE Trans. Circ.
1.371 + // Sys. for Video Tech., April 1997.
1.372 + vp9_clear_system_state();
1.373 + if (var == 0 || n == 0) {
1.374 + *rate = 0;
1.375 + *dist = 0;
1.376 + } else {
1.377 + double D, R;
1.378 + double s2 = (double) var / n;
1.379 + double x = qstep / sqrt(s2);
1.380 + model_rd_norm(x, &R, &D);
1.381 + *rate = (int)((n << 8) * R + 0.5);
1.382 + *dist = (int)(var * D + 0.5);
1.383 + }
1.384 + vp9_clear_system_state();
1.385 +}
1.386 +
1.387 +static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
1.388 + MACROBLOCK *x, MACROBLOCKD *xd,
1.389 + int *out_rate_sum, int64_t *out_dist_sum) {
1.390 + // Note our transform coeffs are 8 times an orthogonal transform.
1.391 + // Hence quantizer step is also 8 times. To get effective quantizer
1.392 + // we need to divide by 8 before sending to modeling function.
1.393 + int i, rate_sum = 0, dist_sum = 0;
1.394 +
1.395 + for (i = 0; i < MAX_MB_PLANE; ++i) {
1.396 + struct macroblock_plane *const p = &x->plane[i];
1.397 + struct macroblockd_plane *const pd = &xd->plane[i];
1.398 + const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1.399 + unsigned int sse;
1.400 + int rate;
1.401 + int64_t dist;
1.402 + (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
1.403 + pd->dst.buf, pd->dst.stride, &sse);
1.404 + // sse works better than var, since there is no dc prediction used
1.405 + model_rd_from_var_lapndz(sse, 1 << num_pels_log2_lookup[bs],
1.406 + pd->dequant[1] >> 3, &rate, &dist);
1.407 +
1.408 + rate_sum += rate;
1.409 + dist_sum += (int)dist;
1.410 + }
1.411 +
1.412 + *out_rate_sum = rate_sum;
1.413 + *out_dist_sum = dist_sum << 4;
1.414 +}
1.415 +
1.416 +static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE bsize,
1.417 + TX_SIZE tx_size,
1.418 + MACROBLOCK *x, MACROBLOCKD *xd,
1.419 + int *out_rate_sum, int64_t *out_dist_sum,
1.420 + int *out_skip) {
1.421 + int j, k;
1.422 + BLOCK_SIZE bs;
1.423 + struct macroblock_plane *const p = &x->plane[0];
1.424 + struct macroblockd_plane *const pd = &xd->plane[0];
1.425 + const int width = 4 << num_4x4_blocks_wide_lookup[bsize];
1.426 + const int height = 4 << num_4x4_blocks_high_lookup[bsize];
1.427 + int rate_sum = 0;
1.428 + int64_t dist_sum = 0;
1.429 + const int t = 4 << tx_size;
1.430 +
1.431 + if (tx_size == TX_4X4) {
1.432 + bs = BLOCK_4X4;
1.433 + } else if (tx_size == TX_8X8) {
1.434 + bs = BLOCK_8X8;
1.435 + } else if (tx_size == TX_16X16) {
1.436 + bs = BLOCK_16X16;
1.437 + } else if (tx_size == TX_32X32) {
1.438 + bs = BLOCK_32X32;
1.439 + } else {
1.440 + assert(0);
1.441 + }
1.442 +
1.443 + *out_skip = 1;
1.444 + for (j = 0; j < height; j += t) {
1.445 + for (k = 0; k < width; k += t) {
1.446 + int rate;
1.447 + int64_t dist;
1.448 + unsigned int sse;
1.449 + cpi->fn_ptr[bs].vf(&p->src.buf[j * p->src.stride + k], p->src.stride,
1.450 + &pd->dst.buf[j * pd->dst.stride + k], pd->dst.stride,
1.451 + &sse);
1.452 + // sse works better than var, since there is no dc prediction used
1.453 + model_rd_from_var_lapndz(sse, t * t, pd->dequant[1] >> 3, &rate, &dist);
1.454 + rate_sum += rate;
1.455 + dist_sum += dist;
1.456 + *out_skip &= (rate < 1024);
1.457 + }
1.458 + }
1.459 +
1.460 + *out_rate_sum = rate_sum;
1.461 + *out_dist_sum = dist_sum << 4;
1.462 +}
1.463 +
1.464 +int64_t vp9_block_error_c(int16_t *coeff, int16_t *dqcoeff,
1.465 + intptr_t block_size, int64_t *ssz) {
1.466 + int i;
1.467 + int64_t error = 0, sqcoeff = 0;
1.468 +
1.469 + for (i = 0; i < block_size; i++) {
1.470 + int this_diff = coeff[i] - dqcoeff[i];
1.471 + error += (unsigned)this_diff * this_diff;
1.472 + sqcoeff += (unsigned) coeff[i] * coeff[i];
1.473 + }
1.474 +
1.475 + *ssz = sqcoeff;
1.476 + return error;
1.477 +}
1.478 +
1.479 +/* The trailing '0' is a terminator which is used inside cost_coeffs() to
1.480 + * decide whether to include cost of a trailing EOB node or not (i.e. we
1.481 + * can skip this if the last coefficient in this transform block, e.g. the
1.482 + * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
1.483 + * were non-zero). */
1.484 +static const int16_t band_counts[TX_SIZES][8] = {
1.485 + { 1, 2, 3, 4, 3, 16 - 13, 0 },
1.486 + { 1, 2, 3, 4, 11, 64 - 21, 0 },
1.487 + { 1, 2, 3, 4, 11, 256 - 21, 0 },
1.488 + { 1, 2, 3, 4, 11, 1024 - 21, 0 },
1.489 +};
1.490 +
1.491 +static INLINE int cost_coeffs(MACROBLOCK *x,
1.492 + int plane, int block,
1.493 + ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
1.494 + TX_SIZE tx_size,
1.495 + const int16_t *scan, const int16_t *nb) {
1.496 + MACROBLOCKD *const xd = &x->e_mbd;
1.497 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.498 + struct macroblockd_plane *pd = &xd->plane[plane];
1.499 + const PLANE_TYPE type = pd->plane_type;
1.500 + const int16_t *band_count = &band_counts[tx_size][1];
1.501 + const int eob = pd->eobs[block];
1.502 + const int16_t *const qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
1.503 + const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
1.504 + unsigned int (*token_costs)[2][PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] =
1.505 + x->token_costs[tx_size][type][ref];
1.506 + const ENTROPY_CONTEXT above_ec = !!*A, left_ec = !!*L;
1.507 + uint8_t *p_tok = x->token_cache;
1.508 + int pt = combine_entropy_contexts(above_ec, left_ec);
1.509 + int c, cost;
1.510 +
1.511 + // Check for consistency of tx_size with mode info
1.512 + assert(type == PLANE_TYPE_Y_WITH_DC ? mbmi->tx_size == tx_size
1.513 + : get_uv_tx_size(mbmi) == tx_size);
1.514 +
1.515 + if (eob == 0) {
1.516 + // single eob token
1.517 + cost = token_costs[0][0][pt][DCT_EOB_TOKEN];
1.518 + c = 0;
1.519 + } else {
1.520 + int band_left = *band_count++;
1.521 +
1.522 + // dc token
1.523 + int v = qcoeff_ptr[0];
1.524 + int prev_t = vp9_dct_value_tokens_ptr[v].token;
1.525 + cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
1.526 + p_tok[0] = vp9_pt_energy_class[prev_t];
1.527 + ++token_costs;
1.528 +
1.529 + // ac tokens
1.530 + for (c = 1; c < eob; c++) {
1.531 + const int rc = scan[c];
1.532 + int t;
1.533 +
1.534 + v = qcoeff_ptr[rc];
1.535 + t = vp9_dct_value_tokens_ptr[v].token;
1.536 + pt = get_coef_context(nb, p_tok, c);
1.537 + cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
1.538 + p_tok[rc] = vp9_pt_energy_class[t];
1.539 + prev_t = t;
1.540 + if (!--band_left) {
1.541 + band_left = *band_count++;
1.542 + ++token_costs;
1.543 + }
1.544 + }
1.545 +
1.546 + // eob token
1.547 + if (band_left) {
1.548 + pt = get_coef_context(nb, p_tok, c);
1.549 + cost += (*token_costs)[0][pt][DCT_EOB_TOKEN];
1.550 + }
1.551 + }
1.552 +
1.553 + // is eob first coefficient;
1.554 + *A = *L = (c > 0);
1.555 +
1.556 + return cost;
1.557 +}
1.558 +
1.559 +static void dist_block(int plane, int block, TX_SIZE tx_size, void *arg) {
1.560 + const int ss_txfrm_size = tx_size << 1;
1.561 + struct rdcost_block_args* args = arg;
1.562 + MACROBLOCK* const x = args->x;
1.563 + MACROBLOCKD* const xd = &x->e_mbd;
1.564 + struct macroblock_plane *const p = &x->plane[plane];
1.565 + struct macroblockd_plane *const pd = &xd->plane[plane];
1.566 + int64_t this_sse;
1.567 + int shift = args->tx_size == TX_32X32 ? 0 : 2;
1.568 + int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
1.569 + int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1.570 + args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
1.571 + &this_sse) >> shift;
1.572 + args->sse = this_sse >> shift;
1.573 +
1.574 + if (x->skip_encode &&
1.575 + xd->mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME) {
1.576 + // TODO(jingning): tune the model to better capture the distortion.
1.577 + int64_t p = (pd->dequant[1] * pd->dequant[1] *
1.578 + (1 << ss_txfrm_size)) >> (shift + 2);
1.579 + args->dist += (p >> 4);
1.580 + args->sse += p;
1.581 + }
1.582 +}
1.583 +
1.584 +static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
1.585 + TX_SIZE tx_size, void *arg) {
1.586 + struct rdcost_block_args* args = arg;
1.587 +
1.588 + int x_idx, y_idx;
1.589 + txfrm_block_to_raster_xy(plane_bsize, args->tx_size, block, &x_idx, &y_idx);
1.590 +
1.591 + args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
1.592 + args->t_left + y_idx, args->tx_size,
1.593 + args->scan, args->nb);
1.594 +}
1.595 +
1.596 +static void block_yrd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
1.597 + TX_SIZE tx_size, void *arg) {
1.598 + struct rdcost_block_args *args = arg;
1.599 + MACROBLOCK *const x = args->x;
1.600 + MACROBLOCKD *const xd = &x->e_mbd;
1.601 + struct encode_b_args encode_args = {x, NULL};
1.602 + int64_t rd1, rd2, rd;
1.603 +
1.604 + if (args->skip)
1.605 + return;
1.606 +
1.607 + if (!is_inter_block(&xd->mi_8x8[0]->mbmi))
1.608 + vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &encode_args);
1.609 + else
1.610 + vp9_xform_quant(plane, block, plane_bsize, tx_size, &encode_args);
1.611 +
1.612 + dist_block(plane, block, tx_size, args);
1.613 + rate_block(plane, block, plane_bsize, tx_size, args);
1.614 + rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
1.615 + rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);
1.616 +
1.617 + // TODO(jingning): temporarily enabled only for luma component
1.618 + rd = MIN(rd1, rd2);
1.619 + if (!xd->lossless && plane == 0)
1.620 + x->zcoeff_blk[tx_size][block] = rd1 > rd2 || !xd->plane[plane].eobs[block];
1.621 +
1.622 + args->this_rate += args->rate;
1.623 + args->this_dist += args->dist;
1.624 + args->this_sse += args->sse;
1.625 + args->this_rd += rd;
1.626 +
1.627 + if (args->this_rd > args->best_rd) {
1.628 + args->skip = 1;
1.629 + return;
1.630 + }
1.631 +}
1.632 +
1.633 +void vp9_get_entropy_contexts(TX_SIZE tx_size,
1.634 + ENTROPY_CONTEXT t_above[16], ENTROPY_CONTEXT t_left[16],
1.635 + const ENTROPY_CONTEXT *above, const ENTROPY_CONTEXT *left,
1.636 + int num_4x4_w, int num_4x4_h) {
1.637 + int i;
1.638 + switch (tx_size) {
1.639 + case TX_4X4:
1.640 + vpx_memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
1.641 + vpx_memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
1.642 + break;
1.643 + case TX_8X8:
1.644 + for (i = 0; i < num_4x4_w; i += 2)
1.645 + t_above[i] = !!*(const uint16_t *)&above[i];
1.646 + for (i = 0; i < num_4x4_h; i += 2)
1.647 + t_left[i] = !!*(const uint16_t *)&left[i];
1.648 + break;
1.649 + case TX_16X16:
1.650 + for (i = 0; i < num_4x4_w; i += 4)
1.651 + t_above[i] = !!*(const uint32_t *)&above[i];
1.652 + for (i = 0; i < num_4x4_h; i += 4)
1.653 + t_left[i] = !!*(const uint32_t *)&left[i];
1.654 + break;
1.655 + case TX_32X32:
1.656 + for (i = 0; i < num_4x4_w; i += 8)
1.657 + t_above[i] = !!*(const uint64_t *)&above[i];
1.658 + for (i = 0; i < num_4x4_h; i += 8)
1.659 + t_left[i] = !!*(const uint64_t *)&left[i];
1.660 + break;
1.661 + default:
1.662 + assert(!"Invalid transform size.");
1.663 + }
1.664 +}
1.665 +
1.666 +static void init_rdcost_stack(MACROBLOCK *x, TX_SIZE tx_size,
1.667 + const int num_4x4_w, const int num_4x4_h,
1.668 + const int64_t ref_rdcost,
1.669 + struct rdcost_block_args *arg) {
1.670 + vpx_memset(arg, 0, sizeof(struct rdcost_block_args));
1.671 + arg->x = x;
1.672 + arg->tx_size = tx_size;
1.673 + arg->bw = num_4x4_w;
1.674 + arg->bh = num_4x4_h;
1.675 + arg->best_rd = ref_rdcost;
1.676 +}
1.677 +
1.678 +static void txfm_rd_in_plane(MACROBLOCK *x,
1.679 + struct rdcost_block_args *rd_stack,
1.680 + int *rate, int64_t *distortion,
1.681 + int *skippable, int64_t *sse,
1.682 + int64_t ref_best_rd, int plane,
1.683 + BLOCK_SIZE bsize, TX_SIZE tx_size) {
1.684 + MACROBLOCKD *const xd = &x->e_mbd;
1.685 + struct macroblockd_plane *const pd = &xd->plane[plane];
1.686 + const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1.687 + const int num_4x4_w = num_4x4_blocks_wide_lookup[bs];
1.688 + const int num_4x4_h = num_4x4_blocks_high_lookup[bs];
1.689 +
1.690 + init_rdcost_stack(x, tx_size, num_4x4_w, num_4x4_h,
1.691 + ref_best_rd, rd_stack);
1.692 + if (plane == 0)
1.693 + xd->mi_8x8[0]->mbmi.tx_size = tx_size;
1.694 +
1.695 + vp9_get_entropy_contexts(tx_size, rd_stack->t_above, rd_stack->t_left,
1.696 + pd->above_context, pd->left_context,
1.697 + num_4x4_w, num_4x4_h);
1.698 +
1.699 + get_scan(xd, tx_size, pd->plane_type, 0, &rd_stack->scan, &rd_stack->nb);
1.700 +
1.701 + foreach_transformed_block_in_plane(xd, bsize, plane,
1.702 + block_yrd_txfm, rd_stack);
1.703 + if (rd_stack->skip) {
1.704 + *rate = INT_MAX;
1.705 + *distortion = INT64_MAX;
1.706 + *sse = INT64_MAX;
1.707 + *skippable = 0;
1.708 + } else {
1.709 + *distortion = rd_stack->this_dist;
1.710 + *rate = rd_stack->this_rate;
1.711 + *sse = rd_stack->this_sse;
1.712 + *skippable = vp9_is_skippable_in_plane(xd, bsize, plane);
1.713 + }
1.714 +}
1.715 +
1.716 +static void choose_largest_txfm_size(VP9_COMP *cpi, MACROBLOCK *x,
1.717 + int *rate, int64_t *distortion,
1.718 + int *skip, int64_t *sse,
1.719 + int64_t ref_best_rd,
1.720 + BLOCK_SIZE bs) {
1.721 + const TX_SIZE max_tx_size = max_txsize_lookup[bs];
1.722 + VP9_COMMON *const cm = &cpi->common;
1.723 + const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
1.724 + MACROBLOCKD *const xd = &x->e_mbd;
1.725 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.726 +
1.727 + mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
1.728 +
1.729 + txfm_rd_in_plane(x, &cpi->rdcost_stack, rate, distortion, skip,
1.730 + &sse[mbmi->tx_size], ref_best_rd, 0, bs,
1.731 + mbmi->tx_size);
1.732 + cpi->tx_stepdown_count[0]++;
1.733 +}
1.734 +
1.735 +static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
1.736 + int (*r)[2], int *rate,
1.737 + int64_t *d, int64_t *distortion,
1.738 + int *s, int *skip,
1.739 + int64_t tx_cache[TX_MODES],
1.740 + BLOCK_SIZE bs) {
1.741 + const TX_SIZE max_tx_size = max_txsize_lookup[bs];
1.742 + VP9_COMMON *const cm = &cpi->common;
1.743 + MACROBLOCKD *const xd = &x->e_mbd;
1.744 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.745 + vp9_prob skip_prob = vp9_get_pred_prob_mbskip(cm, xd);
1.746 + int64_t rd[TX_SIZES][2];
1.747 + int n, m;
1.748 + int s0, s1;
1.749 +
1.750 + const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
1.751 +
1.752 + for (n = TX_4X4; n <= max_tx_size; n++) {
1.753 + r[n][1] = r[n][0];
1.754 + if (r[n][0] == INT_MAX)
1.755 + continue;
1.756 + for (m = 0; m <= n - (n == max_tx_size); m++) {
1.757 + if (m == n)
1.758 + r[n][1] += vp9_cost_zero(tx_probs[m]);
1.759 + else
1.760 + r[n][1] += vp9_cost_one(tx_probs[m]);
1.761 + }
1.762 + }
1.763 +
1.764 + assert(skip_prob > 0);
1.765 + s0 = vp9_cost_bit(skip_prob, 0);
1.766 + s1 = vp9_cost_bit(skip_prob, 1);
1.767 +
1.768 + for (n = TX_4X4; n <= max_tx_size; n++) {
1.769 + if (d[n] == INT64_MAX) {
1.770 + rd[n][0] = rd[n][1] = INT64_MAX;
1.771 + continue;
1.772 + }
1.773 + if (s[n]) {
1.774 + rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
1.775 + } else {
1.776 + rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
1.777 + rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
1.778 + }
1.779 + }
1.780 +
1.781 + if (max_tx_size == TX_32X32 &&
1.782 + (cm->tx_mode == ALLOW_32X32 ||
1.783 + (cm->tx_mode == TX_MODE_SELECT &&
1.784 + rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
1.785 + rd[TX_32X32][1] < rd[TX_4X4][1]))) {
1.786 + mbmi->tx_size = TX_32X32;
1.787 + } else if (max_tx_size >= TX_16X16 &&
1.788 + (cm->tx_mode == ALLOW_16X16 ||
1.789 + cm->tx_mode == ALLOW_32X32 ||
1.790 + (cm->tx_mode == TX_MODE_SELECT &&
1.791 + rd[TX_16X16][1] < rd[TX_8X8][1] &&
1.792 + rd[TX_16X16][1] < rd[TX_4X4][1]))) {
1.793 + mbmi->tx_size = TX_16X16;
1.794 + } else if (cm->tx_mode == ALLOW_8X8 ||
1.795 + cm->tx_mode == ALLOW_16X16 ||
1.796 + cm->tx_mode == ALLOW_32X32 ||
1.797 + (cm->tx_mode == TX_MODE_SELECT && rd[TX_8X8][1] < rd[TX_4X4][1])) {
1.798 + mbmi->tx_size = TX_8X8;
1.799 + } else {
1.800 + mbmi->tx_size = TX_4X4;
1.801 + }
1.802 +
1.803 + *distortion = d[mbmi->tx_size];
1.804 + *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
1.805 + *skip = s[mbmi->tx_size];
1.806 +
1.807 + tx_cache[ONLY_4X4] = rd[TX_4X4][0];
1.808 + tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
1.809 + tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
1.810 + tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];
1.811 + if (max_tx_size == TX_32X32 &&
1.812 + rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
1.813 + rd[TX_32X32][1] < rd[TX_4X4][1])
1.814 + tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
1.815 + else if (max_tx_size >= TX_16X16 &&
1.816 + rd[TX_16X16][1] < rd[TX_8X8][1] && rd[TX_16X16][1] < rd[TX_4X4][1])
1.817 + tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
1.818 + else
1.819 + tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1] < rd[TX_8X8][1] ?
