1 /* 2 * Copyright (c) 2014 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 <limits.h> 13 #include <math.h> 14 #include <stdio.h> 15 16 #include "./vp9_rtcd.h" 17 #include "./vpx_dsp_rtcd.h" 18 19 #include "vpx/vpx_codec.h" 20 #include "vpx_dsp/vpx_dsp_common.h" 21 #include "vpx_mem/vpx_mem.h" 22 #include "vpx_ports/mem.h" 23 24 #include "vp9/common/vp9_blockd.h" 25 #include "vp9/common/vp9_common.h" 26 #include "vp9/common/vp9_mvref_common.h" 27 #include "vp9/common/vp9_pred_common.h" 28 #include "vp9/common/vp9_reconinter.h" 29 #include "vp9/common/vp9_reconintra.h" 30 #include "vp9/common/vp9_scan.h" 31 32 #include "vp9/encoder/vp9_cost.h" 33 #include "vp9/encoder/vp9_encoder.h" 34 #include "vp9/encoder/vp9_pickmode.h" 35 #include "vp9/encoder/vp9_ratectrl.h" 36 #include "vp9/encoder/vp9_rd.h" 37 38 typedef struct { 39 uint8_t *data; 40 int stride; 41 int in_use; 42 } PRED_BUFFER; 43 44 static const int pos_shift_16x16[4][4] = { 45 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 } 46 }; 47 48 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x, 49 const MACROBLOCKD *xd, const TileInfo *const tile, 50 MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame, 51 int_mv *mv_ref_list, int_mv *base_mv, int mi_row, 52 int mi_col, int use_base_mv) { 53 const int *ref_sign_bias = cm->ref_frame_sign_bias; 54 int i, refmv_count = 0; 55 56 const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type]; 57 58 int different_ref_found = 0; 59 int context_counter = 0; 60 int const_motion = 0; 61 62 // Blank the reference vector list 63 memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES); 64 65 // The nearest 2 blocks are treated differently 66 // if the size < 8x8 we get the mv from the bmi substructure, 67 // and we also need to keep a mode count. 68 for (i = 0; i < 2; ++i) { 69 const POSITION *const mv_ref = &mv_ref_search[i]; 70 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) { 71 const MODE_INFO *const candidate_mi = 72 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]; 73 // Keep counts for entropy encoding. 74 context_counter += mode_2_counter[candidate_mi->mode]; 75 different_ref_found = 1; 76 77 if (candidate_mi->ref_frame[0] == ref_frame) 78 ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1), 79 refmv_count, mv_ref_list, Done); 80 } 81 } 82 83 const_motion = 1; 84 85 // Check the rest of the neighbors in much the same way 86 // as before except we don't need to keep track of sub blocks or 87 // mode counts. 88 for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) { 89 const POSITION *const mv_ref = &mv_ref_search[i]; 90 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) { 91 const MODE_INFO *const candidate_mi = 92 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]; 93 different_ref_found = 1; 94 95 if (candidate_mi->ref_frame[0] == ref_frame) 96 ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done); 97 } 98 } 99 100 // Since we couldn't find 2 mvs from the same reference frame 101 // go back through the neighbors and find motion vectors from 102 // different reference frames. 103 if (different_ref_found && !refmv_count) { 104 for (i = 0; i < MVREF_NEIGHBOURS; ++i) { 105 const POSITION *mv_ref = &mv_ref_search[i]; 106 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) { 107 const MODE_INFO *const candidate_mi = 108 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]; 109 110 // If the candidate is INTRA we don't want to consider its mv. 111 IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias, 112 refmv_count, mv_ref_list, Done); 113 } 114 } 115 } 116 if (use_base_mv && 117 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame && 118 ref_frame == LAST_FRAME) { 119 // Get base layer mv. 120 MV_REF *candidate = 121 &cm->prev_frame 122 ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)]; 123 if (candidate->mv[0].as_int != INVALID_MV) { 124 base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2); 125 base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2); 126 clamp_mv_ref(&base_mv->as_mv, xd); 127 } else { 128 base_mv->as_int = INVALID_MV; 129 } 130 } 131 132 Done: 133 134 x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter]; 135 136 // Clamp vectors 137 for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i) 138 clamp_mv_ref(&mv_ref_list[i].as_mv, xd); 139 140 return const_motion; 141 } 142 143 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x, 144 BLOCK_SIZE bsize, int mi_row, int mi_col, 145 int_mv *tmp_mv, int *rate_mv, 146 int64_t best_rd_sofar, int use_base_mv) { 147 MACROBLOCKD *xd = &x->e_mbd; 148 MODE_INFO *mi = xd->mi[0]; 149 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } }; 150 const int step_param = cpi->sf.mv.fullpel_search_step_param; 151 const int sadpb = x->sadperbit16; 152 MV mvp_full; 153 const int ref = mi->ref_frame[0]; 154 const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv; 155 MV center_mv; 156 uint32_t dis; 157 int rate_mode; 158 const MvLimits tmp_mv_limits = x->mv_limits; 159 int rv = 0; 160 int cost_list[5]; 161 int search_subpel = 1; 162 const YV12_BUFFER_CONFIG *scaled_ref_frame = 163 vp9_get_scaled_ref_frame(cpi, ref); 164 if (scaled_ref_frame) { 165 int i; 166 // Swap out the reference frame for a version that's been scaled to 167 // match the resolution of the current frame, allowing the existing 168 // motion search code to be used without additional modifications. 169 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0]; 170 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); 171 } 172 vp9_set_mv_search_range(&x->mv_limits, &ref_mv); 173 174 // Limit motion vector for large lightning change. 