1/* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <assert.h> 12#include <math.h> 13 14#include "./vp9_rtcd.h" 15 16#include "vpx_mem/vpx_mem.h" 17 18#include "vp9/common/vp9_common.h" 19#include "vp9/common/vp9_entropy.h" 20#include "vp9/common/vp9_entropymode.h" 21#include "vp9/common/vp9_idct.h" 22#include "vp9/common/vp9_mvref_common.h" 23#include "vp9/common/vp9_pred_common.h" 24#include "vp9/common/vp9_quant_common.h" 25#include "vp9/common/vp9_reconinter.h" 26#include "vp9/common/vp9_reconintra.h" 27#include "vp9/common/vp9_seg_common.h" 28#include "vp9/common/vp9_systemdependent.h" 29 30#include "vp9/encoder/vp9_cost.h" 31#include "vp9/encoder/vp9_encodemb.h" 32#include "vp9/encoder/vp9_encodemv.h" 33#include "vp9/encoder/vp9_encoder.h" 34#include "vp9/encoder/vp9_mcomp.h" 35#include "vp9/encoder/vp9_quantize.h" 36#include "vp9/encoder/vp9_ratectrl.h" 37#include "vp9/encoder/vp9_rd.h" 38#include "vp9/encoder/vp9_rdopt.h" 39#include "vp9/encoder/vp9_variance.h" 40 41#define RD_THRESH_MAX_FACT 64 42#define RD_THRESH_INC 1 43 44#define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \ 45 (1 << INTRA_FRAME)) 46#define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \ 47 (1 << INTRA_FRAME)) 48#define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \ 49 (1 << INTRA_FRAME)) 50 51#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01) 52 53#define MIN_EARLY_TERM_INDEX 3 54 55typedef struct { 56 PREDICTION_MODE mode; 57 MV_REFERENCE_FRAME ref_frame[2]; 58} MODE_DEFINITION; 59 60typedef struct { 61 MV_REFERENCE_FRAME ref_frame[2]; 62} REF_DEFINITION; 63 64struct rdcost_block_args { 65 MACROBLOCK *x; 66 ENTROPY_CONTEXT t_above[16]; 67 ENTROPY_CONTEXT t_left[16]; 68 int rate; 69 int64_t dist; 70 int64_t sse; 71 int this_rate; 72 int64_t this_dist; 73 int64_t this_sse; 74 int64_t this_rd; 75 int64_t best_rd; 76 int skip; 77 int use_fast_coef_costing; 78 const scan_order *so; 79}; 80 81static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = { 82 {NEARESTMV, {LAST_FRAME, NONE}}, 83 {NEARESTMV, {ALTREF_FRAME, NONE}}, 84 {NEARESTMV, {GOLDEN_FRAME, NONE}}, 85 86 {DC_PRED, {INTRA_FRAME, NONE}}, 87 88 {NEWMV, {LAST_FRAME, NONE}}, 89 {NEWMV, {ALTREF_FRAME, NONE}}, 90 {NEWMV, {GOLDEN_FRAME, NONE}}, 91 92 {NEARMV, {LAST_FRAME, NONE}}, 93 {NEARMV, {ALTREF_FRAME, NONE}}, 94 {NEARMV, {GOLDEN_FRAME, NONE}}, 95 96 {ZEROMV, {LAST_FRAME, NONE}}, 97 {ZEROMV, {GOLDEN_FRAME, NONE}}, 98 {ZEROMV, {ALTREF_FRAME, NONE}}, 99 100 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}}, 101 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}}, 102 103 {TM_PRED, {INTRA_FRAME, NONE}}, 104 105 {NEARMV, {LAST_FRAME, ALTREF_FRAME}}, 106 {NEWMV, {LAST_FRAME, ALTREF_FRAME}}, 107 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}}, 108 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}}, 109 110 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}}, 111 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}}, 112 113 {H_PRED, {INTRA_FRAME, NONE}}, 114 {V_PRED, {INTRA_FRAME, NONE}}, 115 {D135_PRED, {INTRA_FRAME, NONE}}, 116 {D207_PRED, {INTRA_FRAME, NONE}}, 117 {D153_PRED, {INTRA_FRAME, NONE}}, 118 {D63_PRED, {INTRA_FRAME, NONE}}, 119 {D117_PRED, {INTRA_FRAME, NONE}}, 120 {D45_PRED, {INTRA_FRAME, NONE}}, 121}; 122 123static const REF_DEFINITION vp9_ref_order[MAX_REFS] = { 124 {{LAST_FRAME, NONE}}, 125 {{GOLDEN_FRAME, NONE}}, 126 {{ALTREF_FRAME, NONE}}, 127 {{LAST_FRAME, ALTREF_FRAME}}, 128 {{GOLDEN_FRAME, ALTREF_FRAME}}, 129 {{INTRA_FRAME, NONE}}, 130}; 131 132static int raster_block_offset(BLOCK_SIZE plane_bsize, 133 int raster_block, int stride) { 134 const int bw = b_width_log2(plane_bsize); 135 const int y = 4 * (raster_block >> bw); 136 const int x = 4 * (raster_block & ((1 << bw) - 1)); 137 return y * stride + x; 138} 139static int16_t* raster_block_offset_int16(BLOCK_SIZE plane_bsize, 140 int raster_block, int16_t *base) { 141 const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize]; 142 return base + raster_block_offset(plane_bsize, raster_block, stride); 143} 144 145static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, 146 int m, int n, int min_plane, int max_plane) { 147 int i; 148 149 for (i = min_plane; i < max_plane; ++i) { 150 struct macroblock_plane *const p = &x->plane[i]; 151 struct macroblockd_plane *const pd = &x->e_mbd.plane[i]; 152 153 p->coeff = ctx->coeff_pbuf[i][m]; 154 p->qcoeff = ctx->qcoeff_pbuf[i][m]; 155 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m]; 156 p->eobs = ctx->eobs_pbuf[i][m]; 157 158 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n]; 159 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n]; 160 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n]; 161 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n]; 162 163 ctx->coeff_pbuf[i][n] = p->coeff; 164 ctx->qcoeff_pbuf[i][n] = p->qcoeff; 165 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff; 166 ctx->eobs_pbuf[i][n] = p->eobs; 167 } 168} 169 170static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize, 171 MACROBLOCK *x, MACROBLOCKD *xd, 172 int *out_rate_sum, int64_t *out_dist_sum) { 173 // Note our transform coeffs are 8 times an orthogonal transform. 174 // Hence quantizer step is also 8 times. To get effective quantizer 175 // we need to divide by 8 before sending to modeling function. 176 int i; 177 int64_t rate_sum = 0; 178 int64_t dist_sum = 0; 179 const int ref = xd->mi[0].src_mi->mbmi.ref_frame[0]; 180 unsigned int sse; 181 unsigned int var = 0; 182 unsigned int sum_sse = 0; 183 const int shift = 8; 184 int rate; 185 int64_t dist; 186 187 x->pred_sse[ref] = 0; 188 189 for (i = 0; i < MAX_MB_PLANE; ++i) { 190 struct macroblock_plane *const p = &x->plane[i]; 191 struct macroblockd_plane *const pd = &xd->plane[i]; 192 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd); 193 const TX_SIZE max_tx_size = max_txsize_lookup[bs]; 194 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size]; 195 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]); 196 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]); 197 int idx, idy; 198 int lw = b_width_log2_lookup[unit_size] + 2; 199 int lh = b_height_log2_lookup[unit_size] + 2; 200 201 sum_sse = 0; 202 203 for (idy = 0; idy < bh; ++idy) { 204 for (idx = 0; idx < bw; ++idx) { 205 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw); 206 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh); 207 int block_idx = (idy << 1) + idx; 208 209 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, 210 dst, pd->dst.stride, &sse); 211 x->bsse[(i << 2) + block_idx] = sse; 212 sum_sse += sse; 213 214 if (!x->select_tx_size) { 215 if (x->bsse[(i << 2) + block_idx] < p->quant_thred[0] >> shift) 216 x->skip_txfm[(i << 2) + block_idx] = 1; 217 else if (var < p->quant_thred[1] >> shift) 218 x->skip_txfm[(i << 2) + block_idx] = 2; 219 else 220 x->skip_txfm[(i << 2) + block_idx] = 0; 221 } 222 223 if (i == 0) 224 x->pred_sse[ref] += sse; 225 } 226 } 227 228 // Fast approximate the modelling function. 229 if (cpi->oxcf.speed > 4) { 230 int64_t rate; 231 int64_t dist; 232 int64_t square_error = sse; 233 int quantizer = (pd->dequant[1] >> 3); 234 235 if (quantizer < 120) 236 rate = (square_error * (280 - quantizer)) >> 8; 237 else 238 rate = 0; 239 dist = (square_error * quantizer) >> 8; 240 rate_sum += rate; 241 dist_sum += dist; 242 } else { 243 vp9_model_rd_from_var_lapndz(sum_sse, 1 << num_pels_log2_lookup[bs], 244 pd->dequant[1] >> 3, &rate, &dist); 245 rate_sum += rate; 246 dist_sum += dist; 247 } 248 } 249 250 *out_rate_sum = (int)rate_sum; 251 *out_dist_sum = dist_sum << 4; 252} 253 254int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, 255 intptr_t block_size, int64_t *ssz) { 256 int i; 257 int64_t error = 0, sqcoeff = 0; 258 259 for (i = 0; i < block_size; i++) { 260 const int diff = coeff[i] - dqcoeff[i]; 261 error += diff * diff; 262 sqcoeff += coeff[i] * coeff[i]; 263 } 264 265 *ssz = sqcoeff; 266 return error; 267} 268 269/* The trailing '0' is a terminator which is used inside cost_coeffs() to 270 * decide whether to include cost of a trailing EOB node or not (i.e. we 271 * can skip this if the last coefficient in this transform block, e.g. the 272 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block, 273 * were non-zero). */ 274static const int16_t band_counts[TX_SIZES][8] = { 275 { 1, 2, 3, 4, 3, 16 - 13, 0 }, 276 { 1, 2, 3, 4, 11, 64 - 21, 0 }, 277 { 1, 2, 3, 4, 11, 256 - 21, 0 }, 278 { 1, 2, 3, 4, 11, 1024 - 21, 0 }, 279}; 280static INLINE int cost_coeffs(MACROBLOCK *x, 281 int plane, int block, 282 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L, 283 TX_SIZE tx_size, 284 const int16_t *scan, const int16_t *nb, 285 int use_fast_coef_costing) { 286 MACROBLOCKD *const xd = &x->e_mbd; 287 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi; 288 const struct macroblock_plane *p = &x->plane[plane]; 289 const struct macroblockd_plane *pd = &xd->plane[plane]; 290 const PLANE_TYPE type = pd->plane_type; 291 const int16_t *band_count = &band_counts[tx_size][1]; 292 const int eob = p->eobs[block]; 293 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block); 294 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] = 295 x->token_costs[tx_size][type][is_inter_block(mbmi)]; 296 uint8_t token_cache[32 * 32]; 297 int pt = combine_entropy_contexts(*A, *L); 298 int c, cost; 299 // Check for consistency of tx_size with mode info 300 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size 301 : get_uv_tx_size(mbmi, pd) == tx_size); 302 303 if (eob == 0) { 304 // single eob token 305 cost = token_costs[0][0][pt][EOB_TOKEN]; 306 c = 0; 307 } else { 308 int band_left = *band_count++; 309 310 // dc token 311 int v = qcoeff[0]; 312 int prev_t = vp9_dct_value_tokens_ptr[v].token; 313 cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v]; 314 token_cache[0] = vp9_pt_energy_class[prev_t]; 315 ++token_costs; 316 317 // ac tokens 318 for (c = 1; c < eob; c++) { 319 const int rc = scan[c]; 320 int t; 321 322 v = qcoeff[rc]; 323 t = vp9_dct_value_tokens_ptr[v].token; 324 if (use_fast_coef_costing) { 325 cost += (*token_costs)[!prev_t][!prev_t][t] + vp9_dct_value_cost_ptr[v]; 326 } else { 327 pt = get_coef_context(nb, token_cache, c); 328 cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v]; 329 token_cache[rc] = vp9_pt_energy_class[t]; 330 } 331 prev_t = t; 332 if (!--band_left) { 333 band_left = *band_count++; 334 ++token_costs; 335 } 336 } 337 338 // eob token 339 if (band_left) { 340 if (use_fast_coef_costing) { 341 cost += (*token_costs)[0][!prev_t][EOB_TOKEN]; 342 } else { 343 pt = get_coef_context(nb, token_cache, c); 344 cost += (*token_costs)[0][pt][EOB_TOKEN]; 345 } 346 } 347 } 348 349 // is eob first coefficient; 350 *A = *L = (c > 0); 351 352 return cost; 353} 354static void dist_block(int plane, int block, TX_SIZE tx_size, 355 struct rdcost_block_args* args) { 356 const int ss_txfrm_size = tx_size << 1; 357 MACROBLOCK* const x = args->x; 358 MACROBLOCKD* const xd = &x->e_mbd; 359 const struct macroblock_plane *const p = &x->plane[plane]; 360 const struct macroblockd_plane *const pd = &xd->plane[plane]; 361 int64_t this_sse; 362 int shift = tx_size == TX_32X32 ? 0 : 2; 363 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block); 364 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 365 args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, 366 &this_sse) >> shift; 367 args->sse = this_sse >> shift; 368 369 if (x->skip_encode && !is_inter_block(&xd->mi[0].src_mi->mbmi)) { 370 // TODO(jingning): tune the model to better capture the distortion. 371 int64_t p = (pd->dequant[1] * pd->dequant[1] * 372 (1 << ss_txfrm_size)) >> (shift + 2); 373 args->dist += (p >> 4); 374 args->sse += p; 375 } 376} 377 378static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize, 379 TX_SIZE tx_size, struct rdcost_block_args* args) { 380 int x_idx, y_idx; 381 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx); 382 383 args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx, 384 args->t_left + y_idx, tx_size, 385 args->so->scan, args->so->neighbors, 386 args->use_fast_coef_costing); 387} 388 389static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize, 390 TX_SIZE tx_size, void *arg) { 391 struct rdcost_block_args *args = arg; 392 MACROBLOCK *const x = args->x; 393 MACROBLOCKD *const xd = &x->e_mbd; 394 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 395 int64_t rd1, rd2, rd; 396 397 if (args->skip) 398 return; 399 400 if (!is_inter_block(mbmi)) { 401 vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip); 402 dist_block(plane, block, tx_size, args); 403 } else if (max_txsize_lookup[plane_bsize] == tx_size) { 404 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) { 405 // full forward transform and quantization 406 vp9_xform_quant(x, plane, block, plane_bsize, tx_size); 407 dist_block(plane, block, tx_size, args); 408 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) { 409 // compute DC coefficient 410 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block); 411 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block); 412 vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size); 413 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4; 414 args->dist = args->sse; 415 if (!x->plane[plane].eobs[block]) 416 args->dist = args->sse - ((coeff[0] * coeff[0] - 417 (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2); 418 } else { 419 // skip forward transform 420 x->plane[plane].eobs[block] = 0; 421 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4; 422 args->dist = args->sse; 423 } 424 } else { 425 // full forward transform and quantization 426 vp9_xform_quant(x, plane, block, plane_bsize, tx_size); 427 dist_block(plane, block, tx_size, args); 428 } 429 430 rate_block(plane, block, plane_bsize, tx_size, args); 431 rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist); 432 rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse); 433 434 // TODO(jingning): temporarily enabled only for luma component 435 rd = MIN(rd1, rd2); 436 if (plane == 0) 437 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] || 438 (rd1 > rd2 && !xd->lossless); 439 440 args->this_rate += args->rate; 441 args->this_dist += args->dist; 442 args->this_sse += args->sse; 443 args->this_rd += rd; 444 445 if (args->this_rd > args->best_rd) { 446 args->skip = 1; 447 return; 448 } 449} 450 451static void txfm_rd_in_plane(MACROBLOCK *x, 452 int *rate, int64_t *distortion, 453 int *skippable, int64_t *sse, 454 int64_t ref_best_rd, int plane, 455 BLOCK_SIZE bsize, TX_SIZE tx_size, 456 int use_fast_coef_casting) { 457 MACROBLOCKD *const xd = &x->e_mbd; 458 const struct macroblockd_plane *const pd = &xd->plane[plane]; 459 struct rdcost_block_args args; 460 vp9_zero(args); 461 args.x = x; 462 args.best_rd = ref_best_rd; 463 args.use_fast_coef_costing = use_fast_coef_casting; 464 465 if (plane == 0) 466 xd->mi[0].src_mi->mbmi.tx_size = tx_size; 467 468 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left); 469 470 args.so = get_scan(xd, tx_size, pd->plane_type, 0); 471 472 vp9_foreach_transformed_block_in_plane(xd, bsize, plane, 473 block_rd_txfm, &args); 474 if (args.skip) { 475 *rate = INT_MAX; 476 *distortion = INT64_MAX; 477 *sse = INT64_MAX; 478 *skippable = 0; 479 } else { 480 *distortion = args.this_dist; 481 *rate = args.this_rate; 482 *sse = args.this_sse; 483 *skippable = vp9_is_skippable_in_plane(x, bsize, plane); 484 } 485} 486 487static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x, 488 int *rate, int64_t *distortion, 489 int *skip, int64_t *sse, 490 int64_t ref_best_rd, 491 BLOCK_SIZE bs) { 492 const TX_SIZE max_tx_size = max_txsize_lookup[bs]; 493 VP9_COMMON *const cm = &cpi->common; 494 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode]; 495 MACROBLOCKD *const xd = &x->e_mbd; 496 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 497 498 mbmi->tx_size = MIN(max_tx_size, largest_tx_size); 499 500 txfm_rd_in_plane(x, rate, distortion, skip, 501 sse, ref_best_rd, 0, bs, 502 mbmi->tx_size, cpi->sf.use_fast_coef_costing); 503} 504 505static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x, 506 int *rate, 507 int64_t *distortion, 508 int *skip, 509 int64_t *psse, 510 int64_t tx_cache[TX_MODES], 511 int64_t ref_best_rd, 512 BLOCK_SIZE bs) { 513 const TX_SIZE max_tx_size = max_txsize_lookup[bs]; 514 VP9_COMMON *const cm = &cpi->common; 515 MACROBLOCKD *const xd = &x->e_mbd; 516 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 517 vp9_prob skip_prob = vp9_get_skip_prob(cm, xd); 518 int r[TX_SIZES][2], s[TX_SIZES]; 519 int64_t d[TX_SIZES], sse[TX_SIZES]; 520 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX}, 521 {INT64_MAX, INT64_MAX}, 522 {INT64_MAX, INT64_MAX}, 523 {INT64_MAX, INT64_MAX}}; 524 int n, m; 525 int s0, s1; 526 const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode]; 527 int64_t best_rd = INT64_MAX; 528 TX_SIZE best_tx = max_tx_size; 529 530 const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs); 531 assert(skip_prob > 0); 532 s0 = vp9_cost_bit(skip_prob, 0); 533 s1 = vp9_cost_bit(skip_prob, 1); 534 535 for (n = max_tx_size; n >= 0; n--) { 536 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n], 537 &sse[n], ref_best_rd, 0, bs, n, 538 cpi->sf.use_fast_coef_costing); 539 r[n][1] = r[n][0]; 540 if (r[n][0] < INT_MAX) { 541 for (m = 0; m <= n - (n == (int) max_tx_size); m++) { 542 if (m == n) 543 r[n][1] += vp9_cost_zero(tx_probs[m]); 544 else 545 r[n][1] += vp9_cost_one(tx_probs[m]); 546 } 547 } 548 if (d[n] == INT64_MAX) { 549 rd[n][0] = rd[n][1] = INT64_MAX; 550 } else if (s[n]) { 551 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]); 552 } else { 553 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]); 554 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]); 555 } 556 557 // Early termination in transform size search. 