1.820 + rd[TX_4X4][1] : rd[TX_8X8][1];
1.821 +
1.822 + if (max_tx_size == TX_32X32 &&
1.823 + rd[TX_32X32][1] < rd[TX_16X16][1] &&
1.824 + rd[TX_32X32][1] < rd[TX_8X8][1] &&
1.825 + rd[TX_32X32][1] < rd[TX_4X4][1]) {
1.826 + cpi->tx_stepdown_count[0]++;
1.827 + } else if (max_tx_size >= TX_16X16 &&
1.828 + rd[TX_16X16][1] < rd[TX_8X8][1] &&
1.829 + rd[TX_16X16][1] < rd[TX_4X4][1]) {
1.830 + cpi->tx_stepdown_count[max_tx_size - TX_16X16]++;
1.831 + } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
1.832 + cpi->tx_stepdown_count[max_tx_size - TX_8X8]++;
1.833 + } else {
1.834 + cpi->tx_stepdown_count[max_tx_size - TX_4X4]++;
1.835 + }
1.836 +}
1.837 +
1.838 +static void choose_txfm_size_from_modelrd(VP9_COMP *cpi, MACROBLOCK *x,
1.839 + int (*r)[2], int *rate,
1.840 + int64_t *d, int64_t *distortion,
1.841 + int *s, int *skip, int64_t *sse,
1.842 + int64_t ref_best_rd,
1.843 + BLOCK_SIZE bs) {
1.844 + const TX_SIZE max_tx_size = max_txsize_lookup[bs];
1.845 + VP9_COMMON *const cm = &cpi->common;
1.846 + MACROBLOCKD *const xd = &x->e_mbd;
1.847 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.848 + vp9_prob skip_prob = vp9_get_pred_prob_mbskip(cm, xd);
1.849 + int64_t rd[TX_SIZES][2];
1.850 + int n, m;
1.851 + int s0, s1;
1.852 + double scale_rd[TX_SIZES] = {1.73, 1.44, 1.20, 1.00};
1.853 + // double scale_r[TX_SIZES] = {2.82, 2.00, 1.41, 1.00};
1.854 +
1.855 + const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
1.856 +
1.857 + // for (n = TX_4X4; n <= max_txfm_size; n++)
1.858 + // r[n][0] = (r[n][0] * scale_r[n]);
1.859 +
1.860 + for (n = TX_4X4; n <= max_tx_size; n++) {
1.861 + r[n][1] = r[n][0];
1.862 + for (m = 0; m <= n - (n == max_tx_size); m++) {
1.863 + if (m == n)
1.864 + r[n][1] += vp9_cost_zero(tx_probs[m]);
1.865 + else
1.866 + r[n][1] += vp9_cost_one(tx_probs[m]);
1.867 + }
1.868 + }
1.869 +
1.870 + assert(skip_prob > 0);
1.871 + s0 = vp9_cost_bit(skip_prob, 0);
1.872 + s1 = vp9_cost_bit(skip_prob, 1);
1.873 +
1.874 + for (n = TX_4X4; n <= max_tx_size; n++) {
1.875 + if (s[n]) {
1.876 + rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
1.877 + } else {
1.878 + rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
1.879 + rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
1.880 + }
1.881 + }
1.882 + for (n = TX_4X4; n <= max_tx_size; n++) {
1.883 + rd[n][0] = (int64_t)(scale_rd[n] * rd[n][0]);
1.884 + rd[n][1] = (int64_t)(scale_rd[n] * rd[n][1]);
1.885 + }
1.886 +
1.887 + if (max_tx_size == TX_32X32 &&
1.888 + (cm->tx_mode == ALLOW_32X32 ||
1.889 + (cm->tx_mode == TX_MODE_SELECT &&
1.890 + rd[TX_32X32][1] <= rd[TX_16X16][1] &&
1.891 + rd[TX_32X32][1] <= rd[TX_8X8][1] &&
1.892 + rd[TX_32X32][1] <= rd[TX_4X4][1]))) {
1.893 + mbmi->tx_size = TX_32X32;
1.894 + } else if (max_tx_size >= TX_16X16 &&
1.895 + (cm->tx_mode == ALLOW_16X16 ||
1.896 + cm->tx_mode == ALLOW_32X32 ||
1.897 + (cm->tx_mode == TX_MODE_SELECT &&
1.898 + rd[TX_16X16][1] <= rd[TX_8X8][1] &&
1.899 + rd[TX_16X16][1] <= rd[TX_4X4][1]))) {
1.900 + mbmi->tx_size = TX_16X16;
1.901 + } else if (cm->tx_mode == ALLOW_8X8 ||
1.902 + cm->tx_mode == ALLOW_16X16 ||
1.903 + cm->tx_mode == ALLOW_32X32 ||
1.904 + (cm->tx_mode == TX_MODE_SELECT &&
1.905 + rd[TX_8X8][1] <= rd[TX_4X4][1])) {
1.906 + mbmi->tx_size = TX_8X8;
1.907 + } else {
1.908 + mbmi->tx_size = TX_4X4;
1.909 + }
1.910 +
1.911 + // Actually encode using the chosen mode if a model was used, but do not
1.912 + // update the r, d costs
1.913 + txfm_rd_in_plane(x, &cpi->rdcost_stack, rate, distortion, skip,
1.914 + &sse[mbmi->tx_size], ref_best_rd, 0, bs, mbmi->tx_size);
1.915 +
1.916 + if (max_tx_size == TX_32X32 &&
1.917 + rd[TX_32X32][1] <= rd[TX_16X16][1] &&
1.918 + rd[TX_32X32][1] <= rd[TX_8X8][1] &&
1.919 + rd[TX_32X32][1] <= rd[TX_4X4][1]) {
1.920 + cpi->tx_stepdown_count[0]++;
1.921 + } else if (max_tx_size >= TX_16X16 &&
1.922 + rd[TX_16X16][1] <= rd[TX_8X8][1] &&
1.923 + rd[TX_16X16][1] <= rd[TX_4X4][1]) {
1.924 + cpi->tx_stepdown_count[max_tx_size - TX_16X16]++;
1.925 + } else if (rd[TX_8X8][1] <= rd[TX_4X4][1]) {
1.926 + cpi->tx_stepdown_count[max_tx_size - TX_8X8]++;
1.927 + } else {
1.928 + cpi->tx_stepdown_count[max_tx_size - TX_4X4]++;
1.929 + }
1.930 +}
1.931 +
1.932 +static void super_block_yrd(VP9_COMP *cpi,
1.933 + MACROBLOCK *x, int *rate, int64_t *distortion,
1.934 + int *skip, int64_t *psse, BLOCK_SIZE bs,
1.935 + int64_t txfm_cache[TX_MODES],
1.936 + int64_t ref_best_rd) {
1.937 + int r[TX_SIZES][2], s[TX_SIZES];
1.938 + int64_t d[TX_SIZES], sse[TX_SIZES];
1.939 + MACROBLOCKD *xd = &x->e_mbd;
1.940 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.941 + struct rdcost_block_args *rdcost_stack = &cpi->rdcost_stack;
1.942 + const int b_inter_mode = is_inter_block(mbmi);
1.943 +
1.944 + assert(bs == mbmi->sb_type);
1.945 + if (b_inter_mode)
1.946 + vp9_subtract_sby(x, bs);
1.947 +
1.948 + if (cpi->sf.tx_size_search_method == USE_LARGESTALL ||
1.949 + (cpi->sf.tx_size_search_method != USE_FULL_RD &&
1.950 + !b_inter_mode)) {
1.951 + vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
1.952 + choose_largest_txfm_size(cpi, x, rate, distortion, skip, sse,
1.953 + ref_best_rd, bs);
1.954 + if (psse)
1.955 + *psse = sse[mbmi->tx_size];
1.956 + return;
1.957 + }
1.958 +
1.959 + if (cpi->sf.tx_size_search_method == USE_LARGESTINTRA_MODELINTER &&
1.960 + b_inter_mode) {
1.961 + if (bs >= BLOCK_32X32)
1.962 + model_rd_for_sb_y_tx(cpi, bs, TX_32X32, x, xd,
1.963 + &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32]);
1.964 + if (bs >= BLOCK_16X16)
1.965 + model_rd_for_sb_y_tx(cpi, bs, TX_16X16, x, xd,
1.966 + &r[TX_16X16][0], &d[TX_16X16], &s[TX_16X16]);
1.967 +
1.968 + model_rd_for_sb_y_tx(cpi, bs, TX_8X8, x, xd,
1.969 + &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8]);
1.970 +
1.971 + model_rd_for_sb_y_tx(cpi, bs, TX_4X4, x, xd,
1.972 + &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4]);
1.973 +
1.974 + choose_txfm_size_from_modelrd(cpi, x, r, rate, d, distortion, s,
1.975 + skip, sse, ref_best_rd, bs);
1.976 + } else {
1.977 + if (bs >= BLOCK_32X32)
1.978 + txfm_rd_in_plane(x, rdcost_stack, &r[TX_32X32][0], &d[TX_32X32],
1.979 + &s[TX_32X32], &sse[TX_32X32],
1.980 + ref_best_rd, 0, bs, TX_32X32);
1.981 + if (bs >= BLOCK_16X16)
1.982 + txfm_rd_in_plane(x, rdcost_stack, &r[TX_16X16][0], &d[TX_16X16],
1.983 + &s[TX_16X16], &sse[TX_16X16],
1.984 + ref_best_rd, 0, bs, TX_16X16);
1.985 + txfm_rd_in_plane(x, rdcost_stack, &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8],
1.986 + &sse[TX_8X8], ref_best_rd, 0, bs, TX_8X8);
1.987 + txfm_rd_in_plane(x, rdcost_stack, &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4],
1.988 + &sse[TX_4X4], ref_best_rd, 0, bs, TX_4X4);
1.989 + choose_txfm_size_from_rd(cpi, x, r, rate, d, distortion, s,
1.990 + skip, txfm_cache, bs);
1.991 + }
1.992 + if (psse)
1.993 + *psse = sse[mbmi->tx_size];
1.994 +}
1.995 +
1.996 +static int conditional_skipintra(MB_PREDICTION_MODE mode,
1.997 + MB_PREDICTION_MODE best_intra_mode) {
1.998 + if (mode == D117_PRED &&
1.999 + best_intra_mode != V_PRED &&
1.1000 + best_intra_mode != D135_PRED)
1.1001 + return 1;
1.1002 + if (mode == D63_PRED &&
1.1003 + best_intra_mode != V_PRED &&
1.1004 + best_intra_mode != D45_PRED)
1.1005 + return 1;
1.1006 + if (mode == D207_PRED &&
1.1007 + best_intra_mode != H_PRED &&
1.1008 + best_intra_mode != D45_PRED)
1.1009 + return 1;
1.1010 + if (mode == D153_PRED &&
1.1011 + best_intra_mode != H_PRED &&
1.1012 + best_intra_mode != D135_PRED)
1.1013 + return 1;
1.1014 + return 0;
1.1015 +}
1.1016 +
1.1017 +static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
1.1018 + MB_PREDICTION_MODE *best_mode,
1.1019 + int *bmode_costs,
1.1020 + ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
1.1021 + int *bestrate, int *bestratey,
1.1022 + int64_t *bestdistortion,
1.1023 + BLOCK_SIZE bsize, int64_t rd_thresh) {
1.1024 + MB_PREDICTION_MODE mode;
1.1025 + MACROBLOCKD *xd = &x->e_mbd;
1.1026 + int64_t best_rd = rd_thresh;
1.1027 + int rate = 0;
1.1028 + int64_t distortion;
1.1029 + struct macroblock_plane *p = &x->plane[0];
1.1030 + struct macroblockd_plane *pd = &xd->plane[0];
1.1031 + const int src_stride = p->src.stride;
1.1032 + const int dst_stride = pd->dst.stride;
1.1033 + uint8_t *src_init = raster_block_offset_uint8(BLOCK_8X8, ib,
1.1034 + p->src.buf, src_stride);
1.1035 + uint8_t *dst_init = raster_block_offset_uint8(BLOCK_8X8, ib,
1.1036 + pd->dst.buf, dst_stride);
1.1037 + int16_t *src_diff, *coeff;
1.1038 +
1.1039 + ENTROPY_CONTEXT ta[2], tempa[2];
1.1040 + ENTROPY_CONTEXT tl[2], templ[2];
1.1041 +
1.1042 + const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1.1043 + const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1.1044 + int idx, idy;
1.1045 + uint8_t best_dst[8 * 8];
1.1046 +
1.1047 + assert(ib < 4);
1.1048 +
1.1049 + vpx_memcpy(ta, a, sizeof(ta));
1.1050 + vpx_memcpy(tl, l, sizeof(tl));
1.1051 + xd->mi_8x8[0]->mbmi.tx_size = TX_4X4;
1.1052 +
1.1053 + for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1.1054 + int64_t this_rd;
1.1055 + int ratey = 0;
1.1056 +
1.1057 + if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
1.1058 + continue;
1.1059 +
1.1060 + // Only do the oblique modes if the best so far is
1.1061 + // one of the neighboring directional modes
1.1062 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1.1063 + if (conditional_skipintra(mode, *best_mode))
1.1064 + continue;
1.1065 + }
1.1066 +
1.1067 + rate = bmode_costs[mode];
1.1068 + distortion = 0;
1.1069 +
1.1070 + vpx_memcpy(tempa, ta, sizeof(ta));
1.1071 + vpx_memcpy(templ, tl, sizeof(tl));
1.1072 +
1.1073 + for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1.1074 + for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1.1075 + int64_t ssz;
1.1076 + const int16_t *scan;
1.1077 + const int16_t *nb;
1.1078 + uint8_t *src = src_init + idx * 4 + idy * 4 * src_stride;
1.1079 + uint8_t *dst = dst_init + idx * 4 + idy * 4 * dst_stride;
1.1080 + const int block = ib + idy * 2 + idx;
1.1081 + TX_TYPE tx_type;
1.1082 + xd->mi_8x8[0]->bmi[block].as_mode = mode;
1.1083 + src_diff = raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1.1084 + coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1.1085 + vp9_predict_intra_block(xd, block, 1,
1.1086 + TX_4X4, mode,
1.1087 + x->skip_encode ? src : dst,
1.1088 + x->skip_encode ? src_stride : dst_stride,
1.1089 + dst, dst_stride);
1.1090 + vp9_subtract_block(4, 4, src_diff, 8,
1.1091 + src, src_stride,
1.1092 + dst, dst_stride);
1.1093 +
1.1094 + tx_type = get_tx_type_4x4(PLANE_TYPE_Y_WITH_DC, xd, block);
1.1095 + get_scan_nb_4x4(tx_type, &scan, &nb);
1.1096 +
1.1097 + if (tx_type != DCT_DCT)
1.1098 + vp9_short_fht4x4(src_diff, coeff, 8, tx_type);
1.1099 + else
1.1100 + x->fwd_txm4x4(src_diff, coeff, 8);
1.1101 +
1.1102 + vp9_regular_quantize_b_4x4(x, 4, block, scan, get_iscan_4x4(tx_type));
1.1103 +
1.1104 + ratey += cost_coeffs(x, 0, block,
1.1105 + tempa + idx, templ + idy, TX_4X4, scan, nb);
1.1106 + distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
1.1107 + 16, &ssz) >> 2;
1.1108 + if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1.1109 + goto next;
1.1110 +
1.1111 + if (tx_type != DCT_DCT)
1.1112 + vp9_iht4x4_16_add(BLOCK_OFFSET(pd->dqcoeff, block),
1.1113 + dst, pd->dst.stride, tx_type);
1.1114 + else
1.1115 + xd->itxm_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, pd->dst.stride,
1.1116 + 16);
1.1117 + }
1.1118 + }
1.1119 +
1.1120 + rate += ratey;
1.1121 + this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1.1122 +
1.1123 + if (this_rd < best_rd) {
1.1124 + *bestrate = rate;
1.1125 + *bestratey = ratey;
1.1126 + *bestdistortion = distortion;
1.1127 + best_rd = this_rd;
1.1128 + *best_mode = mode;
1.1129 + vpx_memcpy(a, tempa, sizeof(tempa));
1.1130 + vpx_memcpy(l, templ, sizeof(templ));
1.1131 + for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1.1132 + vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
1.1133 + num_4x4_blocks_wide * 4);
1.1134 + }
1.1135 + next:
1.1136 + {}
1.1137 + }
1.1138 +
1.1139 + if (best_rd >= rd_thresh || x->skip_encode)
1.1140 + return best_rd;
1.1141 +
1.1142 + for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1.1143 + vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
1.1144 + num_4x4_blocks_wide * 4);
1.1145 +
1.1146 + return best_rd;
1.1147 +}
1.1148 +
1.1149 +static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP * const cpi,
1.1150 + MACROBLOCK * const mb,
1.1151 + int * const rate,
1.1152 + int * const rate_y,
1.1153 + int64_t * const distortion,
1.1154 + int64_t best_rd) {
1.1155 + int i, j;
1.1156 + MACROBLOCKD *const xd = &mb->e_mbd;
1.1157 + MODE_INFO *const mic = xd->mi_8x8[0];
1.1158 + const MODE_INFO *above_mi = xd->mi_8x8[-xd->mode_info_stride];
1.1159 + const MODE_INFO *left_mi = xd->left_available ? xd->mi_8x8[-1] : NULL;
1.1160 + const BLOCK_SIZE bsize = xd->mi_8x8[0]->mbmi.sb_type;
1.1161 + const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1.1162 + const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1.1163 + int idx, idy;
1.1164 + int cost = 0;
1.1165 + int64_t total_distortion = 0;
1.1166 + int tot_rate_y = 0;
1.1167 + int64_t total_rd = 0;
1.1168 + ENTROPY_CONTEXT t_above[4], t_left[4];
1.1169 + int *bmode_costs;
1.1170 +
1.1171 + vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
1.1172 + vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
1.1173 +
1.1174 + bmode_costs = mb->mbmode_cost;
1.1175 +
1.1176 + // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1.1177 + for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1.1178 + for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1.1179 + MB_PREDICTION_MODE best_mode = DC_PRED;
1.1180 + int r = INT_MAX, ry = INT_MAX;
1.1181 + int64_t d = INT64_MAX, this_rd = INT64_MAX;
1.1182 + i = idy * 2 + idx;
1.1183 + if (cpi->common.frame_type == KEY_FRAME) {
1.1184 + const MB_PREDICTION_MODE A = above_block_mode(mic, above_mi, i);
1.1185 + const MB_PREDICTION_MODE L = left_block_mode(mic, left_mi, i);
1.1186 +
1.1187 + bmode_costs = mb->y_mode_costs[A][L];
1.1188 + }
1.1189 +
1.1190 + this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
1.1191 + t_above + idx, t_left + idy, &r, &ry, &d,
1.1192 + bsize, best_rd - total_rd);
1.1193 + if (this_rd >= best_rd - total_rd)
1.1194 + return INT64_MAX;
1.1195 +
1.1196 + total_rd += this_rd;
1.1197 + cost += r;
1.1198 + total_distortion += d;
1.1199 + tot_rate_y += ry;
1.1200 +
1.1201 + mic->bmi[i].as_mode = best_mode;
1.1202 + for (j = 1; j < num_4x4_blocks_high; ++j)
1.1203 + mic->bmi[i + j * 2].as_mode = best_mode;
1.1204 + for (j = 1; j < num_4x4_blocks_wide; ++j)
1.1205 + mic->bmi[i + j].as_mode = best_mode;
1.1206 +
1.1207 + if (total_rd >= best_rd)
1.1208 + return INT64_MAX;
1.1209 + }
1.1210 + }
1.1211 +
1.1212 + *rate = cost;
1.1213 + *rate_y = tot_rate_y;
1.1214 + *distortion = total_distortion;
1.1215 + mic->mbmi.mode = mic->bmi[3].as_mode;
1.1216 +
1.1217 + return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1.1218 +}
1.1219 +
1.1220 +static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
1.1221 + int *rate, int *rate_tokenonly,
1.1222 + int64_t *distortion, int *skippable,
1.1223 + BLOCK_SIZE bsize,
1.1224 + int64_t tx_cache[TX_MODES],
1.1225 + int64_t best_rd) {
1.1226 + MB_PREDICTION_MODE mode;
1.1227 + MB_PREDICTION_MODE mode_selected = DC_PRED;
1.1228 + MACROBLOCKD *const xd = &x->e_mbd;
1.1229 + MODE_INFO *const mic = xd->mi_8x8[0];
1.1230 + int this_rate, this_rate_tokenonly, s;
1.1231 + int64_t this_distortion, this_rd;
1.1232 + TX_SIZE best_tx = TX_4X4;
1.1233 + int i;
1.1234 + int *bmode_costs = x->mbmode_cost;
1.1235 +
1.1236 + if (cpi->sf.tx_size_search_method == USE_FULL_RD)
1.1237 + for (i = 0; i < TX_MODES; i++)
1.1238 + tx_cache[i] = INT64_MAX;
1.1239 +
1.1240 + /* Y Search for intra prediction mode */
1.1241 + for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1.1242 + int64_t local_tx_cache[TX_MODES];
1.1243 + MODE_INFO *above_mi = xd->mi_8x8[-xd->mode_info_stride];
1.1244 + MODE_INFO *left_mi = xd->left_available ? xd->mi_8x8[-1] : NULL;
1.1245 +
1.1246 + if (!(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]] & (1 << mode)))
1.1247 + continue;
1.1248 +
1.1249 + if (cpi->common.frame_type == KEY_FRAME) {
1.1250 + const MB_PREDICTION_MODE A = above_block_mode(mic, above_mi, 0);
1.1251 + const MB_PREDICTION_MODE L = left_block_mode(mic, left_mi, 0);
1.1252 +
1.1253 + bmode_costs = x->y_mode_costs[A][L];
1.1254 + }
1.1255 + mic->mbmi.mode = mode;
1.1256 +
1.1257 + super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
1.1258 + bsize, local_tx_cache, best_rd);
1.1259 +
1.1260 + if (this_rate_tokenonly == INT_MAX)
1.1261 + continue;
1.1262 +
1.1263 + this_rate = this_rate_tokenonly + bmode_costs[mode];
1.1264 + this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1.1265 +
1.1266 + if (this_rd < best_rd) {
1.1267 + mode_selected = mode;
1.1268 + best_rd = this_rd;
1.1269 + best_tx = mic->mbmi.tx_size;
1.1270 + *rate = this_rate;
1.1271 + *rate_tokenonly = this_rate_tokenonly;
1.1272 + *distortion = this_distortion;
1.1273 + *skippable = s;
1.1274 + }
1.1275 +
1.1276 + if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) {
1.1277 + for (i = 0; i < TX_MODES && local_tx_cache[i] < INT64_MAX; i++) {
1.1278 + const int64_t adj_rd = this_rd + local_tx_cache[i] -
1.1279 + local_tx_cache[cpi->common.tx_mode];
1.1280 + if (adj_rd < tx_cache[i]) {
1.1281 + tx_cache[i] = adj_rd;
1.1282 + }
1.1283 + }
1.1284 + }
1.1285 + }
1.1286 +
1.1287 + mic->mbmi.mode = mode_selected;
1.1288 + mic->mbmi.tx_size = best_tx;
1.1289 +
1.1290 + return best_rd;
1.1291 +}
1.1292 +
1.1293 +static void super_block_uvrd(VP9_COMP *const cpi, MACROBLOCK *x,
1.1294 + int *rate, int64_t *distortion, int *skippable,
1.1295 + int64_t *sse, BLOCK_SIZE bsize,
1.1296 + int64_t ref_best_rd) {
1.1297 + MACROBLOCKD *const xd = &x->e_mbd;
1.1298 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.1299 + TX_SIZE uv_txfm_size = get_uv_tx_size(mbmi);
1.1300 + int plane;
1.1301 + int pnrate = 0, pnskip = 1;
1.1302 + int64_t pndist = 0, pnsse = 0;
1.1303 +
1.1304 + if (ref_best_rd < 0)
1.1305 + goto term;
1.1306 +
1.1307 + if (is_inter_block(mbmi))
1.1308 + vp9_subtract_sbuv(x, bsize);
1.1309 +
1.1310 + *rate = 0;
1.1311 + *distortion = 0;
1.1312 + *sse = 0;
1.1313 + *skippable = 1;
1.1314 +
1.1315 + for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1.1316 + txfm_rd_in_plane(x, &cpi->rdcost_stack, &pnrate, &pndist, &pnskip, &pnsse,
1.1317 + ref_best_rd, plane, bsize, uv_txfm_size);
1.1318 + if (pnrate == INT_MAX)
1.1319 + goto term;
1.1320 + *rate += pnrate;
1.1321 + *distortion += pndist;
1.1322 + *sse += pnsse;
1.1323 + *skippable &= pnskip;
1.1324 + }
1.1325 + return;
1.1326 +
1.1327 + term:
1.1328 + *rate = INT_MAX;
1.1329 + *distortion = INT64_MAX;
1.1330 + *sse = INT64_MAX;
1.1331 + *skippable = 0;
1.1332 + return;
1.1333 +}
1.1334 +
1.1335 +static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1.1336 + PICK_MODE_CONTEXT *ctx,
1.1337 + int *rate, int *rate_tokenonly,
1.1338 + int64_t *distortion, int *skippable,
1.1339 + BLOCK_SIZE bsize) {
1.1340 + MB_PREDICTION_MODE mode;
1.1341 + MB_PREDICTION_MODE mode_selected = DC_PRED;
1.1342 + int64_t best_rd = INT64_MAX, this_rd;
1.1343 + int this_rate_tokenonly, this_rate, s;
1.1344 + int64_t this_distortion, this_sse;
1.1345 +
1.1346 + // int mode_mask = (bsize <= BLOCK_8X8)
1.1347 + // ? ALL_INTRA_MODES : cpi->sf.intra_uv_mode_mask;
1.1348 +
1.1349 + for (mode = DC_PRED; mode <= TM_PRED; mode ++) {
1.1350 + // if (!(mode_mask & (1 << mode)))
1.1351 + if (!(cpi->sf.intra_uv_mode_mask[max_uv_txsize_lookup[bsize]]
1.1352 + & (1 << mode)))
1.1353 + continue;
1.1354 +
1.1355 + x->e_mbd.mi_8x8[0]->mbmi.uv_mode = mode;
1.1356 +
1.1357 + super_block_uvrd(cpi, x, &this_rate_tokenonly,
1.1358 + &this_distortion, &s, &this_sse, bsize, best_rd);
1.1359 + if (this_rate_tokenonly == INT_MAX)
1.1360 + continue;
1.1361 + this_rate = this_rate_tokenonly +
1.1362 + x->intra_uv_mode_cost[cpi->common.frame_type][mode];
1.1363 + this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1.1364 +
1.1365 + if (this_rd < best_rd) {
1.1366 + mode_selected = mode;
1.1367 + best_rd = this_rd;
1.1368 + *rate = this_rate;
1.1369 + *rate_tokenonly = this_rate_tokenonly;
1.1370 + *distortion = this_distortion;
1.1371 + *skippable = s;
1.1372 + if (!x->select_txfm_size) {
1.1373 + int i;
1.1374 + struct macroblock_plane *const p = x->plane;
1.1375 + struct macroblockd_plane *const pd = x->e_mbd.plane;
1.1376 + for (i = 1; i < MAX_MB_PLANE; ++i) {
1.1377 + p[i].coeff = ctx->coeff_pbuf[i][2];
1.1378 + pd[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1.1379 + pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1.1380 + pd[i].eobs = ctx->eobs_pbuf[i][2];
1.1381 +
1.1382 + ctx->coeff_pbuf[i][2] = ctx->coeff_pbuf[i][0];
1.1383 + ctx->qcoeff_pbuf[i][2] = ctx->qcoeff_pbuf[i][0];
1.1384 + ctx->dqcoeff_pbuf[i][2] = ctx->dqcoeff_pbuf[i][0];
1.1385 + ctx->eobs_pbuf[i][2] = ctx->eobs_pbuf[i][0];
1.1386 +
1.1387 + ctx->coeff_pbuf[i][0] = p[i].coeff;
1.1388 + ctx->qcoeff_pbuf[i][0] = pd[i].qcoeff;
1.1389 + ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
1.1390 + ctx->eobs_pbuf[i][0] = pd[i].eobs;
1.1391 + }
1.1392 + }
1.1393 + }
1.1394 + }
1.1395 +
1.1396 + x->e_mbd.mi_8x8[0]->mbmi.uv_mode = mode_selected;
1.1397 +
1.1398 + return best_rd;
1.1399 +}
1.1400 +
1.1401 +static int64_t rd_sbuv_dcpred(VP9_COMP *cpi, MACROBLOCK *x,
1.1402 + int *rate, int *rate_tokenonly,
1.1403 + int64_t *distortion, int *skippable,
1.1404 + BLOCK_SIZE bsize) {
1.1405 + int64_t this_rd;
1.1406 + int64_t this_sse;
1.1407 +
1.1408 + x->e_mbd.mi_8x8[0]->mbmi.uv_mode = DC_PRED;
1.1409 + super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1.1410 + skippable, &this_sse, bsize, INT64_MAX);
1.1411 + *rate = *rate_tokenonly +
1.1412 + x->intra_uv_mode_cost[cpi->common.frame_type][DC_PRED];
1.1413 + this_rd = RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1.1414 +
1.1415 + return this_rd;
1.1416 +}
1.1417 +
1.1418 +static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
1.1419 + BLOCK_SIZE bsize, int *rate_uv,
1.1420 + int *rate_uv_tokenonly,
1.1421 + int64_t *dist_uv, int *skip_uv,
1.1422 + MB_PREDICTION_MODE *mode_uv) {
1.1423 + MACROBLOCK *const x = &cpi->mb;
1.1424 +
1.1425 + // Use an estimated rd for uv_intra based on DC_PRED if the
1.1426 + // appropriate speed flag is set.