175 if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) { 176 x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10); 177 x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10); 178 x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10); 179 x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10); 180 } 181 182 assert(x->mv_best_ref_index[ref] <= 2); 183 if (x->mv_best_ref_index[ref] < 2) 184 mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv; 185 else 186 mvp_full = x->pred_mv[ref]; 187 188 mvp_full.col >>= 3; 189 mvp_full.row >>= 3; 190 191 if (!use_base_mv) 192 center_mv = ref_mv; 193 else 194 center_mv = tmp_mv->as_mv; 195 196 if (x->sb_use_mv_part) { 197 tmp_mv->as_mv.row = x->sb_mvrow_part >> 3; 198 tmp_mv->as_mv.col = x->sb_mvcol_part >> 3; 199 } else { 200 vp9_full_pixel_search( 201 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb, 202 cond_cost_list(cpi, cost_list), ¢er_mv, &tmp_mv->as_mv, INT_MAX, 0); 203 } 204 205 x->mv_limits = tmp_mv_limits; 206 207 // calculate the bit cost on motion vector 208 mvp_full.row = tmp_mv->as_mv.row * 8; 209 mvp_full.col = tmp_mv->as_mv.col * 8; 210 211 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost, 212 MV_COST_WEIGHT); 213 214 rate_mode = 215 cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)]; 216 rv = 217 !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar); 218 219 // For SVC on non-reference frame, avoid subpel for (0, 0) motion. 220 if (cpi->use_svc && cpi->svc.non_reference_frame) { 221 if (mvp_full.row == 0 && mvp_full.col == 0) search_subpel = 0; 222 } 223 224 if (rv && search_subpel) { 225 int subpel_force_stop = cpi->sf.mv.subpel_force_stop; 226 if (use_base_mv && cpi->sf.base_mv_aggressive) subpel_force_stop = 2; 227 cpi->find_fractional_mv_step( 228 x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv, 229 x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop, 230 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list), 231 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0); 232 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost, 233 x->mvcost, MV_COST_WEIGHT); 234 } 235 236 if (scaled_ref_frame) { 237 int i; 238 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i]; 239 } 240 return rv; 241 } 242 243 static void block_variance(const uint8_t *src, int src_stride, 244 const uint8_t *ref, int ref_stride, int w, int h, 245 unsigned int *sse, int *sum, int block_size, 246 #if CONFIG_VP9_HIGHBITDEPTH 247 int use_highbitdepth, vpx_bit_depth_t bd, 248 #endif 249 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) { 250 int i, j, k = 0; 251 252 *sse = 0; 253 *sum = 0; 254 255 for (i = 0; i < h; i += block_size) { 256 for (j = 0; j < w; j += block_size) { 257 #if CONFIG_VP9_HIGHBITDEPTH 258 if (use_highbitdepth) { 259 switch (bd) { 260 case VPX_BITS_8: 261 vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride, 262 ref + ref_stride * i + j, ref_stride, 263 &sse8x8[k], &sum8x8[k]); 264 break; 265 case VPX_BITS_10: 266 vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride, 267 ref + ref_stride * i + j, ref_stride, 268 &sse8x8[k], &sum8x8[k]); 269 break; 270 case VPX_BITS_12: 271 vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride, 272 ref + ref_stride * i + j, ref_stride, 273 &sse8x8[k], &sum8x8[k]); 274 break; 275 } 276 } else { 277 vpx_get8x8var(src + src_stride * i + j, src_stride, 278 ref + ref_stride * i + j, ref_stride, &sse8x8[k], 279 &sum8x8[k]); 280 } 281 #else 282 vpx_get8x8var(src + src_stride * i + j, src_stride, 283 ref + ref_stride * i + j, ref_stride, &sse8x8[k], 284 &sum8x8[k]); 285 #endif 286 *sse += sse8x8[k]; 287 *sum += sum8x8[k]; 288 var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6); 289 k++; 290 } 291 } 292 } 293 294 static void calculate_variance(int bw, int bh, TX_SIZE tx_size, 295 unsigned int *sse_i, int *sum_i, 296 unsigned int *var_o, unsigned int *sse_o, 297 int *sum_o) { 298 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size]; 299 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]); 300 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]); 301 int i, j, k = 0; 302 303 for (i = 0; i < nh; i += 2) { 304 for (j = 0; j < nw; j += 2) { 305 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] + 306 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1]; 307 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] + 308 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1]; 309 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >> 310 (b_width_log2_lookup[unit_size] + 311 b_height_log2_lookup[unit_size] + 6)); 312 k++; 313 } 314 } 315 } 316 317 // Adjust the ac_thr according to speed, width, height and normalized sum 318 static int ac_thr_factor(const int speed, const int width, const int height, 319 const int norm_sum) { 320 if (speed >= 8 && norm_sum < 5) { 321 if (width <= 640 && height <= 480) 322 return 4; 323 else 324 return 2; 325 } 326 return 1; 327 } 328 329 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize, 330 MACROBLOCK *x, MACROBLOCKD *xd, 331 int *out_rate_sum, int64_t *out_dist_sum, 332 unsigned int *var_y, unsigned int *sse_y, 333 int mi_row, int mi_col, int *early_term, 334 int *flag_preduv_computed) { 335 // Note our transform coeffs are 8 times an orthogonal transform. 336 // Hence quantizer step is also 8 times. To get effective quantizer 337 // we need to divide by 8 before sending to modeling function. 