558 if (cpi->sf.tx_size_search_breakout && 559 (rd[n][1] == INT64_MAX || 560 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) || 561 s[n] == 1)) 562 break; 563 564 if (rd[n][1] < best_rd) { 565 best_tx = n; 566 best_rd = rd[n][1]; 567 } 568 } 569 mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ? 570 best_tx : MIN(max_tx_size, max_mode_tx_size); 571 572 573 *distortion = d[mbmi->tx_size]; 574 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT]; 575 *skip = s[mbmi->tx_size]; 576 *psse = sse[mbmi->tx_size]; 577 578 tx_cache[ONLY_4X4] = rd[TX_4X4][0]; 579 tx_cache[ALLOW_8X8] = rd[TX_8X8][0]; 580 tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0]; 581 tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0]; 582 583 if (max_tx_size == TX_32X32 && best_tx == TX_32X32) { 584 tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1]; 585 } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) { 586 tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1]; 587 } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) { 588 tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1]; 589 } else { 590 tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1]; 591 } 592} 593 594static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, 595 int64_t *distortion, int *skip, 596 int64_t *psse, BLOCK_SIZE bs, 597 int64_t txfm_cache[TX_MODES], 598 int64_t ref_best_rd) { 599 MACROBLOCKD *xd = &x->e_mbd; 600 int64_t sse; 601 int64_t *ret_sse = psse ? psse : &sse; 602 603 assert(bs == xd->mi[0].src_mi->mbmi.sb_type); 604 605 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) { 606 vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t)); 607 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd, 608 bs); 609 } else { 610 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, 611 txfm_cache, ref_best_rd, bs); 612 } 613} 614 615static int conditional_skipintra(PREDICTION_MODE mode, 616 PREDICTION_MODE best_intra_mode) { 617 if (mode == D117_PRED && 618 best_intra_mode != V_PRED && 619 best_intra_mode != D135_PRED) 620 return 1; 621 if (mode == D63_PRED && 622 best_intra_mode != V_PRED && 623 best_intra_mode != D45_PRED) 624 return 1; 625 if (mode == D207_PRED && 626 best_intra_mode != H_PRED && 627 best_intra_mode != D45_PRED) 628 return 1; 629 if (mode == D153_PRED && 630 best_intra_mode != H_PRED && 631 best_intra_mode != D135_PRED) 632 return 1; 633 return 0; 634} 635 636static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib, 637 PREDICTION_MODE *best_mode, 638 const int *bmode_costs, 639 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l, 640 int *bestrate, int *bestratey, 641 int64_t *bestdistortion, 642 BLOCK_SIZE bsize, int64_t rd_thresh) { 643 PREDICTION_MODE mode; 644 MACROBLOCKD *const xd = &x->e_mbd; 645 int64_t best_rd = rd_thresh; 646 647 struct macroblock_plane *p = &x->plane[0]; 648 struct macroblockd_plane *pd = &xd->plane[0]; 649 const int src_stride = p->src.stride; 650 const int dst_stride = pd->dst.stride; 651 const uint8_t *src_init = &p->src.buf[raster_block_offset(BLOCK_8X8, ib, 652 src_stride)]; 653 uint8_t *dst_init = &pd->dst.buf[raster_block_offset(BLOCK_8X8, ib, 654 dst_stride)]; 655 ENTROPY_CONTEXT ta[2], tempa[2]; 656 ENTROPY_CONTEXT tl[2], templ[2]; 657 658 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize]; 659 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize]; 660 int idx, idy; 661 uint8_t best_dst[8 * 8]; 662 663 assert(ib < 4); 664 665 vpx_memcpy(ta, a, sizeof(ta)); 666 vpx_memcpy(tl, l, sizeof(tl)); 667 xd->mi[0].src_mi->mbmi.tx_size = TX_4X4; 668 669 for (mode = DC_PRED; mode <= TM_PRED; ++mode) { 670 int64_t this_rd; 671 int ratey = 0; 672 int64_t distortion = 0; 673 int rate = bmode_costs[mode]; 674 675 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) 676 continue; 677 678 // Only do the oblique modes if the best so far is 679 // one of the neighboring directional modes 680 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) { 681 if (conditional_skipintra(mode, *best_mode)) 682 continue; 683 } 684 685 vpx_memcpy(tempa, ta, sizeof(ta)); 686 vpx_memcpy(templ, tl, sizeof(tl)); 687 688 for (idy = 0; idy < num_4x4_blocks_high; ++idy) { 689 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) { 690 const int block = ib + idy * 2 + idx; 691 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride]; 692 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride]; 693 int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block, 694 p->src_diff); 695 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block); 696 xd->mi[0].src_mi->bmi[block].as_mode = mode; 697 vp9_predict_intra_block(xd, block, 1, 698 TX_4X4, mode, 699 x->skip_encode ? src : dst, 700 x->skip_encode ? src_stride : dst_stride, 701 dst, dst_stride, idx, idy, 0); 702 vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride); 703 704 if (xd->lossless) { 705 const scan_order *so = &vp9_default_scan_orders[TX_4X4]; 706 vp9_fwht4x4(src_diff, coeff, 8); 707 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan); 708 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4, 709 so->scan, so->neighbors, 710 cpi->sf.use_fast_coef_costing); 711 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd) 712 goto next; 713 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride, 714 p->eobs[block]); 715 } else { 716 int64_t unused; 717 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block); 718 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type]; 719 vp9_fht4x4(src_diff, coeff, 8, tx_type); 720 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan); 721 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4, 722 so->scan, so->neighbors, 723 cpi->sf.use_fast_coef_costing); 724 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block), 725 16, &unused) >> 2; 726 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd) 727 goto next; 728 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), 729 dst, dst_stride, p->eobs[block]); 730 } 731 } 732 } 733 734 rate += ratey; 735 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion); 736 737 if (this_rd < best_rd) { 738 *bestrate = rate; 739 *bestratey = ratey; 740 *bestdistortion = distortion; 741 best_rd = this_rd; 742 *best_mode = mode; 743 vpx_memcpy(a, tempa, sizeof(tempa)); 744 vpx_memcpy(l, templ, sizeof(templ)); 745 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) 746 vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride, 747 num_4x4_blocks_wide * 4); 748 } 749 next: 750 {} 751 } 752 753 if (best_rd >= rd_thresh || x->skip_encode) 754 return best_rd; 755 756 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) 757 vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8, 758 num_4x4_blocks_wide * 4); 759 760 return best_rd; 761} 762 763static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb, 764 int *rate, int *rate_y, 765 int64_t *distortion, 766 int64_t best_rd) { 767 int i, j; 768 const MACROBLOCKD *const xd = &mb->e_mbd; 769 MODE_INFO *const mic = xd->mi[0].src_mi; 770 const MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi; 771 const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL; 772 const BLOCK_SIZE bsize = xd->mi[0].src_mi->mbmi.sb_type; 773 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize]; 774 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize]; 775 int idx, idy; 776 int cost = 0; 777 int64_t total_distortion = 0; 778 int tot_rate_y = 0; 779 int64_t total_rd = 0; 780 ENTROPY_CONTEXT t_above[4], t_left[4]; 781 const int *bmode_costs = cpi->mbmode_cost; 782 783 vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above)); 784 vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left)); 785 786 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block. 787 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) { 788 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) { 789 PREDICTION_MODE best_mode = DC_PRED; 790 int r = INT_MAX, ry = INT_MAX; 791 int64_t d = INT64_MAX, this_rd = INT64_MAX; 792 i = idy * 2 + idx; 793 if (cpi->common.frame_type == KEY_FRAME) { 794 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i); 795 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i); 796 797 bmode_costs = cpi->y_mode_costs[A][L]; 798 } 799 800 this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs, 801 t_above + idx, t_left + idy, &r, &ry, &d, 802 bsize, best_rd - total_rd); 803 if (this_rd >= best_rd - total_rd) 804 return INT64_MAX; 805 806 total_rd += this_rd; 807 cost += r; 808 total_distortion += d; 809 tot_rate_y += ry; 810 811 mic->bmi[i].as_mode = best_mode; 812 for (j = 1; j < num_4x4_blocks_high; ++j) 813 mic->bmi[i + j * 2].as_mode = best_mode; 814 for (j = 1; j < num_4x4_blocks_wide; ++j) 815 mic->bmi[i + j].as_mode = best_mode; 816 817 if (total_rd >= best_rd) 818 return INT64_MAX; 819 } 820 } 821 822 *rate = cost; 823 *rate_y = tot_rate_y; 824 *distortion = total_distortion; 825 mic->mbmi.mode = mic->bmi[3].as_mode; 826 827 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion); 828} 829 830static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, 831 int *rate, int *rate_tokenonly, 832 int64_t *distortion, int *skippable, 833 BLOCK_SIZE bsize, 834 int64_t tx_cache[TX_MODES], 835 int64_t best_rd) { 836 PREDICTION_MODE mode; 837 PREDICTION_MODE mode_selected = DC_PRED; 838 MACROBLOCKD *const xd = &x->e_mbd; 839 MODE_INFO *const mic = xd->mi[0].src_mi; 840 int this_rate, this_rate_tokenonly, s; 841 int64_t this_distortion, this_rd; 842 TX_SIZE best_tx = TX_4X4; 843 int i; 844 int *bmode_costs = cpi->mbmode_cost; 845 846 if (cpi->sf.tx_size_search_method == USE_FULL_RD) 847 for (i = 0; i < TX_MODES; i++) 848 tx_cache[i] = INT64_MAX; 849 850 /* Y Search for intra prediction mode */ 851 for (mode = DC_PRED; mode <= TM_PRED; mode++) { 852 int64_t local_tx_cache[TX_MODES]; 853 MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi; 854 MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL; 855 856 if (cpi->common.frame_type == KEY_FRAME) { 857 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0); 858 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0); 859 860 bmode_costs = cpi->y_mode_costs[A][L]; 861 } 862 mic->mbmi.mode = mode; 863 864 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, 865 &s, NULL, bsize, local_tx_cache, best_rd); 866 867 if (this_rate_tokenonly == INT_MAX) 868 continue; 869 870 this_rate = this_rate_tokenonly + bmode_costs[mode]; 871 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion); 872 873 if (this_rd < best_rd) { 874 mode_selected = mode; 875 best_rd = this_rd; 876 best_tx = mic->mbmi.tx_size; 877 *rate = this_rate; 878 *rate_tokenonly = this_rate_tokenonly; 879 *distortion = this_distortion; 880 *skippable = s; 881 } 882 883 if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) { 884 for (i = 0; i < TX_MODES && local_tx_cache[i] < INT64_MAX; i++) { 885 const int64_t adj_rd = this_rd + local_tx_cache[i] - 886 local_tx_cache[cpi->common.tx_mode]; 887 if (adj_rd < tx_cache[i]) { 888 tx_cache[i] = adj_rd; 889 } 890 } 891 } 892 } 893 894 mic->mbmi.mode = mode_selected; 895 mic->mbmi.tx_size = best_tx; 896 897 return best_rd; 898} 899 900static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x, 901 int *rate, int64_t *distortion, int *skippable, 902 int64_t *sse, BLOCK_SIZE bsize, 903 int64_t ref_best_rd) { 904 MACROBLOCKD *const xd = &x->e_mbd; 905 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 906 const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]); 907 int plane; 908 int pnrate = 0, pnskip = 1; 909 int64_t pndist = 0, pnsse = 0; 910 911 if (ref_best_rd < 0) 912 goto term; 913 914 if (is_inter_block(mbmi)) { 915 int plane; 916 for (plane = 1; plane < MAX_MB_PLANE; ++plane) 917 vp9_subtract_plane(x, bsize, plane); 918 } 919 920 *rate = 0; 921 *distortion = 0; 922 *sse = 0; 923 *skippable = 1; 924 925 for (plane = 1; plane < MAX_MB_PLANE; ++plane) { 926 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse, 927 ref_best_rd, plane, bsize, uv_tx_size, 928 cpi->sf.use_fast_coef_costing); 929 if (pnrate == INT_MAX) 930 goto term; 931 *rate += pnrate; 932 *distortion += pndist; 933 *sse += pnsse; 934 *skippable &= pnskip; 935 } 936 return; 937 938 term: 939 *rate = INT_MAX; 940 *distortion = INT64_MAX; 941 *sse = INT64_MAX; 942 *skippable = 0; 943 return; 944} 945 946static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x, 947 PICK_MODE_CONTEXT *ctx, 948 int *rate, int *rate_tokenonly, 949 int64_t *distortion, int *skippable, 950 BLOCK_SIZE bsize, TX_SIZE max_tx_size) { 951 MACROBLOCKD *xd = &x->e_mbd; 952 PREDICTION_MODE mode; 953 PREDICTION_MODE mode_selected = DC_PRED; 954 int64_t best_rd = INT64_MAX, this_rd; 955 int this_rate_tokenonly, this_rate, s; 956 int64_t this_distortion, this_sse; 957 958 for (mode = DC_PRED; mode <= TM_PRED; ++mode) { 959 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) 960 continue; 961 962 xd->mi[0].src_mi->mbmi.uv_mode = mode; 963 964 super_block_uvrd(cpi, x, &this_rate_tokenonly, 965 &this_distortion, &s, &this_sse, bsize, best_rd); 966 if (this_rate_tokenonly == INT_MAX) 967 continue; 968 this_rate = this_rate_tokenonly + 969 cpi->intra_uv_mode_cost[cpi->common.frame_type][mode]; 970 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion); 971 972 if (this_rd < best_rd) { 973 mode_selected = mode; 974 best_rd = this_rd; 975 *rate = this_rate; 976 *rate_tokenonly = this_rate_tokenonly; 977 *distortion = this_distortion; 978 *skippable = s; 979 if (!x->select_tx_size) 980 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE); 981 } 982 } 983 984 xd->mi[0].src_mi->mbmi.uv_mode = mode_selected; 985 return best_rd; 986} 987 988static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, 989 int *rate, int *rate_tokenonly, 990 int64_t *distortion, int *skippable, 991 BLOCK_SIZE bsize) { 992 const VP9_COMMON *cm = &cpi->common; 993 int64_t unused; 994 995 x->e_mbd.mi[0].src_mi->mbmi.uv_mode = DC_PRED; 996 super_block_uvrd(cpi, x, rate_tokenonly, distortion, 997 skippable, &unused, bsize, INT64_MAX); 998 *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED]; 999 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion); 1000} 1001 1002static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx, 1003 BLOCK_SIZE bsize, TX_SIZE max_tx_size, 1004 int *rate_uv, int *rate_uv_tokenonly, 1005 int64_t *dist_uv, int *skip_uv, 1006 PREDICTION_MODE *mode_uv) { 1007 MACROBLOCK *const x = &cpi->mb; 1008 1009 // Use an estimated rd for uv_intra based on DC_PRED if the 1010 // appropriate speed flag is set. 1011 if (cpi->sf.use_uv_intra_rd_estimate) { 1012 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, 1013 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize); 1014 // Else do a proper rd search for each possible transform size that may 1015 // be considered in the main rd loop. 1016 } else { 1017 rd_pick_intra_sbuv_mode(cpi, x, ctx, 1018 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv, 1019 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size); 1020 } 1021 *mode_uv = x->e_mbd.mi[0].src_mi->mbmi.uv_mode; 1022} 1023 1024static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode, 1025 int mode_context) { 1026 assert(is_inter_mode(mode)); 1027 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)]; 1028} 1029 1030static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, 1031 BLOCK_SIZE bsize, 1032 int_mv *frame_mv, 1033 int mi_row, int mi_col, 1034 int_mv