1.1427 + if (cpi->sf.use_uv_intra_rd_estimate) {
1.1428 + rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1.1429 + bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1.1430 + // Else do a proper rd search for each possible transform size that may
1.1431 + // be considered in the main rd loop.
1.1432 + } else {
1.1433 + rd_pick_intra_sbuv_mode(cpi, x, ctx,
1.1434 + rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1.1435 + bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1.1436 + }
1.1437 + *mode_uv = x->e_mbd.mi_8x8[0]->mbmi.uv_mode;
1.1438 +}
1.1439 +
1.1440 +static int cost_mv_ref(VP9_COMP *cpi, MB_PREDICTION_MODE mode,
1.1441 + int mode_context) {
1.1442 + MACROBLOCK *const x = &cpi->mb;
1.1443 + MACROBLOCKD *const xd = &x->e_mbd;
1.1444 + const int segment_id = xd->mi_8x8[0]->mbmi.segment_id;
1.1445 +
1.1446 + // Don't account for mode here if segment skip is enabled.
1.1447 + if (!vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
1.1448 + assert(is_inter_mode(mode));
1.1449 + return x->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1.1450 + } else {
1.1451 + return 0;
1.1452 + }
1.1453 +}
1.1454 +
1.1455 +void vp9_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv) {
1.1456 + x->e_mbd.mi_8x8[0]->mbmi.mode = mb;
1.1457 + x->e_mbd.mi_8x8[0]->mbmi.mv[0].as_int = mv->as_int;
1.1458 +}
1.1459 +
1.1460 +static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1.1461 + BLOCK_SIZE bsize,
1.1462 + int_mv *frame_mv,
1.1463 + int mi_row, int mi_col,
1.1464 + int_mv single_newmv[MAX_REF_FRAMES],
1.1465 + int *rate_mv);
1.1466 +
1.1467 +static int labels2mode(MACROBLOCK *x, int i,
1.1468 + MB_PREDICTION_MODE this_mode,
1.1469 + int_mv *this_mv, int_mv *this_second_mv,
1.1470 + int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1.1471 + int_mv seg_mvs[MAX_REF_FRAMES],
1.1472 + int_mv *best_ref_mv,
1.1473 + int_mv *second_best_ref_mv,
1.1474 + int *mvjcost, int *mvcost[2], VP9_COMP *cpi) {
1.1475 + MACROBLOCKD *const xd = &x->e_mbd;
1.1476 + MODE_INFO *const mic = xd->mi_8x8[0];
1.1477 + MB_MODE_INFO *mbmi = &mic->mbmi;
1.1478 + int cost = 0, thismvcost = 0;
1.1479 + int idx, idy;
1.1480 + const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1.1481 + const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1.1482 + const int has_second_rf = has_second_ref(mbmi);
1.1483 +
1.1484 + /* We have to be careful retrieving previously-encoded motion vectors.
1.1485 + Ones from this macroblock have to be pulled from the BLOCKD array
1.1486 + as they have not yet made it to the bmi array in our MB_MODE_INFO. */
1.1487 + MB_PREDICTION_MODE m;
1.1488 +
1.1489 + // the only time we should do costing for new motion vector or mode
1.1490 + // is when we are on a new label (jbb May 08, 2007)
1.1491 + switch (m = this_mode) {
1.1492 + case NEWMV:
1.1493 + this_mv->as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1.1494 + thismvcost = vp9_mv_bit_cost(&this_mv->as_mv, &best_ref_mv->as_mv,
1.1495 + mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1.1496 + if (has_second_rf) {
1.1497 + this_second_mv->as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1.1498 + thismvcost += vp9_mv_bit_cost(&this_second_mv->as_mv,
1.1499 + &second_best_ref_mv->as_mv,
1.1500 + mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1.1501 + }
1.1502 + break;
1.1503 + case NEARESTMV:
1.1504 + this_mv->as_int = frame_mv[NEARESTMV][mbmi->ref_frame[0]].as_int;
1.1505 + if (has_second_rf)
1.1506 + this_second_mv->as_int =
1.1507 + frame_mv[NEARESTMV][mbmi->ref_frame[1]].as_int;
1.1508 + break;
1.1509 + case NEARMV:
1.1510 + this_mv->as_int = frame_mv[NEARMV][mbmi->ref_frame[0]].as_int;
1.1511 + if (has_second_rf)
1.1512 + this_second_mv->as_int =
1.1513 + frame_mv[NEARMV][mbmi->ref_frame[1]].as_int;
1.1514 + break;
1.1515 + case ZEROMV:
1.1516 + this_mv->as_int = 0;
1.1517 + if (has_second_rf)
1.1518 + this_second_mv->as_int = 0;
1.1519 + break;
1.1520 + default:
1.1521 + break;
1.1522 + }
1.1523 +
1.1524 + cost = cost_mv_ref(cpi, this_mode,
1.1525 + mbmi->mode_context[mbmi->ref_frame[0]]);
1.1526 +
1.1527 + mic->bmi[i].as_mv[0].as_int = this_mv->as_int;
1.1528 + if (has_second_rf)
1.1529 + mic->bmi[i].as_mv[1].as_int = this_second_mv->as_int;
1.1530 +
1.1531 + mic->bmi[i].as_mode = m;
1.1532 +
1.1533 + for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1.1534 + for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1.1535 + vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
1.1536 + &mic->bmi[i], sizeof(mic->bmi[i]));
1.1537 +
1.1538 + cost += thismvcost;
1.1539 + return cost;
1.1540 +}
1.1541 +
1.1542 +static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1.1543 + MACROBLOCK *x,
1.1544 + int64_t best_yrd,
1.1545 + int i,
1.1546 + int *labelyrate,
1.1547 + int64_t *distortion, int64_t *sse,
1.1548 + ENTROPY_CONTEXT *ta,
1.1549 + ENTROPY_CONTEXT *tl) {
1.1550 + int k;
1.1551 + MACROBLOCKD *xd = &x->e_mbd;
1.1552 + struct macroblockd_plane *const pd = &xd->plane[0];
1.1553 + struct macroblock_plane *const p = &x->plane[0];
1.1554 + MODE_INFO *const mi = xd->mi_8x8[0];
1.1555 + const BLOCK_SIZE bsize = mi->mbmi.sb_type;
1.1556 + const int width = plane_block_width(bsize, pd);
1.1557 + const int height = plane_block_height(bsize, pd);
1.1558 + int idx, idy;
1.1559 +
1.1560 + uint8_t *const src = raster_block_offset_uint8(BLOCK_8X8, i,
1.1561 + p->src.buf, p->src.stride);
1.1562 + uint8_t *const dst = raster_block_offset_uint8(BLOCK_8X8, i,
1.1563 + pd->dst.buf, pd->dst.stride);
1.1564 + int64_t thisdistortion = 0, thissse = 0;
1.1565 + int thisrate = 0, ref;
1.1566 + const int is_compound = has_second_ref(&mi->mbmi);
1.1567 + for (ref = 0; ref < 1 + is_compound; ++ref) {
1.1568 + const uint8_t *pre = raster_block_offset_uint8(BLOCK_8X8, i,
1.1569 + pd->pre[ref].buf, pd->pre[ref].stride);
1.1570 + vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1.1571 + dst, pd->dst.stride,
1.1572 + &mi->bmi[i].as_mv[ref].as_mv,
1.1573 + &xd->scale_factor[ref],
1.1574 + width, height, ref, &xd->subpix, MV_PRECISION_Q3);
1.1575 + }
1.1576 +
1.1577 + vp9_subtract_block(height, width,
1.1578 + raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
1.1579 + src, p->src.stride,
1.1580 + dst, pd->dst.stride);
1.1581 +
1.1582 + k = i;
1.1583 + for (idy = 0; idy < height / 4; ++idy) {
1.1584 + for (idx = 0; idx < width / 4; ++idx) {
1.1585 + int64_t ssz, rd, rd1, rd2;
1.1586 + int16_t* coeff;
1.1587 +
1.1588 + k += (idy * 2 + idx);
1.1589 + coeff = BLOCK_OFFSET(p->coeff, k);
1.1590 + x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1.1591 + coeff, 8);
1.1592 + vp9_regular_quantize_b_4x4(x, 4, k, get_scan_4x4(DCT_DCT),
1.1593 + get_iscan_4x4(DCT_DCT));
1.1594 + thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1.1595 + 16, &ssz);
1.1596 + thissse += ssz;
1.1597 + thisrate += cost_coeffs(x, 0, k,
1.1598 + ta + (k & 1),
1.1599 + tl + (k >> 1), TX_4X4,
1.1600 + vp9_default_scan_4x4,
1.1601 + vp9_default_scan_4x4_neighbors);
1.1602 + rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1.1603 + rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1.1604 + rd = MIN(rd1, rd2);
1.1605 + if (rd >= best_yrd)
1.1606 + return INT64_MAX;
1.1607 + }
1.1608 + }
1.1609 +
1.1610 + *distortion = thisdistortion >> 2;
1.1611 + *labelyrate = thisrate;
1.1612 + *sse = thissse >> 2;
1.1613 +
1.1614 + return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1.1615 +}
1.1616 +
1.1617 +typedef struct {
1.1618 + int eobs;
1.1619 + int brate;
1.1620 + int byrate;
1.1621 + int64_t bdist;
1.1622 + int64_t bsse;
1.1623 + int64_t brdcost;
1.1624 + int_mv mvs[2];
1.1625 + ENTROPY_CONTEXT ta[2];
1.1626 + ENTROPY_CONTEXT tl[2];
1.1627 +} SEG_RDSTAT;
1.1628 +
1.1629 +typedef struct {
1.1630 + int_mv *ref_mv, *second_ref_mv;
1.1631 + int_mv mvp;
1.1632 +
1.1633 + int64_t segment_rd;
1.1634 + int r;
1.1635 + int64_t d;
1.1636 + int64_t sse;
1.1637 + int segment_yrate;
1.1638 + MB_PREDICTION_MODE modes[4];
1.1639 + SEG_RDSTAT rdstat[4][INTER_MODES];
1.1640 + int mvthresh;
1.1641 +} BEST_SEG_INFO;
1.1642 +
1.1643 +static INLINE int mv_check_bounds(MACROBLOCK *x, int_mv *mv) {
1.1644 + int r = 0;
1.1645 + r |= (mv->as_mv.row >> 3) < x->mv_row_min;
1.1646 + r |= (mv->as_mv.row >> 3) > x->mv_row_max;
1.1647 + r |= (mv->as_mv.col >> 3) < x->mv_col_min;
1.1648 + r |= (mv->as_mv.col >> 3) > x->mv_col_max;
1.1649 + return r;
1.1650 +}
1.1651 +
1.1652 +static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1.1653 + MB_MODE_INFO *const mbmi = &x->e_mbd.mi_8x8[0]->mbmi;
1.1654 + struct macroblock_plane *const p = &x->plane[0];
1.1655 + struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1.1656 +
1.1657 + p->src.buf = raster_block_offset_uint8(BLOCK_8X8, i, p->src.buf,
1.1658 + p->src.stride);
1.1659 + assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1.1660 + pd->pre[0].buf = raster_block_offset_uint8(BLOCK_8X8, i, pd->pre[0].buf,
1.1661 + pd->pre[0].stride);
1.1662 + if (has_second_ref(mbmi))
1.1663 + pd->pre[1].buf = raster_block_offset_uint8(BLOCK_8X8, i, pd->pre[1].buf,
1.1664 + pd->pre[1].stride);
1.1665 +}
1.1666 +
1.1667 +static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1.1668 + struct buf_2d orig_pre[2]) {
1.1669 + MB_MODE_INFO *mbmi = &x->e_mbd.mi_8x8[0]->mbmi;
1.1670 + x->plane[0].src = orig_src;
1.1671 + x->e_mbd.plane[0].pre[0] = orig_pre[0];
1.1672 + if (has_second_ref(mbmi))
1.1673 + x->e_mbd.plane[0].pre[1] = orig_pre[1];
1.1674 +}
1.1675 +
1.1676 +static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
1.1677 + const TileInfo *const tile,
1.1678 + BEST_SEG_INFO *bsi_buf, int filter_idx,
1.1679 + int_mv seg_mvs[4][MAX_REF_FRAMES],
1.1680 + int mi_row, int mi_col) {
1.1681 + int i, br = 0, idx, idy;
1.1682 + int64_t bd = 0, block_sse = 0;
1.1683 + MB_PREDICTION_MODE this_mode;
1.1684 + MODE_INFO *mi = x->e_mbd.mi_8x8[0];
1.1685 + MB_MODE_INFO *const mbmi = &mi->mbmi;
1.1686 + struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1.1687 + const int label_count = 4;
1.1688 + int64_t this_segment_rd = 0;
1.1689 + int label_mv_thresh;
1.1690 + int segmentyrate = 0;
1.1691 + const BLOCK_SIZE bsize = mbmi->sb_type;
1.1692 + const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1.1693 + const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1.1694 + vp9_variance_fn_ptr_t *v_fn_ptr;
1.1695 + ENTROPY_CONTEXT t_above[2], t_left[2];
1.1696 + BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1.1697 + int mode_idx;
1.1698 + int subpelmv = 1, have_ref = 0;
1.1699 + const int has_second_rf = has_second_ref(mbmi);
1.1700 +
1.1701 + vpx_memcpy(t_above, pd->above_context, sizeof(t_above));
1.1702 + vpx_memcpy(t_left, pd->left_context, sizeof(t_left));
1.1703 +
1.1704 + v_fn_ptr = &cpi->fn_ptr[bsize];
1.1705 +
1.1706 + // 64 makes this threshold really big effectively
1.1707 + // making it so that we very rarely check mvs on
1.1708 + // segments. setting this to 1 would make mv thresh
1.1709 + // roughly equal to what it is for macroblocks
1.1710 + label_mv_thresh = 1 * bsi->mvthresh / label_count;
1.1711 +
1.1712 + // Segmentation method overheads
1.1713 + for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1.1714 + for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1.1715 + // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1.1716 + // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1.1717 + int_mv mode_mv[MB_MODE_COUNT], second_mode_mv[MB_MODE_COUNT];
1.1718 + int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1.1719 + MB_PREDICTION_MODE mode_selected = ZEROMV;
1.1720 + int64_t best_rd = INT64_MAX;
1.1721 + i = idy * 2 + idx;
1.1722 +
1.1723 + frame_mv[ZEROMV][mbmi->ref_frame[0]].as_int = 0;
1.1724 + vp9_append_sub8x8_mvs_for_idx(&cpi->common, &x->e_mbd, tile,
1.1725 + &frame_mv[NEARESTMV][mbmi->ref_frame[0]],
1.1726 + &frame_mv[NEARMV][mbmi->ref_frame[0]],
1.1727 + i, 0, mi_row, mi_col);
1.1728 + if (has_second_rf) {
1.1729 + frame_mv[ZEROMV][mbmi->ref_frame[1]].as_int = 0;
1.1730 + vp9_append_sub8x8_mvs_for_idx(&cpi->common, &x->e_mbd, tile,
1.1731 + &frame_mv[NEARESTMV][mbmi->ref_frame[1]],
1.1732 + &frame_mv[NEARMV][mbmi->ref_frame[1]],
1.1733 + i, 1, mi_row, mi_col);
1.1734 + }
1.1735 + // search for the best motion vector on this segment
1.1736 + for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1.1737 + const struct buf_2d orig_src = x->plane[0].src;
1.1738 + struct buf_2d orig_pre[2];
1.1739 +
1.1740 + mode_idx = INTER_OFFSET(this_mode);
1.1741 + bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1.1742 +
1.1743 + // if we're near/nearest and mv == 0,0, compare to zeromv
1.1744 + if ((this_mode == NEARMV || this_mode == NEARESTMV ||
1.1745 + this_mode == ZEROMV) &&
1.1746 + frame_mv[this_mode][mbmi->ref_frame[0]].as_int == 0 &&
1.1747 + (!has_second_rf ||
1.1748 + frame_mv[this_mode][mbmi->ref_frame[1]].as_int == 0)) {
1.1749 + int rfc = mbmi->mode_context[mbmi->ref_frame[0]];
1.1750 + int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1.1751 + int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1.1752 + int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1.1753 +
1.1754 + if (this_mode == NEARMV) {
1.1755 + if (c1 > c3)
1.1756 + continue;
1.1757 + } else if (this_mode == NEARESTMV) {
1.1758 + if (c2 > c3)
1.1759 + continue;
1.1760 + } else {
1.1761 + assert(this_mode == ZEROMV);
1.1762 + if (!has_second_rf) {
1.1763 + if ((c3 >= c2 &&
1.1764 + frame_mv[NEARESTMV][mbmi->ref_frame[0]].as_int == 0) ||
1.1765 + (c3 >= c1 &&
1.1766 + frame_mv[NEARMV][mbmi->ref_frame[0]].as_int == 0))
1.1767 + continue;
1.1768 + } else {
1.1769 + if ((c3 >= c2 &&
1.1770 + frame_mv[NEARESTMV][mbmi->ref_frame[0]].as_int == 0 &&
1.1771 + frame_mv[NEARESTMV][mbmi->ref_frame[1]].as_int == 0) ||
1.1772 + (c3 >= c1 &&
1.1773 + frame_mv[NEARMV][mbmi->ref_frame[0]].as_int == 0 &&
1.1774 + frame_mv[NEARMV][mbmi->ref_frame[1]].as_int == 0))
1.1775 + continue;
1.1776 + }
1.1777 + }
1.1778 + }
1.1779 +
1.1780 + vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1.1781 + vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1.1782 + sizeof(bsi->rdstat[i][mode_idx].ta));
1.1783 + vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1.1784 + sizeof(bsi->rdstat[i][mode_idx].tl));
1.1785 +
1.1786 + // motion search for newmv (single predictor case only)
1.1787 + if (!has_second_rf && this_mode == NEWMV &&
1.1788 + seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1.1789 + int step_param = 0;
1.1790 + int further_steps;
1.1791 + int thissme, bestsme = INT_MAX;
1.1792 + int sadpb = x->sadperbit4;
1.1793 + int_mv mvp_full;
1.1794 + int max_mv;
1.1795 +
1.1796 + /* Is the best so far sufficiently good that we cant justify doing
1.1797 + * and new motion search. */
1.1798 + if (best_rd < label_mv_thresh)
1.1799 + break;
1.1800 +
1.1801 + if (cpi->compressor_speed) {
1.1802 + // use previous block's result as next block's MV predictor.
1.1803 + if (i > 0) {
1.1804 + bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1.1805 + if (i == 2)
1.1806 + bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1.1807 + }
1.1808 + }
1.1809 + if (i == 0)
1.1810 + max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1.1811 + else
1.1812 + max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1.1813 +
1.1814 + if (cpi->sf.auto_mv_step_size && cpi->common.show_frame) {
1.1815 + // Take wtd average of the step_params based on the last frame's
1.1816 + // max mv magnitude and the best ref mvs of the current block for
1.1817 + // the given reference.