338 unsigned int sse; 339 int rate; 340 int64_t dist; 341 struct macroblock_plane *const p = &x->plane[0]; 342 struct macroblockd_plane *const pd = &xd->plane[0]; 343 const uint32_t dc_quant = pd->dequant[0]; 344 const uint32_t ac_quant = pd->dequant[1]; 345 const int64_t dc_thr = dc_quant * dc_quant >> 6; 346 int64_t ac_thr = ac_quant * ac_quant >> 6; 347 unsigned int var; 348 int sum; 349 int skip_dc = 0; 350 351 const int bw = b_width_log2_lookup[bsize]; 352 const int bh = b_height_log2_lookup[bsize]; 353 const int num8x8 = 1 << (bw + bh - 2); 354 unsigned int sse8x8[64] = { 0 }; 355 int sum8x8[64] = { 0 }; 356 unsigned int var8x8[64] = { 0 }; 357 TX_SIZE tx_size; 358 int i, k; 359 #if CONFIG_VP9_HIGHBITDEPTH 360 const vpx_bit_depth_t bd = cpi->common.bit_depth; 361 #endif 362 // Calculate variance for whole partition, and also save 8x8 blocks' variance 363 // to be used in following transform skipping test. 364 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, 365 4 << bw, 4 << bh, &sse, &sum, 8, 366 #if CONFIG_VP9_HIGHBITDEPTH 367 cpi->common.use_highbitdepth, bd, 368 #endif 369 sse8x8, sum8x8, var8x8); 370 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4)); 371 372 *var_y = var; 373 *sse_y = sse; 374 375 #if CONFIG_VP9_TEMPORAL_DENOISING 376 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) && 377 cpi->oxcf.speed > 5) 378 ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level, 379 (abs(sum) >> (bw + bh)), 380 cpi->svc.temporal_layer_id); 381 else 382 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width, 383 cpi->common.height, abs(sum) >> (bw + bh)); 384 #else 385 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width, 386 cpi->common.height, abs(sum) >> (bw + bh)); 387 #endif 388 389 if (cpi->common.tx_mode == TX_MODE_SELECT) { 390 if (sse > (var << 2)) 391 tx_size = VPXMIN(max_txsize_lookup[bsize], 392 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 393 else 394 tx_size = TX_8X8; 395 396 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && 397 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id)) 398 tx_size = TX_8X8; 399 else if (tx_size > TX_16X16) 400 tx_size = TX_16X16; 401 } else { 402 tx_size = VPXMIN(max_txsize_lookup[bsize], 403 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 404 } 405 406 assert(tx_size >= TX_8X8); 407 xd->mi[0]->tx_size = tx_size; 408 409 // Evaluate if the partition block is a skippable block in Y plane. 410 { 411 unsigned int sse16x16[16] = { 0 }; 412 int sum16x16[16] = { 0 }; 413 unsigned int var16x16[16] = { 0 }; 414 const int num16x16 = num8x8 >> 2; 415 416 unsigned int sse32x32[4] = { 0 }; 417 int sum32x32[4] = { 0 }; 418 unsigned int var32x32[4] = { 0 }; 419 const int num32x32 = num8x8 >> 4; 420 421 int ac_test = 1; 422 int dc_test = 1; 423 const int num = (tx_size == TX_8X8) 424 ? num8x8 425 : ((tx_size == TX_16X16) ? num16x16 : num32x32); 426 const unsigned int *sse_tx = 427 (tx_size == TX_8X8) ? sse8x8 428 : ((tx_size == TX_16X16) ? sse16x16 : sse32x32); 429 const unsigned int *var_tx = 430 (tx_size == TX_8X8) ? var8x8 431 : ((tx_size == TX_16X16) ? var16x16 : var32x32); 432 433 // Calculate variance if tx_size > TX_8X8 434 if (tx_size >= TX_16X16) 435 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16, 436 sum16x16); 437 if (tx_size == TX_32X32) 438 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32, 439 sse32x32, sum32x32); 440 441 // Skipping test 442 x->skip_txfm[0] = SKIP_TXFM_NONE; 443 for (k = 0; k < num; k++) 444 // Check if all ac coefficients can be quantized to zero. 445 if (!(var_tx[k] < ac_thr || var == 0)) { 446 ac_test = 0; 447 break; 448 } 449 450 for (k = 0; k < num; k++) 451 // Check if dc coefficient can be quantized to zero. 452 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) { 453 dc_test = 0; 454 break; 455 } 456 457 if (ac_test) { 458 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY; 459 460 if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC; 461 } else if (dc_test) { 462 skip_dc = 1; 463 } 464 } 465 466 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) { 467 int skip_uv[2] = { 0 }; 468 unsigned int var_uv[2]; 469 unsigned int sse_uv[2]; 470 471 *out_rate_sum = 0; 472 *out_dist_sum = sse << 4; 473 474 // Transform skipping test in UV planes. 475 for (i = 1; i <= 2; i++) { 476 if (cpi->oxcf.speed < 8 || x->color_sensitivity[i - 1]) { 477 struct macroblock_plane *const p = &x->plane[i]; 478 struct macroblockd_plane *const pd = &xd->plane[i]; 479 const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd); 480 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size]; 481 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd); 482 const int uv_bw = b_width_log2_lookup[uv_bsize]; 483 const int uv_bh = b_height_log2_lookup[uv_bsize]; 484 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) + 485 (uv_bh - b_height_log2_lookup[unit_size]); 486 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf); 487 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf); 488 int j = i - 1; 489 490 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i); 491 flag_preduv_computed[i - 1] = 1; 492 var_uv[j] = cpi->fn_ptr[uv_bsize].vf( 493 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]); 494 495 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) && 496 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j])) 497 skip_uv[j] = 1; 498 else 499 break; 500 } else { 501 skip_uv[i - 1] = 1; 502 } 503 } 504 505 // If the transform in YUV planes are skippable, the mode search checks 506 // fewer inter modes and doesn't check intra modes. 