single_newmv[MAX_REF_FRAMES], 1035 int *rate_mv); 1036 1037static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i, 1038 PREDICTION_MODE mode, int_mv this_mv[2], 1039 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], 1040 int_mv seg_mvs[MAX_REF_FRAMES], 1041 int_mv *best_ref_mv[2], const int *mvjcost, 1042 int *mvcost[2]) { 1043 MODE_INFO *const mic = xd->mi[0].src_mi; 1044 const MB_MODE_INFO *const mbmi = &mic->mbmi; 1045 int thismvcost = 0; 1046 int idx, idy; 1047 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type]; 1048 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type]; 1049 const int is_compound = has_second_ref(mbmi); 1050 1051 switch (mode) { 1052 case NEWMV: 1053 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int; 1054 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv, 1055 mvjcost, mvcost, MV_COST_WEIGHT_SUB); 1056 if (is_compound) { 1057 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int; 1058 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv, 1059 mvjcost, mvcost, MV_COST_WEIGHT_SUB); 1060 } 1061 break; 1062 case NEARMV: 1063 case NEARESTMV: 1064 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int; 1065 if (is_compound) 1066 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int; 1067 break; 1068 case ZEROMV: 1069 this_mv[0].as_int = 0; 1070 if (is_compound) 1071 this_mv[1].as_int = 0; 1072 break; 1073 default: 1074 break; 1075 } 1076 1077 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int; 1078 if (is_compound) 1079 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int; 1080 1081 mic->bmi[i].as_mode = mode; 1082 1083 for (idy = 0; idy < num_4x4_blocks_high; ++idy) 1084 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) 1085 vpx_memcpy(&mic->bmi[i + idy * 2 + idx], 1086 &mic->bmi[i], sizeof(mic->bmi[i])); 1087 1088 return cost_mv_ref(cpi, mode, mbmi->mode_context[mbmi->ref_frame[0]]) + 1089 thismvcost; 1090} 1091 1092static int64_t encode_inter_mb_segment(VP9_COMP *cpi, 1093 MACROBLOCK *x, 1094 int64_t best_yrd, 1095 int i, 1096 int *labelyrate, 1097 int64_t *distortion, int64_t *sse, 1098 ENTROPY_CONTEXT *ta, 1099 ENTROPY_CONTEXT *tl, 1100 int mi_row, int mi_col) { 1101 int k; 1102 MACROBLOCKD *xd = &x->e_mbd; 1103 struct macroblockd_plane *const pd = &xd->plane[0]; 1104 struct macroblock_plane *const p = &x->plane[0]; 1105 MODE_INFO *const mi = xd->mi[0].src_mi; 1106 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd); 1107 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize]; 1108 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize]; 1109 int idx, idy; 1110 1111 const uint8_t *const src = &p->src.buf[raster_block_offset(BLOCK_8X8, i, 1112 p->src.stride)]; 1113 uint8_t *const dst = &pd->dst.buf[raster_block_offset(BLOCK_8X8, i, 1114 pd->dst.stride)]; 1115 int64_t thisdistortion = 0, thissse = 0; 1116 int thisrate = 0, ref; 1117 const scan_order *so = &vp9_default_scan_orders[TX_4X4]; 1118 const int is_compound = has_second_ref(&mi->mbmi); 1119 const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter); 1120 1121 for (ref = 0; ref < 1 + is_compound; ++ref) { 1122 const uint8_t *pre = &pd->pre[ref].buf[raster_block_offset(BLOCK_8X8, i, 1123 pd->pre[ref].stride)]; 1124 vp9_build_inter_predictor(pre, pd->pre[ref].stride, 1125 dst, pd->dst.stride, 1126 &mi->bmi[i].as_mv[ref].as_mv, 1127 &xd->block_refs[ref]->sf, width, height, ref, 1128 kernel, MV_PRECISION_Q3, 1129 mi_col * MI_SIZE + 4 * (i % 2), 1130 mi_row * MI_SIZE + 4 * (i / 2)); 1131 } 1132 1133 vp9_subtract_block(height, width, 1134 raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8, 1135 src, p->src.stride, 1136 dst, pd->dst.stride); 1137 1138 k = i; 1139 for (idy = 0; idy < height / 4; ++idy) { 1140 for (idx = 0; idx < width / 4; ++idx) { 1141 int64_t ssz, rd, rd1, rd2; 1142 tran_low_t* coeff; 1143 1144 k += (idy * 2 + idx); 1145 coeff = BLOCK_OFFSET(p->coeff, k); 1146 x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff), 1147 coeff, 8); 1148 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan); 1149 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 1150 16, &ssz); 1151 thissse += ssz; 1152 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4, 1153 so->scan, so->neighbors, 1154 cpi->sf.use_fast_coef_costing); 1155 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2); 1156 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2); 1157 rd = MIN(rd1, rd2); 1158 if (rd >= best_yrd) 1159 return INT64_MAX; 1160 } 1161 } 1162 1163 *distortion = thisdistortion >> 2; 1164 *labelyrate = thisrate; 1165 *sse = thissse >> 2; 1166 1167 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion); 1168} 1169 1170typedef struct { 1171 int eobs; 1172 int brate; 1173 int byrate; 1174 int64_t bdist; 1175 int64_t bsse; 1176 int64_t brdcost; 1177 int_mv mvs[2]; 1178 ENTROPY_CONTEXT ta[2]; 1179 ENTROPY_CONTEXT tl[2]; 1180} SEG_RDSTAT; 1181 1182typedef struct { 1183 int_mv *ref_mv[2]; 1184 int_mv mvp; 1185 1186 int64_t segment_rd; 1187 int r; 1188 int64_t d; 1189 int64_t sse; 1190 int segment_yrate; 1191 PREDICTION_MODE modes[4]; 1192 SEG_RDSTAT rdstat[4][INTER_MODES]; 1193 int mvthresh; 1194} BEST_SEG_INFO; 1195 1196static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) { 1197 return (mv->row >> 3) < x->mv_row_min || 1198 (mv->row >> 3) > x->mv_row_max || 1199 (mv->col >> 3) < x->mv_col_min || 1200 (mv->col >> 3) > x->mv_col_max; 1201} 1202 1203static INLINE void mi_buf_shift(MACROBLOCK *x, int i) { 1204 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0].src_mi->mbmi; 1205 struct macroblock_plane *const p = &x->plane[0]; 1206 struct macroblockd_plane *const pd = &x->e_mbd.plane[0]; 1207 1208 p->src.buf = &p->src.buf[raster_block_offset(BLOCK_8X8, i, p->src.stride)]; 1209 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0); 1210 pd->pre[0].buf = &pd->pre[0].buf[raster_block_offset(BLOCK_8X8, i, 1211 pd->pre[0].stride)]; 1212 if (has_second_ref(mbmi)) 1213 pd->pre[1].buf = &pd->pre[1].buf[raster_block_offset(BLOCK_8X8, i, 1214 pd->pre[1].stride)]; 1215} 1216 1217static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src, 1218 struct buf_2d orig_pre[2]) { 1219 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0].src_mi->mbmi; 1220 x->plane[0].src = orig_src; 1221 x->e_mbd.plane[0].pre[0] = orig_pre[0]; 1222 if (has_second_ref(mbmi)) 1223 x->e_mbd.plane[0].pre[1] = orig_pre[1]; 1224} 1225 1226static INLINE int mv_has_subpel(const MV *mv) { 1227 return (mv->row & 0x0F) || (mv->col & 0x0F); 1228} 1229 1230// Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion. 1231// TODO(aconverse): Find out if this is still productive then clean up or remove 1232static int check_best_zero_mv( 1233 const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES], 1234 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode, 1235 const MV_REFERENCE_FRAME ref_frames[2]) { 1236 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) && 1237 frame_mv[this_mode][ref_frames[0]].as_int == 0 && 1238 (ref_frames[1] == NONE || 1239 frame_mv[this_mode][ref_frames[1]].as_int == 0)) { 1240 int rfc = mode_context[ref_frames[0]]; 1241 int c1 = cost_mv_ref(cpi, NEARMV, rfc); 1242 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc); 1243 int c3 = cost_mv_ref(cpi, ZEROMV, rfc); 1244 1245 if (this_mode == NEARMV) { 1246 if (c1 > c3) return 0; 1247 } else if (this_mode == NEARESTMV) { 1248 if (c2 > c3) return 0; 1249 } else { 1250 assert(this_mode == ZEROMV); 1251 if (ref_frames[1] == NONE) { 1252 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) || 1253 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0)) 1254 return 0; 1255 } else { 1256 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 && 1257 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) || 1258 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 && 1259 frame_mv[NEARMV][ref_frames[1]].as_int == 0)) 1260 return 0; 1261 } 1262 } 1263 } 1264 return 1; 1265} 1266 1267static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x, 1268 const TileInfo * const tile, 1269 int_mv *best_ref_mv, 1270 int_mv *second_best_ref_mv, 1271 int64_t best_rd, int *returntotrate, 1272 int *returnyrate, 1273 int64_t *returndistortion, 1274 int *skippable, int64_t *psse, 1275 int mvthresh, 1276 int_mv seg_mvs[4][MAX_REF_FRAMES], 1277 BEST_SEG_INFO *bsi_buf, int filter_idx, 1278 int mi_row, int mi_col) { 1279 int i; 1280 BEST_SEG_INFO *bsi = bsi_buf + filter_idx; 1281 MACROBLOCKD *xd = &x->e_mbd; 1282 MODE_INFO *mi = xd->mi[0].src_mi; 1283 MB_MODE_INFO *mbmi = &mi->mbmi; 1284 int mode_idx; 1285 int k, br = 0, idx, idy; 1286 int64_t bd = 0, block_sse = 0; 1287 PREDICTION_MODE this_mode; 1288 VP9_COMMON *cm = &cpi->common; 1289 struct macroblock_plane *const p = &x->plane[0]; 1290 struct macroblockd_plane *const pd = &xd->plane[0]; 1291 const int label_count = 4; 1292 int64_t this_segment_rd = 0; 1293 int label_mv_thresh; 1294 int segmentyrate = 0; 1295 const BLOCK_SIZE bsize = mbmi->sb_type; 1296 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize]; 1297 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize]; 1298 ENTROPY_CONTEXT t_above[2], t_left[2]; 1299 int subpelmv = 1, have_ref = 0; 1300 const int has_second_rf = has_second_ref(mbmi); 1301 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize]; 1302 1303 vp9_zero(*bsi); 1304 1305 bsi->segment_rd = best_rd; 1306 bsi->ref_mv[0] = best_ref_mv; 1307 bsi->ref_mv[1] = second_best_ref_mv; 1308 bsi->mvp.as_int = best_ref_mv->as_int; 1309 bsi->mvthresh = mvthresh; 1310 1311 for (i = 0; i < 4; i++) 1312 bsi->modes[i] = ZEROMV; 1313 1314 vpx_memcpy(t_above, pd->above_context, sizeof(t_above)); 1315 vpx_memcpy(t_left, pd->left_context, sizeof(t_left)); 1316 1317 // 64 makes this threshold really big effectively 1318 // making it so that we very rarely check mvs on 1319 // segments. setting this to 1 would make mv thresh 1320 // roughly equal to what it is for macroblocks 1321 label_mv_thresh = 1 * bsi->mvthresh / label_count; 1322 1323 // Segmentation method overheads 1324 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) { 1325 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) { 1326 // TODO(jingning,rbultje): rewrite the rate-distortion optimization 1327 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop 1328 int_mv mode_mv[MB_MODE_COUNT][2]; 1329 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; 1330 PREDICTION_MODE mode_selected = ZEROMV; 1331 int64_t best_rd = INT64_MAX; 1332 const int i = idy * 2 + idx; 1333 int ref; 1334 1335 for (ref = 0; ref < 1 + has_second_rf; ++ref) { 1336 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref]; 1337 frame_mv[ZEROMV][frame].as_int = 0; 1338 vp9_append_sub8x8_mvs_for_idx(cm, xd, tile, i, ref, mi_row, mi_col, 1339 &frame_mv[NEARESTMV][frame], 1340 &frame_mv[NEARMV][frame]); 1341 } 1342 1343 // search for the best motion vector on this segment 1344 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) { 1345 const struct buf_2d orig_src = x->plane[0].src; 1346 struct buf_2d orig_pre[2]; 1347 1348 mode_idx = INTER_OFFSET(this_mode); 1349 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX; 1350 if (!(inter_mode_mask & (1 << this_mode))) 1351 continue; 1352 1353 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv, 1354 this_mode, mbmi->ref_frame)) 1355 continue; 1356 1357 vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre)); 1358 vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above, 1359 sizeof(bsi->rdstat[i][mode_idx].ta)); 1360 vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left, 1361 sizeof(bsi->rdstat[i][mode_idx].tl)); 1362 1363 // motion search for newmv (single predictor case only) 1364 if (!has_second_rf && this_mode == NEWMV && 1365 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) { 1366 MV *const new_mv = &mode_mv[NEWMV][0].as_mv; 1367 int step_param = 0; 1368 int thissme, bestsme = INT_MAX; 1369 int sadpb = x->sadperbit4; 1370 MV mvp_full; 1371 int max_mv; 1372 int sad_list[5]; 1373 1374 /* Is the best so far sufficiently good that we cant justify doing 1375 * and new motion search. */ 1376 if (best_rd < label_mv_thresh) 1377 break; 1378 1379 if (cpi->oxcf.mode != BEST) { 1380 // use previous block's result as next block's MV predictor. 1381 if (i > 0) { 1382 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int; 1383 if (i == 2) 1384 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int; 1385 } 1386 } 1387 if (i == 0) 1388 max_mv = x->max_mv_context[mbmi->ref_frame[0]]; 1389 else 1390 max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3; 1391 1392 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) { 1393 // Take wtd average of the step_params based on the last frame's 1394 // max mv magnitude and the best ref mvs of the current block for 1395 // the given reference. 1396 step_param = (vp9_init_search_range(max_mv) + 1397 cpi->mv_step_param) / 2; 1398 } else { 1399 step_param = cpi->mv_step_param; 1400 } 1401 1402 mvp_full.row = bsi->mvp.as_mv.row >> 3; 1403 mvp_full.col = bsi->mvp.as_mv.col >> 3; 1404 1405 if (cpi->sf.adaptive_motion_search) { 1406 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3; 1407 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3; 1408 step_param = MAX(step_param, 8); 1409 } 1410 1411 // adjust src pointer for this block 1412 mi_buf_shift(x, i); 1413 1414 vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv); 1415 1416 bestsme = vp9_full_pixel_search( 1417 cpi, x, bsize, &mvp_full, step_param, sadpb, 1418 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? sad_list : NULL, 1419 &bsi->ref_mv[0]->as_mv, new_mv, 1420 INT_MAX, 1); 1421 1422 // Should we do a full search (best quality only) 1423 if (cpi->oxcf.mode == BEST) { 1424 int_mv *const best_mv = &mi->bmi[i].as_mv[0]; 1425 /* Check if mvp_full is within the range. */ 1426 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max, 1427 x->mv_row_min, x->mv_row_max); 1428 thissme = cpi->full_search_sad(x, &mvp_full, 1429 sadpb, 16, &cpi->fn_ptr[bsize], 1430 &bsi->ref_mv[0]->as_mv, 1431 &best_mv->as_mv); 1432 sad_list[1] = sad_list[2] = sad_list[3] = sad_list[4] = INT_MAX; 1433 if (thissme < bestsme) { 1434 bestsme = thissme; 1435 *new_mv = best_mv->as_mv; 1436 } else { 1437 // The full search result is actually worse so re-instate the 1438 // previous best vector 1439 best_mv->as_mv = *new_mv; 1440 } 1441 } 1442 1443 if (bestsme < INT_MAX) { 1444 int distortion; 1445 cpi->find_fractional_mv_step( 1446 x, 1447 new_mv, 1448 &bsi->ref_mv[0]->as_mv, 1449 cm->allow_high_precision_mv, 1450 x->errorperbit, &cpi->fn_ptr[bsize], 1451 cpi->sf.mv.subpel_force_stop, 1452 cpi->sf.mv.subpel_iters_per_step, 1453 cond_sad_list(cpi, sad_list), 1454 x->nmvjointcost, x->mvcost, 1455 &distortion, 1456 &x->pred_sse[mbmi->ref_frame[0]], 1457 NULL, 0, 0); 1458 1459 // save motion search result for use in compound prediction 1460 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv; 1461 } 1462 1463 if (cpi->sf.adaptive_motion_search) 1464 x->pred_mv[mbmi->ref_frame[0]] = *new_mv; 1465 1466 // restore src pointers 1467 mi_buf_restore(x, orig_src, orig_pre); 1468 } 1469 1470 if (has_second_rf) { 1471 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV || 1472 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) 1473 continue; 1474 } 1475 1476 if (has_second_rf && this_mode == NEWMV && 1477 mbmi->interp_filter == EIGHTTAP) { 1478 // adjust src pointers 1479 mi_buf_shift(x, i); 1480 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) { 1481 int rate_mv; 1482 joint_motion_search(cpi, x, bsize, frame_mv[this_mode], 1483 mi_row, mi_col, seg_mvs[i], 1484 &rate_mv); 1485 seg_mvs[i][mbmi->ref_frame[0]].as_int = 1486 frame_mv[this_mode][mbmi->ref_frame[0]].as_int; 1487 seg_mvs[i][mbmi->ref_frame[1]].as_int = 1488 frame_mv[this_mode][mbmi->ref_frame[1]].as_int; 1489 } 1490 // restore src pointers 1491 mi_buf_restore(x, orig_src, orig_pre); 1492 } 1493 1494 bsi->rdstat[i][mode_idx].brate = 1495 set_and_cost_bmi_mvs(cpi, xd, i, this_mode, mode_mv[this_mode], 1496 frame_mv, seg_mvs[i], bsi->ref_mv, 1497 x->nmvjointcost, x->mvcost); 1498 1499 for (ref = 0; ref < 1 + has_second_rf; ++ref) { 1500 bsi->rdstat[i][mode_idx].mvs[ref].as_int = 1501 mode_mv[this_mode][ref].as_int; 1502 if (num_4x4_blocks_wide > 1) 1503 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int = 1504 mode_mv[this_mode][ref].as_int; 1505 if (num_4x4_blocks_high > 1) 1506 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int = 1507 mode_mv[this_mode][ref].as_int; 1508 } 1509 1510 // Trap vectors that reach beyond the UMV borders 1511 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) || 1512 (has_second_rf && 1513 