1.1818 + step_param = (vp9_init_search_range(cpi, max_mv) +
1.1819 + cpi->mv_step_param) >> 1;
1.1820 + } else {
1.1821 + step_param = cpi->mv_step_param;
1.1822 + }
1.1823 +
1.1824 + mvp_full.as_mv.row = bsi->mvp.as_mv.row >> 3;
1.1825 + mvp_full.as_mv.col = bsi->mvp.as_mv.col >> 3;
1.1826 +
1.1827 + if (cpi->sf.adaptive_motion_search && cpi->common.show_frame) {
1.1828 + mvp_full.as_mv.row = x->pred_mv[mbmi->ref_frame[0]].as_mv.row >> 3;
1.1829 + mvp_full.as_mv.col = x->pred_mv[mbmi->ref_frame[0]].as_mv.col >> 3;
1.1830 + step_param = MAX(step_param, 8);
1.1831 + }
1.1832 +
1.1833 + further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
1.1834 + // adjust src pointer for this block
1.1835 + mi_buf_shift(x, i);
1.1836 + if (cpi->sf.search_method == HEX) {
1.1837 + bestsme = vp9_hex_search(x, &mvp_full.as_mv,
1.1838 + step_param,
1.1839 + sadpb, 1, v_fn_ptr, 1,
1.1840 + &bsi->ref_mv->as_mv,
1.1841 + &mode_mv[NEWMV].as_mv);
1.1842 + } else if (cpi->sf.search_method == SQUARE) {
1.1843 + bestsme = vp9_square_search(x, &mvp_full.as_mv,
1.1844 + step_param,
1.1845 + sadpb, 1, v_fn_ptr, 1,
1.1846 + &bsi->ref_mv->as_mv,
1.1847 + &mode_mv[NEWMV].as_mv);
1.1848 + } else if (cpi->sf.search_method == BIGDIA) {
1.1849 + bestsme = vp9_bigdia_search(x, &mvp_full.as_mv,
1.1850 + step_param,
1.1851 + sadpb, 1, v_fn_ptr, 1,
1.1852 + &bsi->ref_mv->as_mv,
1.1853 + &mode_mv[NEWMV].as_mv);
1.1854 + } else {
1.1855 + bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
1.1856 + sadpb, further_steps, 0, v_fn_ptr,
1.1857 + bsi->ref_mv, &mode_mv[NEWMV]);
1.1858 + }
1.1859 +
1.1860 + // Should we do a full search (best quality only)
1.1861 + if (cpi->compressor_speed == 0) {
1.1862 + /* Check if mvp_full is within the range. */
1.1863 + clamp_mv(&mvp_full.as_mv, x->mv_col_min, x->mv_col_max,
1.1864 + x->mv_row_min, x->mv_row_max);
1.1865 +
1.1866 + thissme = cpi->full_search_sad(x, &mvp_full,
1.1867 + sadpb, 16, v_fn_ptr,
1.1868 + x->nmvjointcost, x->mvcost,
1.1869 + bsi->ref_mv, i);
1.1870 +
1.1871 + if (thissme < bestsme) {
1.1872 + bestsme = thissme;
1.1873 + mode_mv[NEWMV].as_int = mi->bmi[i].as_mv[0].as_int;
1.1874 + } else {
1.1875 + /* The full search result is actually worse so re-instate the
1.1876 + * previous best vector */
1.1877 + mi->bmi[i].as_mv[0].as_int = mode_mv[NEWMV].as_int;
1.1878 + }
1.1879 + }
1.1880 +
1.1881 + if (bestsme < INT_MAX) {
1.1882 + int distortion;
1.1883 + unsigned int sse;
1.1884 + cpi->find_fractional_mv_step(x,
1.1885 + &mode_mv[NEWMV].as_mv,
1.1886 + &bsi->ref_mv->as_mv,
1.1887 + cpi->common.allow_high_precision_mv,
1.1888 + x->errorperbit, v_fn_ptr,
1.1889 + 0, cpi->sf.subpel_iters_per_step,
1.1890 + x->nmvjointcost, x->mvcost,
1.1891 + &distortion, &sse);
1.1892 +
1.1893 + // save motion search result for use in compound prediction
1.1894 + seg_mvs[i][mbmi->ref_frame[0]].as_int = mode_mv[NEWMV].as_int;
1.1895 + }
1.1896 +
1.1897 + if (cpi->sf.adaptive_motion_search)
1.1898 + x->pred_mv[mbmi->ref_frame[0]].as_int = mode_mv[NEWMV].as_int;
1.1899 +
1.1900 + // restore src pointers
1.1901 + mi_buf_restore(x, orig_src, orig_pre);
1.1902 + }
1.1903 +
1.1904 + if (has_second_rf) {
1.1905 + if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1.1906 + seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1.1907 + continue;
1.1908 + }
1.1909 +
1.1910 + if (has_second_rf && this_mode == NEWMV &&
1.1911 + mbmi->interp_filter == EIGHTTAP) {
1.1912 + // adjust src pointers
1.1913 + mi_buf_shift(x, i);
1.1914 + if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1.1915 + int rate_mv;
1.1916 + joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1.1917 + mi_row, mi_col, seg_mvs[i],
1.1918 + &rate_mv);
1.1919 + seg_mvs[i][mbmi->ref_frame[0]].as_int =
1.1920 + frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1.1921 + seg_mvs[i][mbmi->ref_frame[1]].as_int =
1.1922 + frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1.1923 + }
1.1924 + // restore src pointers
1.1925 + mi_buf_restore(x, orig_src, orig_pre);
1.1926 + }
1.1927 +
1.1928 + bsi->rdstat[i][mode_idx].brate =
1.1929 + labels2mode(x, i, this_mode, &mode_mv[this_mode],
1.1930 + &second_mode_mv[this_mode], frame_mv, seg_mvs[i],
1.1931 + bsi->ref_mv, bsi->second_ref_mv, x->nmvjointcost,
1.1932 + x->mvcost, cpi);
1.1933 +
1.1934 +
1.1935 + bsi->rdstat[i][mode_idx].mvs[0].as_int = mode_mv[this_mode].as_int;
1.1936 + if (num_4x4_blocks_wide > 1)
1.1937 + bsi->rdstat[i + 1][mode_idx].mvs[0].as_int =
1.1938 + mode_mv[this_mode].as_int;
1.1939 + if (num_4x4_blocks_high > 1)
1.1940 + bsi->rdstat[i + 2][mode_idx].mvs[0].as_int =
1.1941 + mode_mv[this_mode].as_int;
1.1942 + if (has_second_rf) {
1.1943 + bsi->rdstat[i][mode_idx].mvs[1].as_int =
1.1944 + second_mode_mv[this_mode].as_int;
1.1945 + if (num_4x4_blocks_wide > 1)
1.1946 + bsi->rdstat[i + 1][mode_idx].mvs[1].as_int =
1.1947 + second_mode_mv[this_mode].as_int;
1.1948 + if (num_4x4_blocks_high > 1)
1.1949 + bsi->rdstat[i + 2][mode_idx].mvs[1].as_int =
1.1950 + second_mode_mv[this_mode].as_int;
1.1951 + }
1.1952 +
1.1953 + // Trap vectors that reach beyond the UMV borders
1.1954 + if (mv_check_bounds(x, &mode_mv[this_mode]))
1.1955 + continue;
1.1956 + if (has_second_rf &&
1.1957 + mv_check_bounds(x, &second_mode_mv[this_mode]))
1.1958 + continue;
1.1959 +
1.1960 + if (filter_idx > 0) {
1.1961 + BEST_SEG_INFO *ref_bsi = bsi_buf;
1.1962 + subpelmv = (mode_mv[this_mode].as_mv.row & 0x0f) ||
1.1963 + (mode_mv[this_mode].as_mv.col & 0x0f);
1.1964 + have_ref = mode_mv[this_mode].as_int ==
1.1965 + ref_bsi->rdstat[i][mode_idx].mvs[0].as_int;
1.1966 + if (has_second_rf) {
1.1967 + subpelmv |= (second_mode_mv[this_mode].as_mv.row & 0x0f) ||
1.1968 + (second_mode_mv[this_mode].as_mv.col & 0x0f);
1.1969 + have_ref &= second_mode_mv[this_mode].as_int ==
1.1970 + ref_bsi->rdstat[i][mode_idx].mvs[1].as_int;
1.1971 + }
1.1972 +
1.1973 + if (filter_idx > 1 && !subpelmv && !have_ref) {
1.1974 + ref_bsi = bsi_buf + 1;
1.1975 + have_ref = mode_mv[this_mode].as_int ==
1.1976 + ref_bsi->rdstat[i][mode_idx].mvs[0].as_int;
1.1977 + if (has_second_rf) {
1.1978 + have_ref &= second_mode_mv[this_mode].as_int ==
1.1979 + ref_bsi->rdstat[i][mode_idx].mvs[1].as_int;
1.1980 + }
1.1981 + }
1.1982 +
1.1983 + if (!subpelmv && have_ref &&
1.1984 + ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1.1985 + vpx_memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1.1986 + sizeof(SEG_RDSTAT));
1.1987 + if (num_4x4_blocks_wide > 1)
1.1988 + bsi->rdstat[i + 1][mode_idx].eobs =
1.1989 + ref_bsi->rdstat[i + 1][mode_idx].eobs;
1.1990 + if (num_4x4_blocks_high > 1)
1.1991 + bsi->rdstat[i + 2][mode_idx].eobs =
1.1992 + ref_bsi->rdstat[i + 2][mode_idx].eobs;
1.1993 +
1.1994 + if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1.1995 + mode_selected = this_mode;
1.1996 + best_rd = bsi->rdstat[i][mode_idx].brdcost;
1.1997 + }
1.1998 + continue;
1.1999 + }
1.2000 + }
1.2001 +
1.2002 + bsi->rdstat[i][mode_idx].brdcost =
1.2003 + encode_inter_mb_segment(cpi, x,
1.2004 + bsi->segment_rd - this_segment_rd, i,
1.2005 + &bsi->rdstat[i][mode_idx].byrate,
1.2006 + &bsi->rdstat[i][mode_idx].bdist,
1.2007 + &bsi->rdstat[i][mode_idx].bsse,
1.2008 + bsi->rdstat[i][mode_idx].ta,
1.2009 + bsi->rdstat[i][mode_idx].tl);
1.2010 + if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1.2011 + bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1.2012 + bsi->rdstat[i][mode_idx].brate, 0);
1.2013 + bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1.2014 + bsi->rdstat[i][mode_idx].eobs = pd->eobs[i];
1.2015 + if (num_4x4_blocks_wide > 1)
1.2016 + bsi->rdstat[i + 1][mode_idx].eobs = pd->eobs[i + 1];
1.2017 + if (num_4x4_blocks_high > 1)
1.2018 + bsi->rdstat[i + 2][mode_idx].eobs = pd->eobs[i + 2];
1.2019 + }
1.2020 +
1.2021 + if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1.2022 + mode_selected = this_mode;
1.2023 + best_rd = bsi->rdstat[i][mode_idx].brdcost;
1.2024 + }
1.2025 + } /*for each 4x4 mode*/
1.2026 +
1.2027 + if (best_rd == INT64_MAX) {
1.2028 + int iy, midx;
1.2029 + for (iy = i + 1; iy < 4; ++iy)
1.2030 + for (midx = 0; midx < INTER_MODES; ++midx)
1.2031 + bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1.2032 + bsi->segment_rd = INT64_MAX;
1.2033 + return;
1.2034 + }
1.2035 +
1.2036 + mode_idx = INTER_OFFSET(mode_selected);
1.2037 + vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
1.2038 + vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
1.2039 +
1.2040 + labels2mode(x, i, mode_selected, &mode_mv[mode_selected],
1.2041 + &second_mode_mv[mode_selected], frame_mv, seg_mvs[i],
1.2042 + bsi->ref_mv, bsi->second_ref_mv, x->nmvjointcost,
1.2043 + x->mvcost, cpi);
1.2044 +
1.2045 + br += bsi->rdstat[i][mode_idx].brate;
1.2046 + bd += bsi->rdstat[i][mode_idx].bdist;
1.2047 + block_sse += bsi->rdstat[i][mode_idx].bsse;
1.2048 + segmentyrate += bsi->rdstat[i][mode_idx].byrate;
1.2049 + this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
1.2050 +
1.2051 + if (this_segment_rd > bsi->segment_rd) {
1.2052 + int iy, midx;
1.2053 + for (iy = i + 1; iy < 4; ++iy)
1.2054 + for (midx = 0; midx < INTER_MODES; ++midx)
1.2055 + bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1.2056 + bsi->segment_rd = INT64_MAX;
1.2057 + return;
1.2058 + }
1.2059 + }
1.2060 + } /* for each label */
1.2061 +
1.2062 + bsi->r = br;
1.2063 + bsi->d = bd;
1.2064 + bsi->segment_yrate = segmentyrate;
1.2065 + bsi->segment_rd = this_segment_rd;
1.2066 + bsi->sse = block_sse;
1.2067 +
1.2068 + // update the coding decisions
1.2069 + for (i = 0; i < 4; ++i)
1.2070 + bsi->modes[i] = mi->bmi[i].as_mode;
1.2071 +}
1.2072 +
1.2073 +static int64_t rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x,
1.2074 + const TileInfo *const tile,
1.2075 + int_mv *best_ref_mv,
1.2076 + int_mv *second_best_ref_mv,
1.2077 + int64_t best_rd,
1.2078 + int *returntotrate,
1.2079 + int *returnyrate,
1.2080 + int64_t *returndistortion,
1.2081 + int *skippable, int64_t *psse,
1.2082 + int mvthresh,
1.2083 + int_mv seg_mvs[4][MAX_REF_FRAMES],
1.2084 + BEST_SEG_INFO *bsi_buf,
1.2085 + int filter_idx,
1.2086 + int mi_row, int mi_col) {
1.2087 + int i;
1.2088 + BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1.2089 + MACROBLOCKD *xd = &x->e_mbd;
1.2090 + MODE_INFO *mi = xd->mi_8x8[0];
1.2091 + MB_MODE_INFO *mbmi = &mi->mbmi;
1.2092 + int mode_idx;
1.2093 +
1.2094 + vp9_zero(*bsi);
1.2095 +
1.2096 + bsi->segment_rd = best_rd;
1.2097 + bsi->ref_mv = best_ref_mv;
1.2098 + bsi->second_ref_mv = second_best_ref_mv;
1.2099 + bsi->mvp.as_int = best_ref_mv->as_int;
1.2100 + bsi->mvthresh = mvthresh;
1.2101 +
1.2102 + for (i = 0; i < 4; i++)
1.2103 + bsi->modes[i] = ZEROMV;
1.2104 +
1.2105 + rd_check_segment_txsize(cpi, x, tile, bsi_buf, filter_idx, seg_mvs,
1.2106 + mi_row, mi_col);
1.2107 +
1.2108 + if (bsi->segment_rd > best_rd)
1.2109 + return INT64_MAX;
1.2110 + /* set it to the best */
1.2111 + for (i = 0; i < 4; i++) {
1.2112 + mode_idx = INTER_OFFSET(bsi->modes[i]);
1.2113 + mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
1.2114 + if (has_second_ref(mbmi))
1.2115 + mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
1.2116 + xd->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
1.2117 + mi->bmi[i].as_mode = bsi->modes[i];
1.2118 + }
1.2119 +
1.2120 + /*
1.2121 + * used to set mbmi->mv.as_int
1.2122 + */
1.2123 + *returntotrate = bsi->r;
1.2124 + *returndistortion = bsi->d;
1.2125 + *returnyrate = bsi->segment_yrate;
1.2126 + *skippable = vp9_is_skippable_in_plane(&x->e_mbd, BLOCK_8X8, 0);
1.2127 + *psse = bsi->sse;
1.2128 + mbmi->mode = bsi->modes[3];
1.2129 +
1.2130 + return bsi->segment_rd;
1.2131 +}
1.2132 +
1.2133 +static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
1.2134 + uint8_t *ref_y_buffer, int ref_y_stride,
1.2135 + int ref_frame, BLOCK_SIZE block_size ) {
1.2136 + MACROBLOCKD *xd = &x->e_mbd;
1.2137 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.2138 + int_mv this_mv;
1.2139 + int i;
1.2140 + int zero_seen = 0;
1.2141 + int best_index = 0;
1.2142 + int best_sad = INT_MAX;
1.2143 + int this_sad = INT_MAX;
1.2144 + unsigned int max_mv = 0;
1.2145 +
1.2146 + uint8_t *src_y_ptr = x->plane[0].src.buf;
1.2147 + uint8_t *ref_y_ptr;
1.2148 + int row_offset, col_offset;
1.2149 + int num_mv_refs = MAX_MV_REF_CANDIDATES +
1.2150 + (cpi->sf.adaptive_motion_search &&
1.2151 + cpi->common.show_frame &&
1.2152 + block_size < cpi->sf.max_partition_size);
1.2153 +
1.2154 + // Get the sad for each candidate reference mv
1.2155 + for (i = 0; i < num_mv_refs; i++) {
1.2156 + this_mv.as_int = (i < MAX_MV_REF_CANDIDATES) ?
1.2157 + mbmi->ref_mvs[ref_frame][i].as_int : x->pred_mv[ref_frame].as_int;
1.2158 +
1.2159 + max_mv = MAX(max_mv,
1.2160 + MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3);
1.2161 + // The list is at an end if we see 0 for a second time.
1.2162 + if (!this_mv.as_int && zero_seen)
1.2163 + break;
1.2164 + zero_seen = zero_seen || !this_mv.as_int;
1.2165 +
1.2166 + row_offset = this_mv.as_mv.row >> 3;
1.2167 + col_offset = this_mv.as_mv.col >> 3;
1.2168 + ref_y_ptr = ref_y_buffer + (ref_y_stride * row_offset) + col_offset;
1.2169 +
1.2170 + // Find sad for current vector.
1.2171 + this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
1.2172 + ref_y_ptr, ref_y_stride,
1.2173 + 0x7fffffff);
1.2174 +
1.2175 + // Note if it is the best so far.
1.2176 + if (this_sad < best_sad) {
1.2177 + best_sad = this_sad;
1.2178 + best_index = i;
1.2179 + }
1.2180 + }
1.2181 +
1.2182 + // Note the index of the mv that worked best in the reference list.
1.2183 + x->mv_best_ref_index[ref_frame] = best_index;
1.2184 + x->max_mv_context[ref_frame] = max_mv;
1.2185 +}
1.2186 +
1.2187 +static void estimate_ref_frame_costs(VP9_COMP *cpi, int segment_id,
1.2188 + unsigned int *ref_costs_single,
1.2189 + unsigned int *ref_costs_comp,
1.2190 + vp9_prob *comp_mode_p) {
1.2191 + VP9_COMMON *const cm = &cpi->common;
1.2192 + MACROBLOCKD *const xd = &cpi->mb.e_mbd;
1.2193 + int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
1.2194 + SEG_LVL_REF_FRAME);
1.2195 + if (seg_ref_active) {
1.2196 + vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
1.2197 + vpx_memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
1.2198 + *comp_mode_p = 128;
1.2199 + } else {
1.2200 + vp9_prob intra_inter_p = vp9_get_pred_prob_intra_inter(cm, xd);
1.2201 + vp9_prob comp_inter_p = 128;
1.2202 +
1.2203 + if (cm->comp_pred_mode == HYBRID_PREDICTION) {
1.2204 + comp_inter_p = vp9_get_pred_prob_comp_inter_inter(cm, xd);
1.2205 + *comp_mode_p = comp_inter_p;
1.2206 + } else {
1.2207 + *comp_mode_p = 128;
1.2208 + }
1.2209 +
1.2210 + ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1.2211 +
1.2212 + if (cm->comp_pred_mode != COMP_PREDICTION_ONLY) {
1.2213 + vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1.2214 + vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1.2215 + unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1.2216 +
1.2217 + if (cm->comp_pred_mode == HYBRID_PREDICTION)
1.2218 + base_cost += vp9_cost_bit(comp_inter_p, 0);
1.2219 +
1.2220 + ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
1.2221 + ref_costs_single[ALTREF_FRAME] = base_cost;
1.2222 + ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1.2223 + ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1.2224 + ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1.2225 + ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1.2226 + ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1.2227 + } else {
1.2228 + ref_costs_single[LAST_FRAME] = 512;
1.2229 + ref_costs_single[GOLDEN_FRAME] = 512;
1.2230 + ref_costs_single[ALTREF_FRAME] = 512;
1.2231 + }
1.2232 + if (cm->comp_pred_mode != SINGLE_PREDICTION_ONLY) {
1.2233 + vp9_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
1.2234 + unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1.2235 +
1.2236 + if (cm->comp_pred_mode == HYBRID_PREDICTION)
1.2237 + base_cost += vp9_cost_bit(comp_inter_p, 1);
1.2238 +
1.2239 + ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
1.2240 + ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
1.2241 + } else {
1.2242 + ref_costs_comp[LAST_FRAME] = 512;
1.2243 + ref_costs_comp[GOLDEN_FRAME] = 512;
1.2244 + }
1.2245 + }
1.2246 +}
1.2247 +
1.2248 +static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
1.2249 + int mode_index,
1.2250 + int_mv *ref_mv,
1.2251 + int_mv *second_ref_mv,
1.2252 + int64_t comp_pred_diff[NB_PREDICTION_TYPES],
1.2253 + int64_t tx_size_diff[TX_MODES],
1.2254 + int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS]) {
1.2255 + MACROBLOCKD *const xd = &x->e_mbd;
1.2256 +
1.2257 + // Take a snapshot of the coding context so it can be
1.2258 + // restored if we decide to encode this way
1.2259 + ctx->skip = x->skip;
1.2260 + ctx->best_mode_index = mode_index;
1.2261 + ctx->mic = *xd->mi_8x8[0];
1.2262 +
1.2263 + ctx->best_ref_mv.as_int = ref_mv->as_int;
1.2264 + ctx->second_best_ref_mv.as_int = second_ref_mv->as_int;
1.2265 +
1.2266 + ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_PREDICTION_ONLY];
1.2267 + ctx->comp_pred_diff = (int)comp_pred_diff[COMP_PREDICTION_ONLY];
1.2268 + ctx->hybrid_pred_diff = (int)comp_pred_diff[HYBRID_PREDICTION];
1.2269 +
1.2270 + vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
1.2271 + vpx_memcpy(ctx->best_filter_diff, best_filter_diff,
1.2272 + sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
1.2273 +}
1.2274 +
1.2275 +static void setup_pred_block(const MACROBLOCKD *xd,
1.2276 + struct buf_2d dst[MAX_MB_PLANE],
1.2277 + const YV12_BUFFER_CONFIG *src,
1.2278 + int mi_row, int mi_col,
1.2279 + const struct scale_factors *scale,
1.2280 + const struct scale_factors *scale_uv) {
1.2281 + int i;
1.2282 +
1.2283 + dst[0].buf = src->y_buffer;
1.2284 + dst[0].stride = src->y_stride;
1.2285 + dst[1].buf = src->u_buffer;
1.2286 + dst[2].buf = src->v_buffer;
1.2287 + dst[1].stride = dst[2].stride = src->uv_stride;
1.2288 +#if CONFIG_ALPHA
1.2289 + dst[3].buf = src->alpha_buffer;
1.2290 + dst[3].stride = src->alpha_stride;
1.2291 +#endif
1.2292 +
1.2293 + // TODO(jkoleszar): Make scale factors per-plane data
1.2294 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2295 + setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
1.2296 + i ? scale_uv : scale,
1.2297 + xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
1.2298 + }
1.2299 +}
1.2300 +
1.2301 +static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
1.2302 + const TileInfo *const tile,
1.2303 + int idx, MV_REFERENCE_FRAME frame_type,
1.2304 + BLOCK_SIZE block_size,
1.2305 + int mi_row, int mi_col,
1.2306 + int_mv frame_nearest_mv[MAX_REF_FRAMES],
1.2307 + int_mv frame_near_mv[MAX_REF_FRAMES],
1.2308 + struct buf_2d yv12_mb[4][MAX_MB_PLANE],
1.2309 + struct scale_factors scale[MAX_REF_FRAMES]) {
1.2310 + VP9_COMMON *cm = &cpi->common;
1.2311 + YV12_BUFFER_CONFIG *yv12 = &cm->yv12_fb[cpi->common.ref_frame_map[idx]];
1.2312 + MACROBLOCKD *const xd = &x->e_mbd;
1.2313 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.2314 +
1.2315 + // set up scaling factors
1.2316 + scale[frame_type] = cpi->common.active_ref_scale[frame_type - 1];
1.2317 +
1.2318 + scale[frame_type].sfc->set_scaled_offsets(&scale[frame_type],
1.2319 + mi_row * MI_SIZE, mi_col * MI_SIZE);
1.2320 +
1.2321 + // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
1.2322 + // use the UV scaling factors.
1.2323 + setup_pred_block(xd, yv12_mb[frame_type], yv12, mi_row, mi_col,
1.2324 + &scale[frame_type], &scale[frame_type]);
1.2325 +
1.2326 + // Gets an initial list of candidate vectors from neighbours and orders them
1.2327 + vp9_find_mv_refs(cm, xd, tile, xd->mi_8x8[0],
1.2328 + xd->last_mi,
1.2329 + frame_type,
1.2330 + mbmi->ref_mvs[frame_type], mi_row, mi_col);
1.2331 +
1.2332 + // Candidate refinement carried out at encoder and decoder
1.2333 + vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv,
1.2334 + mbmi->ref_mvs[frame_type],
1.2335 + &frame_nearest_mv[frame_type],
1.2336 + &frame_near_mv[frame_type]);
1.2337 +
1.2338 + // Further refinement that is encode side only to test the top few candidates
1.2339 + // in full and choose the best as the centre point for subsequent searches.
1.2340 + // The current implementation doesn't support scaling.
1.2341 + if (!vp9_is_scaled(scale[frame_type].sfc) && block_size >= BLOCK_8X8)
1.2342 + mv_pred(cpi, x, yv12_mb[frame_type][0].buf, yv12->y_stride,
1.2343 + frame_type, block_size);
1.2344 +}
1.2345 +
1.2346 +static YV12_BUFFER_CONFIG *get_scaled_ref_frame(VP9_COMP *cpi, int ref_frame) {
1.2347 + YV12_BUFFER_CONFIG *scaled_ref_frame = NULL;
1.2348 + int fb = get_ref_frame_idx(cpi, ref_frame);
1.2349 + int fb_scale = get_scale_ref_frame_idx(cpi, ref_frame);
1.2350 + if (cpi->scaled_ref_idx[fb_scale] != cpi->common.ref_frame_map[fb])
1.2351 + scaled_ref_frame = &cpi->common.yv12_fb[cpi->scaled_ref_idx[fb_scale]];
1.2352 + return scaled_ref_frame;
1.2353 +}
1.2354 +
1.2355 +static INLINE int get_switchable_rate(const MACROBLOCK *x) {
1.2356 + const MACROBLOCKD *const xd = &x->e_mbd;
1.2357 + const MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.2358 + const int ctx = vp9_get_pred_context_switchable_interp(xd);
1.2359 + return SWITCHABLE_INTERP_RATE_FACTOR *
1.2360 + x->switchable_interp_costs[ctx][mbmi->interp_filter];
1.2361 +}
1.2362 +
1.2363 +static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1.2364 + const TileInfo *const tile,
1.2365 + BLOCK_SIZE bsize,
1.2366 + int mi_row, int mi_col,
1.2367 + int_mv *tmp_mv, int *rate_mv) {
1.2368 + MACROBLOCKD *xd = &x->e_mbd;
1.2369 + VP9_COMMON *cm = &cpi->common;
1.2370 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.2371 + struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
1.2372 + int bestsme = INT_MAX;
1.2373 + int further_steps, step_param;
1.2374 + int sadpb = x->sadperbit16;
1.2375 + int_mv mvp_full;
1.2376 + int ref = mbmi->ref_frame[0];
1.2377 + int_mv ref_mv = mbmi->ref_mvs[ref][0];
1.2378 + const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
1.2379 +
1.2380 + int tmp_col_min = x->mv_col_min;
1.2381 + int tmp_col_max = x->mv_col_max;
1.2382 + int tmp_row_min = x->mv_row_min;
1.2383 + int tmp_row_max = x->mv_row_max;
1.2384 +
1.2385 + YV12_BUFFER_CONFIG *scaled_ref_frame = get_scaled_ref_frame(cpi, ref);
1.2386 +
1.2387 + if (scaled_ref_frame) {
1.2388 + int i;
1.2389 + // Swap out the reference frame for a version that's been scaled to
1.2390 + // match the resolution of the current frame, allowing the existing
1.2391 + // motion search code to be used without additional modifications.
1.2392 + for (i = 0; i < MAX_MB_PLANE; i++)
1.2393 + backup_yv12[i] = xd->plane[i].pre[0];
1.2394 +
1.2395 + setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
1.2396 + }
1.2397 +
1.2398 + vp9_clamp_mv_min_max(x, &ref_mv.as_mv);
1.2399 +
1.2400 + // Adjust search parameters based on small partitions' result.
1.2401 + if (x->fast_ms) {
1.2402 + // && abs(mvp_full.as_mv.row - x->pred_mv.as_mv.row) < 24 &&
1.2403 + // abs(mvp_full.as_mv.col - x->pred_mv.as_mv.col) < 24) {
1.2404 + // adjust search range
1.2405 + step_param = 6;
1.2406 + if (x->fast_ms > 1)
1.2407 + step_param = 8;
1.2408 +
1.2409 + // Get prediction MV.
1.2410 + mvp_full.as_int = x->pred_mv[ref].as_int;
1.2411 +
1.2412 + // Adjust MV sign if needed.
1.2413 + if (cm->ref_frame_sign_bias[ref]) {
1.2414 + mvp_full.as_mv.col *= -1;
1.2415 + mvp_full.as_mv.row *= -1;
1.2416 + }
1.2417 + } else {
1.2418 + // Work out the size of the first step in the mv step search.
1.2419 + // 0 here is maximum length first step. 1 is MAX >> 1 etc.
1.2420 + if (cpi->sf.auto_mv_step_size && cpi->common.show_frame) {
1.2421 + // Take wtd average of the step_params based on the last frame's
1.2422 + // max mv magnitude and that based on the best ref mvs of the current
1.2423 + // block for the given reference.
1.2424 + step_param = (vp9_init_search_range(cpi, x->max_mv_context[ref]) +
1.2425 + cpi->mv_step_param) >> 1;
1.2426 + } else {
1.2427 + step_param = cpi->mv_step_param;
1.2428 + }
1.2429 + }
1.2430 +
1.2431 + if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64 &&
1.2432 + cpi->common.show_frame) {
1.2433 + int boffset = 2 * (b_width_log2(BLOCK_64X64) - MIN(b_height_log2(bsize),
1.2434 + b_width_log2(bsize)));
1.2435 + step_param = MAX(step_param, boffset);
1.2436 + }
1.2437 +
1.2438 + mvp_full.as_int = x->mv_best_ref_index[ref] < MAX_MV_REF_CANDIDATES ?