507 if (skip_uv[0] & skip_uv[1]) { 508 *early_term = 1; 509 } 510 return; 511 } 512 513 if (!skip_dc) { 514 #if CONFIG_VP9_HIGHBITDEPTH 515 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize], 516 dc_quant >> (xd->bd - 5), &rate, &dist); 517 #else 518 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize], 519 dc_quant >> 3, &rate, &dist); 520 #endif // CONFIG_VP9_HIGHBITDEPTH 521 } 522 523 if (!skip_dc) { 524 *out_rate_sum = rate >> 1; 525 *out_dist_sum = dist << 3; 526 } else { 527 *out_rate_sum = 0; 528 *out_dist_sum = (sse - var) << 4; 529 } 530 531 #if CONFIG_VP9_HIGHBITDEPTH 532 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], 533 ac_quant >> (xd->bd - 5), &rate, &dist); 534 #else 535 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3, 536 &rate, &dist); 537 #endif // CONFIG_VP9_HIGHBITDEPTH 538 539 *out_rate_sum += rate; 540 *out_dist_sum += dist << 4; 541 } 542 543 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x, 544 MACROBLOCKD *xd, int *out_rate_sum, 545 int64_t *out_dist_sum, unsigned int *var_y, 546 unsigned int *sse_y) { 547 // Note our transform coeffs are 8 times an orthogonal transform. 548 // Hence quantizer step is also 8 times. To get effective quantizer 549 // we need to divide by 8 before sending to modeling function. 550 unsigned int sse; 551 int rate; 552 int64_t dist; 553 struct macroblock_plane *const p = &x->plane[0]; 554 struct macroblockd_plane *const pd = &xd->plane[0]; 555 const int64_t dc_thr = p->quant_thred[0] >> 6; 556 const int64_t ac_thr = p->quant_thred[1] >> 6; 557 const uint32_t dc_quant = pd->dequant[0]; 558 const uint32_t ac_quant = pd->dequant[1]; 559 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride, 560 pd->dst.buf, pd->dst.stride, &sse); 561 int skip_dc = 0; 562 563 *var_y = var; 564 *sse_y = sse; 565 566 if (cpi->common.tx_mode == TX_MODE_SELECT) { 567 if (sse > (var << 2)) 568 xd->mi[0]->tx_size = 569 VPXMIN(max_txsize_lookup[bsize], 570 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 571 else 572 xd->mi[0]->tx_size = TX_8X8; 573 574 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && 575 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id)) 576 xd->mi[0]->tx_size = TX_8X8; 577 else if (xd->mi[0]->tx_size > TX_16X16) 578 xd->mi[0]->tx_size = TX_16X16; 579 } else { 580 xd->mi[0]->tx_size = 581 VPXMIN(max_txsize_lookup[bsize], 582 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 583 } 584 585 // Evaluate if the partition block is a skippable block in Y plane. 586 { 587 const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size]; 588 const unsigned int num_blk_log2 = 589 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) + 590 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]); 591 const unsigned int sse_tx = sse >> num_blk_log2; 592 const unsigned int var_tx = var >> num_blk_log2; 593 594 x->skip_txfm[0] = SKIP_TXFM_NONE; 595 // Check if all ac coefficients can be quantized to zero. 596 if (var_tx < ac_thr || var == 0) { 597 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY; 598 // Check if dc coefficient can be quantized to zero. 599 if (sse_tx - var_tx < dc_thr || sse == var) 600 x->skip_txfm[0] = SKIP_TXFM_AC_DC; 601 } else { 602 if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1; 603 } 604 } 605 606 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) { 607 *out_rate_sum = 0; 608 *out_dist_sum = sse << 4; 609 return; 610 } 611 612 if (!skip_dc) { 613 #if CONFIG_VP9_HIGHBITDEPTH 614 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize], 615 dc_quant >> (xd->bd - 5), &rate, &dist); 616 #else 617 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize], 618 dc_quant >> 3, &rate, &dist); 619 #endif // CONFIG_VP9_HIGHBITDEPTH 620 } 621 622 if (!skip_dc) { 623 *out_rate_sum = rate >> 1; 624 *out_dist_sum = dist << 3; 625 } else { 626 *out_rate_sum = 0; 627 *out_dist_sum = (sse - var) << 4; 628 } 629 630 #if CONFIG_VP9_HIGHBITDEPTH 631 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], 632 ac_quant >> (xd->bd - 5), &rate, &dist); 633 #else 634 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3, 635 &rate, &dist); 636 #endif // CONFIG_VP9_HIGHBITDEPTH 637 638 *out_rate_sum += rate; 639 *out_dist_sum += dist << 4; 640 } 641 642 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc, 643 int *skippable, int64_t *sse, BLOCK_SIZE bsize, 644 TX_SIZE tx_size, int rd_computed) { 645 MACROBLOCKD *xd = &x->e_mbd; 646 const struct macroblockd_plane *pd = &xd->plane[0]; 647 struct macroblock_plane *const p = &x->plane[0]; 648 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize]; 649 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize]; 650 const int step = 1 << (tx_size << 1); 651 const int block_step = (1 << tx_size); 652 int block = 0, r, c; 653 const int max_blocks_wide = 654 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5); 655 const int max_blocks_high = 656 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5); 657 int eob_cost = 0; 658 const int bw = 4 * num_4x4_w; 659 const int bh = 4 * num_4x4_h; 660 661 #if CONFIG_VP9_HIGHBITDEPTH 662 // TODO(jingning): Implement the high bit-depth Hadamard transforms and 663 // remove this check condition. 