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv))) 1514 continue; 1515 1516 if (filter_idx > 0) { 1517 BEST_SEG_INFO *ref_bsi = bsi_buf; 1518 subpelmv = 0; 1519 have_ref = 1; 1520 1521 for (ref = 0; ref < 1 + has_second_rf; ++ref) { 1522 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv); 1523 have_ref &= mode_mv[this_mode][ref].as_int == 1524 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int; 1525 } 1526 1527 if (filter_idx > 1 && !subpelmv && !have_ref) { 1528 ref_bsi = bsi_buf + 1; 1529 have_ref = 1; 1530 for (ref = 0; ref < 1 + has_second_rf; ++ref) 1531 have_ref &= mode_mv[this_mode][ref].as_int == 1532 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int; 1533 } 1534 1535 if (!subpelmv && have_ref && 1536 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) { 1537 vpx_memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx], 1538 sizeof(SEG_RDSTAT)); 1539 if (num_4x4_blocks_wide > 1) 1540 bsi->rdstat[i + 1][mode_idx].eobs = 1541 ref_bsi->rdstat[i + 1][mode_idx].eobs; 1542 if (num_4x4_blocks_high > 1) 1543 bsi->rdstat[i + 2][mode_idx].eobs = 1544 ref_bsi->rdstat[i + 2][mode_idx].eobs; 1545 1546 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) { 1547 mode_selected = this_mode; 1548 best_rd = bsi->rdstat[i][mode_idx].brdcost; 1549 } 1550 continue; 1551 } 1552 } 1553 1554 bsi->rdstat[i][mode_idx].brdcost = 1555 encode_inter_mb_segment(cpi, x, 1556 bsi->segment_rd - this_segment_rd, i, 1557 &bsi->rdstat[i][mode_idx].byrate, 1558 &bsi->rdstat[i][mode_idx].bdist, 1559 &bsi->rdstat[i][mode_idx].bsse, 1560 bsi->rdstat[i][mode_idx].ta, 1561 bsi->rdstat[i][mode_idx].tl, 1562 mi_row, mi_col); 1563 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) { 1564 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv, 1565 bsi->rdstat[i][mode_idx].brate, 0); 1566 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate; 1567 bsi->rdstat[i][mode_idx].eobs = p->eobs[i]; 1568 if (num_4x4_blocks_wide > 1) 1569 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1]; 1570 if (num_4x4_blocks_high > 1) 1571 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2]; 1572 } 1573 1574 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) { 1575 mode_selected = this_mode; 1576 best_rd = bsi->rdstat[i][mode_idx].brdcost; 1577 } 1578 } /*for each 4x4 mode*/ 1579 1580 if (best_rd == INT64_MAX) { 1581 int iy, midx; 1582 for (iy = i + 1; iy < 4; ++iy) 1583 for (midx = 0; midx < INTER_MODES; ++midx) 1584 bsi->rdstat[iy][midx].brdcost = INT64_MAX; 1585 bsi->segment_rd = INT64_MAX; 1586 return INT64_MAX;; 1587 } 1588 1589 mode_idx = INTER_OFFSET(mode_selected); 1590 vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above)); 1591 vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left)); 1592 1593 set_and_cost_bmi_mvs(cpi, xd, i, mode_selected, mode_mv[mode_selected], 1594 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost, 1595 x->mvcost); 1596 1597 br += bsi->rdstat[i][mode_idx].brate; 1598 bd += bsi->rdstat[i][mode_idx].bdist; 1599 block_sse += bsi->rdstat[i][mode_idx].bsse; 1600 segmentyrate += bsi->rdstat[i][mode_idx].byrate; 1601 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost; 1602 1603 if (this_segment_rd > bsi->segment_rd) { 1604 int iy, midx; 1605 for (iy = i + 1; iy < 4; ++iy) 1606 for (midx = 0; midx < INTER_MODES; ++midx) 1607 bsi->rdstat[iy][midx].brdcost = INT64_MAX; 1608 bsi->segment_rd = INT64_MAX; 1609 return INT64_MAX;; 1610 } 1611 } 1612 } /* for each label */ 1613 1614 bsi->r = br; 1615 bsi->d = bd; 1616 bsi->segment_yrate = segmentyrate; 1617 bsi->segment_rd = this_segment_rd; 1618 bsi->sse = block_sse; 1619 1620 // update the coding decisions 1621 for (k = 0; k < 4; ++k) 1622 bsi->modes[k] = mi->bmi[k].as_mode; 1623 1624 if (bsi->segment_rd > best_rd) 1625 return INT64_MAX; 1626 /* set it to the best */ 1627 for (i = 0; i < 4; i++) { 1628 mode_idx = INTER_OFFSET(bsi->modes[i]); 1629 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int; 1630 if (has_second_ref(mbmi)) 1631 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int; 1632 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs; 1633 mi->bmi[i].as_mode = bsi->modes[i]; 1634 } 1635 1636 /* 1637 * used to set mbmi->mv.as_int 1638 */ 1639 *returntotrate = bsi->r; 1640 *returndistortion = bsi->d; 1641 *returnyrate = bsi->segment_yrate; 1642 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0); 1643 *psse = bsi->sse; 1644 mbmi->mode = bsi->modes[3]; 1645 1646 return bsi->segment_rd; 1647} 1648 1649static void estimate_ref_frame_costs(const VP9_COMMON *cm, 1650 const MACROBLOCKD *xd, 1651 int segment_id, 1652 unsigned int *ref_costs_single, 1653 unsigned int *ref_costs_comp, 1654 vp9_prob *comp_mode_p) { 1655 int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id, 1656 SEG_LVL_REF_FRAME); 1657 if (seg_ref_active) { 1658 vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single)); 1659 vpx_memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp)); 1660 *comp_mode_p = 128; 1661 } else { 1662 vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd); 1663 vp9_prob comp_inter_p = 128; 1664 1665 if (cm->reference_mode == REFERENCE_MODE_SELECT) { 1666 comp_inter_p = vp9_get_reference_mode_prob(cm, xd); 1667 *comp_mode_p = comp_inter_p; 1668 } else { 1669 *comp_mode_p = 128; 1670 } 1671 1672 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0); 1673 1674 if (cm->reference_mode != COMPOUND_REFERENCE) { 1675 vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd); 1676 vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd); 1677 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1); 1678 1679 if (cm->reference_mode == REFERENCE_MODE_SELECT) 1680 base_cost += vp9_cost_bit(comp_inter_p, 0); 1681 1682 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] = 1683 ref_costs_single[ALTREF_FRAME] = base_cost; 1684 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0); 1685 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1); 1686 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1); 1687 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0); 1688 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1); 1689 } else { 1690 ref_costs_single[LAST_FRAME] = 512; 1691 ref_costs_single[GOLDEN_FRAME] = 512; 1692 ref_costs_single[ALTREF_FRAME] = 512; 1693 } 1694 if (cm->reference_mode != SINGLE_REFERENCE) { 1695 vp9_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd); 1696 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1); 1697 1698 if (cm->reference_mode == REFERENCE_MODE_SELECT) 1699 base_cost += vp9_cost_bit(comp_inter_p, 1); 1700 1701 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0); 1702 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1); 1703 } else { 1704 ref_costs_comp[LAST_FRAME] = 512; 1705 ref_costs_comp[GOLDEN_FRAME] = 512; 1706 } 1707 } 1708} 1709 1710static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, 1711 int mode_index, 1712 int64_t comp_pred_diff[REFERENCE_MODES], 1713 const int64_t tx_size_diff[TX_MODES], 1714 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS], 1715 int skippable) { 1716 MACROBLOCKD *const xd = &x->e_mbd; 1717 1718 // Take a snapshot of the coding context so it can be 1719 // restored if we decide to encode this way 1720 ctx->skip = x->skip; 1721 ctx->skippable = skippable; 1722 ctx->best_mode_index = mode_index; 1723 ctx->mic = *xd->mi[0].src_mi; 1724 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE]; 1725 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE]; 1726 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT]; 1727 1728 vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff)); 1729 vpx_memcpy(ctx->best_filter_diff, best_filter_diff, 1730 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS); 1731} 1732 1733static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x, 1734 const TileInfo *const tile, 1735 MV_REFERENCE_FRAME ref_frame, 1736 BLOCK_SIZE block_size, 1737 int mi_row, int mi_col, 1738 int_mv frame_nearest_mv[MAX_REF_FRAMES], 1739 int_mv frame_near_mv[MAX_REF_FRAMES], 1740 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) { 1741 const VP9_COMMON *cm = &cpi->common; 1742 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame); 1743 MACROBLOCKD *const xd = &x->e_mbd; 1744 MODE_INFO *const mi = xd->mi[0].src_mi; 1745 int_mv *const candidates = mi->mbmi.ref_mvs[ref_frame]; 1746 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf; 1747 1748 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this 1749 // use the UV scaling factors. 1750 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf); 1751 1752 // Gets an initial list of candidate vectors from neighbours and orders them 1753 vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col); 1754 1755 // Candidate refinement carried out at encoder and decoder 1756 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates, 1757 &frame_nearest_mv[ref_frame], 1758 &frame_near_mv[ref_frame]); 1759 1760 // Further refinement that is encode side only to test the top few candidates 1761 // in full and choose the best as the centre point for subsequent searches. 1762 // The current implementation doesn't support scaling. 1763 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8) 1764 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, 1765 ref_frame, block_size); 1766} 1767 1768static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, 1769 BLOCK_SIZE bsize, 1770 int mi_row, int mi_col, 1771 int_mv *tmp_mv, int *rate_mv) { 1772 MACROBLOCKD *xd = &x->e_mbd; 1773 const VP9_COMMON *cm = &cpi->common; 1774 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi; 1775 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}}; 1776 int bestsme = INT_MAX; 1777 int step_param; 1778 int sadpb = x->sadperbit16; 1779 MV mvp_full; 1780 int ref = mbmi->ref_frame[0]; 1781 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; 1782 1783 int tmp_col_min = x->mv_col_min; 1784 int tmp_col_max = x->mv_col_max; 1785 int tmp_row_min = x->mv_row_min; 1786 int tmp_row_max = x->mv_row_max; 1787 int sad_list[5]; 1788 1789 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, 1790 ref); 1791 1792 MV pred_mv[3]; 1793 pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv; 1794 pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv; 1795 pred_mv[2] = x->pred_mv[ref]; 1796 1797 if (scaled_ref_frame) { 1798 int i; 1799 // Swap out the reference frame for a version that's been scaled to 1800 // match the resolution of the current frame, allowing the existing 1801 // motion search code to be used without additional modifications. 1802 for (i = 0; i < MAX_MB_PLANE; i++) 1803 backup_yv12[i] = xd->plane[i].pre[0]; 1804 1805 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); 1806 } 1807 1808 vp9_set_mv_search_range(x, &ref_mv); 1809 1810 // Work out the size of the first step in the mv step search. 1811 // 0 here is maximum length first step. 1 is MAX >> 1 etc. 1812 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) { 1813 // Take wtd average of the step_params based on the last frame's 1814 // max mv magnitude and that based on the best ref mvs of the current 1815 // block for the given reference. 1816 step_param = (vp9_init_search_range(x->max_mv_context[ref]) + 1817 cpi->mv_step_param) / 2; 1818 } else { 1819 step_param = cpi->mv_step_param; 1820 } 1821 1822 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) { 1823 int boffset = 2 * (b_width_log2(BLOCK_64X64) - MIN(b_height_log2(bsize), 1824 b_width_log2(bsize))); 1825 step_param = MAX(step_param, boffset); 1826 } 1827 1828 if (cpi->sf.adaptive_motion_search) { 1829 int bwl = b_width_log2(bsize); 1830 int bhl = b_height_log2(bsize); 1831 int i; 1832 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4); 1833 1834 if (tlevel < 5) 1835 step_param += 2; 1836 1837 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) { 1838 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) { 1839 x->pred_mv[ref].row = 0; 1840 x->pred_mv[ref].col = 0; 1841 tmp_mv->as_int = INVALID_MV; 1842 1843 if (scaled_ref_frame) { 1844 int i; 1845 for (i = 0; i < MAX_MB_PLANE; i++) 1846 xd->plane[i].pre[0] = backup_yv12[i]; 1847 } 1848 return; 1849 } 1850 } 1851 } 1852 1853 mvp_full = pred_mv[x->mv_best_ref_index[ref]]; 1854 1855 mvp_full.col >>= 3; 1856 mvp_full.row >>= 3; 1857 1858 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb, 1859 cond_sad_list(cpi, sad_list), 1860 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1); 1861 1862 x->mv_col_min = tmp_col_min; 1863 x->mv_col_max = tmp_col_max; 1864 x->mv_row_min = tmp_row_min; 1865 x->mv_row_max = tmp_row_max; 1866 1867 if (bestsme < INT_MAX) { 1868 int dis; /* TODO: use dis in distortion calculation later. */ 1869 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv, 1870 cm->allow_high_precision_mv, 1871 x->errorperbit, 1872 &cpi->fn_ptr[bsize], 1873 cpi->sf.mv.subpel_force_stop, 1874 cpi->sf.mv.subpel_iters_per_step, 1875 cond_sad_list(cpi, sad_list), 1876 x->nmvjointcost, x->mvcost, 1877 &dis, &x->pred_sse[ref], NULL, 0, 0); 1878 } 1879 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, 1880 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 1881 1882 if (cpi->sf.adaptive_motion_search) 1883 x->pred_mv[ref] = tmp_mv->as_mv; 1884 1885 if (scaled_ref_frame) { 1886 int i; 1887 for (i = 0; i < MAX_MB_PLANE; i++) 1888 xd->plane[i].pre[0] = backup_yv12[i]; 1889 } 1890} 1891 1892static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, 1893 BLOCK_SIZE bsize, 1894 int_mv *frame_mv, 1895 int mi_row, int mi_col, 1896 int_mv single_newmv[MAX_REF_FRAMES], 1897 int *rate_mv) { 1898 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize]; 1899 const int ph = 4 * num_4x4_blocks_high_lookup[bsize]; 1900 MACROBLOCKD *xd = &x->e_mbd; 1901 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi; 1902 const int refs[2] = { mbmi->ref_frame[0], 1903 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] }; 1904 int_mv ref_mv[2]; 1905 int ite, ref; 1906 // Prediction buffer from second frame. 1907 uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t)); 1908 const InterpKernel *kernel = vp9_get_interp_kernel(mbmi->interp_filter); 1909 1910 // Do joint motion search in compound mode to get more accurate mv. 1911 struct buf_2d backup_yv12[2][MAX_MB_PLANE]; 1912 struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0]; 1913 int last_besterr[2] = {INT_MAX, INT_MAX}; 1914 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = { 1915 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]), 1916 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1]) 1917 }; 1918 1919 for (ref = 0; ref < 2; ++ref) { 1920 ref_mv[ref] = mbmi->ref_mvs[refs[ref]][0]; 1921 1922 if (scaled_ref_frame[ref]) { 1923 int i; 1924 // Swap out the reference frame for a version that's been scaled to 1925 // match the resolution of the current frame, allowing the existing 1926 // motion search code to be used without additional modifications. 1927 for (i = 0; i < MAX_MB_PLANE; i++) 1928 backup_yv12[ref][i] = xd->plane[i].pre[ref]; 1929 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col, 1930 NULL); 1931 } 1932 1933 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int; 1934 } 1935 1936 // Allow joint search multiple times iteratively for each ref frame 1937 // and break out the search loop if it couldn't find better mv. 1938 for (ite = 0; ite < 4; ite++) { 1939 struct buf_2d ref_yv12[2]; 1940 int bestsme = INT_MAX; 1941 int sadpb = x->sadperbit16; 1942 MV tmp_mv; 1943 int search_range = 3; 1944 1945 int tmp_col_min = x->mv_col_min; 1946 int tmp_col_max = x->mv_col_max; 1947 int tmp_row_min = x->mv_row_min; 1948 int tmp_row_max = x->mv_row_max; 1949 int id = ite % 2; 1950 1951 // Initialized here because of compiler problem in Visual Studio. 1952 ref_yv12[0] = xd->plane[0].pre[0]; 1953 ref_yv12[1] = xd->plane[0].pre[1]; 1954 1955 // Get pred block from second frame. 1956 vp9_build_inter_predictor(ref_yv12[!id].buf, 1957 ref_yv12[!id].stride, 1958 second_pred, pw, 1959 &frame_mv[refs[!id]].as_mv, 1960 &xd->block_refs[!id]->sf, 1961 pw, ph, 0, 1962 kernel, MV_PRECISION_Q3, 1963 mi_col * MI_SIZE, mi_row * MI_SIZE); 1964 1965 // Compound motion search on first ref frame. 