1.2439 + mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_int :
1.2440 + x->pred_mv[ref].as_int;
1.2441 +
1.2442 + mvp_full.as_mv.col >>= 3;
1.2443 + mvp_full.as_mv.row >>= 3;
1.2444 +
1.2445 + // Further step/diamond searches as necessary
1.2446 + further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
1.2447 +
1.2448 + if (cpi->sf.search_method == HEX) {
1.2449 + bestsme = vp9_hex_search(x, &mvp_full.as_mv,
1.2450 + step_param,
1.2451 + sadpb, 1,
1.2452 + &cpi->fn_ptr[block_size], 1,
1.2453 + &ref_mv.as_mv, &tmp_mv->as_mv);
1.2454 + } else if (cpi->sf.search_method == SQUARE) {
1.2455 + bestsme = vp9_square_search(x, &mvp_full.as_mv,
1.2456 + step_param,
1.2457 + sadpb, 1,
1.2458 + &cpi->fn_ptr[block_size], 1,
1.2459 + &ref_mv.as_mv, &tmp_mv->as_mv);
1.2460 + } else if (cpi->sf.search_method == BIGDIA) {
1.2461 + bestsme = vp9_bigdia_search(x, &mvp_full.as_mv,
1.2462 + step_param,
1.2463 + sadpb, 1,
1.2464 + &cpi->fn_ptr[block_size], 1,
1.2465 + &ref_mv.as_mv, &tmp_mv->as_mv);
1.2466 + } else {
1.2467 + bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
1.2468 + sadpb, further_steps, 1,
1.2469 + &cpi->fn_ptr[block_size],
1.2470 + &ref_mv, tmp_mv);
1.2471 + }
1.2472 +
1.2473 + x->mv_col_min = tmp_col_min;
1.2474 + x->mv_col_max = tmp_col_max;
1.2475 + x->mv_row_min = tmp_row_min;
1.2476 + x->mv_row_max = tmp_row_max;
1.2477 +
1.2478 + if (bestsme < INT_MAX) {
1.2479 + int dis; /* TODO: use dis in distortion calculation later. */
1.2480 + unsigned int sse;
1.2481 + cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv.as_mv,
1.2482 + cm->allow_high_precision_mv,
1.2483 + x->errorperbit,
1.2484 + &cpi->fn_ptr[block_size],
1.2485 + 0, cpi->sf.subpel_iters_per_step,
1.2486 + x->nmvjointcost, x->mvcost,
1.2487 + &dis, &sse);
1.2488 + }
1.2489 + *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv.as_mv,
1.2490 + x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1.2491 +
1.2492 + if (cpi->sf.adaptive_motion_search && cpi->common.show_frame)
1.2493 + x->pred_mv[ref].as_int = tmp_mv->as_int;
1.2494 +
1.2495 + if (scaled_ref_frame) {
1.2496 + int i;
1.2497 + for (i = 0; i < MAX_MB_PLANE; i++)
1.2498 + xd->plane[i].pre[0] = backup_yv12[i];
1.2499 + }
1.2500 +}
1.2501 +
1.2502 +static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1.2503 + BLOCK_SIZE bsize,
1.2504 + int_mv *frame_mv,
1.2505 + int mi_row, int mi_col,
1.2506 + int_mv single_newmv[MAX_REF_FRAMES],
1.2507 + int *rate_mv) {
1.2508 + int pw = 4 << b_width_log2(bsize), ph = 4 << b_height_log2(bsize);
1.2509 + MACROBLOCKD *xd = &x->e_mbd;
1.2510 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.2511 + const int refs[2] = { mbmi->ref_frame[0],
1.2512 + mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
1.2513 + int_mv ref_mv[2];
1.2514 + const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
1.2515 + int ite, ref;
1.2516 + // Prediction buffer from second frame.
1.2517 + uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
1.2518 +
1.2519 + // Do joint motion search in compound mode to get more accurate mv.
1.2520 + struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1.2521 + struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0];
1.2522 + int last_besterr[2] = {INT_MAX, INT_MAX};
1.2523 + YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1.2524 + get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
1.2525 + get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
1.2526 + };
1.2527 +
1.2528 + for (ref = 0; ref < 2; ++ref) {
1.2529 + ref_mv[ref] = mbmi->ref_mvs[refs[ref]][0];
1.2530 +
1.2531 + if (scaled_ref_frame[ref]) {
1.2532 + int i;
1.2533 + // Swap out the reference frame for a version that's been scaled to
1.2534 + // match the resolution of the current frame, allowing the existing
1.2535 + // motion search code to be used without additional modifications.
1.2536 + for (i = 0; i < MAX_MB_PLANE; i++)
1.2537 + backup_yv12[ref][i] = xd->plane[i].pre[ref];
1.2538 + setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col, NULL);
1.2539 + }
1.2540 +
1.2541 + xd->scale_factor[ref].sfc->set_scaled_offsets(&xd->scale_factor[ref],
1.2542 + mi_row, mi_col);
1.2543 + frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1.2544 + }
1.2545 +
1.2546 + // Allow joint search multiple times iteratively for each ref frame
1.2547 + // and break out the search loop if it couldn't find better mv.
1.2548 + for (ite = 0; ite < 4; ite++) {
1.2549 + struct buf_2d ref_yv12[2];
1.2550 + int bestsme = INT_MAX;
1.2551 + int sadpb = x->sadperbit16;
1.2552 + int_mv tmp_mv;
1.2553 + int search_range = 3;
1.2554 +
1.2555 + int tmp_col_min = x->mv_col_min;
1.2556 + int tmp_col_max = x->mv_col_max;
1.2557 + int tmp_row_min = x->mv_row_min;
1.2558 + int tmp_row_max = x->mv_row_max;
1.2559 + int id = ite % 2;
1.2560 +
1.2561 + // Initialized here because of compiler problem in Visual Studio.
1.2562 + ref_yv12[0] = xd->plane[0].pre[0];
1.2563 + ref_yv12[1] = xd->plane[0].pre[1];
1.2564 +
1.2565 + // Get pred block from second frame.
1.2566 + vp9_build_inter_predictor(ref_yv12[!id].buf,
1.2567 + ref_yv12[!id].stride,
1.2568 + second_pred, pw,
1.2569 + &frame_mv[refs[!id]].as_mv,
1.2570 + &xd->scale_factor[!id],
1.2571 + pw, ph, 0,
1.2572 + &xd->subpix, MV_PRECISION_Q3);
1.2573 +
1.2574 + // Compound motion search on first ref frame.
1.2575 + if (id)
1.2576 + xd->plane[0].pre[0] = ref_yv12[id];
1.2577 + vp9_clamp_mv_min_max(x, &ref_mv[id].as_mv);
1.2578 +
1.2579 + // Use mv result from single mode as mvp.
1.2580 + tmp_mv.as_int = frame_mv[refs[id]].as_int;
1.2581 +
1.2582 + tmp_mv.as_mv.col >>= 3;
1.2583 + tmp_mv.as_mv.row >>= 3;
1.2584 +
1.2585 + // Small-range full-pixel motion search
1.2586 + bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
1.2587 + search_range,
1.2588 + &cpi->fn_ptr[block_size],
1.2589 + x->nmvjointcost, x->mvcost,
1.2590 + &ref_mv[id], second_pred,
1.2591 + pw, ph);
1.2592 +
1.2593 + x->mv_col_min = tmp_col_min;
1.2594 + x->mv_col_max = tmp_col_max;
1.2595 + x->mv_row_min = tmp_row_min;
1.2596 + x->mv_row_max = tmp_row_max;
1.2597 +
1.2598 + if (bestsme < INT_MAX) {
1.2599 + int dis; /* TODO: use dis in distortion calculation later. */
1.2600 + unsigned int sse;
1.2601 +
1.2602 + bestsme = cpi->find_fractional_mv_step_comp(
1.2603 + x, &tmp_mv.as_mv,
1.2604 + &ref_mv[id].as_mv,
1.2605 + cpi->common.allow_high_precision_mv,
1.2606 + x->errorperbit,
1.2607 + &cpi->fn_ptr[block_size],
1.2608 + 0, cpi->sf.subpel_iters_per_step,
1.2609 + x->nmvjointcost, x->mvcost,
1.2610 + &dis, &sse, second_pred,
1.2611 + pw, ph);
1.2612 + }
1.2613 +
1.2614 + if (id)
1.2615 + xd->plane[0].pre[0] = scaled_first_yv12;
1.2616 +
1.2617 + if (bestsme < last_besterr[id]) {
1.2618 + frame_mv[refs[id]].as_int = tmp_mv.as_int;
1.2619 + last_besterr[id] = bestsme;
1.2620 + } else {
1.2621 + break;
1.2622 + }
1.2623 + }
1.2624 +
1.2625 + *rate_mv = 0;
1.2626 +
1.2627 + for (ref = 0; ref < 2; ++ref) {
1.2628 + if (scaled_ref_frame[ref]) {
1.2629 + // restore the predictor
1.2630 + int i;
1.2631 + for (i = 0; i < MAX_MB_PLANE; i++)
1.2632 + xd->plane[i].pre[ref] = backup_yv12[ref][i];
1.2633 + }
1.2634 +
1.2635 + *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1.2636 + &mbmi->ref_mvs[refs[ref]][0].as_mv,
1.2637 + x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1.2638 + }
1.2639 +
1.2640 + vpx_free(second_pred);
1.2641 +}
1.2642 +
1.2643 +static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
1.2644 + const TileInfo *const tile,
1.2645 + BLOCK_SIZE bsize,
1.2646 + int64_t txfm_cache[],
1.2647 + int *rate2, int64_t *distortion,
1.2648 + int *skippable,
1.2649 + int *rate_y, int64_t *distortion_y,
1.2650 + int *rate_uv, int64_t *distortion_uv,
1.2651 + int *mode_excluded, int *disable_skip,
1.2652 + INTERPOLATION_TYPE *best_filter,
1.2653 + int_mv (*mode_mv)[MAX_REF_FRAMES],
1.2654 + int mi_row, int mi_col,
1.2655 + int_mv single_newmv[MAX_REF_FRAMES],
1.2656 + int64_t *psse,
1.2657 + const int64_t ref_best_rd) {
1.2658 + VP9_COMMON *cm = &cpi->common;
1.2659 + MACROBLOCKD *xd = &x->e_mbd;
1.2660 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.2661 + const int is_comp_pred = has_second_ref(mbmi);
1.2662 + const int num_refs = is_comp_pred ? 2 : 1;
1.2663 + const int this_mode = mbmi->mode;
1.2664 + int_mv *frame_mv = mode_mv[this_mode];
1.2665 + int i;
1.2666 + int refs[2] = { mbmi->ref_frame[0],
1.2667 + (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
1.2668 + int_mv cur_mv[2];
1.2669 + int64_t this_rd = 0;
1.2670 + DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64);
1.2671 + int pred_exists = 0;
1.2672 + int intpel_mv;
1.2673 + int64_t rd, best_rd = INT64_MAX;
1.2674 + int best_needs_copy = 0;
1.2675 + uint8_t *orig_dst[MAX_MB_PLANE];
1.2676 + int orig_dst_stride[MAX_MB_PLANE];
1.2677 + int rs = 0;
1.2678 +
1.2679 + if (is_comp_pred) {
1.2680 + if (frame_mv[refs[0]].as_int == INVALID_MV ||
1.2681 + frame_mv[refs[1]].as_int == INVALID_MV)
1.2682 + return INT64_MAX;
1.2683 + }
1.2684 +
1.2685 + if (this_mode == NEWMV) {
1.2686 + int rate_mv;
1.2687 + if (is_comp_pred) {
1.2688 + // Initialize mv using single prediction mode result.
1.2689 + frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
1.2690 + frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
1.2691 +
1.2692 + if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1.2693 + joint_motion_search(cpi, x, bsize, frame_mv,
1.2694 + mi_row, mi_col, single_newmv, &rate_mv);
1.2695 + } else {
1.2696 + rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
1.2697 + &mbmi->ref_mvs[refs[0]][0].as_mv,
1.2698 + x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1.2699 + rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
1.2700 + &mbmi->ref_mvs[refs[1]][0].as_mv,
1.2701 + x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1.2702 + }
1.2703 + *rate2 += rate_mv;
1.2704 + } else {
1.2705 + int_mv tmp_mv;
1.2706 + single_motion_search(cpi, x, tile, bsize, mi_row, mi_col,
1.2707 + &tmp_mv, &rate_mv);
1.2708 + *rate2 += rate_mv;
1.2709 + frame_mv[refs[0]].as_int =
1.2710 + xd->mi_8x8[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
1.2711 + single_newmv[refs[0]].as_int = tmp_mv.as_int;
1.2712 + }
1.2713 + }
1.2714 +
1.2715 + // if we're near/nearest and mv == 0,0, compare to zeromv
1.2716 + if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1.2717 + frame_mv[refs[0]].as_int == 0 &&
1.2718 + !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
1.2719 + (num_refs == 1 || frame_mv[refs[1]].as_int == 0)) {
1.2720 + int rfc = mbmi->mode_context[mbmi->ref_frame[0]];
1.2721 + int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1.2722 + int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1.2723 + int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1.2724 +
1.2725 + if (this_mode == NEARMV) {
1.2726 + if (c1 > c3)
1.2727 + return INT64_MAX;
1.2728 + } else if (this_mode == NEARESTMV) {
1.2729 + if (c2 > c3)
1.2730 + return INT64_MAX;
1.2731 + } else {
1.2732 + assert(this_mode == ZEROMV);
1.2733 + if (num_refs == 1) {
1.2734 + if ((c3 >= c2 &&
1.2735 + mode_mv[NEARESTMV][mbmi->ref_frame[0]].as_int == 0) ||
1.2736 + (c3 >= c1 &&
1.2737 + mode_mv[NEARMV][mbmi->ref_frame[0]].as_int == 0))
1.2738 + return INT64_MAX;
1.2739 + } else {
1.2740 + if ((c3 >= c2 &&
1.2741 + mode_mv[NEARESTMV][mbmi->ref_frame[0]].as_int == 0 &&
1.2742 + mode_mv[NEARESTMV][mbmi->ref_frame[1]].as_int == 0) ||
1.2743 + (c3 >= c1 &&
1.2744 + mode_mv[NEARMV][mbmi->ref_frame[0]].as_int == 0 &&
1.2745 + mode_mv[NEARMV][mbmi->ref_frame[1]].as_int == 0))
1.2746 + return INT64_MAX;
1.2747 + }
1.2748 + }
1.2749 + }
1.2750 +
1.2751 + for (i = 0; i < num_refs; ++i) {
1.2752 + cur_mv[i] = frame_mv[refs[i]];
1.2753 + // Clip "next_nearest" so that it does not extend to far out of image
1.2754 + if (this_mode != NEWMV)
1.2755 + clamp_mv2(&cur_mv[i].as_mv, xd);
1.2756 +
1.2757 + if (mv_check_bounds(x, &cur_mv[i]))
1.2758 + return INT64_MAX;
1.2759 + mbmi->mv[i].as_int = cur_mv[i].as_int;
1.2760 + }
1.2761 +
1.2762 + // do first prediction into the destination buffer. Do the next
1.2763 + // prediction into a temporary buffer. Then keep track of which one
1.2764 + // of these currently holds the best predictor, and use the other
1.2765 + // one for future predictions. In the end, copy from tmp_buf to
1.2766 + // dst if necessary.
1.2767 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2768 + orig_dst[i] = xd->plane[i].dst.buf;
1.2769 + orig_dst_stride[i] = xd->plane[i].dst.stride;
1.2770 + }
1.2771 +
1.2772 + /* We don't include the cost of the second reference here, because there
1.2773 + * are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
1.2774 + * words if you present them in that order, the second one is always known
1.2775 + * if the first is known */
1.2776 + *rate2 += cost_mv_ref(cpi, this_mode,
1.2777 + mbmi->mode_context[mbmi->ref_frame[0]]);
1.2778 +
1.2779 + if (!(*mode_excluded)) {
1.2780 + if (is_comp_pred) {
1.2781 + *mode_excluded = (cpi->common.comp_pred_mode == SINGLE_PREDICTION_ONLY);
1.2782 + } else {
1.2783 + *mode_excluded = (cpi->common.comp_pred_mode == COMP_PREDICTION_ONLY);
1.2784 + }
1.2785 + }
1.2786 +
1.2787 + pred_exists = 0;
1.2788 + // Are all MVs integer pel for Y and UV
1.2789 + intpel_mv = (mbmi->mv[0].as_mv.row & 15) == 0 &&
1.2790 + (mbmi->mv[0].as_mv.col & 15) == 0;
1.2791 + if (is_comp_pred)
1.2792 + intpel_mv &= (mbmi->mv[1].as_mv.row & 15) == 0 &&
1.2793 + (mbmi->mv[1].as_mv.col & 15) == 0;
1.2794 + // Search for best switchable filter by checking the variance of
1.2795 + // pred error irrespective of whether the filter will be used
1.2796 + if (cm->mcomp_filter_type != BILINEAR) {
1.2797 + *best_filter = EIGHTTAP;
1.2798 + if (x->source_variance <
1.2799 + cpi->sf.disable_filter_search_var_thresh) {
1.2800 + *best_filter = EIGHTTAP;
1.2801 + vp9_zero(cpi->rd_filter_cache);
1.2802 + } else {
1.2803 + int i, newbest;
1.2804 + int tmp_rate_sum = 0;
1.2805 + int64_t tmp_dist_sum = 0;
1.2806 +
1.2807 + cpi->rd_filter_cache[SWITCHABLE_FILTERS] = INT64_MAX;
1.2808 + for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
1.2809 + int j;
1.2810 + int64_t rs_rd;
1.2811 + mbmi->interp_filter = i;
1.2812 + vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
1.2813 + rs = get_switchable_rate(x);
1.2814 + rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
1.2815 +
1.2816 + if (i > 0 && intpel_mv) {
1.2817 + cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
1.2818 + tmp_rate_sum, tmp_dist_sum);
1.2819 + cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
1.2820 + MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
1.2821 + cpi->rd_filter_cache[i] + rs_rd);
1.2822 + rd = cpi->rd_filter_cache[i];
1.2823 + if (cm->mcomp_filter_type == SWITCHABLE)
1.2824 + rd += rs_rd;
1.2825 + } else {
1.2826 + int rate_sum = 0;
1.2827 + int64_t dist_sum = 0;
1.2828 + if ((cm->mcomp_filter_type == SWITCHABLE &&
1.2829 + (!i || best_needs_copy)) ||
1.2830 + (cm->mcomp_filter_type != SWITCHABLE &&
1.2831 + (cm->mcomp_filter_type == mbmi->interp_filter ||
1.2832 + (i == 0 && intpel_mv)))) {
1.2833 + for (j = 0; j < MAX_MB_PLANE; j++) {
1.2834 + xd->plane[j].dst.buf = orig_dst[j];
1.2835 + xd->plane[j].dst.stride = orig_dst_stride[j];
1.2836 + }
1.2837 + } else {
1.2838 + for (j = 0; j < MAX_MB_PLANE; j++) {
1.2839 + xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
1.2840 + xd->plane[j].dst.stride = 64;
1.2841 + }
1.2842 + }
1.2843 + vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
1.2844 + model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum);
1.2845 + cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
1.2846 + rate_sum, dist_sum);
1.2847 + cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
1.2848 + MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
1.2849 + cpi->rd_filter_cache[i] + rs_rd);
1.2850 + rd = cpi->rd_filter_cache[i];
1.2851 + if (cm->mcomp_filter_type == SWITCHABLE)
1.2852 + rd += rs_rd;
1.2853 + if (i == 0 && intpel_mv) {
1.2854 + tmp_rate_sum = rate_sum;
1.2855 + tmp_dist_sum = dist_sum;
1.2856 + }
1.2857 + }
1.2858 + if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
1.2859 + if (rd / 2 > ref_best_rd) {
1.2860 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2861 + xd->plane[i].dst.buf = orig_dst[i];
1.2862 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.2863 + }
1.2864 + return INT64_MAX;
1.2865 + }
1.2866 + }
1.2867 + newbest = i == 0 || rd < best_rd;
1.2868 +
1.2869 + if (newbest) {
1.2870 + best_rd = rd;
1.2871 + *best_filter = mbmi->interp_filter;
1.2872 + if (cm->mcomp_filter_type == SWITCHABLE && i && !intpel_mv)
1.2873 + best_needs_copy = !best_needs_copy;
1.2874 + }
1.2875 +
1.2876 + if ((cm->mcomp_filter_type == SWITCHABLE && newbest) ||
1.2877 + (cm->mcomp_filter_type != SWITCHABLE &&
1.2878 + cm->mcomp_filter_type == mbmi->interp_filter)) {
1.2879 + pred_exists = 1;
1.2880 + }
1.2881 + }
1.2882 +
1.2883 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2884 + xd->plane[i].dst.buf = orig_dst[i];
1.2885 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.2886 + }
1.2887 + }
1.2888 + }
1.2889 + // Set the appropriate filter
1.2890 + mbmi->interp_filter = cm->mcomp_filter_type != SWITCHABLE ?
1.2891 + cm->mcomp_filter_type : *best_filter;
1.2892 + vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
1.2893 + rs = cm->mcomp_filter_type == SWITCHABLE ? get_switchable_rate(x) : 0;
1.2894 +
1.2895 + if (pred_exists) {
1.2896 + if (best_needs_copy) {
1.2897 + // again temporarily set the buffers to local memory to prevent a memcpy
1.2898 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2899 + xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
1.2900 + xd->plane[i].dst.stride = 64;
1.2901 + }
1.2902 + }
1.2903 + } else {
1.2904 + // Handles the special case when a filter that is not in the
1.2905 + // switchable list (ex. bilinear, 6-tap) is indicated at the frame level
1.2906 + vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
1.2907 + }
1.2908 +
1.2909 +
1.2910 + if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
1.2911 + int tmp_rate;
1.2912 + int64_t tmp_dist;
1.2913 + model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist);
1.2914 + rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
1.2915 + // if current pred_error modeled rd is substantially more than the best
1.2916 + // so far, do not bother doing full rd
1.2917 + if (rd / 2 > ref_best_rd) {
1.2918 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.2919 + xd->plane[i].dst.buf = orig_dst[i];
1.2920 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.2921 + }
1.2922 + return INT64_MAX;
1.2923 + }
1.2924 + }
1.2925 +
1.2926 + if (cpi->common.mcomp_filter_type == SWITCHABLE)
1.2927 + *rate2 += get_switchable_rate(x);
1.2928 +
1.2929 + if (!is_comp_pred && cpi->enable_encode_breakout) {
1.2930 + if (cpi->active_map_enabled && x->active_ptr[0] == 0)
1.2931 + x->skip = 1;
1.2932 + else if (x->encode_breakout) {
1.2933 + const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
1.2934 + const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
1.2935 + unsigned int var, sse;
1.2936 + // Skipping threshold for ac.
1.2937 + unsigned int thresh_ac;
1.2938 + // The encode_breakout input
1.2939 + unsigned int encode_breakout = x->encode_breakout << 4;
1.2940 + unsigned int max_thresh = 36000;
1.2941 +
1.2942 + // Use extreme low threshold for static frames to limit skipping.
1.2943 + if (cpi->enable_encode_breakout == 2)
1.2944 + max_thresh = 128;
1.2945 +
1.2946 + // Calculate threshold according to dequant value.
1.2947 + thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
1.2948 +
1.2949 + // Use encode_breakout input if it is bigger than internal threshold.
1.2950 + if (thresh_ac < encode_breakout)
1.2951 + thresh_ac = encode_breakout;
1.2952 +
1.2953 + // Set a maximum for threshold to avoid big PSNR loss in low bitrate case.
1.2954 + if (thresh_ac > max_thresh)
1.2955 + thresh_ac = max_thresh;
1.2956 +
1.2957 + var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
1.2958 + xd->plane[0].dst.buf,
1.2959 + xd->plane[0].dst.stride, &sse);
1.2960 +
1.2961 + // Adjust threshold according to partition size.
1.2962 + thresh_ac >>= 8 - (b_width_log2_lookup[bsize] +
1.2963 + b_height_log2_lookup[bsize]);
1.2964 +
1.2965 + // Y skipping condition checking
1.2966 + if (sse < thresh_ac || sse == 0) {
1.2967 + // Skipping threshold for dc
1.2968 + unsigned int thresh_dc;
1.2969 +
1.2970 + thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
1.2971 +
1.2972 + // dc skipping checking
1.2973 + if ((sse - var) < thresh_dc || sse == var) {
1.2974 + unsigned int sse_u, sse_v;
1.2975 + unsigned int var_u, var_v;
1.2976 +
1.2977 + var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
1.2978 + x->plane[1].src.stride,
1.2979 + xd->plane[1].dst.buf,
1.2980 + xd->plane[1].dst.stride, &sse_u);
1.2981 +
1.2982 + // U skipping condition checking
1.2983 + if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
1.2984 + (sse_u - var_u < thresh_dc || sse_u == var_u)) {
1.2985 + var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
1.2986 + x->plane[2].src.stride,
1.2987 + xd->plane[2].dst.buf,
1.2988 + xd->plane[2].dst.stride, &sse_v);
1.2989 +
1.2990 + // V skipping condition checking
1.2991 + if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
1.2992 + (sse_v - var_v < thresh_dc || sse_v == var_v)) {
1.2993 + x->skip = 1;
1.2994 +
1.2995 + // The cost of skip bit needs to be added.