664 // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions 665 // for now, as the vp9_quantize_fp below for highbitdepth build is slow. 666 if (xd->bd != 8 || 667 (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME && 668 bsize < BLOCK_32X32)) { 669 unsigned int var_y, sse_y; 670 (void)tx_size; 671 if (!rd_computed) 672 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist, 673 &var_y, &sse_y); 674 *sse = INT_MAX; 675 *skippable = 0; 676 return; 677 } 678 #endif 679 680 if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME && 681 (bsize < BLOCK_32X32 || 682 (cpi->use_svc && 683 (bsize < BLOCK_32X32 || cpi->svc.temporal_layer_id > 0)))) { 684 unsigned int var_y, sse_y; 685 (void)tx_size; 686 if (!rd_computed) 687 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist, 688 &var_y, &sse_y); 689 *sse = INT_MAX; 690 *skippable = 0; 691 return; 692 } 693 694 (void)cpi; 695 696 // The max tx_size passed in is TX_16X16. 697 assert(tx_size != TX_32X32); 698 699 vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride, 700 pd->dst.buf, pd->dst.stride); 701 *skippable = 1; 702 // Keep track of the row and column of the blocks we use so that we know 703 // if we are in the unrestricted motion border. 704 for (r = 0; r < max_blocks_high; r += block_step) { 705 for (c = 0; c < num_4x4_w; c += block_step) { 706 if (c < max_blocks_wide) { 707 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size]; 708 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block); 709 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block); 710 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 711 uint16_t *const eob = &p->eobs[block]; 712 const int diff_stride = bw; 713 const int16_t *src_diff; 714 src_diff = &p->src_diff[(r * diff_stride + c) << 2]; 715 716 switch (tx_size) { 717 case TX_16X16: 718 vpx_hadamard_16x16(src_diff, diff_stride, coeff); 719 vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp, 720 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan, 721 scan_order->iscan); 722 break; 723 case TX_8X8: 724 vpx_hadamard_8x8(src_diff, diff_stride, coeff); 725 vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp, 726 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan, 727 scan_order->iscan); 728 break; 729 case TX_4X4: 730 x->fwd_txfm4x4(src_diff, coeff, diff_stride); 731 vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp, 732 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan, 733 scan_order->iscan); 734 break; 735 default: assert(0); break; 736 } 737 *skippable &= (*eob == 0); 738 eob_cost += 1; 739 } 740 block += step; 741 } 742 } 743 744 this_rdc->rate = 0; 745 if (*sse < INT64_MAX) { 746 *sse = (*sse << 6) >> 2; 747 if (*skippable) { 748 this_rdc->dist = *sse; 749 return; 750 } 751 } 752 753 block = 0; 754 this_rdc->dist = 0; 755 for (r = 0; r < max_blocks_high; r += block_step) { 756 for (c = 0; c < num_4x4_w; c += block_step) { 757 if (c < max_blocks_wide) { 758 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block); 759 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block); 760 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); 761 uint16_t *const eob = &p->eobs[block]; 762 763 if (*eob == 1) 764 this_rdc->rate += (int)abs(qcoeff[0]); 765 else if (*eob > 1) 766 this_rdc->rate += vpx_satd(qcoeff, step << 4); 767 768 this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2; 769 } 770 block += step; 771 } 772 } 773 774 // If skippable is set, rate gets clobbered later. 775 this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT); 776 this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT); 777 } 778 779 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize, 780 MACROBLOCK *x, MACROBLOCKD *xd, 781 RD_COST *this_rdc, unsigned int *var_y, 782 unsigned int *sse_y, int start_plane, 783 int stop_plane) { 784 // Note our transform coeffs are 8 times an orthogonal transform. 785 // Hence quantizer step is also 8 times. To get effective quantizer 786 // we need to divide by 8 before sending to modeling function. 787 unsigned int sse; 788 int rate; 789 int64_t dist; 790 int i; 791 #if CONFIG_VP9_HIGHBITDEPTH 792 uint64_t tot_var = *var_y; 793 uint64_t tot_sse = *sse_y; 794 #else 795 uint32_t tot_var = *var_y; 796 uint32_t tot_sse = *sse_y; 797 #endif 798 799 this_rdc->rate = 0; 800 this_rdc->dist = 0; 801 802 for (i = start_plane; i <= stop_plane; ++i) { 803 struct macroblock_plane *const p = &x->plane[i]; 804 struct macroblockd_plane *const pd = &xd->plane[i]; 805 const uint32_t dc_quant = pd->dequant[0]; 806 const uint32_t ac_quant = pd->dequant[1]; 807 const BLOCK_SIZE bs = plane_bsize; 808 unsigned int var; 809 if (!x->color_sensitivity[i - 1]) continue; 810 811 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf, 812 pd->dst.stride, &sse); 813 assert(sse >= var); 814 tot_var += var; 815 tot_sse += sse; 816 817 #if CONFIG_VP9_HIGHBITDEPTH 818 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs], 819 dc_quant >> (xd->bd - 5), &rate, &dist); 820 #else 821 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs], 822 dc_quant >> 3, &rate, &dist); 823 #endif // CONFIG_VP9_HIGHBITDEPTH 824 825 this_rdc->rate += rate >> 1; 826 this_rdc->dist += dist << 3; 827 828 #if CONFIG_VP9_HIGHBITDEPTH 829 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], 830 ac_quant >> (xd->bd - 5), &rate, &dist); 831 #else 832 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3, 833 &rate, &dist); 834 #endif // CONFIG_VP9_HIGHBITDEPTH 835 836 this_rdc->rate += rate; 837 this_rdc->dist += dist << 4; 838 } 839 840 #if CONFIG_VP9_HIGHBITDEPTH 841 *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var; 842 *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse; 843 #else 844 *var_y = tot_var; 845 *sse_y = tot_sse; 846 #endif 847 } 848 849 static int get_pred_buffer(PRED_BUFFER *p, int len) { 850 int i; 851 852 for (i = 0; i < len; i++) { 853 if (!p[i].in_use) { 