1966 if (id) 1967 xd->plane[0].pre[0] = ref_yv12[id]; 1968 vp9_set_mv_search_range(x, &ref_mv[id].as_mv); 1969 1970 // Use mv result from single mode as mvp. 1971 tmp_mv = frame_mv[refs[id]].as_mv; 1972 1973 tmp_mv.col >>= 3; 1974 tmp_mv.row >>= 3; 1975 1976 // Small-range full-pixel motion search 1977 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, 1978 search_range, 1979 &cpi->fn_ptr[bsize], 1980 &ref_mv[id].as_mv, second_pred); 1981 if (bestsme < INT_MAX) 1982 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv, 1983 second_pred, &cpi->fn_ptr[bsize], 1); 1984 1985 x->mv_col_min = tmp_col_min; 1986 x->mv_col_max = tmp_col_max; 1987 x->mv_row_min = tmp_row_min; 1988 x->mv_row_max = tmp_row_max; 1989 1990 if (bestsme < INT_MAX) { 1991 int dis; /* TODO: use dis in distortion calculation later. */ 1992 unsigned int sse; 1993 bestsme = cpi->find_fractional_mv_step( 1994 x, &tmp_mv, 1995 &ref_mv[id].as_mv, 1996 cpi->common.allow_high_precision_mv, 1997 x->errorperbit, 1998 &cpi->fn_ptr[bsize], 1999 0, cpi->sf.mv.subpel_iters_per_step, 2000 NULL, 2001 x->nmvjointcost, x->mvcost, 2002 &dis, &sse, second_pred, 2003 pw, ph); 2004 } 2005 2006 if (id) 2007 xd->plane[0].pre[0] = scaled_first_yv12; 2008 2009 if (bestsme < last_besterr[id]) { 2010 frame_mv[refs[id]].as_mv = tmp_mv; 2011 last_besterr[id] = bestsme; 2012 } else { 2013 break; 2014 } 2015 } 2016 2017 *rate_mv = 0; 2018 2019 for (ref = 0; ref < 2; ++ref) { 2020 if (scaled_ref_frame[ref]) { 2021 // restore the predictor 2022 int i; 2023 for (i = 0; i < MAX_MB_PLANE; i++) 2024 xd->plane[i].pre[ref] = backup_yv12[ref][i]; 2025 } 2026 2027 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv, 2028 &mbmi->ref_mvs[refs[ref]][0].as_mv, 2029 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 2030 } 2031 2032 vpx_free(second_pred); 2033} 2034 2035static INLINE void restore_dst_buf(MACROBLOCKD *xd, 2036 uint8_t *orig_dst[MAX_MB_PLANE], 2037 int orig_dst_stride[MAX_MB_PLANE]) { 2038 int i; 2039 for (i = 0; i < MAX_MB_PLANE; i++) { 2040 xd->plane[i].dst.buf = orig_dst[i]; 2041 xd->plane[i].dst.stride = orig_dst_stride[i]; 2042 } 2043} 2044 2045static void rd_encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, 2046 BLOCK_SIZE bsize, int *rate2, 2047 int64_t *distortion, int64_t *distortion_uv, 2048 int *disable_skip) { 2049 VP9_COMMON *cm = &cpi->common; 2050 MACROBLOCKD *xd = &x->e_mbd; 2051 const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]); 2052 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]); 2053 unsigned int var, sse; 2054 // Skipping threshold for ac. 2055 unsigned int thresh_ac; 2056 // Skipping threshold for dc 2057 unsigned int thresh_dc; 2058 2059 var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride, 2060 xd->plane[0].dst.buf, 2061 xd->plane[0].dst.stride, &sse); 2062 2063 if (x->encode_breakout > 0) { 2064 // Set a maximum for threshold to avoid big PSNR loss in low bitrate 2065 // case. Use extreme low threshold for static frames to limit skipping. 2066 const unsigned int max_thresh = (cpi->allow_encode_breakout == 2067 ENCODE_BREAKOUT_LIMITED) ? 128 : 36000; 2068 // The encode_breakout input 2069 const unsigned int min_thresh = 2070 MIN(((unsigned int)x->encode_breakout << 4), max_thresh); 2071 2072 // Calculate threshold according to dequant value. 2073 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9; 2074 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh); 2075 2076 // Adjust threshold according to partition size. 2077 thresh_ac >>= 8 - (b_width_log2(bsize) + 2078 b_height_log2(bsize)); 2079 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6); 2080 } else { 2081 thresh_ac = 0; 2082 thresh_dc = 0; 2083 } 2084 2085 // Y skipping condition checking 2086 if (sse < thresh_ac || sse == 0) { 2087 // dc skipping checking 2088 if ((sse - var) < thresh_dc || sse == var) { 2089 unsigned int sse_u, sse_v; 2090 unsigned int var_u, var_v; 2091 2092 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, 2093 x->plane[1].src.stride, 2094 xd->plane[1].dst.buf, 2095 xd->plane[1].dst.stride, &sse_u); 2096 2097 // U skipping condition checking 2098 if ((sse_u * 4 < thresh_ac || sse_u == 0) && 2099 (sse_u - var_u < thresh_dc || sse_u == var_u)) { 2100 var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf, 2101 x->plane[2].src.stride, 2102 xd->plane[2].dst.buf, 2103 xd->plane[2].dst.stride, &sse_v); 2104 2105 // V skipping condition checking 2106 if ((sse_v * 4 < thresh_ac || sse_v == 0) && 2107 (sse_v - var_v < thresh_dc || sse_v == var_v)) { 2108 x->skip = 1; 2109 2110 // The cost of skip bit needs to be added. 2111 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2112 2113 // Scaling factor for SSE from spatial domain to frequency domain 2114 // is 16. Adjust distortion accordingly. 2115 *distortion_uv = (sse_u + sse_v) << 4; 2116 *distortion = (sse << 4) + *distortion_uv; 2117 2118 *disable_skip = 1; 2119 } 2120 } 2121 } 2122 } 2123} 2124 2125static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, 2126 BLOCK_SIZE bsize, 2127 int64_t txfm_cache[], 2128 int *rate2, int64_t *distortion, 2129 int *skippable, 2130 int *rate_y, int64_t *distortion_y, 2131 int *rate_uv, int64_t *distortion_uv, 2132 int *disable_skip, 2133 int_mv (*mode_mv)[MAX_REF_FRAMES], 2134 int mi_row, int mi_col, 2135 int_mv single_newmv[MAX_REF_FRAMES], 2136 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES], 2137 int (*single_skippable)[MAX_REF_FRAMES], 2138 int64_t *psse, 2139 const int64_t ref_best_rd) { 2140 VP9_COMMON *cm = &cpi->common; 2141 RD_OPT *rd_opt = &cpi->rd; 2142 MACROBLOCKD *xd = &x->e_mbd; 2143 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi; 2144 const int is_comp_pred = has_second_ref(mbmi); 2145 const int this_mode = mbmi->mode; 2146 int_mv *frame_mv = mode_mv[this_mode]; 2147 int i; 2148 int refs[2] = { mbmi->ref_frame[0], 2149 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) }; 2150 int_mv cur_mv[2]; 2151 int64_t this_rd = 0; 2152 DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64); 2153 int pred_exists = 0; 2154 int intpel_mv; 2155 int64_t rd, tmp_rd, best_rd = INT64_MAX; 2156 int best_needs_copy = 0; 2157 uint8_t *orig_dst[MAX_MB_PLANE]; 2158 int orig_dst_stride[MAX_MB_PLANE]; 2159 int rs = 0; 2160 INTERP_FILTER best_filter = SWITCHABLE; 2161 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0}; 2162 int64_t bsse[MAX_MB_PLANE << 2] = {0}; 2163 2164 int bsl = mi_width_log2_lookup[bsize]; 2165 int pred_filter_search = cpi->sf.cb_pred_filter_search ? 2166 (((mi_row + mi_col) >> bsl) + 2167 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0; 2168 2169 if (pred_filter_search) { 2170 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE; 2171 if (xd->up_available) 2172 af = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter; 2173 if (xd->left_available) 2174 lf = xd->mi[-1].src_mi->mbmi.interp_filter; 2175 2176 if ((this_mode != NEWMV) || (af == lf)) 2177 best_filter = af; 2178 } 2179 2180 if (is_comp_pred) { 2181 if (frame_mv[refs[0]].as_int == INVALID_MV || 2182 frame_mv[refs[1]].as_int == INVALID_MV) 2183 return INT64_MAX; 2184 2185 if (cpi->sf.adaptive_mode_search) { 2186 if (single_filter[this_mode][refs[0]] == 2187 single_filter[this_mode][refs[1]]) 2188 best_filter = single_filter[this_mode][refs[0]]; 2189 } 2190 } 2191 2192 if (this_mode == NEWMV) { 2193 int rate_mv; 2194 if (is_comp_pred) { 2195 // Initialize mv using single prediction mode result. 2196 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int; 2197 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int; 2198 2199 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) { 2200 joint_motion_search(cpi, x, bsize, frame_mv, 2201 mi_row, mi_col, single_newmv, &rate_mv); 2202 } else { 2203 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv, 2204 &mbmi->ref_mvs[refs[0]][0].as_mv, 2205 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 2206 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv, 2207 &mbmi->ref_mvs[refs[1]][0].as_mv, 2208 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 2209 } 2210 *rate2 += rate_mv; 2211 } else { 2212 int_mv tmp_mv; 2213 single_motion_search(cpi, x, bsize, mi_row, mi_col, 2214 &tmp_mv, &rate_mv); 2215 if (tmp_mv.as_int == INVALID_MV) 2216 return INT64_MAX; 2217 *rate2 += rate_mv; 2218 frame_mv[refs[0]].as_int = 2219 xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = tmp_mv.as_int; 2220 single_newmv[refs[0]].as_int = tmp_mv.as_int; 2221 } 2222 } 2223 2224 for (i = 0; i < is_comp_pred + 1; ++i) { 2225 cur_mv[i] = frame_mv[refs[i]]; 2226 // Clip "next_nearest" so that it does not extend to far out of image 2227 if (this_mode != NEWMV) 2228 clamp_mv2(&cur_mv[i].as_mv, xd); 2229 2230 if (mv_check_bounds(x, &cur_mv[i].as_mv)) 2231 return INT64_MAX; 2232 mbmi->mv[i].as_int = cur_mv[i].as_int; 2233 } 2234 2235 // do first prediction into the destination buffer. Do the next 2236 // prediction into a temporary buffer. Then keep track of which one 2237 // of these currently holds the best predictor, and use the other 2238 // one for future predictions. In the end, copy from tmp_buf to 2239 // dst if necessary. 2240 for (i = 0; i < MAX_MB_PLANE; i++) { 2241 orig_dst[i] = xd->plane[i].dst.buf; 2242 orig_dst_stride[i] = xd->plane[i].dst.stride; 2243 } 2244 2245 /* We don't include the cost of the second reference here, because there 2246 * are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other 2247 * words if you present them in that order, the second one is always known 2248 * if the first is known */ 2249 *rate2 += cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]); 2250 2251 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd && 2252 mbmi->mode != NEARESTMV) 2253 return INT64_MAX; 2254 2255 pred_exists = 0; 2256 // Are all MVs integer pel for Y and UV 2257 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv); 2258 if (is_comp_pred) 2259 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv); 2260 2261 // Search for best switchable filter by checking the variance of 2262 // pred error irrespective of whether the filter will be used 2263 rd_opt->mask_filter = 0; 2264 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) 2265 rd_opt->filter_cache[i] = INT64_MAX; 2266 2267 if (cm->interp_filter != BILINEAR) { 2268 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) { 2269 best_filter = EIGHTTAP; 2270 } else if (best_filter == SWITCHABLE) { 2271 int newbest; 2272 int tmp_rate_sum = 0; 2273 int64_t tmp_dist_sum = 0; 2274 2275 for (i = 0; i < SWITCHABLE_FILTERS; ++i) { 2276 int j; 2277 int64_t rs_rd; 2278 mbmi->interp_filter = i; 2279 rs = vp9_get_switchable_rate(cpi); 2280 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0); 2281 2282 if (i > 0 && intpel_mv) { 2283 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum); 2284 rd_opt->filter_cache[i] = rd; 2285 rd_opt->filter_cache[SWITCHABLE_FILTERS] = 2286 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd); 2287 if (cm->interp_filter == SWITCHABLE) 2288 rd += rs_rd; 2289 rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd); 2290 } else { 2291 int rate_sum = 0; 2292 int64_t dist_sum = 0; 2293 if (i > 0 && cpi->sf.adaptive_interp_filter_search && 2294 (cpi->sf.interp_filter_search_mask & (1 << i))) { 2295 rate_sum = INT_MAX; 2296 dist_sum = INT64_MAX; 2297 continue; 2298 } 2299 2300 if ((cm->interp_filter == SWITCHABLE && 2301 (!i || best_needs_copy)) || 2302 (cm->interp_filter != SWITCHABLE && 2303 (cm->interp_filter == mbmi->interp_filter || 2304 (i == 0 && intpel_mv)))) { 2305 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2306 } else { 2307 for (j = 0; j < MAX_MB_PLANE; j++) { 2308 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64; 2309 xd->plane[j].dst.stride = 64; 2310 } 2311 } 2312 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize); 2313 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum); 2314 2315 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum); 2316 rd_opt->filter_cache[i] = rd; 2317 rd_opt->filter_cache[SWITCHABLE_FILTERS] = 2318 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd); 2319 if (cm->interp_filter == SWITCHABLE) 2320 rd += rs_rd; 2321 rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd); 2322 2323 if (i == 0 && intpel_mv) { 2324 tmp_rate_sum = rate_sum; 2325 tmp_dist_sum = dist_sum; 2326 } 2327 } 2328 2329 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) { 2330 if (rd / 2 > ref_best_rd) { 2331 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2332 return INT64_MAX; 2333 } 2334 } 2335 newbest = i == 0 || rd < best_rd; 2336 2337 if (newbest) { 2338 best_rd = rd; 2339 best_filter = mbmi->interp_filter; 2340 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv) 2341 best_needs_copy = !best_needs_copy; 2342 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm)); 2343 vpx_memcpy(bsse, x->bsse, sizeof(bsse)); 2344 } 2345 2346 if ((cm->interp_filter == SWITCHABLE && newbest) || 2347 (cm->interp_filter != SWITCHABLE && 2348 cm->interp_filter == mbmi->interp_filter)) { 2349 pred_exists = 1; 2350 tmp_rd = best_rd; 2351 } 2352 } 2353 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2354 } 2355 } 2356 // Set the appropriate filter 2357 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ? 2358 cm->interp_filter : best_filter; 2359 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi) : 0; 2360 2361 if (pred_exists) { 2362 if (best_needs_copy) { 2363 // again temporarily set the buffers to local memory to prevent a memcpy 2364 for (i = 0; i < MAX_MB_PLANE; i++) { 2365 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64; 2366 xd->plane[i].dst.stride = 64; 2367 } 2368 } 2369 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0); 2370 } else { 2371 int tmp_rate; 2372 int64_t tmp_dist; 2373 // Handles the special case when a filter that is not in the 2374 // switchable list (ex. bilinear) is indicated at the frame level, or 2375 // skip condition holds. 2376 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize); 2377 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist); 2378 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist); 2379 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm)); 2380 vpx_memcpy(bsse, x->bsse, sizeof(bsse)); 2381 } 2382 2383 if (!is_comp_pred) 2384 single_filter[this_mode][refs[0]] = mbmi->interp_filter; 2385 2386 if (cpi->sf.adaptive_mode_search) 2387 if (is_comp_pred) 2388 if (single_skippable[this_mode][refs[0]] && 2389 single_skippable[this_mode][refs[1]]) 2390 vpx_memset(skip_txfm, 1, sizeof(skip_txfm)); 2391 2392 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) { 2393 // if current pred_error modeled rd is substantially more than the best 2394 // so far, do not bother doing full rd 2395 if (rd / 2 > ref_best_rd) { 2396 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2397 return INT64_MAX; 2398 } 2399 } 2400 2401 if (cm->interp_filter == SWITCHABLE) 2402 *rate2 += rs; 2403 2404 if (!is_comp_pred) { 2405 if (cpi->allow_encode_breakout) 2406 rd_encode_breakout_test(cpi, x, bsize, rate2, distortion, distortion_uv, 2407 disable_skip); 2408 } 2409 2410 vpx_memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm)); 2411 vpx_memcpy(x->bsse, bsse, sizeof(bsse)); 2412 2413 if (!x->skip) { 2414 int skippable_y, skippable_uv; 2415 int64_t sseuv = INT64_MAX; 2416 int64_t rdcosty = INT64_MAX; 2417 2418 // Y cost and distortion 2419 vp9_subtract_plane(x, bsize, 0); 2420 super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, psse, 2421 bsize, txfm_cache, ref_best_rd); 2422 2423 if (*rate_y == INT_MAX) { 2424 *rate2 = INT_MAX; 2425 *distortion = INT64_MAX; 2426 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2427 return INT64_MAX; 2428 } 2429 2430 *rate2 += *rate_y; 2431 *distortion += *distortion_y; 2432 2433 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion); 2434 rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse)); 2435 2436 super_block_uvrd(cpi, x, rate_uv, distortion_uv, &skippable_uv, &sseuv, 2437 bsize, ref_best_rd - rdcosty); 2438 if (*rate_uv == INT_MAX) { 2439 *rate2 = INT_MAX; 2440 *distortion = INT64_MAX; 2441 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2442 return INT64_MAX; 2443 } 2444 2445 *psse += sseuv; 2446 *rate2 += *rate_uv; 2447 *distortion += *distortion_uv; 2448 *skippable = skippable_y && skippable_uv; 2449 } 2450 2451 if (!is_comp_pred) 2452 single_skippable[this_mode][refs[0]] = *skippable; 2453 2454 restore_dst_buf(xd, orig_dst, orig_dst_stride); 2455 return this_rd; // if 0, this will be re-calculated by caller 2456} 2457 2458void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, 2459 int *returnrate, int64_t *returndist, 2460 BLOCK_SIZE bsize, 2461 PICK_MODE_CONTEXT *ctx, int64_t best_rd) { 2462 VP9_COMMON *const cm = &cpi->common; 2463 MACROBLOCKD *const xd = &x->e_mbd; 2464 struct macroblockd_plane *const pd = xd->plane; 2465 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0; 2466 int y_skip = 0, uv_skip = 0; 2467 int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 }; 2468 TX_SIZE max_uv_tx_size; 2469 x->skip_encode = 0; 2470 ctx->skip = 0; 2471 xd->mi[0].src_mi->mbmi.ref_frame[0] = INTRA_FRAME; 2472 2473 