1.2996 + *rate2 += vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd), 1);
1.2997 +
1.2998 + // Scaling factor for SSE from spatial domain to frequency domain
1.2999 + // is 16. Adjust distortion accordingly.
1.3000 + *distortion_uv = (sse_u + sse_v) << 4;
1.3001 + *distortion = (sse << 4) + *distortion_uv;
1.3002 +
1.3003 + *disable_skip = 1;
1.3004 + this_rd = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
1.3005 + }
1.3006 + }
1.3007 + }
1.3008 + }
1.3009 + }
1.3010 + }
1.3011 +
1.3012 + if (!x->skip) {
1.3013 + int skippable_y, skippable_uv;
1.3014 + int64_t sseuv = INT64_MAX;
1.3015 + int64_t rdcosty = INT64_MAX;
1.3016 +
1.3017 + // Y cost and distortion
1.3018 + super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, psse,
1.3019 + bsize, txfm_cache, ref_best_rd);
1.3020 +
1.3021 + if (*rate_y == INT_MAX) {
1.3022 + *rate2 = INT_MAX;
1.3023 + *distortion = INT64_MAX;
1.3024 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.3025 + xd->plane[i].dst.buf = orig_dst[i];
1.3026 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.3027 + }
1.3028 + return INT64_MAX;
1.3029 + }
1.3030 +
1.3031 + *rate2 += *rate_y;
1.3032 + *distortion += *distortion_y;
1.3033 +
1.3034 + rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
1.3035 + rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
1.3036 +
1.3037 + super_block_uvrd(cpi, x, rate_uv, distortion_uv, &skippable_uv, &sseuv,
1.3038 + bsize, ref_best_rd - rdcosty);
1.3039 + if (*rate_uv == INT_MAX) {
1.3040 + *rate2 = INT_MAX;
1.3041 + *distortion = INT64_MAX;
1.3042 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.3043 + xd->plane[i].dst.buf = orig_dst[i];
1.3044 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.3045 + }
1.3046 + return INT64_MAX;
1.3047 + }
1.3048 +
1.3049 + *psse += sseuv;
1.3050 + *rate2 += *rate_uv;
1.3051 + *distortion += *distortion_uv;
1.3052 + *skippable = skippable_y && skippable_uv;
1.3053 + }
1.3054 +
1.3055 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.3056 + xd->plane[i].dst.buf = orig_dst[i];
1.3057 + xd->plane[i].dst.stride = orig_dst_stride[i];
1.3058 + }
1.3059 +
1.3060 + return this_rd; // if 0, this will be re-calculated by caller
1.3061 +}
1.3062 +
1.3063 +static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
1.3064 + int max_plane) {
1.3065 + struct macroblock_plane *const p = x->plane;
1.3066 + struct macroblockd_plane *const pd = x->e_mbd.plane;
1.3067 + int i;
1.3068 +
1.3069 + for (i = 0; i < max_plane; ++i) {
1.3070 + p[i].coeff = ctx->coeff_pbuf[i][1];
1.3071 + pd[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1.3072 + pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1.3073 + pd[i].eobs = ctx->eobs_pbuf[i][1];
1.3074 +
1.3075 + ctx->coeff_pbuf[i][1] = ctx->coeff_pbuf[i][0];
1.3076 + ctx->qcoeff_pbuf[i][1] = ctx->qcoeff_pbuf[i][0];
1.3077 + ctx->dqcoeff_pbuf[i][1] = ctx->dqcoeff_pbuf[i][0];
1.3078 + ctx->eobs_pbuf[i][1] = ctx->eobs_pbuf[i][0];
1.3079 +
1.3080 + ctx->coeff_pbuf[i][0] = p[i].coeff;
1.3081 + ctx->qcoeff_pbuf[i][0] = pd[i].qcoeff;
1.3082 + ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
1.3083 + ctx->eobs_pbuf[i][0] = pd[i].eobs;
1.3084 + }
1.3085 +}
1.3086 +
1.3087 +void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
1.3088 + int *returnrate, int64_t *returndist,
1.3089 + BLOCK_SIZE bsize,
1.3090 + PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
1.3091 + VP9_COMMON *const cm = &cpi->common;
1.3092 + MACROBLOCKD *const xd = &x->e_mbd;
1.3093 + int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
1.3094 + int y_skip = 0, uv_skip = 0;
1.3095 + int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
1.3096 + x->skip_encode = 0;
1.3097 + ctx->skip = 0;
1.3098 + xd->mi_8x8[0]->mbmi.ref_frame[0] = INTRA_FRAME;
1.3099 + if (bsize >= BLOCK_8X8) {
1.3100 + if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
1.3101 + &dist_y, &y_skip, bsize, tx_cache,
1.3102 + best_rd) >= best_rd) {
1.3103 + *returnrate = INT_MAX;
1.3104 + return;
1.3105 + }
1.3106 + rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
1.3107 + &dist_uv, &uv_skip, bsize);
1.3108 + } else {
1.3109 + y_skip = 0;
1.3110 + if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
1.3111 + &dist_y, best_rd) >= best_rd) {
1.3112 + *returnrate = INT_MAX;
1.3113 + return;
1.3114 + }
1.3115 + rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
1.3116 + &dist_uv, &uv_skip, BLOCK_8X8);
1.3117 + }
1.3118 +
1.3119 + if (y_skip && uv_skip) {
1.3120 + *returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
1.3121 + vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd), 1);
1.3122 + *returndist = dist_y + dist_uv;
1.3123 + vp9_zero(ctx->tx_rd_diff);
1.3124 + } else {
1.3125 + int i;
1.3126 + *returnrate = rate_y + rate_uv +
1.3127 + vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd), 0);
1.3128 + *returndist = dist_y + dist_uv;
1.3129 + if (cpi->sf.tx_size_search_method == USE_FULL_RD)
1.3130 + for (i = 0; i < TX_MODES; i++) {
1.3131 + if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX)
1.3132 + ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
1.3133 + else
1.3134 + ctx->tx_rd_diff[i] = 0;
1.3135 + }
1.3136 + }
1.3137 +
1.3138 + ctx->mic = *xd->mi_8x8[0];
1.3139 +}
1.3140 +
1.3141 +int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
1.3142 + const TileInfo *const tile,
1.3143 + int mi_row, int mi_col,
1.3144 + int *returnrate,
1.3145 + int64_t *returndistortion,
1.3146 + BLOCK_SIZE bsize,
1.3147 + PICK_MODE_CONTEXT *ctx,
1.3148 + int64_t best_rd_so_far) {
1.3149 + VP9_COMMON *cm = &cpi->common;
1.3150 + MACROBLOCKD *xd = &x->e_mbd;
1.3151 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.3152 + const struct segmentation *seg = &cm->seg;
1.3153 + const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
1.3154 + MB_PREDICTION_MODE this_mode;
1.3155 + MV_REFERENCE_FRAME ref_frame, second_ref_frame;
1.3156 + unsigned char segment_id = mbmi->segment_id;
1.3157 + int comp_pred, i;
1.3158 + int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1.3159 + struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1.3160 + int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
1.3161 + static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1.3162 + VP9_ALT_FLAG };
1.3163 + int idx_list[4] = {0,
1.3164 + cpi->lst_fb_idx,
1.3165 + cpi->gld_fb_idx,
1.3166 + cpi->alt_fb_idx};
1.3167 + int64_t best_rd = best_rd_so_far;
1.3168 + int64_t best_tx_rd[TX_MODES];
1.3169 + int64_t best_tx_diff[TX_MODES];
1.3170 + int64_t best_pred_diff[NB_PREDICTION_TYPES];
1.3171 + int64_t best_pred_rd[NB_PREDICTION_TYPES];
1.3172 + int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
1.3173 + int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
1.3174 + MB_MODE_INFO best_mbmode = { 0 };
1.3175 + int j;
1.3176 + int mode_index, best_mode_index = 0;
1.3177 + unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
1.3178 + vp9_prob comp_mode_p;
1.3179 + int64_t best_intra_rd = INT64_MAX;
1.3180 + int64_t best_inter_rd = INT64_MAX;
1.3181 + MB_PREDICTION_MODE best_intra_mode = DC_PRED;
1.3182 + MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
1.3183 + INTERPOLATION_TYPE tmp_best_filter = SWITCHABLE;
1.3184 + int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
1.3185 + int64_t dist_uv[TX_SIZES];
1.3186 + int skip_uv[TX_SIZES];
1.3187 + MB_PREDICTION_MODE mode_uv[TX_SIZES];
1.3188 + struct scale_factors scale_factor[4];
1.3189 + unsigned int ref_frame_mask = 0;
1.3190 + unsigned int mode_mask = 0;
1.3191 + int64_t mode_distortions[MB_MODE_COUNT] = {-1};
1.3192 + int64_t frame_distortions[MAX_REF_FRAMES] = {-1};
1.3193 + int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
1.3194 + const int bws = num_8x8_blocks_wide_lookup[bsize] / 2;
1.3195 + const int bhs = num_8x8_blocks_high_lookup[bsize] / 2;
1.3196 + int best_skip2 = 0;
1.3197 +
1.3198 + x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1.3199 +
1.3200 + // Everywhere the flag is set the error is much higher than its neighbors.
1.3201 + ctx->frames_with_high_error = 0;
1.3202 + ctx->modes_with_high_error = 0;
1.3203 +
1.3204 + estimate_ref_frame_costs(cpi, segment_id, ref_costs_single, ref_costs_comp,
1.3205 + &comp_mode_p);
1.3206 +
1.3207 + for (i = 0; i < NB_PREDICTION_TYPES; ++i)
1.3208 + best_pred_rd[i] = INT64_MAX;
1.3209 + for (i = 0; i < TX_MODES; i++)
1.3210 + best_tx_rd[i] = INT64_MAX;
1.3211 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
1.3212 + best_filter_rd[i] = INT64_MAX;
1.3213 + for (i = 0; i < TX_SIZES; i++)
1.3214 + rate_uv_intra[i] = INT_MAX;
1.3215 +
1.3216 + *returnrate = INT_MAX;
1.3217 +
1.3218 + // Create a mask set to 1 for each reference frame used by a smaller
1.3219 + // resolution.
1.3220 + if (cpi->sf.use_avoid_tested_higherror) {
1.3221 + switch (block_size) {
1.3222 + case BLOCK_64X64:
1.3223 + for (i = 0; i < 4; i++) {
1.3224 + for (j = 0; j < 4; j++) {
1.3225 + ref_frame_mask |= x->mb_context[i][j].frames_with_high_error;
1.3226 + mode_mask |= x->mb_context[i][j].modes_with_high_error;
1.3227 + }
1.3228 + }
1.3229 + for (i = 0; i < 4; i++) {
1.3230 + ref_frame_mask |= x->sb32_context[i].frames_with_high_error;
1.3231 + mode_mask |= x->sb32_context[i].modes_with_high_error;
1.3232 + }
1.3233 + break;
1.3234 + case BLOCK_32X32:
1.3235 + for (i = 0; i < 4; i++) {
1.3236 + ref_frame_mask |=
1.3237 + x->mb_context[x->sb_index][i].frames_with_high_error;
1.3238 + mode_mask |= x->mb_context[x->sb_index][i].modes_with_high_error;
1.3239 + }
1.3240 + break;
1.3241 + default:
1.3242 + // Until we handle all block sizes set it to present;
1.3243 + ref_frame_mask = 0;
1.3244 + mode_mask = 0;
1.3245 + break;
1.3246 + }
1.3247 + ref_frame_mask = ~ref_frame_mask;
1.3248 + mode_mask = ~mode_mask;
1.3249 + }
1.3250 +
1.3251 + for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
1.3252 + if (cpi->ref_frame_flags & flag_list[ref_frame]) {
1.3253 + setup_buffer_inter(cpi, x, tile, idx_list[ref_frame], ref_frame,
1.3254 + block_size, mi_row, mi_col,
1.3255 + frame_mv[NEARESTMV], frame_mv[NEARMV],
1.3256 + yv12_mb, scale_factor);
1.3257 + }
1.3258 + frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1.3259 + frame_mv[ZEROMV][ref_frame].as_int = 0;
1.3260 + }
1.3261 +
1.3262 + for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
1.3263 + int mode_excluded = 0;
1.3264 + int64_t this_rd = INT64_MAX;
1.3265 + int disable_skip = 0;
1.3266 + int compmode_cost = 0;
1.3267 + int rate2 = 0, rate_y = 0, rate_uv = 0;
1.3268 + int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
1.3269 + int skippable = 0;
1.3270 + int64_t tx_cache[TX_MODES];
1.3271 + int i;
1.3272 + int this_skip2 = 0;
1.3273 + int64_t total_sse = INT_MAX;
1.3274 + int early_term = 0;
1.3275 +
1.3276 + for (i = 0; i < TX_MODES; ++i)
1.3277 + tx_cache[i] = INT64_MAX;
1.3278 +
1.3279 + x->skip = 0;
1.3280 + this_mode = vp9_mode_order[mode_index].mode;
1.3281 + ref_frame = vp9_mode_order[mode_index].ref_frame;
1.3282 + second_ref_frame = vp9_mode_order[mode_index].second_ref_frame;
1.3283 +
1.3284 + // Look at the reference frame of the best mode so far and set the
1.3285 + // skip mask to look at a subset of the remaining modes.
1.3286 + if (mode_index > cpi->sf.mode_skip_start) {
1.3287 + if (mode_index == (cpi->sf.mode_skip_start + 1)) {
1.3288 + switch (vp9_mode_order[best_mode_index].ref_frame) {
1.3289 + case INTRA_FRAME:
1.3290 + cpi->mode_skip_mask = 0;
1.3291 + break;
1.3292 + case LAST_FRAME:
1.3293 + cpi->mode_skip_mask = LAST_FRAME_MODE_MASK;
1.3294 + break;
1.3295 + case GOLDEN_FRAME:
1.3296 + cpi->mode_skip_mask = GOLDEN_FRAME_MODE_MASK;
1.3297 + break;
1.3298 + case ALTREF_FRAME:
1.3299 + cpi->mode_skip_mask = ALT_REF_MODE_MASK;
1.3300 + break;
1.3301 + case NONE:
1.3302 + case MAX_REF_FRAMES:
1.3303 + assert(!"Invalid Reference frame");
1.3304 + }
1.3305 + }
1.3306 + if (cpi->mode_skip_mask & ((int64_t)1 << mode_index))
1.3307 + continue;
1.3308 + }
1.3309 +
1.3310 + // Skip if the current reference frame has been masked off
1.3311 + if (cpi->sf.reference_masking && !cpi->set_ref_frame_mask &&
1.3312 + (cpi->ref_frame_mask & (1 << ref_frame)))
1.3313 + continue;
1.3314 +
1.3315 + // Test best rd so far against threshold for trying this mode.
1.3316 + if ((best_rd < ((int64_t)cpi->rd_threshes[segment_id][bsize][mode_index] *
1.3317 + cpi->rd_thresh_freq_fact[bsize][mode_index] >> 5)) ||
1.3318 + cpi->rd_threshes[segment_id][bsize][mode_index] == INT_MAX)
1.3319 + continue;
1.3320 +
1.3321 + // Do not allow compound prediction if the segment level reference
1.3322 + // frame feature is in use as in this case there can only be one reference.
1.3323 + if ((second_ref_frame > INTRA_FRAME) &&
1.3324 + vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
1.3325 + continue;
1.3326 +
1.3327 + // Skip some checking based on small partitions' result.
1.3328 + if (x->fast_ms > 1 && !ref_frame)
1.3329 + continue;
1.3330 + if (x->fast_ms > 2 && ref_frame != x->subblock_ref)
1.3331 + continue;
1.3332 +
1.3333 + if (cpi->sf.use_avoid_tested_higherror && bsize >= BLOCK_8X8) {
1.3334 + if (!(ref_frame_mask & (1 << ref_frame))) {
1.3335 + continue;
1.3336 + }
1.3337 + if (!(mode_mask & (1 << this_mode))) {
1.3338 + continue;
1.3339 + }
1.3340 + if (second_ref_frame != NONE
1.3341 + && !(ref_frame_mask & (1 << second_ref_frame))) {
1.3342 + continue;
1.3343 + }
1.3344 + }
1.3345 +
1.3346 + mbmi->ref_frame[0] = ref_frame;
1.3347 + mbmi->ref_frame[1] = second_ref_frame;
1.3348 +
1.3349 + if (!(ref_frame == INTRA_FRAME
1.3350 + || (cpi->ref_frame_flags & flag_list[ref_frame]))) {
1.3351 + continue;
1.3352 + }
1.3353 + if (!(second_ref_frame == NONE
1.3354 + || (cpi->ref_frame_flags & flag_list[second_ref_frame]))) {
1.3355 + continue;
1.3356 + }
1.3357 +
1.3358 + comp_pred = second_ref_frame > INTRA_FRAME;
1.3359 + if (comp_pred) {
1.3360 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA)
1.3361 + if (vp9_mode_order[best_mode_index].ref_frame == INTRA_FRAME)
1.3362 + continue;
1.3363 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH)
1.3364 + if (ref_frame != best_inter_ref_frame &&
1.3365 + second_ref_frame != best_inter_ref_frame)
1.3366 + continue;
1.3367 + }
1.3368 +
1.3369 + set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
1.3370 + mbmi->uv_mode = DC_PRED;
1.3371 +
1.3372 + // Evaluate all sub-pel filters irrespective of whether we can use
1.3373 + // them for this frame.
1.3374 + mbmi->interp_filter = cm->mcomp_filter_type;
1.3375 + vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
1.3376 +
1.3377 + if (comp_pred) {
1.3378 + if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
1.3379 + continue;
1.3380 + set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
1.3381 +
1.3382 + mode_excluded = mode_excluded
1.3383 + ? mode_excluded
1.3384 + : cm->comp_pred_mode == SINGLE_PREDICTION_ONLY;
1.3385 + } else {
1.3386 + if (ref_frame != INTRA_FRAME && second_ref_frame != INTRA_FRAME) {
1.3387 + mode_excluded =
1.3388 + mode_excluded ?
1.3389 + mode_excluded : cm->comp_pred_mode == COMP_PREDICTION_ONLY;
1.3390 + }
1.3391 + }
1.3392 +
1.3393 + // Select prediction reference frames.
1.3394 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.3395 + xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1.3396 + if (comp_pred)
1.3397 + xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
1.3398 + }
1.3399 +
1.3400 + // If the segment reference frame feature is enabled....
1.3401 + // then do nothing if the current ref frame is not allowed..
1.3402 + if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
1.3403 + vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) !=
1.3404 + (int)ref_frame) {
1.3405 + continue;
1.3406 + // If the segment skip feature is enabled....
1.3407 + // then do nothing if the current mode is not allowed..
1.3408 + } else if (vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) &&
1.3409 + (this_mode != ZEROMV && ref_frame != INTRA_FRAME)) {
1.3410 + continue;
1.3411 + // Disable this drop out case if the ref frame
1.3412 + // segment level feature is enabled for this segment. This is to
1.3413 + // prevent the possibility that we end up unable to pick any mode.
1.3414 + } else if (!vp9_segfeature_active(seg, segment_id,
1.3415 + SEG_LVL_REF_FRAME)) {
1.3416 + // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
1.3417 + // unless ARNR filtering is enabled in which case we want
1.3418 + // an unfiltered alternative. We allow near/nearest as well
1.3419 + // because they may result in zero-zero MVs but be cheaper.
1.3420 + if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
1.3421 + if ((this_mode != ZEROMV &&
1.3422 + !(this_mode == NEARMV &&
1.3423 + frame_mv[NEARMV][ALTREF_FRAME].as_int == 0) &&
1.3424 + !(this_mode == NEARESTMV &&
1.3425 + frame_mv[NEARESTMV][ALTREF_FRAME].as_int == 0)) ||
1.3426 + ref_frame != ALTREF_FRAME) {
1.3427 + continue;
1.3428 + }
1.3429 + }
1.3430 + }
1.3431 + // TODO(JBB): This is to make up for the fact that we don't have sad
1.3432 + // functions that work when the block size reads outside the umv. We
1.3433 + // should fix this either by making the motion search just work on
1.3434 + // a representative block in the boundary ( first ) and then implement a
1.3435 + // function that does sads when inside the border..
1.3436 + if (((mi_row + bhs) > cm->mi_rows || (mi_col + bws) > cm->mi_cols) &&
1.3437 + this_mode == NEWMV) {
1.3438 + continue;
1.3439 + }
1.3440 +
1.3441 +#ifdef MODE_TEST_HIT_STATS
1.3442 + // TEST/DEBUG CODE
1.3443 + // Keep a rcord of the number of test hits at each size
1.3444 + cpi->mode_test_hits[bsize]++;
1.3445 +#endif
1.3446 +
1.3447 +
1.3448 + if (ref_frame == INTRA_FRAME) {
1.3449 + TX_SIZE uv_tx;
1.3450 + // Disable intra modes other than DC_PRED for blocks with low variance
1.3451 + // Threshold for intra skipping based on source variance
1.3452 + // TODO(debargha): Specialize the threshold for super block sizes
1.3453 + static const unsigned int skip_intra_var_thresh[BLOCK_SIZES] = {
1.3454 + 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
1.3455 + };
1.3456 + if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
1.3457 + this_mode != DC_PRED &&
1.3458 + x->source_variance < skip_intra_var_thresh[mbmi->sb_type])
1.3459 + continue;
1.3460 + // Only search the oblique modes if the best so far is
1.3461 + // one of the neighboring directional modes
1.3462 + if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
1.3463 + (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
1.3464 + if (vp9_mode_order[best_mode_index].ref_frame > INTRA_FRAME)
1.3465 + continue;
1.3466 + }
1.3467 + mbmi->mode = this_mode;
1.3468 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1.3469 + if (conditional_skipintra(mbmi->mode, best_intra_mode))
1.3470 + continue;
1.3471 + }
1.3472 +
1.3473 + super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL,
1.3474 + bsize, tx_cache, best_rd);
1.3475 +
1.3476 + if (rate_y == INT_MAX)
1.3477 + continue;
1.3478 +
1.3479 + uv_tx = MIN(mbmi->tx_size, max_uv_txsize_lookup[bsize]);
1.3480 + if (rate_uv_intra[uv_tx] == INT_MAX) {
1.3481 + choose_intra_uv_mode(cpi, ctx, bsize, &rate_uv_intra[uv_tx],
1.3482 + &rate_uv_tokenonly[uv_tx],
1.3483 + &dist_uv[uv_tx], &skip_uv[uv_tx],
1.3484 + &mode_uv[uv_tx]);
1.3485 + }
1.3486 +
1.3487 + rate_uv = rate_uv_tokenonly[uv_tx];
1.3488 + distortion_uv = dist_uv[uv_tx];
1.3489 + skippable = skippable && skip_uv[uv_tx];
1.3490 + mbmi->uv_mode = mode_uv[uv_tx];
1.3491 +
1.3492 + rate2 = rate_y + x->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
1.3493 + if (this_mode != DC_PRED && this_mode != TM_PRED)
1.3494 + rate2 += intra_cost_penalty;
1.3495 + distortion2 = distortion_y + distortion_uv;
1.3496 + } else {
1.3497 + mbmi->mode = this_mode;
1.3498 + compmode_cost = vp9_cost_bit(comp_mode_p, second_ref_frame > INTRA_FRAME);
1.3499 + this_rd = handle_inter_mode(cpi, x, tile, bsize,
1.3500 + tx_cache,
1.3501 + &rate2, &distortion2, &skippable,
1.3502 + &rate_y, &distortion_y,
1.3503 + &rate_uv, &distortion_uv,
1.3504 + &mode_excluded, &disable_skip,
1.3505 + &tmp_best_filter, frame_mv,
1.3506 + mi_row, mi_col,
1.3507 + single_newmv, &total_sse, best_rd);
1.3508 + if (this_rd == INT64_MAX)
1.3509 + continue;
1.3510 + }
1.3511 +
1.3512 + if (cm->comp_pred_mode == HYBRID_PREDICTION) {
1.3513 + rate2 += compmode_cost;
1.3514 + }
1.3515 +
1.3516 + // Estimate the reference frame signaling cost and add it
1.3517 + // to the rolling cost variable.
1.3518 + if (second_ref_frame > INTRA_FRAME) {
1.3519 + rate2 += ref_costs_comp[ref_frame];
1.3520 + } else {
1.3521 + rate2 += ref_costs_single[ref_frame];
1.3522 + }
1.3523 +
1.3524 + if (!disable_skip) {
1.3525 + // Test for the condition where skip block will be activated
1.3526 + // because there are no non zero coefficients and make any
1.3527 + // necessary adjustment for rate. Ignore if skip is coded at
1.3528 + // segment level as the cost wont have been added in.