854 p[i].in_use = 1; 855 return i; 856 } 857 } 858 return -1; 859 } 860 861 static void free_pred_buffer(PRED_BUFFER *p) { 862 if (p != NULL) p->in_use = 0; 863 } 864 865 static void encode_breakout_test( 866 VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col, 867 MV_REFERENCE_FRAME ref_frame, PREDICTION_MODE this_mode, unsigned int var_y, 868 unsigned int sse_y, struct buf_2d yv12_mb[][MAX_MB_PLANE], int *rate, 869 int64_t *dist, int *flag_preduv_computed) { 870 MACROBLOCKD *xd = &x->e_mbd; 871 MODE_INFO *const mi = xd->mi[0]; 872 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]); 873 unsigned int var = var_y, sse = sse_y; 874 // Skipping threshold for ac. 875 unsigned int thresh_ac; 876 // Skipping threshold for dc. 877 unsigned int thresh_dc; 878 int motion_low = 1; 879 if (cpi->use_svc && ref_frame == GOLDEN_FRAME) return; 880 if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 || 881 mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64) 882 motion_low = 0; 883 if (x->encode_breakout > 0 && motion_low == 1) { 884 // Set a maximum for threshold to avoid big PSNR loss in low bit rate 885 // case. Use extreme low threshold for static frames to limit 886 // skipping. 887 const unsigned int max_thresh = 36000; 888 // The encode_breakout input 889 const unsigned int min_thresh = 890 VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh); 891 #if CONFIG_VP9_HIGHBITDEPTH 892 const int shift = (xd->bd << 1) - 16; 893 #endif 894 895 // Calculate threshold according to dequant value. 896 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3; 897 #if CONFIG_VP9_HIGHBITDEPTH 898 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) { 899 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift); 900 } 901 #endif // CONFIG_VP9_HIGHBITDEPTH 902 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh); 903 904 // Adjust ac threshold according to partition size. 905 thresh_ac >>= 906 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]); 907 908 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6); 909 #if CONFIG_VP9_HIGHBITDEPTH 910 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) { 911 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift); 912 } 913 #endif // CONFIG_VP9_HIGHBITDEPTH 914 } else { 915 thresh_ac = 0; 916 thresh_dc = 0; 917 } 918 919 // Y skipping condition checking for ac and dc. 920 if (var <= thresh_ac && (sse - var) <= thresh_dc) { 921 unsigned int sse_u, sse_v; 922 unsigned int var_u, var_v; 923 unsigned int thresh_ac_uv = thresh_ac; 924 unsigned int thresh_dc_uv = thresh_dc; 925 if (x->sb_is_skin) { 926 thresh_ac_uv = 0; 927 thresh_dc_uv = 0; 928 } 929 930 if (!flag_preduv_computed[0] || !flag_preduv_computed[1]) { 931 xd->plane[1].pre[0] = yv12_mb[ref_frame][1]; 932 xd->plane[2].pre[0] = yv12_mb[ref_frame][2]; 933 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize); 934 } 935 936 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride, 937 xd->plane[1].dst.buf, 938 xd->plane[1].dst.stride, &sse_u); 939 940 // U skipping condition checking 941 if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) { 942 var_v = cpi->fn_ptr[uv_size].vf( 943 x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf, 944 xd->plane[2].dst.stride, &sse_v); 945 946 // V skipping condition checking 947 if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) { 948 x->skip = 1; 949 950 // The cost of skip bit needs to be added. 951 *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]] 952 [INTER_OFFSET(this_mode)]; 953 954 // More on this part of rate 955 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 956 957 // Scaling factor for SSE from spatial domain to frequency 958 // domain is 16. Adjust distortion accordingly. 959 // TODO(yunqingwang): In this function, only y-plane dist is 960 // calculated. 961 *dist = (sse << 4); // + ((sse_u + sse_v) << 4); 962 963 // *disable_skip = 1; 964 } 965 } 966 } 967 } 968 969 struct estimate_block_intra_args { 970 VP9_COMP *cpi; 971 MACROBLOCK *x; 972 PREDICTION_MODE mode; 973 int skippable; 974 RD_COST *rdc; 975 }; 976 977 static void estimate_block_intra(int plane, int block, int row, int col, 978 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, 979 void *arg) { 980 struct estimate_block_intra_args *const args = arg; 981 VP9_COMP *const cpi = args->cpi; 982 MACROBLOCK *const x = args->x; 983 MACROBLOCKD *const xd = &x->e_mbd; 984 struct macroblock_plane *const p = &x->plane[0]; 985 struct macroblockd_plane *const pd = &xd->plane[0]; 986 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size]; 987 uint8_t *const src_buf_base = p->src.buf; 988 uint8_t *const dst_buf_base = pd->dst.buf; 989 const int src_stride = p->src.stride; 990 const int dst_stride = pd->dst.stride; 991 RD_COST this_rdc; 992 993 (void)block; 994 995 p->src.buf = &src_buf_base[4 * (row * src_stride + col)]; 996 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)]; 997 // Use source buffer as an approximation for the fully reconstructed buffer. 998 vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size, 999 args->mode, x->skip_encode ? p->src.buf : pd->dst.buf, 1000 x->skip_encode ? src_stride : dst_stride, pd->dst.buf, 1001 dst_stride, col, row, plane); 1002 1003 if (plane == 0) { 1004 int64_t this_sse = INT64_MAX; 1005 // TODO(jingning): This needs further refactoring. 