if (bsize >= BLOCK_8X8) { 2474 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, 2475 &dist_y, &y_skip, bsize, tx_cache, 2476 best_rd) >= best_rd) { 2477 *returnrate = INT_MAX; 2478 return; 2479 } 2480 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize, 2481 pd[1].subsampling_x, 2482 pd[1].subsampling_y); 2483 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, 2484 &dist_uv, &uv_skip, bsize, max_uv_tx_size); 2485 } else { 2486 y_skip = 0; 2487 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly, 2488 &dist_y, best_rd) >= best_rd) { 2489 *returnrate = INT_MAX; 2490 return; 2491 } 2492 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize, 2493 pd[1].subsampling_x, 2494 pd[1].subsampling_y); 2495 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, 2496 &dist_uv, &uv_skip, BLOCK_8X8, max_uv_tx_size); 2497 } 2498 2499 if (y_skip && uv_skip) { 2500 *returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly + 2501 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2502 *returndist = dist_y + dist_uv; 2503 vp9_zero(ctx->tx_rd_diff); 2504 } else { 2505 int i; 2506 *returnrate = rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 2507 *returndist = dist_y + dist_uv; 2508 if (cpi->sf.tx_size_search_method == USE_FULL_RD) 2509 for (i = 0; i < TX_MODES; i++) { 2510 if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX) 2511 ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode]; 2512 else 2513 ctx->tx_rd_diff[i] = 0; 2514 } 2515 } 2516 2517 ctx->mic = *xd->mi[0].src_mi; 2518} 2519 2520// Updating rd_thresh_freq_fact[] here means that the different 2521// partition/block sizes are handled independently based on the best 2522// choice for the current partition. It may well be better to keep a scaled 2523// best rd so far value and update rd_thresh_freq_fact based on the mode/size 2524// combination that wins out. 2525static void update_rd_thresh_fact(VP9_COMP *cpi, int bsize, 2526 int best_mode_index) { 2527 if (cpi->sf.adaptive_rd_thresh > 0) { 2528 const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES; 2529 int mode; 2530 for (mode = 0; mode < top_mode; ++mode) { 2531 int *const fact = &cpi->rd.thresh_freq_fact[bsize][mode]; 2532 2533 if (mode == best_mode_index) { 2534 *fact -= (*fact >> 3); 2535 } else { 2536 *fact = MIN(*fact + RD_THRESH_INC, 2537 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT); 2538 } 2539 } 2540 } 2541} 2542 2543int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, 2544 const TileInfo *const tile, 2545 int mi_row, int mi_col, 2546 int *returnrate, 2547 int64_t *returndistortion, 2548 BLOCK_SIZE bsize, 2549 PICK_MODE_CONTEXT *ctx, 2550 int64_t best_rd_so_far) { 2551 VP9_COMMON *const cm = &cpi->common; 2552 RD_OPT *const rd_opt = &cpi->rd; 2553 MACROBLOCKD *const xd = &x->e_mbd; 2554 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 2555 const struct segmentation *const seg = &cm->seg; 2556 struct macroblockd_plane *const pd = xd->plane; 2557 PREDICTION_MODE this_mode; 2558 MV_REFERENCE_FRAME ref_frame, second_ref_frame; 2559 unsigned char segment_id = mbmi->segment_id; 2560 int comp_pred, i, k; 2561 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; 2562 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; 2563 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } }; 2564 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES]; 2565 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES]; 2566 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, 2567 VP9_ALT_FLAG }; 2568 int64_t best_rd = best_rd_so_far; 2569 int64_t best_tx_rd[TX_MODES]; 2570 int64_t best_tx_diff[TX_MODES]; 2571 int64_t best_pred_diff[REFERENCE_MODES]; 2572 int64_t best_pred_rd[REFERENCE_MODES]; 2573 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS]; 2574 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS]; 2575 MB_MODE_INFO best_mbmode; 2576 int best_mode_skippable = 0; 2577 int midx, best_mode_index = -1; 2578 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES]; 2579 vp9_prob comp_mode_p; 2580 int64_t best_intra_rd = INT64_MAX; 2581 unsigned int best_pred_sse = UINT_MAX; 2582 PREDICTION_MODE best_intra_mode = DC_PRED; 2583 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES]; 2584 int64_t dist_uv[TX_SIZES]; 2585 int skip_uv[TX_SIZES]; 2586 PREDICTION_MODE mode_uv[TX_SIZES]; 2587 const int intra_cost_penalty = 2588 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth); 2589 int best_skip2 = 0; 2590 uint8_t ref_frame_skip_mask[2] = { 0 }; 2591 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 }; 2592 int mode_skip_start = cpi->sf.mode_skip_start + 1; 2593 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize]; 2594 const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize]; 2595 int mode_threshold[MAX_MODES]; 2596 int *mode_map = rd_opt->mode_map[bsize]; 2597 const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags; 2598 vp9_zero(best_mbmode); 2599 2600 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; 2601 2602 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp, 2603 &comp_mode_p); 2604 2605 for (i = 0; i < REFERENCE_MODES; ++i) 2606 best_pred_rd[i] = INT64_MAX; 2607 for (i = 0; i < TX_MODES; i++) 2608 best_tx_rd[i] = INT64_MAX; 2609 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) 2610 best_filter_rd[i] = INT64_MAX; 2611 for (i = 0; i < TX_SIZES; i++) 2612 rate_uv_intra[i] = INT_MAX; 2613 for (i = 0; i < MAX_REF_FRAMES; ++i) 2614 x->pred_sse[i] = INT_MAX; 2615 for (i = 0; i < MB_MODE_COUNT; ++i) { 2616 for (k = 0; k < MAX_REF_FRAMES; ++k) { 2617 single_inter_filter[i][k] = SWITCHABLE; 2618 single_skippable[i][k] = 0; 2619 } 2620 } 2621 2622 *returnrate = INT_MAX; 2623 2624 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { 2625 x->pred_mv_sad[ref_frame] = INT_MAX; 2626 if (cpi->ref_frame_flags & flag_list[ref_frame]) { 2627 setup_buffer_inter(cpi, x, tile, ref_frame, bsize, mi_row, mi_col, 2628 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb); 2629 } 2630 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV; 2631 frame_mv[ZEROMV][ref_frame].as_int = 0; 2632 } 2633 2634 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { 2635 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) { 2636 // Skip checking missing references in both single and compound reference 2637 // modes. Note that a mode will be skipped iff both reference frames 2638 // are masked out. 2639 ref_frame_skip_mask[0] |= (1 << ref_frame); 2640 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 2641 } else if (cpi->sf.reference_masking) { 2642 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) { 2643 // Skip fixed mv modes for poor references 2644 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) { 2645 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO; 2646 break; 2647 } 2648 } 2649 } 2650 // If the segment reference frame feature is enabled.... 2651 // then do nothing if the current ref frame is not allowed.. 2652 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) && 2653 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) { 2654 ref_frame_skip_mask[0] |= (1 << ref_frame); 2655 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 2656 } 2657 } 2658 2659 // Disable this drop out case if the ref frame 2660 // segment level feature is enabled for this segment. This is to 2661 // prevent the possibility that we end up unable to pick any mode. 2662 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) { 2663 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame, 2664 // unless ARNR filtering is enabled in which case we want 2665 // an unfiltered alternative. We allow near/nearest as well 2666 // because they may result in zero-zero MVs but be cheaper. 2667 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) { 2668 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME); 2669 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK; 2670 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO; 2671 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0) 2672 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV); 2673 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0) 2674 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV); 2675 } 2676 } 2677 2678 if (cpi->rc.is_src_frame_alt_ref) { 2679 if (cpi->sf.alt_ref_search_fp) { 2680 mode_skip_mask[ALTREF_FRAME] = 0; 2681 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME); 2682 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK; 2683 } 2684 } 2685 2686 if (bsize > cpi->sf.max_intra_bsize) { 2687 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME); 2688 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME); 2689 } 2690 2691 mode_skip_mask[INTRA_FRAME] |= 2692 ~(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]]); 2693 2694 for (i = 0; i < MAX_MODES; ++i) 2695 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5; 2696 2697 midx = cpi->sf.schedule_mode_search ? mode_skip_start : 0; 2698 while (midx > 4) { 2699 uint8_t end_pos = 0; 2700 for (i = 5; i < midx; ++i) { 2701 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) { 2702 uint8_t tmp = mode_map[i]; 2703 mode_map[i] = mode_map[i - 1]; 2704 mode_map[i - 1] = tmp; 2705 end_pos = i; 2706 } 2707 } 2708 midx = end_pos; 2709 } 2710 2711 for (midx = 0; midx < MAX_MODES; ++midx) { 2712 int mode_index = mode_map[midx]; 2713 int mode_excluded = 0; 2714 int64_t this_rd = INT64_MAX; 2715 int disable_skip = 0; 2716 int compmode_cost = 0; 2717 int rate2 = 0, rate_y = 0, rate_uv = 0; 2718 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0; 2719 int skippable = 0; 2720 int64_t tx_cache[TX_MODES]; 2721 int this_skip2 = 0; 2722 int64_t total_sse = INT64_MAX; 2723 int early_term = 0; 2724 2725 this_mode = vp9_mode_order[mode_index].mode; 2726 ref_frame = vp9_mode_order[mode_index].ref_frame[0]; 2727 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1]; 2728 2729 // Look at the reference frame of the best mode so far and set the 2730 // skip mask to look at a subset of the remaining modes. 2731 if (midx == mode_skip_start && best_mode_index >= 0) { 2732 switch (best_mbmode.ref_frame[0]) { 2733 case INTRA_FRAME: 2734 break; 2735 case LAST_FRAME: 2736 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK; 2737 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 2738 break; 2739 case GOLDEN_FRAME: 2740 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK; 2741 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 2742 break; 2743 case ALTREF_FRAME: 2744 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK; 2745 break; 2746 case NONE: 2747 case MAX_REF_FRAMES: 2748 assert(0 && "Invalid Reference frame"); 2749 break; 2750 } 2751 } 2752 2753 if (ref_frame_skip_mask[0] & (1 << ref_frame) && 2754 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame))) 2755 continue; 2756 2757 if (mode_skip_mask[ref_frame] & (1 << this_mode)) 2758 continue; 2759 2760 // Test best rd so far against threshold for trying this mode. 2761 if (best_mode_skippable && cpi->sf.schedule_mode_search) 2762 mode_threshold[mode_index] <<= 1; 2763 2764 if (best_rd < mode_threshold[mode_index]) 2765 continue; 2766 2767 if (cpi->sf.motion_field_mode_search) { 2768 const int mi_width = MIN(num_8x8_blocks_wide_lookup[bsize], 2769 tile->mi_col_end - mi_col); 2770 const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize], 2771 tile->mi_row_end - mi_row); 2772 const int bsl = mi_width_log2(bsize); 2773 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl) 2774 + get_chessboard_index(cm->current_video_frame)) & 0x1; 2775 MB_MODE_INFO *ref_mbmi; 2776 int const_motion = 1; 2777 int skip_ref_frame = !cb_partition_search_ctrl; 2778 MV_REFERENCE_FRAME rf = NONE; 2779 int_mv ref_mv; 2780 ref_mv.as_int = INVALID_MV; 2781 2782 if ((mi_row - 1) >= tile->mi_row_start) { 2783 ref_mv = xd->mi[-xd->mi_stride].src_mi->mbmi.mv[0]; 2784 rf = xd->mi[-xd->mi_stride].src_mi->mbmi.ref_frame[0]; 2785 for (i = 0; i < mi_width; ++i) { 2786 ref_mbmi = &xd->mi[-xd->mi_stride + i].src_mi->mbmi; 2787 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) && 2788 (ref_frame == ref_mbmi->ref_frame[0]); 2789 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]); 2790 } 2791 } 2792 2793 if ((mi_col - 1) >= tile->mi_col_start) { 2794 if (ref_mv.as_int == INVALID_MV) 2795 ref_mv = xd->mi[-1].src_mi->mbmi.mv[0]; 2796 if (rf == NONE) 2797 rf = xd->mi[-1].src_mi->mbmi.ref_frame[0]; 2798 for (i = 0; i < mi_height; ++i) { 2799 ref_mbmi = &xd->mi[i * xd->mi_stride - 1].src_mi->mbmi; 2800 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) && 2801 (ref_frame == ref_mbmi->ref_frame[0]); 2802 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]); 2803 } 2804 } 2805 2806 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV) 2807 if (rf > INTRA_FRAME) 2808 if (ref_frame != rf) 2809 continue; 2810 2811 if (const_motion) 2812 if (this_mode == NEARMV || this_mode == ZEROMV) 2813 continue; 2814 } 2815 2816 comp_pred = second_ref_frame > INTRA_FRAME; 2817 if (comp_pred) { 2818 if (!cm->allow_comp_inter_inter) 2819 continue; 2820 2821 // Skip compound inter modes if ARF is not available. 2822 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) 2823 continue; 2824 2825 // Do not allow compound prediction if the segment level reference frame 2826 // feature is in use as in this case there can only be one reference. 2827 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) 2828 continue; 2829 2830 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) && 2831 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME) 2832 continue; 2833 2834 mode_excluded = cm->reference_mode == SINGLE_REFERENCE; 2835 } else { 2836 if (ref_frame != INTRA_FRAME) 2837 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE; 2838 } 2839 2840 if (ref_frame == INTRA_FRAME) { 2841 if (cpi->sf.adaptive_mode_search) 2842 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse) 2843 continue; 2844 2845 if (this_mode != DC_PRED) { 2846 // Disable intra modes other than DC_PRED for blocks with low variance 2847 // Threshold for intra skipping based on source variance 2848 // TODO(debargha): Specialize the threshold for super block sizes 2849 const unsigned int skip_intra_var_thresh = 64; 2850 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) && 2851 x->source_variance < skip_intra_var_thresh) 2852 continue; 2853 // Only search the oblique modes if the best so far is 2854 // one of the neighboring directional modes 2855 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) && 2856 (this_mode >= D45_PRED && this_mode <= TM_PRED)) { 2857 if (best_mode_index >= 0 && 2858 best_mbmode.ref_frame[0] > INTRA_FRAME) 2859 continue; 2860 } 2861 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) { 2862 if (conditional_skipintra(this_mode, best_intra_mode)) 2863 continue; 2864 } 2865 } 2866 } else { 2867 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame}; 2868 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv, 2869 this_mode, ref_frames)) 2870 continue; 2871 } 2872 2873 mbmi->mode = this_mode; 2874 mbmi->uv_mode = DC_PRED; 2875 mbmi->ref_frame[0] = ref_frame; 2876 mbmi->ref_frame[1] = second_ref_frame; 2877 // Evaluate all sub-pel filters irrespective of whether we can use 2878 // them for this frame. 2879 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP 2880 : cm->interp_filter; 2881 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0; 2882 2883 x->skip = 0; 2884 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame); 2885 2886 // Select prediction reference frames. 2887 for (i = 0; i < MAX_MB_PLANE; i++) { 2888 xd->plane[i].pre[0] = yv12_mb[ref_frame][i]; 2889 if (comp_pred) 2890 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i]; 2891 } 2892 2893 for (i = 0; i < TX_MODES; ++i) 2894 tx_cache[i] = INT64_MAX; 2895 2896 if (ref_frame == INTRA_FRAME) { 2897 TX_SIZE uv_tx; 2898 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, 2899 NULL, bsize, tx_cache, best_rd); 2900 2901 if (rate_y == INT_MAX) 2902 continue; 2903 2904 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd[1].subsampling_x, 2905 pd[1].subsampling_y); 2906 if (rate_uv_intra[uv_tx] == INT_MAX) { 2907 choose_intra_uv_mode(cpi, ctx, bsize, uv_tx, 2908 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx], 2909 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]); 2910 } 2911 2912 rate_uv = rate_uv_tokenonly[uv_tx]; 2913 distortion_uv = dist_uv[uv_tx]; 2914 skippable = skippable && skip_uv[uv_tx]; 2915 mbmi->uv_mode = mode_uv[uv_tx]; 2916 2917 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx]; 2918 if (this_mode != DC_PRED && this_mode != TM_PRED) 2919 rate2 += intra_cost_penalty; 2920 distortion2 = distortion_y + distortion_uv; 2921 } else { 2922 this_rd = handle_inter_mode(cpi, x, bsize, 2923 tx_cache, 2924 &rate2, &distortion2, &skippable, 2925 &rate_y, &distortion_y, 2926 &rate_uv, &distortion_uv, 2927 &disable_skip, frame_mv, 2928 mi_row, mi_col, 2929 single_newmv, single_inter_filter, 2930 single_skippable, &total_sse, best_rd); 2931 if (this_rd == INT64_MAX) 2932 continue; 2933 2934 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred); 2935 2936 if (cm->reference_mode == REFERENCE_MODE_SELECT) 2937 rate2 += compmode_cost; 2938 } 2939 2940 // Estimate the reference frame signaling cost and add it 2941 // to the rolling cost variable. 