1.3529 + // Is Mb level skip allowed (i.e. not coded at segment level).
1.3530 + const int mb_skip_allowed = !vp9_segfeature_active(seg, segment_id,
1.3531 + SEG_LVL_SKIP);
1.3532 +
1.3533 + if (skippable) {
1.3534 + // Back out the coefficient coding costs
1.3535 + rate2 -= (rate_y + rate_uv);
1.3536 + // for best yrd calculation
1.3537 + rate_uv = 0;
1.3538 +
1.3539 + if (mb_skip_allowed) {
1.3540 + int prob_skip_cost;
1.3541 +
1.3542 + // Cost the skip mb case
1.3543 + vp9_prob skip_prob =
1.3544 + vp9_get_pred_prob_mbskip(cm, xd);
1.3545 +
1.3546 + if (skip_prob) {
1.3547 + prob_skip_cost = vp9_cost_bit(skip_prob, 1);
1.3548 + rate2 += prob_skip_cost;
1.3549 + }
1.3550 + }
1.3551 + } else if (mb_skip_allowed && ref_frame != INTRA_FRAME && !xd->lossless) {
1.3552 + if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
1.3553 + RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
1.3554 + // Add in the cost of the no skip flag.
1.3555 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.3556 + 0);
1.3557 + rate2 += prob_skip_cost;
1.3558 + } else {
1.3559 + // FIXME(rbultje) make this work for splitmv also
1.3560 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.3561 + 1);
1.3562 + rate2 += prob_skip_cost;
1.3563 + distortion2 = total_sse;
1.3564 + assert(total_sse >= 0);
1.3565 + rate2 -= (rate_y + rate_uv);
1.3566 + rate_y = 0;
1.3567 + rate_uv = 0;
1.3568 + this_skip2 = 1;
1.3569 + }
1.3570 + } else if (mb_skip_allowed) {
1.3571 + // Add in the cost of the no skip flag.
1.3572 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.3573 + 0);
1.3574 + rate2 += prob_skip_cost;
1.3575 + }
1.3576 +
1.3577 + // Calculate the final RD estimate for this mode.
1.3578 + this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
1.3579 + }
1.3580 +
1.3581 + // Keep record of best intra rd
1.3582 + if (!is_inter_block(&xd->mi_8x8[0]->mbmi) &&
1.3583 + this_rd < best_intra_rd) {
1.3584 + best_intra_rd = this_rd;
1.3585 + best_intra_mode = xd->mi_8x8[0]->mbmi.mode;
1.3586 + }
1.3587 +
1.3588 + // Keep record of best inter rd with single reference
1.3589 + if (is_inter_block(&xd->mi_8x8[0]->mbmi) &&
1.3590 + !has_second_ref(&xd->mi_8x8[0]->mbmi) &&
1.3591 + !mode_excluded && this_rd < best_inter_rd) {
1.3592 + best_inter_rd = this_rd;
1.3593 + best_inter_ref_frame = ref_frame;
1.3594 + }
1.3595 +
1.3596 + if (!disable_skip && ref_frame == INTRA_FRAME) {
1.3597 + for (i = 0; i < NB_PREDICTION_TYPES; ++i)
1.3598 + best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
1.3599 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
1.3600 + best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
1.3601 + }
1.3602 +
1.3603 + // Store the respective mode distortions for later use.
1.3604 + if (mode_distortions[this_mode] == -1
1.3605 + || distortion2 < mode_distortions[this_mode]) {
1.3606 + mode_distortions[this_mode] = distortion2;
1.3607 + }
1.3608 + if (frame_distortions[ref_frame] == -1
1.3609 + || distortion2 < frame_distortions[ref_frame]) {
1.3610 + frame_distortions[ref_frame] = distortion2;
1.3611 + }
1.3612 +
1.3613 + // Did this mode help.. i.e. is it the new best mode
1.3614 + if (this_rd < best_rd || x->skip) {
1.3615 + int max_plane = MAX_MB_PLANE;
1.3616 + if (!mode_excluded) {
1.3617 + // Note index of best mode so far
1.3618 + best_mode_index = mode_index;
1.3619 +
1.3620 + if (ref_frame == INTRA_FRAME) {
1.3621 + /* required for left and above block mv */
1.3622 + mbmi->mv[0].as_int = 0;
1.3623 + max_plane = 1;
1.3624 + }
1.3625 +
1.3626 + *returnrate = rate2;
1.3627 + *returndistortion = distortion2;
1.3628 + best_rd = this_rd;
1.3629 + best_mbmode = *mbmi;
1.3630 + best_skip2 = this_skip2;
1.3631 + if (!x->select_txfm_size)
1.3632 + swap_block_ptr(x, ctx, max_plane);
1.3633 + vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
1.3634 + sizeof(uint8_t) * ctx->num_4x4_blk);
1.3635 +
1.3636 + // TODO(debargha): enhance this test with a better distortion prediction
1.3637 + // based on qp, activity mask and history
1.3638 + if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
1.3639 + (mode_index > MIN_EARLY_TERM_INDEX)) {
1.3640 + const int qstep = xd->plane[0].dequant[1];
1.3641 + // TODO(debargha): Enhance this by specializing for each mode_index
1.3642 + int scale = 4;
1.3643 + if (x->source_variance < UINT_MAX) {
1.3644 + const int var_adjust = (x->source_variance < 16);
1.3645 + scale -= var_adjust;
1.3646 + }
1.3647 + if (ref_frame > INTRA_FRAME &&
1.3648 + distortion2 * scale < qstep * qstep) {
1.3649 + early_term = 1;
1.3650 + }
1.3651 + }
1.3652 + }
1.3653 + }
1.3654 +
1.3655 + /* keep record of best compound/single-only prediction */
1.3656 + if (!disable_skip && ref_frame != INTRA_FRAME) {
1.3657 + int single_rd, hybrid_rd, single_rate, hybrid_rate;
1.3658 +
1.3659 + if (cm->comp_pred_mode == HYBRID_PREDICTION) {
1.3660 + single_rate = rate2 - compmode_cost;
1.3661 + hybrid_rate = rate2;
1.3662 + } else {
1.3663 + single_rate = rate2;
1.3664 + hybrid_rate = rate2 + compmode_cost;
1.3665 + }
1.3666 +
1.3667 + single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
1.3668 + hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
1.3669 +
1.3670 + if (second_ref_frame <= INTRA_FRAME &&
1.3671 + single_rd < best_pred_rd[SINGLE_PREDICTION_ONLY]) {
1.3672 + best_pred_rd[SINGLE_PREDICTION_ONLY] = single_rd;
1.3673 + } else if (second_ref_frame > INTRA_FRAME &&
1.3674 + single_rd < best_pred_rd[COMP_PREDICTION_ONLY]) {
1.3675 + best_pred_rd[COMP_PREDICTION_ONLY] = single_rd;
1.3676 + }
1.3677 + if (hybrid_rd < best_pred_rd[HYBRID_PREDICTION])
1.3678 + best_pred_rd[HYBRID_PREDICTION] = hybrid_rd;
1.3679 + }
1.3680 +
1.3681 + /* keep record of best filter type */
1.3682 + if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
1.3683 + cm->mcomp_filter_type != BILINEAR) {
1.3684 + int64_t ref = cpi->rd_filter_cache[cm->mcomp_filter_type == SWITCHABLE ?
1.3685 + SWITCHABLE_FILTERS : cm->mcomp_filter_type];
1.3686 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
1.3687 + int64_t adj_rd;
1.3688 + // In cases of poor prediction, filter_cache[] can contain really big
1.3689 + // values, which actually are bigger than this_rd itself. This can
1.3690 + // cause negative best_filter_rd[] values, which is obviously silly.
1.3691 + // Therefore, if filter_cache < ref, we do an adjusted calculation.
1.3692 + if (cpi->rd_filter_cache[i] >= ref) {
1.3693 + adj_rd = this_rd + cpi->rd_filter_cache[i] - ref;
1.3694 + } else {
1.3695 + // FIXME(rbultje) do this for comppsred also
1.3696 + //
1.3697 + // To prevent out-of-range computation in
1.3698 + // adj_rd = cpi->rd_filter_cache[i] * this_rd / ref
1.3699 + // cpi->rd_filter_cache[i] / ref is converted to a 256 based ratio.
1.3700 + int tmp = cpi->rd_filter_cache[i] * 256 / ref;
1.3701 + adj_rd = (this_rd * tmp) >> 8;
1.3702 + }
1.3703 + best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
1.3704 + }
1.3705 + }
1.3706 +
1.3707 + /* keep record of best txfm size */
1.3708 + if (bsize < BLOCK_32X32) {
1.3709 + if (bsize < BLOCK_16X16)
1.3710 + tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
1.3711 +
1.3712 + tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
1.3713 + }
1.3714 + if (!mode_excluded && this_rd != INT64_MAX) {
1.3715 + for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
1.3716 + int64_t adj_rd = INT64_MAX;
1.3717 + adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
1.3718 +
1.3719 + if (adj_rd < best_tx_rd[i])
1.3720 + best_tx_rd[i] = adj_rd;
1.3721 + }
1.3722 + }
1.3723 +
1.3724 + if (early_term)
1.3725 + break;
1.3726 +
1.3727 + if (x->skip && !comp_pred)
1.3728 + break;
1.3729 + }
1.3730 +
1.3731 + if (best_rd >= best_rd_so_far)
1.3732 + return INT64_MAX;
1.3733 +
1.3734 + // If we used an estimate for the uv intra rd in the loop above...
1.3735 + if (cpi->sf.use_uv_intra_rd_estimate) {
1.3736 + // Do Intra UV best rd mode selection if best mode choice above was intra.
1.3737 + if (vp9_mode_order[best_mode_index].ref_frame == INTRA_FRAME) {
1.3738 + TX_SIZE uv_tx_size = get_uv_tx_size(mbmi);
1.3739 + rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
1.3740 + &rate_uv_tokenonly[uv_tx_size],
1.3741 + &dist_uv[uv_tx_size],
1.3742 + &skip_uv[uv_tx_size],
1.3743 + bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1.3744 + }
1.3745 + }
1.3746 +
1.3747 + // If we are using reference masking and the set mask flag is set then
1.3748 + // create the reference frame mask.
1.3749 + if (cpi->sf.reference_masking && cpi->set_ref_frame_mask)
1.3750 + cpi->ref_frame_mask = ~(1 << vp9_mode_order[best_mode_index].ref_frame);
1.3751 +
1.3752 + // Flag all modes that have a distortion thats > 2x the best we found at
1.3753 + // this level.
1.3754 + for (mode_index = 0; mode_index < MB_MODE_COUNT; ++mode_index) {
1.3755 + if (mode_index == NEARESTMV || mode_index == NEARMV || mode_index == NEWMV)
1.3756 + continue;
1.3757 +
1.3758 + if (mode_distortions[mode_index] > 2 * *returndistortion) {
1.3759 + ctx->modes_with_high_error |= (1 << mode_index);
1.3760 + }
1.3761 + }
1.3762 +
1.3763 + // Flag all ref frames that have a distortion thats > 2x the best we found at
1.3764 + // this level.
1.3765 + for (ref_frame = INTRA_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
1.3766 + if (frame_distortions[ref_frame] > 2 * *returndistortion) {
1.3767 + ctx->frames_with_high_error |= (1 << ref_frame);
1.3768 + }
1.3769 + }
1.3770 +
1.3771 + assert((cm->mcomp_filter_type == SWITCHABLE) ||
1.3772 + (cm->mcomp_filter_type == best_mbmode.interp_filter) ||
1.3773 + (best_mbmode.ref_frame[0] == INTRA_FRAME));
1.3774 +
1.3775 + // Updating rd_thresh_freq_fact[] here means that the different
1.3776 + // partition/block sizes are handled independently based on the best
1.3777 + // choice for the current partition. It may well be better to keep a scaled
1.3778 + // best rd so far value and update rd_thresh_freq_fact based on the mode/size
1.3779 + // combination that wins out.
1.3780 + if (cpi->sf.adaptive_rd_thresh) {
1.3781 + for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
1.3782 + if (mode_index == best_mode_index) {
1.3783 + cpi->rd_thresh_freq_fact[bsize][mode_index] -=
1.3784 + (cpi->rd_thresh_freq_fact[bsize][mode_index] >> 3);
1.3785 + } else {
1.3786 + cpi->rd_thresh_freq_fact[bsize][mode_index] += RD_THRESH_INC;
1.3787 + if (cpi->rd_thresh_freq_fact[bsize][mode_index] >
1.3788 + (cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT)) {
1.3789 + cpi->rd_thresh_freq_fact[bsize][mode_index] =
1.3790 + cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT;
1.3791 + }
1.3792 + }
1.3793 + }
1.3794 + }
1.3795 +
1.3796 + // macroblock modes
1.3797 + *mbmi = best_mbmode;
1.3798 + x->skip |= best_skip2;
1.3799 +
1.3800 + for (i = 0; i < NB_PREDICTION_TYPES; ++i) {
1.3801 + if (best_pred_rd[i] == INT64_MAX)
1.3802 + best_pred_diff[i] = INT_MIN;
1.3803 + else
1.3804 + best_pred_diff[i] = best_rd - best_pred_rd[i];
1.3805 + }
1.3806 +
1.3807 + if (!x->skip) {
1.3808 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
1.3809 + if (best_filter_rd[i] == INT64_MAX)
1.3810 + best_filter_diff[i] = 0;
1.3811 + else
1.3812 + best_filter_diff[i] = best_rd - best_filter_rd[i];
1.3813 + }
1.3814 + if (cm->mcomp_filter_type == SWITCHABLE)
1.3815 + assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
1.3816 + } else {
1.3817 + vp9_zero(best_filter_diff);
1.3818 + }
1.3819 +
1.3820 + if (!x->skip) {
1.3821 + for (i = 0; i < TX_MODES; i++) {
1.3822 + if (best_tx_rd[i] == INT64_MAX)
1.3823 + best_tx_diff[i] = 0;
1.3824 + else
1.3825 + best_tx_diff[i] = best_rd - best_tx_rd[i];
1.3826 + }
1.3827 + } else {
1.3828 + vp9_zero(best_tx_diff);
1.3829 + }
1.3830 +
1.3831 + set_scale_factors(xd, mbmi->ref_frame[0], mbmi->ref_frame[1],
1.3832 + scale_factor);
1.3833 + store_coding_context(x, ctx, best_mode_index,
1.3834 + &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
1.3835 + &mbmi->ref_mvs[mbmi->ref_frame[1] < 0 ? 0 :
1.3836 + mbmi->ref_frame[1]][0],
1.3837 + best_pred_diff, best_tx_diff, best_filter_diff);
1.3838 +
1.3839 + return best_rd;
1.3840 +}
1.3841 +
1.3842 +
1.3843 +int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
1.3844 + const TileInfo *const tile,
1.3845 + int mi_row, int mi_col,
1.3846 + int *returnrate,
1.3847 + int64_t *returndistortion,
1.3848 + BLOCK_SIZE bsize,
1.3849 + PICK_MODE_CONTEXT *ctx,
1.3850 + int64_t best_rd_so_far) {
1.3851 + VP9_COMMON *cm = &cpi->common;
1.3852 + MACROBLOCKD *xd = &x->e_mbd;
1.3853 + MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
1.3854 + const struct segmentation *seg = &cm->seg;
1.3855 + const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
1.3856 + MV_REFERENCE_FRAME ref_frame, second_ref_frame;
1.3857 + unsigned char segment_id = mbmi->segment_id;
1.3858 + int comp_pred, i;
1.3859 + int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1.3860 + struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1.3861 + static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1.3862 + VP9_ALT_FLAG };
1.3863 + int idx_list[4] = {0,
1.3864 + cpi->lst_fb_idx,
1.3865 + cpi->gld_fb_idx,
1.3866 + cpi->alt_fb_idx};
1.3867 + int64_t best_rd = best_rd_so_far;
1.3868 + int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
1.3869 + int64_t best_tx_rd[TX_MODES];
1.3870 + int64_t best_tx_diff[TX_MODES];
1.3871 + int64_t best_pred_diff[NB_PREDICTION_TYPES];
1.3872 + int64_t best_pred_rd[NB_PREDICTION_TYPES];
1.3873 + int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
1.3874 + int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
1.3875 + MB_MODE_INFO best_mbmode = { 0 };
1.3876 + int mode_index, best_mode_index = 0;
1.3877 + unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
1.3878 + vp9_prob comp_mode_p;
1.3879 + int64_t best_inter_rd = INT64_MAX;
1.3880 + MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
1.3881 + INTERPOLATION_TYPE tmp_best_filter = SWITCHABLE;
1.3882 + int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
1.3883 + int64_t dist_uv[TX_SIZES];
1.3884 + int skip_uv[TX_SIZES];
1.3885 + MB_PREDICTION_MODE mode_uv[TX_SIZES] = { 0 };
1.3886 + struct scale_factors scale_factor[4];
1.3887 + unsigned int ref_frame_mask = 0;
1.3888 + unsigned int mode_mask = 0;
1.3889 + int intra_cost_penalty = 20 * vp9_dc_quant(cpi->common.base_qindex,
1.3890 + cpi->common.y_dc_delta_q);
1.3891 + int_mv seg_mvs[4][MAX_REF_FRAMES];
1.3892 + b_mode_info best_bmodes[4];
1.3893 + int best_skip2 = 0;
1.3894 +
1.3895 + x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1.3896 + vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
1.3897 +
1.3898 + for (i = 0; i < 4; i++) {
1.3899 + int j;
1.3900 + for (j = 0; j < MAX_REF_FRAMES; j++)
1.3901 + seg_mvs[i][j].as_int = INVALID_MV;
1.3902 + }
1.3903 +
1.3904 + estimate_ref_frame_costs(cpi, segment_id, ref_costs_single, ref_costs_comp,
1.3905 + &comp_mode_p);
1.3906 +
1.3907 + for (i = 0; i < NB_PREDICTION_TYPES; ++i)
1.3908 + best_pred_rd[i] = INT64_MAX;
1.3909 + for (i = 0; i < TX_MODES; i++)
1.3910 + best_tx_rd[i] = INT64_MAX;
1.3911 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
1.3912 + best_filter_rd[i] = INT64_MAX;
1.3913 + for (i = 0; i < TX_SIZES; i++)
1.3914 + rate_uv_intra[i] = INT_MAX;
1.3915 +
1.3916 + *returnrate = INT_MAX;
1.3917 +
1.3918 + // Create a mask set to 1 for each reference frame used by a smaller
1.3919 + // resolution.
1.3920 + if (cpi->sf.use_avoid_tested_higherror) {
1.3921 + ref_frame_mask = 0;
1.3922 + mode_mask = 0;
1.3923 + ref_frame_mask = ~ref_frame_mask;
1.3924 + mode_mask = ~mode_mask;
1.3925 + }
1.3926 +
1.3927 + for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
1.3928 + if (cpi->ref_frame_flags & flag_list[ref_frame]) {
1.3929 + setup_buffer_inter(cpi, x, tile, idx_list[ref_frame], ref_frame,
1.3930 + block_size, mi_row, mi_col,
1.3931 + frame_mv[NEARESTMV], frame_mv[NEARMV],
1.3932 + yv12_mb, scale_factor);
1.3933 + }
1.3934 + frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1.3935 + frame_mv[ZEROMV][ref_frame].as_int = 0;
1.3936 + }
1.3937 +
1.3938 + for (mode_index = 0; mode_index < MAX_REFS; ++mode_index) {
1.3939 + int mode_excluded = 0;
1.3940 + int64_t this_rd = INT64_MAX;
1.3941 + int disable_skip = 0;
1.3942 + int compmode_cost = 0;
1.3943 + int rate2 = 0, rate_y = 0, rate_uv = 0;
1.3944 + int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
1.3945 + int skippable = 0;
1.3946 + int64_t tx_cache[TX_MODES];
1.3947 + int i;
1.3948 + int this_skip2 = 0;
1.3949 + int64_t total_sse = INT_MAX;
1.3950 + int early_term = 0;
1.3951 +
1.3952 + for (i = 0; i < TX_MODES; ++i)
1.3953 + tx_cache[i] = INT64_MAX;
1.3954 +
1.3955 + x->skip = 0;
1.3956 + ref_frame = vp9_ref_order[mode_index].ref_frame;
1.3957 + second_ref_frame = vp9_ref_order[mode_index].second_ref_frame;
1.3958 +
1.3959 + // Look at the reference frame of the best mode so far and set the
1.3960 + // skip mask to look at a subset of the remaining modes.
1.3961 + if (mode_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
1.3962 + if (mode_index == 3) {
1.3963 + switch (vp9_ref_order[best_mode_index].ref_frame) {
1.3964 + case INTRA_FRAME:
1.3965 + cpi->mode_skip_mask = 0;
1.3966 + break;
1.3967 + case LAST_FRAME:
1.3968 + cpi->mode_skip_mask = 0x0010;
1.3969 + break;
1.3970 + case GOLDEN_FRAME:
1.3971 + cpi->mode_skip_mask = 0x0008;
1.3972 + break;
1.3973 + case ALTREF_FRAME:
1.3974 + cpi->mode_skip_mask = 0x0000;
1.3975 + break;
1.3976 + case NONE:
1.3977 + case MAX_REF_FRAMES:
1.3978 + assert(!"Invalid Reference frame");
1.3979 + }
1.3980 + }
1.3981 + if (cpi->mode_skip_mask & ((int64_t)1 << mode_index))
1.3982 + continue;
1.3983 + }
1.3984 +
1.3985 + // Skip if the current reference frame has been masked off
1.3986 + if (cpi->sf.reference_masking && !cpi->set_ref_frame_mask &&
1.3987 + (cpi->ref_frame_mask & (1 << ref_frame)))
1.3988 + continue;
1.3989 +
1.3990 + // Test best rd so far against threshold for trying this mode.
1.3991 + if ((best_rd <
1.3992 + ((int64_t)cpi->rd_thresh_sub8x8[segment_id][bsize][mode_index] *
1.3993 + cpi->rd_thresh_freq_sub8x8[bsize][mode_index] >> 5)) ||
1.3994 + cpi->rd_thresh_sub8x8[segment_id][bsize][mode_index] == INT_MAX)
1.3995 + continue;
1.3996 +
1.3997 + // Do not allow compound prediction if the segment level reference
1.3998 + // frame feature is in use as in this case there can only be one reference.
1.3999 + if ((second_ref_frame > INTRA_FRAME) &&
1.4000 + vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
1.4001 + continue;
1.4002 +
1.4003 + mbmi->ref_frame[0] = ref_frame;
1.4004 + mbmi->ref_frame[1] = second_ref_frame;
1.4005 +
1.4006 + if (!(ref_frame == INTRA_FRAME
1.4007 + || (cpi->ref_frame_flags & flag_list[ref_frame]))) {
1.4008 + continue;
1.4009 + }
1.4010 + if (!(second_ref_frame == NONE
1.4011 + || (cpi->ref_frame_flags & flag_list[second_ref_frame]))) {
1.4012 + continue;
1.4013 + }
1.4014 +
1.4015 + comp_pred = second_ref_frame > INTRA_FRAME;
1.4016 + if (comp_pred) {
1.4017 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA)
1.4018 + if (vp9_ref_order[best_mode_index].ref_frame == INTRA_FRAME)
1.4019 + continue;
1.4020 + if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH)
1.4021 + if (ref_frame != best_inter_ref_frame &&
1.4022 + second_ref_frame != best_inter_ref_frame)
1.4023 + continue;
1.4024 + }
1.4025 +
1.4026 + // TODO(jingning, jkoleszar): scaling reference frame not supported for
1.4027 + // sub8x8 blocks.
1.4028 + if (ref_frame > 0 &&
1.4029 + vp9_is_scaled(scale_factor[ref_frame].sfc))
1.4030 + continue;
1.4031 +
1.4032 + if (second_ref_frame > 0 &&
1.4033 + vp9_is_scaled(scale_factor[second_ref_frame].sfc))
1.4034 + continue;
1.4035 +
1.4036 + set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
1.4037 + mbmi->uv_mode = DC_PRED;
1.4038 +
1.4039 + // Evaluate all sub-pel filters irrespective of whether we can use
1.4040 + // them for this frame.
1.4041 + mbmi->interp_filter = cm->mcomp_filter_type;
1.4042 + vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
1.4043 +
1.4044 + if (comp_pred) {
1.4045 + if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
1.4046 + continue;
1.4047 + set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
1.4048 +
1.4049 + mode_excluded = mode_excluded
1.4050 + ? mode_excluded
1.4051 + : cm->comp_pred_mode == SINGLE_PREDICTION_ONLY;
1.4052 + } else {
1.4053 + if (ref_frame != INTRA_FRAME && second_ref_frame != INTRA_FRAME) {
1.4054 + mode_excluded =
1.4055 + mode_excluded ?
1.4056 + mode_excluded : cm->comp_pred_mode == COMP_PREDICTION_ONLY;
1.4057 + }
1.4058 + }
1.4059 +
1.4060 + // Select prediction reference frames.