1006 block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx, 1007 VPXMIN(tx_size, TX_16X16), 0); 1008 } else { 1009 unsigned int var = 0; 1010 unsigned int sse = 0; 1011 model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane, 1012 plane); 1013 } 1014 1015 p->src.buf = src_buf_base; 1016 pd->dst.buf = dst_buf_base; 1017 args->rdc->rate += this_rdc.rate; 1018 args->rdc->dist += this_rdc.dist; 1019 } 1020 1021 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = { 1022 { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM }, 1023 { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV }, 1024 { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG }, 1025 { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA }, 1026 }; 1027 1028 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED, 1029 TM_PRED }; 1030 1031 static int mode_offset(const PREDICTION_MODE mode) { 1032 if (mode >= NEARESTMV) { 1033 return INTER_OFFSET(mode); 1034 } else { 1035 switch (mode) { 1036 case DC_PRED: return 0; 1037 case V_PRED: return 1; 1038 case H_PRED: return 2; 1039 case TM_PRED: return 3; 1040 default: return -1; 1041 } 1042 } 1043 } 1044 1045 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh, 1046 const int *const thresh_fact) { 1047 int is_rd_less_than_thresh; 1048 is_rd_less_than_thresh = 1049 best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX; 1050 return is_rd_less_than_thresh; 1051 } 1052 1053 static INLINE void update_thresh_freq_fact_row_mt( 1054 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance, 1055 int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame, 1056 THR_MODES best_mode_idx, PREDICTION_MODE mode) { 1057 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)]; 1058 int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx; 1059 int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx]; 1060 if (thr_mode_idx == best_mode_idx) 1061 *freq_fact -= (*freq_fact >> 4); 1062 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV && 1063 ref_frame == LAST_FRAME && source_variance < 5) { 1064 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32); 1065 } else { 1066 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 1067 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT); 1068 } 1069 } 1070 1071 static INLINE void update_thresh_freq_fact( 1072 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance, 1073 BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx, 1074 PREDICTION_MODE mode) { 1075 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)]; 1076 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx]; 1077 if (thr_mode_idx == best_mode_idx) 1078 *freq_fact -= (*freq_fact >> 4); 1079 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV && 1080 ref_frame == LAST_FRAME && source_variance < 5) { 1081 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32); 1082 } else { 1083 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 1084 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT); 1085 } 1086 } 1087 1088 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost, 1089 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) { 1090 MACROBLOCKD *const xd = &x->e_mbd; 1091 MODE_INFO *const mi = xd->mi[0]; 1092 RD_COST this_rdc, best_rdc; 1093 PREDICTION_MODE this_mode; 1094 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 }; 1095 const TX_SIZE intra_tx_size = 1096 VPXMIN(max_txsize_lookup[bsize], 1097 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 1098 MODE_INFO *const mic = xd->mi[0]; 1099 int *bmode_costs; 1100 const MODE_INFO *above_mi = xd->above_mi; 1101 const MODE_INFO *left_mi = xd->left_mi; 1102 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0); 1103 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0); 1104 bmode_costs = cpi->y_mode_costs[A][L]; 1105 1106 (void)ctx; 1107 vp9_rd_cost_reset(&best_rdc); 1108 vp9_rd_cost_reset(&this_rdc); 1109 1110 mi->ref_frame[0] = INTRA_FRAME; 1111 // Initialize interp_filter here so we do not have to check for inter block 1112 // modes in get_pred_context_switchable_interp() 1113 mi->interp_filter = SWITCHABLE_FILTERS; 1114 1115 mi->mv[0].as_int = INVALID_MV; 1116 mi->uv_mode = DC_PRED; 1117 memset(x->skip_txfm, 0, sizeof(x->skip_txfm)); 1118 1119 // Change the limit of this loop to add other intra prediction 1120 // mode tests. 1121 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) { 1122 this_rdc.dist = this_rdc.rate = 0; 1123 args.mode = this_mode; 1124 args.skippable = 1; 1125 args.rdc = &this_rdc; 1126 mi->tx_size = intra_tx_size; 1127 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra, 1128 &args); 1129 if (args.skippable) { 1130 x->skip_txfm[0] = SKIP_TXFM_AC_DC; 1131 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1); 1132 } else { 1133 x->skip_txfm[0] = SKIP_TXFM_NONE; 1134 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0); 1135 } 1136 this_rdc.rate += bmode_costs[this_mode]; 1137 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist); 1138 1139 if (this_rdc.rdcost < best_rdc.rdcost) { 1140 best_rdc = this_rdc; 1141 mi->mode = this_mode; 1142 } 1143 } 1144 1145 *rd_cost = best_rdc; 1146 } 1147 1148 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd, 1149 int ref_frame_cost[MAX_REF_FRAMES]) { 1150 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd); 1151 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd); 1152 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd); 1153 1154 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0); 1155 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] = 1156 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1); 1157 1158 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0); 1159 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1); 1160 