2942 if (comp_pred) { 2943 rate2 += ref_costs_comp[ref_frame]; 2944 } else { 2945 rate2 += ref_costs_single[ref_frame]; 2946 } 2947 2948 if (!disable_skip) { 2949 if (skippable) { 2950 // Back out the coefficient coding costs 2951 rate2 -= (rate_y + rate_uv); 2952 2953 // Cost the skip mb case 2954 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2955 } else if (ref_frame != INTRA_FRAME && !xd->lossless) { 2956 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) < 2957 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) { 2958 // Add in the cost of the no skip flag. 2959 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 2960 } else { 2961 // FIXME(rbultje) make this work for splitmv also 2962 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2963 distortion2 = total_sse; 2964 assert(total_sse >= 0); 2965 rate2 -= (rate_y + rate_uv); 2966 this_skip2 = 1; 2967 } 2968 } else { 2969 // Add in the cost of the no skip flag. 2970 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 2971 } 2972 2973 // Calculate the final RD estimate for this mode. 2974 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2); 2975 } 2976 2977 if (ref_frame == INTRA_FRAME) { 2978 // Keep record of best intra rd 2979 if (this_rd < best_intra_rd) { 2980 best_intra_rd = this_rd; 2981 best_intra_mode = mbmi->mode; 2982 } 2983 } 2984 2985 if (!disable_skip && ref_frame == INTRA_FRAME) { 2986 for (i = 0; i < REFERENCE_MODES; ++i) 2987 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd); 2988 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) 2989 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd); 2990 } 2991 2992 // Did this mode help.. i.e. is it the new best mode 2993 if (this_rd < best_rd || x->skip) { 2994 int max_plane = MAX_MB_PLANE; 2995 if (!mode_excluded) { 2996 // Note index of best mode so far 2997 best_mode_index = mode_index; 2998 2999 if (ref_frame == INTRA_FRAME) { 3000 /* required for left and above block mv */ 3001 mbmi->mv[0].as_int = 0; 3002 max_plane = 1; 3003 } else { 3004 best_pred_sse = x->pred_sse[ref_frame]; 3005 } 3006 3007 *returnrate = rate2; 3008 *returndistortion = distortion2; 3009 best_rd = this_rd; 3010 best_mbmode = *mbmi; 3011 best_skip2 = this_skip2; 3012 best_mode_skippable = skippable; 3013 3014 if (!x->select_tx_size) 3015 swap_block_ptr(x, ctx, 1, 0, 0, max_plane); 3016 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size], 3017 sizeof(uint8_t) * ctx->num_4x4_blk); 3018 3019 // TODO(debargha): enhance this test with a better distortion prediction 3020 // based on qp, activity mask and history 3021 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) && 3022 (mode_index > MIN_EARLY_TERM_INDEX)) { 3023 const int qstep = xd->plane[0].dequant[1]; 3024 // TODO(debargha): Enhance this by specializing for each mode_index 3025 int scale = 4; 3026 if (x->source_variance < UINT_MAX) { 3027 const int var_adjust = (x->source_variance < 16); 3028 scale -= var_adjust; 3029 } 3030 if (ref_frame > INTRA_FRAME && 3031 distortion2 * scale < qstep * qstep) { 3032 early_term = 1; 3033 } 3034 } 3035 } 3036 } 3037 3038 /* keep record of best compound/single-only prediction */ 3039 if (!disable_skip && ref_frame != INTRA_FRAME) { 3040 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate; 3041 3042 if (cm->reference_mode == REFERENCE_MODE_SELECT) { 3043 single_rate = rate2 - compmode_cost; 3044 hybrid_rate = rate2; 3045 } else { 3046 single_rate = rate2; 3047 hybrid_rate = rate2 + compmode_cost; 3048 } 3049 3050 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2); 3051 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2); 3052 3053 if (!comp_pred) { 3054 if (single_rd < best_pred_rd[SINGLE_REFERENCE]) 3055 best_pred_rd[SINGLE_REFERENCE] = single_rd; 3056 } else { 3057 if (single_rd < best_pred_rd[COMPOUND_REFERENCE]) 3058 best_pred_rd[COMPOUND_REFERENCE] = single_rd; 3059 } 3060 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT]) 3061 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd; 3062 3063 /* keep record of best filter type */ 3064 if (!mode_excluded && cm->interp_filter != BILINEAR) { 3065 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ? 3066 SWITCHABLE_FILTERS : cm->interp_filter]; 3067 3068 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) { 3069 int64_t adj_rd; 3070 if (ref == INT64_MAX) 3071 adj_rd = 0; 3072 else if (rd_opt->filter_cache[i] == INT64_MAX) 3073 // when early termination is triggered, the encoder does not have 3074 // access to the rate-distortion cost. it only knows that the cost 3075 // should be above the maximum valid value. hence it takes the known 3076 // maximum plus an arbitrary constant as the rate-distortion cost. 3077 adj_rd = rd_opt->mask_filter - ref + 10; 3078 else 3079 adj_rd = rd_opt->filter_cache[i] - ref; 3080 3081 adj_rd += this_rd; 3082 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd); 3083 } 3084 } 3085 } 3086 3087 /* keep record of best txfm size */ 3088 if (bsize < BLOCK_32X32) { 3089 if (bsize < BLOCK_16X16) 3090 tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8]; 3091 3092 tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16]; 3093 } 3094 if (!mode_excluded && this_rd != INT64_MAX) { 3095 for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) { 3096 int64_t adj_rd = INT64_MAX; 3097 adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode]; 3098 3099 if (adj_rd < best_tx_rd[i]) 3100 best_tx_rd[i] = adj_rd; 3101 } 3102 } 3103 3104 if (early_term) 3105 break; 3106 3107 if (x->skip && !comp_pred) 3108 break; 3109 } 3110 3111 // The inter modes' rate costs are not calculated precisely in some cases. 3112 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and 3113 // ZEROMV. Here, checks are added for those cases, and the mode decisions 3114 // are corrected. 3115 if (best_mbmode.mode == NEWMV) { 3116 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0], 3117 best_mbmode.ref_frame[1]}; 3118 int comp_pred_mode = refs[1] > INTRA_FRAME; 3119 3120 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int && 3121 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int == 3122 best_mbmode.mv[1].as_int) || !comp_pred_mode)) 3123 best_mbmode.mode = NEARESTMV; 3124 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int && 3125 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int == 3126 best_mbmode.mv[1].as_int) || !comp_pred_mode)) 3127 best_mbmode.mode = NEARMV; 3128 else if (best_mbmode.mv[0].as_int == 0 && 3129 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode)) 3130 best_mbmode.mode = ZEROMV; 3131 } 3132 3133 if (best_mode_index < 0 || best_rd >= best_rd_so_far) 3134 return INT64_MAX; 3135 3136 // If we used an estimate for the uv intra rd in the loop above... 3137 if (cpi->sf.use_uv_intra_rd_estimate) { 3138 // Do Intra UV best rd mode selection if best mode choice above was intra. 3139 if (best_mbmode.ref_frame[0] == INTRA_FRAME) { 3140 TX_SIZE uv_tx_size; 3141 *mbmi = best_mbmode; 3142 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]); 3143 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size], 3144 &rate_uv_tokenonly[uv_tx_size], 3145 &dist_uv[uv_tx_size], 3146 &skip_uv[uv_tx_size], 3147 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, 3148 uv_tx_size); 3149 } 3150 } 3151 3152 assert((cm->interp_filter == SWITCHABLE) || 3153 (cm->interp_filter == best_mbmode.interp_filter) || 3154 !is_inter_block(&best_mbmode)); 3155 3156 if (!cpi->rc.is_src_frame_alt_ref) 3157 update_rd_thresh_fact(cpi, bsize, best_mode_index); 3158 3159 // macroblock modes 3160 *mbmi = best_mbmode; 3161 x->skip |= best_skip2; 3162 3163 for (i = 0; i < REFERENCE_MODES; ++i) { 3164 if (best_pred_rd[i] == INT64_MAX) 3165 best_pred_diff[i] = INT_MIN; 3166 else 3167 best_pred_diff[i] = best_rd - best_pred_rd[i]; 3168 } 3169 3170 if (!x->skip) { 3171 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) { 3172 if (best_filter_rd[i] == INT64_MAX) 3173 best_filter_diff[i] = 0; 3174 else 3175 best_filter_diff[i] = best_rd - best_filter_rd[i]; 3176 } 3177 if (cm->interp_filter == SWITCHABLE) 3178 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0); 3179 for (i = 0; i < TX_MODES; i++) { 3180 if (best_tx_rd[i] == INT64_MAX) 3181 best_tx_diff[i] = 0; 3182 else 3183 best_tx_diff[i] = best_rd - best_tx_rd[i]; 3184 } 3185 } else { 3186 vp9_zero(best_filter_diff); 3187 vp9_zero(best_tx_diff); 3188 } 3189 3190 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff 3191 // updating code causes PSNR loss. Need to figure out the confliction. 3192 x->skip |= best_mode_skippable; 3193 3194 store_coding_context(x, ctx, best_mode_index, best_pred_diff, 3195 best_tx_diff, best_filter_diff, best_mode_skippable); 3196 3197 return best_rd; 3198} 3199 3200int64_t vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, MACROBLOCK *x, 3201 int *returnrate, 3202 int64_t *returndistortion, 3203 BLOCK_SIZE bsize, 3204 PICK_MODE_CONTEXT *ctx, 3205 int64_t best_rd_so_far) { 3206 VP9_COMMON *const cm = &cpi->common; 3207 RD_OPT *const rd_opt = &cpi->rd; 3208 MACROBLOCKD *const xd = &x->e_mbd; 3209 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 3210 unsigned char segment_id = mbmi->segment_id; 3211 const int comp_pred = 0; 3212 int i; 3213 int64_t best_tx_diff[TX_MODES]; 3214 int64_t best_pred_diff[REFERENCE_MODES]; 3215 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS]; 3216 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES]; 3217 vp9_prob comp_mode_p; 3218 INTERP_FILTER best_filter = SWITCHABLE; 3219 int64_t this_rd = INT64_MAX; 3220 int rate2 = 0; 3221 const int64_t distortion2 = 0; 3222 3223 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; 3224 3225 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp, 3226 &comp_mode_p); 3227 3228 for (i = 0; i < MAX_REF_FRAMES; ++i) 3229 x->pred_sse[i] = INT_MAX; 3230 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i) 3231 x->pred_mv_sad[i] = INT_MAX; 3232 3233 *returnrate = INT_MAX; 3234 3235 assert(vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)); 3236 3237 mbmi->mode = ZEROMV; 3238 mbmi->uv_mode = DC_PRED; 3239 mbmi->ref_frame[0] = LAST_FRAME; 3240 mbmi->ref_frame[1] = NONE; 3241 mbmi->mv[0].as_int = 0; 3242 x->skip = 1; 3243 3244 // Search for best switchable filter by checking the variance of 3245 // pred error irrespective of whether the filter will be used 3246 rd_opt->mask_filter = 0; 3247 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) 3248 rd_opt->filter_cache[i] = INT64_MAX; 3249 3250 if (cm->interp_filter != BILINEAR) { 3251 best_filter = EIGHTTAP; 3252 if (cm->interp_filter == SWITCHABLE && 3253 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) { 3254 int rs; 3255 int best_rs = INT_MAX; 3256 for (i = 0; i < SWITCHABLE_FILTERS; ++i) { 3257 mbmi->interp_filter = i; 3258 rs = vp9_get_switchable_rate(cpi); 3259 if (rs < best_rs) { 3260 best_rs = rs; 3261 best_filter = mbmi->interp_filter; 3262 } 3263 } 3264 } 3265 } 3266 // Set the appropriate filter 3267 if (cm->interp_filter == SWITCHABLE) { 3268 mbmi->interp_filter = best_filter; 3269 rate2 += vp9_get_switchable_rate(cpi); 3270 } else { 3271 mbmi->interp_filter = cm->interp_filter; 3272 } 3273 3274 if (cm->reference_mode == REFERENCE_MODE_SELECT) 3275 rate2 += vp9_cost_bit(comp_mode_p, comp_pred); 3276 3277 // Estimate the reference frame signaling cost and add it 3278 // to the rolling cost variable. 3279 rate2 += ref_costs_single[LAST_FRAME]; 3280 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2); 3281 3282 *returnrate = rate2; 3283 *returndistortion = distortion2; 3284 3285 if (this_rd >= best_rd_so_far) 3286 return INT64_MAX; 3287 3288 assert((cm->interp_filter == SWITCHABLE) || 3289 (cm->interp_filter == mbmi->interp_filter)); 3290 3291 update_rd_thresh_fact(cpi, bsize, THR_ZEROMV); 3292 3293 vp9_zero(best_pred_diff); 3294 vp9_zero(best_filter_diff); 3295 vp9_zero(best_tx_diff); 3296 3297 if (!x->select_tx_size) 3298 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE); 3299 store_coding_context(x, ctx, THR_ZEROMV, 3300 best_pred_diff, best_tx_diff, best_filter_diff, 0); 3301 3302 return this_rd; 3303} 3304 3305int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, 3306 const TileInfo *const tile, 3307 int mi_row, int mi_col, 3308 int *returnrate, 3309 int64_t *returndistortion, 3310 BLOCK_SIZE bsize, 3311 PICK_MODE_CONTEXT *ctx, 3312 int64_t best_rd_so_far) { 3313 VP9_COMMON *const cm = &cpi->common; 3314 RD_OPT *const rd_opt = &cpi->rd; 3315 MACROBLOCKD *const xd = &x->e_mbd; 3316 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi; 3317 const struct segmentation *const seg = &cm->seg; 3318 MV_REFERENCE_FRAME ref_frame, second_ref_frame; 3319 unsigned char segment_id = mbmi->segment_id; 3320 int comp_pred, i; 3321 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; 3322 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; 3323 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, 3324 VP9_ALT_FLAG }; 3325 int64_t best_rd = best_rd_so_far; 3326 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise 3327 static const int64_t best_tx_diff[TX_MODES] = { 0 }; 3328 int64_t best_pred_diff[REFERENCE_MODES]; 3329 int64_t best_pred_rd[REFERENCE_MODES]; 3330 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS]; 3331 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS]; 3332 MB_MODE_INFO best_mbmode; 3333 int ref_index, best_ref_index = 0; 3334 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES]; 3335 vp9_prob comp_mode_p; 3336 INTERP_FILTER tmp_best_filter = SWITCHABLE; 3337 int rate_uv_intra, rate_uv_tokenonly; 3338 int64_t dist_uv; 3339 int skip_uv; 3340 PREDICTION_MODE mode_uv = DC_PRED; 3341 const int intra_cost_penalty = 3342 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth); 3343 int_mv seg_mvs[4][MAX_REF_FRAMES]; 3344 b_mode_info best_bmodes[4]; 3345 int best_skip2 = 0; 3346 int ref_frame_skip_mask[2] = { 0 }; 3347 3348 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; 3349 vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4); 3350 vp9_zero(best_mbmode); 3351 3352 for (i = 0; i < 4; i++) { 3353 int j; 3354 for (j = 0; j < MAX_REF_FRAMES; j++) 3355 seg_mvs[i][j].as_int = INVALID_MV; 3356 } 3357 3358 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp, 3359 &comp_mode_p); 3360 3361 for (i = 0; i < REFERENCE_MODES; ++i) 3362 best_pred_rd[i] = INT64_MAX; 3363 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) 3364 best_filter_rd[i] = INT64_MAX; 3365 rate_uv_intra = INT_MAX; 3366 3367 *returnrate = INT_MAX; 3368 3369 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) { 3370 if (cpi->ref_frame_flags & flag_list[ref_frame]) { 3371 setup_buffer_inter(cpi, x, tile, 3372 ref_frame, bsize, mi_row, mi_col, 3373 frame_mv[NEARESTMV], frame_mv[NEARMV], 3374 yv12_mb); 3375 } else { 3376 ref_frame_skip_mask[0] |= (1 << ref_frame); 3377 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 3378 } 3379 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV; 3380 frame_mv[ZEROMV][ref_frame].as_int = 0; 3381 } 3382 3383 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) { 3384 int mode_excluded = 0; 3385 int64_t this_rd = INT64_MAX; 3386 int disable_skip = 0; 3387 int compmode_cost = 0; 3388 int rate2 = 0, rate_y = 0, rate_uv = 0; 3389 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0; 3390 int skippable = 0; 3391 int i; 3392 int this_skip2 = 0; 3393 int64_t total_sse = INT_MAX; 3394 int early_term = 0; 3395 3396 ref_frame = vp9_ref_order[ref_index].ref_frame[0]; 3397 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1]; 3398 3399 // Look at the reference frame of the best mode so far and set the 3400 // skip mask to look at a subset of the remaining modes. 