1.4061 + for (i = 0; i < MAX_MB_PLANE; i++) {
1.4062 + xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1.4063 + if (comp_pred)
1.4064 + xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
1.4065 + }
1.4066 +
1.4067 + // If the segment reference frame feature is enabled....
1.4068 + // then do nothing if the current ref frame is not allowed..
1.4069 + if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
1.4070 + vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) !=
1.4071 + (int)ref_frame) {
1.4072 + continue;
1.4073 + // If the segment skip feature is enabled....
1.4074 + // then do nothing if the current mode is not allowed..
1.4075 + } else if (vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) &&
1.4076 + ref_frame != INTRA_FRAME) {
1.4077 + continue;
1.4078 + // Disable this drop out case if the ref frame
1.4079 + // segment level feature is enabled for this segment. This is to
1.4080 + // prevent the possibility that we end up unable to pick any mode.
1.4081 + } else if (!vp9_segfeature_active(seg, segment_id,
1.4082 + SEG_LVL_REF_FRAME)) {
1.4083 + // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
1.4084 + // unless ARNR filtering is enabled in which case we want
1.4085 + // an unfiltered alternative. We allow near/nearest as well
1.4086 + // because they may result in zero-zero MVs but be cheaper.
1.4087 + if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
1.4088 + continue;
1.4089 + }
1.4090 +
1.4091 +#ifdef MODE_TEST_HIT_STATS
1.4092 + // TEST/DEBUG CODE
1.4093 + // Keep a rcord of the number of test hits at each size
1.4094 + cpi->mode_test_hits[bsize]++;
1.4095 +#endif
1.4096 +
1.4097 + if (ref_frame == INTRA_FRAME) {
1.4098 + int rate;
1.4099 + mbmi->tx_size = TX_4X4;
1.4100 + if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
1.4101 + &distortion_y, best_rd) >= best_rd)
1.4102 + continue;
1.4103 + rate2 += rate;
1.4104 + rate2 += intra_cost_penalty;
1.4105 + distortion2 += distortion_y;
1.4106 +
1.4107 + if (rate_uv_intra[TX_4X4] == INT_MAX) {
1.4108 + choose_intra_uv_mode(cpi, ctx, bsize, &rate_uv_intra[TX_4X4],
1.4109 + &rate_uv_tokenonly[TX_4X4],
1.4110 + &dist_uv[TX_4X4], &skip_uv[TX_4X4],
1.4111 + &mode_uv[TX_4X4]);
1.4112 + }
1.4113 + rate2 += rate_uv_intra[TX_4X4];
1.4114 + rate_uv = rate_uv_tokenonly[TX_4X4];
1.4115 + distortion2 += dist_uv[TX_4X4];
1.4116 + distortion_uv = dist_uv[TX_4X4];
1.4117 + mbmi->uv_mode = mode_uv[TX_4X4];
1.4118 + tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
1.4119 + for (i = 0; i < TX_MODES; ++i)
1.4120 + tx_cache[i] = tx_cache[ONLY_4X4];
1.4121 + } else {
1.4122 + int rate;
1.4123 + int64_t distortion;
1.4124 + int64_t this_rd_thresh;
1.4125 + int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
1.4126 + int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
1.4127 + int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
1.4128 + int tmp_best_skippable = 0;
1.4129 + int switchable_filter_index;
1.4130 + int_mv *second_ref = comp_pred ?
1.4131 + &mbmi->ref_mvs[second_ref_frame][0] : NULL;
1.4132 + b_mode_info tmp_best_bmodes[16];
1.4133 + MB_MODE_INFO tmp_best_mbmode;
1.4134 + BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
1.4135 + int pred_exists = 0;
1.4136 + int uv_skippable;
1.4137 +
1.4138 + this_rd_thresh = (ref_frame == LAST_FRAME) ?
1.4139 + cpi->rd_thresh_sub8x8[segment_id][bsize][THR_LAST] :
1.4140 + cpi->rd_thresh_sub8x8[segment_id][bsize][THR_ALTR];
1.4141 + this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
1.4142 + cpi->rd_thresh_sub8x8[segment_id][bsize][THR_GOLD] : this_rd_thresh;
1.4143 + xd->mi_8x8[0]->mbmi.tx_size = TX_4X4;
1.4144 +
1.4145 + cpi->rd_filter_cache[SWITCHABLE_FILTERS] = INT64_MAX;
1.4146 + if (cm->mcomp_filter_type != BILINEAR) {
1.4147 + tmp_best_filter = EIGHTTAP;
1.4148 + if (x->source_variance <
1.4149 + cpi->sf.disable_filter_search_var_thresh) {
1.4150 + tmp_best_filter = EIGHTTAP;
1.4151 + vp9_zero(cpi->rd_filter_cache);
1.4152 + } else {
1.4153 + for (switchable_filter_index = 0;
1.4154 + switchable_filter_index < SWITCHABLE_FILTERS;
1.4155 + ++switchable_filter_index) {
1.4156 + int newbest, rs;
1.4157 + int64_t rs_rd;
1.4158 + mbmi->interp_filter = switchable_filter_index;
1.4159 + vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
1.4160 +
1.4161 + tmp_rd = rd_pick_best_mbsegmentation(cpi, x, tile,
1.4162 + &mbmi->ref_mvs[ref_frame][0],
1.4163 + second_ref,
1.4164 + best_yrd,
1.4165 + &rate, &rate_y, &distortion,
1.4166 + &skippable, &total_sse,
1.4167 + (int)this_rd_thresh, seg_mvs,
1.4168 + bsi, switchable_filter_index,
1.4169 + mi_row, mi_col);
1.4170 +
1.4171 + if (tmp_rd == INT64_MAX)
1.4172 + continue;
1.4173 + cpi->rd_filter_cache[switchable_filter_index] = tmp_rd;
1.4174 + rs = get_switchable_rate(x);
1.4175 + rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
1.4176 + cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
1.4177 + MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
1.4178 + tmp_rd + rs_rd);
1.4179 + if (cm->mcomp_filter_type == SWITCHABLE)
1.4180 + tmp_rd += rs_rd;
1.4181 +
1.4182 + newbest = (tmp_rd < tmp_best_rd);
1.4183 + if (newbest) {
1.4184 + tmp_best_filter = mbmi->interp_filter;
1.4185 + tmp_best_rd = tmp_rd;
1.4186 + }
1.4187 + if ((newbest && cm->mcomp_filter_type == SWITCHABLE) ||
1.4188 + (mbmi->interp_filter == cm->mcomp_filter_type &&
1.4189 + cm->mcomp_filter_type != SWITCHABLE)) {
1.4190 + tmp_best_rdu = tmp_rd;
1.4191 + tmp_best_rate = rate;
1.4192 + tmp_best_ratey = rate_y;
1.4193 + tmp_best_distortion = distortion;
1.4194 + tmp_best_sse = total_sse;
1.4195 + tmp_best_skippable = skippable;
1.4196 + tmp_best_mbmode = *mbmi;
1.4197 + for (i = 0; i < 4; i++) {
1.4198 + tmp_best_bmodes[i] = xd->mi_8x8[0]->bmi[i];
1.4199 + x->zcoeff_blk[TX_4X4][i] = !xd->plane[0].eobs[i];
1.4200 + }
1.4201 + pred_exists = 1;
1.4202 + if (switchable_filter_index == 0 &&
1.4203 + cpi->sf.use_rd_breakout &&
1.4204 + best_rd < INT64_MAX) {
1.4205 + if (tmp_best_rdu / 2 > best_rd) {
1.4206 + // skip searching the other filters if the first is
1.4207 + // already substantially larger than the best so far
1.4208 + tmp_best_filter = mbmi->interp_filter;
1.4209 + tmp_best_rdu = INT64_MAX;
1.4210 + break;
1.4211 + }
1.4212 + }
1.4213 + }
1.4214 + } // switchable_filter_index loop
1.4215 + }
1.4216 + }
1.4217 +
1.4218 + if (tmp_best_rdu == INT64_MAX)
1.4219 + continue;
1.4220 +
1.4221 + mbmi->interp_filter = (cm->mcomp_filter_type == SWITCHABLE ?
1.4222 + tmp_best_filter : cm->mcomp_filter_type);
1.4223 + vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
1.4224 + if (!pred_exists) {
1.4225 + // Handles the special case when a filter that is not in the
1.4226 + // switchable list (bilinear, 6-tap) is indicated at the frame level
1.4227 + tmp_rd = rd_pick_best_mbsegmentation(cpi, x, tile,
1.4228 + &mbmi->ref_mvs[ref_frame][0],
1.4229 + second_ref,
1.4230 + best_yrd,
1.4231 + &rate, &rate_y, &distortion,
1.4232 + &skippable, &total_sse,
1.4233 + (int)this_rd_thresh, seg_mvs,
1.4234 + bsi, 0,
1.4235 + mi_row, mi_col);
1.4236 + if (tmp_rd == INT64_MAX)
1.4237 + continue;
1.4238 + } else {
1.4239 + if (cpi->common.mcomp_filter_type == SWITCHABLE) {
1.4240 + int rs = get_switchable_rate(x);
1.4241 + tmp_best_rdu -= RDCOST(x->rdmult, x->rddiv, rs, 0);
1.4242 + }
1.4243 + tmp_rd = tmp_best_rdu;
1.4244 + total_sse = tmp_best_sse;
1.4245 + rate = tmp_best_rate;
1.4246 + rate_y = tmp_best_ratey;
1.4247 + distortion = tmp_best_distortion;
1.4248 + skippable = tmp_best_skippable;
1.4249 + *mbmi = tmp_best_mbmode;
1.4250 + for (i = 0; i < 4; i++)
1.4251 + xd->mi_8x8[0]->bmi[i] = tmp_best_bmodes[i];
1.4252 + }
1.4253 +
1.4254 + rate2 += rate;
1.4255 + distortion2 += distortion;
1.4256 +
1.4257 + if (cpi->common.mcomp_filter_type == SWITCHABLE)
1.4258 + rate2 += get_switchable_rate(x);
1.4259 +
1.4260 + if (!mode_excluded) {
1.4261 + if (comp_pred)
1.4262 + mode_excluded = cpi->common.comp_pred_mode == SINGLE_PREDICTION_ONLY;
1.4263 + else
1.4264 + mode_excluded = cpi->common.comp_pred_mode == COMP_PREDICTION_ONLY;
1.4265 + }
1.4266 + compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
1.4267 +
1.4268 + tmp_best_rdu = best_rd -
1.4269 + MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
1.4270 + RDCOST(x->rdmult, x->rddiv, 0, total_sse));
1.4271 +
1.4272 + if (tmp_best_rdu > 0) {
1.4273 + // If even the 'Y' rd value of split is higher than best so far
1.4274 + // then dont bother looking at UV
1.4275 + vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
1.4276 + BLOCK_8X8);
1.4277 + super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
1.4278 + &uv_sse, BLOCK_8X8, tmp_best_rdu);
1.4279 + if (rate_uv == INT_MAX)
1.4280 + continue;
1.4281 + rate2 += rate_uv;
1.4282 + distortion2 += distortion_uv;
1.4283 + skippable = skippable && uv_skippable;
1.4284 + total_sse += uv_sse;
1.4285 +
1.4286 + tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
1.4287 + for (i = 0; i < TX_MODES; ++i)
1.4288 + tx_cache[i] = tx_cache[ONLY_4X4];
1.4289 + }
1.4290 + }
1.4291 +
1.4292 + if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
1.4293 + rate2 += compmode_cost;
1.4294 + }
1.4295 +
1.4296 + // Estimate the reference frame signaling cost and add it
1.4297 + // to the rolling cost variable.
1.4298 + if (second_ref_frame > INTRA_FRAME) {
1.4299 + rate2 += ref_costs_comp[ref_frame];
1.4300 + } else {
1.4301 + rate2 += ref_costs_single[ref_frame];
1.4302 + }
1.4303 +
1.4304 + if (!disable_skip) {
1.4305 + // Test for the condition where skip block will be activated
1.4306 + // because there are no non zero coefficients and make any
1.4307 + // necessary adjustment for rate. Ignore if skip is coded at
1.4308 + // segment level as the cost wont have been added in.
1.4309 + // Is Mb level skip allowed (i.e. not coded at segment level).
1.4310 + const int mb_skip_allowed = !vp9_segfeature_active(seg, segment_id,
1.4311 + SEG_LVL_SKIP);
1.4312 +
1.4313 + if (mb_skip_allowed && ref_frame != INTRA_FRAME && !xd->lossless) {
1.4314 + if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
1.4315 + RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
1.4316 + // Add in the cost of the no skip flag.
1.4317 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.4318 + 0);
1.4319 + rate2 += prob_skip_cost;
1.4320 + } else {
1.4321 + // FIXME(rbultje) make this work for splitmv also
1.4322 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.4323 + 1);
1.4324 + rate2 += prob_skip_cost;
1.4325 + distortion2 = total_sse;
1.4326 + assert(total_sse >= 0);
1.4327 + rate2 -= (rate_y + rate_uv);
1.4328 + rate_y = 0;
1.4329 + rate_uv = 0;
1.4330 + this_skip2 = 1;
1.4331 + }
1.4332 + } else if (mb_skip_allowed) {
1.4333 + // Add in the cost of the no skip flag.
1.4334 + int prob_skip_cost = vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd),
1.4335 + 0);
1.4336 + rate2 += prob_skip_cost;
1.4337 + }
1.4338 +
1.4339 + // Calculate the final RD estimate for this mode.
1.4340 + this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
1.4341 + }
1.4342 +
1.4343 + // Keep record of best inter rd with single reference
1.4344 + if (xd->mi_8x8[0]->mbmi.ref_frame[0] > INTRA_FRAME &&
1.4345 + xd->mi_8x8[0]->mbmi.ref_frame[1] == NONE &&
1.4346 + !mode_excluded &&
1.4347 + this_rd < best_inter_rd) {
1.4348 + best_inter_rd = this_rd;
1.4349 + best_inter_ref_frame = ref_frame;
1.4350 + }
1.4351 +
1.4352 + if (!disable_skip && ref_frame == INTRA_FRAME) {
1.4353 + for (i = 0; i < NB_PREDICTION_TYPES; ++i)
1.4354 + best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
1.4355 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
1.4356 + best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
1.4357 + }
1.4358 +
1.4359 + // Did this mode help.. i.e. is it the new best mode
1.4360 + if (this_rd < best_rd || x->skip) {
1.4361 + if (!mode_excluded) {
1.4362 + int max_plane = MAX_MB_PLANE;
1.4363 + // Note index of best mode so far
1.4364 + best_mode_index = mode_index;
1.4365 +
1.4366 + if (ref_frame == INTRA_FRAME) {
1.4367 + /* required for left and above block mv */
1.4368 + mbmi->mv[0].as_int = 0;
1.4369 + max_plane = 1;
1.4370 + }
1.4371 +
1.4372 + *returnrate = rate2;
1.4373 + *returndistortion = distortion2;
1.4374 + best_rd = this_rd;
1.4375 + best_yrd = best_rd -
1.4376 + RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
1.4377 + best_mbmode = *mbmi;
1.4378 + best_skip2 = this_skip2;
1.4379 + if (!x->select_txfm_size)
1.4380 + swap_block_ptr(x, ctx, max_plane);
1.4381 + vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
1.4382 + sizeof(uint8_t) * ctx->num_4x4_blk);
1.4383 +
1.4384 + for (i = 0; i < 4; i++)
1.4385 + best_bmodes[i] = xd->mi_8x8[0]->bmi[i];
1.4386 +
1.4387 + // TODO(debargha): enhance this test with a better distortion prediction
1.4388 + // based on qp, activity mask and history
1.4389 + if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
1.4390 + (mode_index > MIN_EARLY_TERM_INDEX)) {
1.4391 + const int qstep = xd->plane[0].dequant[1];
1.4392 + // TODO(debargha): Enhance this by specializing for each mode_index
1.4393 + int scale = 4;
1.4394 + if (x->source_variance < UINT_MAX) {
1.4395 + const int var_adjust = (x->source_variance < 16);
1.4396 + scale -= var_adjust;
1.4397 + }
1.4398 + if (ref_frame > INTRA_FRAME &&
1.4399 + distortion2 * scale < qstep * qstep) {
1.4400 + early_term = 1;
1.4401 + }
1.4402 + }
1.4403 + }
1.4404 + }
1.4405 +
1.4406 + /* keep record of best compound/single-only prediction */
1.4407 + if (!disable_skip && ref_frame != INTRA_FRAME) {
1.4408 + int single_rd, hybrid_rd, single_rate, hybrid_rate;
1.4409 +
1.4410 + if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
1.4411 + single_rate = rate2 - compmode_cost;
1.4412 + hybrid_rate = rate2;
1.4413 + } else {
1.4414 + single_rate = rate2;
1.4415 + hybrid_rate = rate2 + compmode_cost;
1.4416 + }
1.4417 +
1.4418 + single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
1.4419 + hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
1.4420 +
1.4421 + if (second_ref_frame <= INTRA_FRAME &&
1.4422 + single_rd < best_pred_rd[SINGLE_PREDICTION_ONLY]) {
1.4423 + best_pred_rd[SINGLE_PREDICTION_ONLY] = single_rd;
1.4424 + } else if (second_ref_frame > INTRA_FRAME &&
1.4425 + single_rd < best_pred_rd[COMP_PREDICTION_ONLY]) {
1.4426 + best_pred_rd[COMP_PREDICTION_ONLY] = single_rd;
1.4427 + }
1.4428 + if (hybrid_rd < best_pred_rd[HYBRID_PREDICTION])
1.4429 + best_pred_rd[HYBRID_PREDICTION] = hybrid_rd;
1.4430 + }
1.4431 +
1.4432 + /* keep record of best filter type */
1.4433 + if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
1.4434 + cm->mcomp_filter_type != BILINEAR) {
1.4435 + int64_t ref = cpi->rd_filter_cache[cm->mcomp_filter_type == SWITCHABLE ?
1.4436 + SWITCHABLE_FILTERS : cm->mcomp_filter_type];
1.4437 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
1.4438 + int64_t adj_rd;
1.4439 + // In cases of poor prediction, filter_cache[] can contain really big
1.4440 + // values, which actually are bigger than this_rd itself. This can
1.4441 + // cause negative best_filter_rd[] values, which is obviously silly.
1.4442 + // Therefore, if filter_cache < ref, we do an adjusted calculation.
1.4443 + if (cpi->rd_filter_cache[i] >= ref)
1.4444 + adj_rd = this_rd + cpi->rd_filter_cache[i] - ref;
1.4445 + else // FIXME(rbultje) do this for comppred also
1.4446 + adj_rd = this_rd - (ref - cpi->rd_filter_cache[i]) * this_rd / ref;
1.4447 + best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
1.4448 + }
1.4449 + }
1.4450 +
1.4451 + /* keep record of best txfm size */
1.4452 + if (bsize < BLOCK_32X32) {
1.4453 + if (bsize < BLOCK_16X16) {
1.4454 + tx_cache[ALLOW_8X8] = tx_cache[ONLY_4X4];
1.4455 + tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
1.4456 + }
1.4457 + tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
1.4458 + }
1.4459 + if (!mode_excluded && this_rd != INT64_MAX) {
1.4460 + for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
1.4461 + int64_t adj_rd = INT64_MAX;
1.4462 + if (ref_frame > INTRA_FRAME)
1.4463 + adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
1.4464 + else
1.4465 + adj_rd = this_rd;
1.4466 +
1.4467 + if (adj_rd < best_tx_rd[i])
1.4468 + best_tx_rd[i] = adj_rd;
1.4469 + }
1.4470 + }
1.4471 +
1.4472 + if (early_term)
1.4473 + break;
1.4474 +
1.4475 + if (x->skip && !comp_pred)
1.4476 + break;
1.4477 + }
1.4478 +
1.4479 + if (best_rd >= best_rd_so_far)
1.4480 + return INT64_MAX;
1.4481 +
1.4482 + // If we used an estimate for the uv intra rd in the loop above...
1.4483 + if (cpi->sf.use_uv_intra_rd_estimate) {
1.4484 + // Do Intra UV best rd mode selection if best mode choice above was intra.
1.4485 + if (vp9_ref_order[best_mode_index].ref_frame == INTRA_FRAME) {
1.4486 + TX_SIZE uv_tx_size = get_uv_tx_size(mbmi);
1.4487 + rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
1.4488 + &rate_uv_tokenonly[uv_tx_size],
1.4489 + &dist_uv[uv_tx_size],
1.4490 + &skip_uv[uv_tx_size],
1.4491 + BLOCK_8X8);
1.4492 + }
1.4493 + }
1.4494 +
1.4495 + // If we are using reference masking and the set mask flag is set then
1.4496 + // create the reference frame mask.
1.4497 + if (cpi->sf.reference_masking && cpi->set_ref_frame_mask)
1.4498 + cpi->ref_frame_mask = ~(1 << vp9_ref_order[best_mode_index].ref_frame);
1.4499 +
1.4500 + if (best_rd == INT64_MAX && bsize < BLOCK_8X8) {
1.4501 + *returnrate = INT_MAX;
1.4502 + *returndistortion = INT_MAX;
1.4503 + return best_rd;
1.4504 + }
1.4505 +
1.4506 + assert((cm->mcomp_filter_type == SWITCHABLE) ||
1.4507 + (cm->mcomp_filter_type == best_mbmode.interp_filter) ||
1.4508 + (best_mbmode.ref_frame[0] == INTRA_FRAME));
1.4509 +
1.4510 + // Updating rd_thresh_freq_fact[] here means that the different
1.4511 + // partition/block sizes are handled independently based on the best
1.4512 + // choice for the current partition. It may well be better to keep a scaled
1.4513 + // best rd so far value and update rd_thresh_freq_fact based on the mode/size
1.4514 + // combination that wins out.
1.4515 + if (cpi->sf.adaptive_rd_thresh) {
1.4516 + for (mode_index = 0; mode_index < MAX_REFS; ++mode_index) {
1.4517 + if (mode_index == best_mode_index) {
1.4518 + cpi->rd_thresh_freq_sub8x8[bsize][mode_index] -=
1.4519 + (cpi->rd_thresh_freq_sub8x8[bsize][mode_index] >> 3);
1.4520 + } else {
1.4521 + cpi->rd_thresh_freq_sub8x8[bsize][mode_index] += RD_THRESH_INC;
1.4522 + if (cpi->rd_thresh_freq_sub8x8[bsize][mode_index] >
1.4523 + (cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT)) {
1.4524 + cpi->rd_thresh_freq_sub8x8[bsize][mode_index] =
1.4525 + cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT;
1.4526 + }
1.4527 + }
1.4528 + }
1.4529 + }
1.4530 +
1.4531 + // macroblock modes
1.4532 + *mbmi = best_mbmode;
1.4533 + x->skip |= best_skip2;
1.4534 + if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
1.4535 + for (i = 0; i < 4; i++)
1.4536 + xd->mi_8x8[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
1.4537 + } else {
1.4538 + for (i = 0; i < 4; ++i)
1.4539 + vpx_memcpy(&xd->mi_8x8[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
1.4540 +
1.4541 + mbmi->mv[0].as_int = xd->mi_8x8[0]->bmi[3].as_mv[0].as_int;
1.4542 + mbmi->mv[1].as_int = xd->mi_8x8[0]->bmi[3].as_mv[1].as_int;
1.4543 + }
1.4544 +
1.4545 + for (i = 0; i < NB_PREDICTION_TYPES; ++i) {
1.4546 + if (best_pred_rd[i] == INT64_MAX)
1.4547 + best_pred_diff[i] = INT_MIN;
1.4548 + else
1.4549 + best_pred_diff[i] = best_rd - best_pred_rd[i];
1.4550 + }
1.4551 +
1.4552 + if (!x->skip) {
1.4553 + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
1.4554 + if (best_filter_rd[i] == INT64_MAX)
1.4555 + best_filter_diff[i] = 0;
1.4556 + else
1.4557 + best_filter_diff[i] = best_rd - best_filter_rd[i];
1.4558 + }
1.4559 + if (cm->mcomp_filter_type == SWITCHABLE)
1.4560 + assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
1.4561 + } else {
1.4562 + vp9_zero(best_filter_diff);
1.4563 + }
1.4564 +
1.4565 + if (!x->skip) {
1.4566 + for (i = 0; i < TX_MODES; i++) {
1.4567 + if (best_tx_rd[i] == INT64_MAX)
1.4568 + best_tx_diff[i] = 0;
1.4569 + else
1.4570 + best_tx_diff[i] = best_rd - best_tx_rd[i];
1.4571 + }
1.4572 + } else {
1.4573 + vp9_zero(best_tx_diff);
1.4574 + }
1.4575 +
1.4576 + set_scale_factors(xd, mbmi->ref_frame[0], mbmi->ref_frame[1],
1.4577 + scale_factor);
1.4578 + store_coding_context(x, ctx, best_mode_index,
1.4579 + &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
1.4580 + &mbmi->ref_mvs[mbmi->ref_frame[1] < 0 ? 0 :
1.4581 + mbmi->ref_frame[1]][0],
1.4582 + best_pred_diff, best_tx_diff, best_filter_diff);
1.4583 +
1.4584 + return best_rd;
1.4585 +}