3401 if (ref_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) { 3402 if (ref_index == 3) { 3403 switch (best_mbmode.ref_frame[0]) { 3404 case INTRA_FRAME: 3405 break; 3406 case LAST_FRAME: 3407 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME); 3408 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 3409 break; 3410 case GOLDEN_FRAME: 3411 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME); 3412 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK; 3413 break; 3414 case ALTREF_FRAME: 3415 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME); 3416 break; 3417 case NONE: 3418 case MAX_REF_FRAMES: 3419 assert(0 && "Invalid Reference frame"); 3420 break; 3421 } 3422 } 3423 } 3424 3425 if (ref_frame_skip_mask[0] & (1 << ref_frame) && 3426 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame))) 3427 continue; 3428 3429 // Test best rd so far against threshold for trying this mode. 3430 if (rd_less_than_thresh(best_rd, 3431 rd_opt->threshes[segment_id][bsize][ref_index], 3432 rd_opt->thresh_freq_fact[bsize][ref_index])) 3433 continue; 3434 3435 comp_pred = second_ref_frame > INTRA_FRAME; 3436 if (comp_pred) { 3437 if (!cm->allow_comp_inter_inter) 3438 continue; 3439 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) 3440 continue; 3441 // Do not allow compound prediction if the segment level reference frame 3442 // feature is in use as in this case there can only be one reference. 3443 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) 3444 continue; 3445 3446 if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) && 3447 best_mbmode.ref_frame[0] == INTRA_FRAME) 3448 continue; 3449 } 3450 3451 // TODO(jingning, jkoleszar): scaling reference frame not supported for 3452 // sub8x8 blocks. 3453 if (ref_frame > INTRA_FRAME && 3454 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf)) 3455 continue; 3456 3457 if (second_ref_frame > INTRA_FRAME && 3458 vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf)) 3459 continue; 3460 3461 if (comp_pred) 3462 mode_excluded = cm->reference_mode == SINGLE_REFERENCE; 3463 else if (ref_frame != INTRA_FRAME) 3464 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE; 3465 3466 // If the segment reference frame feature is enabled.... 3467 // then do nothing if the current ref frame is not allowed.. 3468 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) && 3469 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) { 3470 continue; 3471 // Disable this drop out case if the ref frame 3472 // segment level feature is enabled for this segment. This is to 3473 // prevent the possibility that we end up unable to pick any mode. 3474 } else if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) { 3475 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame, 3476 // unless ARNR filtering is enabled in which case we want 3477 // an unfiltered alternative. We allow near/nearest as well 3478 // because they may result in zero-zero MVs but be cheaper. 3479 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) 3480 continue; 3481 } 3482 3483 mbmi->tx_size = TX_4X4; 3484 mbmi->uv_mode = DC_PRED; 3485 mbmi->ref_frame[0] = ref_frame; 3486 mbmi->ref_frame[1] = second_ref_frame; 3487 // Evaluate all sub-pel filters irrespective of whether we can use 3488 // them for this frame. 3489 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP 3490 : cm->interp_filter; 3491 x->skip = 0; 3492 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame); 3493 3494 // Select prediction reference frames. 3495 for (i = 0; i < MAX_MB_PLANE; i++) { 3496 xd->plane[i].pre[0] = yv12_mb[ref_frame][i]; 3497 if (comp_pred) 3498 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i]; 3499 } 3500 3501 if (ref_frame == INTRA_FRAME) { 3502 int rate; 3503 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, 3504 &distortion_y, best_rd) >= best_rd) 3505 continue; 3506 rate2 += rate; 3507 rate2 += intra_cost_penalty; 3508 distortion2 += distortion_y; 3509 3510 if (rate_uv_intra == INT_MAX) { 3511 choose_intra_uv_mode(cpi, ctx, bsize, TX_4X4, 3512 &rate_uv_intra, 3513 &rate_uv_tokenonly, 3514 &dist_uv, &skip_uv, 3515 &mode_uv); 3516 } 3517 rate2 += rate_uv_intra; 3518 rate_uv = rate_uv_tokenonly; 3519 distortion2 += dist_uv; 3520 distortion_uv = dist_uv; 3521 mbmi->uv_mode = mode_uv; 3522 } else { 3523 int rate; 3524 int64_t distortion; 3525 int64_t this_rd_thresh; 3526 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX; 3527 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX; 3528 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse; 3529 int tmp_best_skippable = 0; 3530 int switchable_filter_index; 3531 int_mv *second_ref = comp_pred ? 3532 &mbmi->ref_mvs[second_ref_frame][0] : NULL; 3533 b_mode_info tmp_best_bmodes[16]; 3534 MB_MODE_INFO tmp_best_mbmode; 3535 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS]; 3536 int pred_exists = 0; 3537 int uv_skippable; 3538 3539 this_rd_thresh = (ref_frame == LAST_FRAME) ? 3540 rd_opt->threshes[segment_id][bsize][THR_LAST] : 3541 rd_opt->threshes[segment_id][bsize][THR_ALTR]; 3542 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ? 3543 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh; 3544 rd_opt->mask_filter = 0; 3545 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) 3546 rd_opt->filter_cache[i] = INT64_MAX; 3547 3548 if (cm->interp_filter != BILINEAR) { 3549 tmp_best_filter = EIGHTTAP; 3550 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) { 3551 tmp_best_filter = EIGHTTAP; 3552 } else if (cpi->sf.adaptive_pred_interp_filter == 1 && 3553 ctx->pred_interp_filter < SWITCHABLE) { 3554 tmp_best_filter = ctx->pred_interp_filter; 3555 } else if (cpi->sf.adaptive_pred_interp_filter == 2) { 3556 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ? 3557 ctx->pred_interp_filter : 0; 3558 } else { 3559 for (switchable_filter_index = 0; 3560 switchable_filter_index < SWITCHABLE_FILTERS; 3561 ++switchable_filter_index) { 3562 int newbest, rs; 3563 int64_t rs_rd; 3564 mbmi->interp_filter = switchable_filter_index; 3565 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile, 3566 &mbmi->ref_mvs[ref_frame][0], 3567 second_ref, best_yrd, &rate, 3568 &rate_y, &distortion, 3569 &skippable, &total_sse, 3570 (int) this_rd_thresh, seg_mvs, 3571 bsi, switchable_filter_index, 3572 mi_row, mi_col); 3573 3574 if (tmp_rd == INT64_MAX) 3575 continue; 3576 rs = vp9_get_switchable_rate(cpi); 3577 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0); 3578 rd_opt->filter_cache[switchable_filter_index] = tmp_rd; 3579 rd_opt->filter_cache[SWITCHABLE_FILTERS] = 3580 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], 3581 tmp_rd + rs_rd); 3582 if (cm->interp_filter == SWITCHABLE) 3583 tmp_rd += rs_rd; 3584 3585 rd_opt->mask_filter = MAX(rd_opt->mask_filter, tmp_rd); 3586 3587 newbest = (tmp_rd < tmp_best_rd); 3588 if (newbest) { 3589 tmp_best_filter = mbmi->interp_filter; 3590 tmp_best_rd = tmp_rd; 3591 } 3592 if ((newbest && cm->interp_filter == SWITCHABLE) || 3593 (mbmi->interp_filter == cm->interp_filter && 3594 cm->interp_filter != SWITCHABLE)) { 3595 tmp_best_rdu = tmp_rd; 3596 tmp_best_rate = rate; 3597 tmp_best_ratey = rate_y; 3598 tmp_best_distortion = distortion; 3599 tmp_best_sse = total_sse; 3600 tmp_best_skippable = skippable; 3601 tmp_best_mbmode = *mbmi; 3602 for (i = 0; i < 4; i++) { 3603 tmp_best_bmodes[i] = xd->mi[0].src_mi->bmi[i]; 3604 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i]; 3605 } 3606 pred_exists = 1; 3607 if (switchable_filter_index == 0 && 3608 cpi->sf.use_rd_breakout && 3609 best_rd < INT64_MAX) { 3610 if (tmp_best_rdu / 2 > best_rd) { 3611 // skip searching the other filters if the first is 3612 // already substantially larger than the best so far 3613 tmp_best_filter = mbmi->interp_filter; 3614 tmp_best_rdu = INT64_MAX; 3615 break; 3616 } 3617 } 3618 } 3619 } // switchable_filter_index loop 3620 } 3621 } 3622 3623 if (tmp_best_rdu == INT64_MAX && pred_exists) 3624 continue; 3625 3626 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ? 3627 tmp_best_filter : cm->interp_filter); 3628 if (!pred_exists) { 3629 // Handles the special case when a filter that is not in the 3630 // switchable list (bilinear, 6-tap) is indicated at the frame level 3631 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile, 3632 &mbmi->ref_mvs[ref_frame][0], 3633 second_ref, best_yrd, &rate, &rate_y, 3634 &distortion, &skippable, &total_sse, 3635 (int) this_rd_thresh, seg_mvs, bsi, 0, 3636 mi_row, mi_col); 3637 if (tmp_rd == INT64_MAX) 3638 continue; 3639 } else { 3640 total_sse = tmp_best_sse; 3641 rate = tmp_best_rate; 3642 rate_y = tmp_best_ratey; 3643 distortion = tmp_best_distortion; 3644 skippable = tmp_best_skippable; 3645 *mbmi = tmp_best_mbmode; 3646 for (i = 0; i < 4; i++) 3647 xd->mi[0].src_mi->bmi[i] = tmp_best_bmodes[i]; 3648 } 3649 3650 rate2 += rate; 3651 distortion2 += distortion; 3652 3653 if (cm->interp_filter == SWITCHABLE) 3654 rate2 += vp9_get_switchable_rate(cpi); 3655 3656 if (!mode_excluded) 3657 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE 3658 : cm->reference_mode == COMPOUND_REFERENCE; 3659 3660 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred); 3661 3662 tmp_best_rdu = best_rd - 3663 MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2), 3664 RDCOST(x->rdmult, x->rddiv, 0, total_sse)); 3665 3666 if (tmp_best_rdu > 0) { 3667 // If even the 'Y' rd value of split is higher than best so far 3668 // then dont bother looking at UV 3669 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, 3670 BLOCK_8X8); 3671 super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable, 3672 &uv_sse, BLOCK_8X8, tmp_best_rdu); 3673 if (rate_uv == INT_MAX) 3674 continue; 3675 rate2 += rate_uv; 3676 distortion2 += distortion_uv; 3677 skippable = skippable && uv_skippable; 3678 total_sse += uv_sse; 3679 } 3680 } 3681 3682 if (cm->reference_mode == REFERENCE_MODE_SELECT) 3683 rate2 += compmode_cost; 3684 3685 // Estimate the reference frame signaling cost and add it 3686 // to the rolling cost variable. 3687 if (second_ref_frame > INTRA_FRAME) { 3688 rate2 += ref_costs_comp[ref_frame]; 3689 } else { 3690 rate2 += ref_costs_single[ref_frame]; 3691 } 3692 3693 if (!disable_skip) { 3694 // Skip is never coded at the segment level for sub8x8 blocks and instead 3695 // always coded in the bitstream at the mode info level. 3696 3697 if (ref_frame != INTRA_FRAME && !xd->lossless) { 3698 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) < 3699 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) { 3700 // Add in the cost of the no skip flag. 3701 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 3702 } else { 3703 // FIXME(rbultje) make this work for splitmv also 3704 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 3705 distortion2 = total_sse; 3706 assert(total_sse >= 0); 3707 rate2 -= (rate_y + rate_uv); 3708 rate_y = 0; 3709 rate_uv = 0; 3710 this_skip2 = 1; 3711 } 3712 } else { 3713 // Add in the cost of the no skip flag. 3714 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 3715 } 3716 3717 // Calculate the final RD estimate for this mode. 3718 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2); 3719 } 3720 3721 if (!disable_skip && ref_frame == INTRA_FRAME) { 3722 for (i = 0; i < REFERENCE_MODES; ++i) 3723 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd); 3724 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) 3725 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd); 3726 } 3727 3728 // Did this mode help.. i.e. is it the new best mode 3729 if (this_rd < best_rd || x->skip) { 3730 if (!mode_excluded) { 3731 int max_plane = MAX_MB_PLANE; 3732 // Note index of best mode so far 3733 best_ref_index = ref_index; 3734 3735 if (ref_frame == INTRA_FRAME) { 3736 /* required for left and above block mv */ 3737 mbmi->mv[0].as_int = 0; 3738 max_plane = 1; 3739 } 3740 3741 *returnrate = rate2; 3742 *returndistortion = distortion2; 3743 best_rd = this_rd; 3744 best_yrd = best_rd - 3745 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv); 3746 best_mbmode = *mbmi; 3747 best_skip2 = this_skip2; 3748 if (!x->select_tx_size) 3749 swap_block_ptr(x, ctx, 1, 0, 0, max_plane); 3750 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4], 3751 sizeof(uint8_t) * ctx->num_4x4_blk); 3752 3753 for (i = 0; i < 4; i++) 3754 best_bmodes[i] = xd->mi[0].src_mi->bmi[i]; 3755 3756 // TODO(debargha): enhance this test with a better distortion prediction 3757 // based on qp, activity mask and history 3758 if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) && 3759 (ref_index > MIN_EARLY_TERM_INDEX)) { 3760 const int qstep = xd->plane[0].dequant[1]; 3761 // TODO(debargha): Enhance this by specializing for each mode_index 3762 int scale = 4; 3763 if (x->source_variance < UINT_MAX) { 3764 const int var_adjust = (x->source_variance < 16); 3765 scale -= var_adjust; 3766 } 3767 if (ref_frame > INTRA_FRAME && 3768 distortion2 * scale < qstep * qstep) { 3769 early_term = 1; 3770 } 3771 } 3772 } 3773 } 3774 3775 /* keep record of best compound/single-only prediction */ 3776 if (!disable_skip && ref_frame != INTRA_FRAME) { 3777 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate; 3778 3779 if (cm->reference_mode == REFERENCE_MODE_SELECT) { 3780 single_rate = rate2 - compmode_cost; 3781 hybrid_rate = rate2; 3782 } else { 3783 single_rate = rate2; 3784 hybrid_rate = rate2 + compmode_cost; 3785 } 3786 3787 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2); 3788 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2); 3789 3790 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE]) 3791 best_pred_rd[SINGLE_REFERENCE] = single_rd; 3792 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE]) 3793 best_pred_rd[COMPOUND_REFERENCE] = single_rd; 3794 3795 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT]) 3796 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd; 3797 } 3798 3799 /* keep record of best filter type */ 3800 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME && 3801 cm->interp_filter != BILINEAR) { 3802 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ? 3803 SWITCHABLE_FILTERS : cm->interp_filter]; 3804 int64_t adj_rd; 3805 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) { 3806 if (ref == INT64_MAX) 3807 adj_rd = 0; 3808 else if (rd_opt->filter_cache[i] == INT64_MAX) 3809 // when early termination is triggered, the encoder does not have 3810 // access to the rate-distortion cost. it only knows that the cost 3811 // should be above the maximum valid value. hence it takes the known 3812 // maximum plus an arbitrary constant as the rate-distortion cost. 3813 adj_rd = rd_opt->mask_filter - ref + 10; 3814 else 3815 adj_rd = rd_opt->filter_cache[i] - ref; 3816 3817 adj_rd += this_rd; 3818 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd); 3819 } 3820 } 3821 3822 if (early_term) 3823 break; 3824 3825 if (x->skip && !comp_pred) 3826 break; 3827 } 3828 3829 if (best_rd >= best_rd_so_far) 3830 return INT64_MAX; 3831 3832 // If we used an estimate for the uv intra rd in the loop above... 3833 if (cpi->sf.use_uv_intra_rd_estimate) { 3834 // Do Intra UV best rd mode selection if best mode choice above was intra. 3835 if (best_mbmode.ref_frame[0] == INTRA_FRAME) { 3836 *mbmi = best_mbmode; 3837 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra, 3838 &rate_uv_tokenonly, 3839 &dist_uv, 3840 &skip_uv, 3841 BLOCK_8X8, TX_4X4); 3842 } 3843 } 3844 3845 if (best_rd == INT64_MAX) { 3846 *returnrate = INT_MAX; 3847 *returndistortion = INT64_MAX; 3848 return best_rd; 3849 } 3850 3851 assert((cm->interp_filter == SWITCHABLE) || 3852 (cm->interp_filter == best_mbmode.interp_filter) || 3853 !is_inter_block(&best_mbmode)); 3854 3855 update_rd_thresh_fact(cpi, bsize, best_ref_index); 3856 3857 // macroblock modes 3858 *mbmi = best_mbmode; 3859 x->skip |= best_skip2; 3860 if (!is_inter_block(&best_mbmode)) { 3861 for (i = 0; i < 4; i++) 3862 xd->mi[0].src_mi->bmi[i].as_mode = best_bmodes[i].as_mode; 3863 } else { 3864 for (i = 0; i < 4; ++i) 3865 vpx_memcpy(&xd->mi[0].src_mi->bmi[i], &best_bmodes[i], 3866 sizeof(b_mode_info)); 3867 3868 mbmi->mv[0].as_int = xd->mi[0].src_mi->bmi[3].as_mv[0].as_int; 3869 mbmi->mv[1].as_int = xd->mi[0].src_mi->bmi[3].as_mv[1].as_int; 3870 } 3871 3872 for (i = 0; i < REFERENCE_MODES; ++i) { 3873 if (best_pred_rd[i] == INT64_MAX) 3874 best_pred_diff[i] = INT_MIN; 3875 else 3876 best_pred_diff[i] = best_rd - best_pred_rd[i]; 3877 } 3878 3879 if (!x->skip) { 3880 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) { 3881 if (best_filter_rd[i] == INT64_MAX) 3882 best_filter_diff[i] = 0; 3883 else 3884 best_filter_diff[i] = best_rd - best_filter_rd[i]; 3885 } 3886 if (cm->interp_filter == SWITCHABLE) 3887 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0); 3888 } else { 3889 vp9_zero(best_filter_diff); 3890 } 3891 3892 store_coding_context(x, ctx, best_ref_index, 3893 best_pred_diff, best_tx_diff, best_filter_diff, 0); 3894 3895 return best_rd; 3896} 3897 3898