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1); 1161 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0); 1162 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1); 1163 } 1164 1165 typedef struct { 1166 MV_REFERENCE_FRAME ref_frame; 1167 PREDICTION_MODE pred_mode; 1168 } REF_MODE; 1169 1170 #define RT_INTER_MODES 12 1171 static const REF_MODE ref_mode_set[RT_INTER_MODES] = { 1172 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV }, 1173 { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV }, 1174 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV }, 1175 { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV }, 1176 { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV }, 1177 { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV } 1178 }; 1179 1180 #define RT_INTER_MODES_SVC 8 1181 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES_SVC] = { 1182 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV }, 1183 { LAST_FRAME, NEARMV }, { GOLDEN_FRAME, ZEROMV }, 1184 { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV }, 1185 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV } 1186 }; 1187 1188 static INLINE void find_predictors( 1189 VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, 1190 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], 1191 int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask, 1192 const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col, 1193 struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize, 1194 int force_skip_low_temp_var, int comp_pred_allowed) { 1195 VP9_COMMON *const cm = &cpi->common; 1196 MACROBLOCKD *const xd = &x->e_mbd; 1197 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame); 1198 TileInfo *const tile_info = &tile_data->tile_info; 1199 // TODO(jingning) placeholder for inter-frame non-RD mode decision. 1200 x->pred_mv_sad[ref_frame] = INT_MAX; 1201 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV; 1202 frame_mv[ZEROMV][ref_frame].as_int = 0; 1203 // this needs various further optimizations. to be continued.. 1204 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) { 1205 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame]; 1206 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf; 1207 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf); 1208 if (cm->use_prev_frame_mvs || comp_pred_allowed) { 1209 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col, 1210 x->mbmi_ext->mode_context); 1211 } else { 1212 const_motion[ref_frame] = 1213 mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame, 1214 candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col, 1215 (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id)); 1216 } 1217 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates, 1218 &frame_mv[NEARESTMV][ref_frame], 1219 &frame_mv[NEARMV][ref_frame]); 1220 // Early exit for golden frame if force_skip_low_temp_var is set. 1221 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 && 1222 !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) { 1223 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame, 1224 bsize); 1225 } 1226 } else { 1227 *ref_frame_skip_mask |= (1 << ref_frame); 1228 } 1229 } 1230 1231 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd, 1232 PREDICTION_MODE this_mode, RD_COST *this_rdc, 1233 BLOCK_SIZE bsize, int mv_row, int mv_col, 1234 int is_last_frame, int lowvar_highsumdiff, 1235 int is_skin) { 1236 // Bias against MVs associated with NEWMV mode that are very different from 1237 // top/left neighbors. 1238 if (this_mode == NEWMV) { 1239 int al_mv_average_row; 1240 int al_mv_average_col; 1241 int left_row, left_col; 1242 int row_diff, col_diff; 1243 int above_mv_valid = 0; 1244 int left_mv_valid = 0; 1245 int above_row = 0; 1246 int above_col = 0; 1247 1248 if (xd->above_mi) { 1249 above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV; 1250 above_row = xd->above_mi->mv[0].as_mv.row; 1251 above_col = xd->above_mi->mv[0].as_mv.col; 1252 } 1253 if (xd->left_mi) { 1254 left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV; 1255 left_row = xd->left_mi->mv[0].as_mv.row; 1256 left_col = xd->left_mi->mv[0].as_mv.col; 1257 } 1258 if (above_mv_valid && left_mv_valid) { 1259 al_mv_average_row = (above_row + left_row + 1) >> 1; 1260 al_mv_average_col = (above_col + left_col + 1) >> 1; 1261 } else if (above_mv_valid) { 1262 al_mv_average_row = above_row; 1263 al_mv_average_col = above_col; 1264 } else if (left_mv_valid) { 1265 al_mv_average_row = left_row; 1266 al_mv_average_col = left_col; 1267 } else { 1268 al_mv_average_row = al_mv_average_col = 0; 1269 } 1270 row_diff = (al_mv_average_row - mv_row); 1271 col_diff = (al_mv_average_col - mv_col); 1272 if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) { 1273 if (bsize > BLOCK_32X32) 1274 this_rdc->rdcost = this_rdc->rdcost << 1; 1275 else 1276 this_rdc->rdcost = 3 * this_rdc->rdcost >> 1; 1277 } 1278 } 1279 // If noise estimation is enabled, and estimated level is above threshold, 1280 // add a bias to LAST reference with small motion, for large blocks. 1281 if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 && 1282 is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8) 1283 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3); 1284 else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 && 1285 is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 && 1286 mv_col > -16) 1287 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3); 1288 } 1289 1290 #if CONFIG_VP9_TEMPORAL_DENOISING 1291 static void vp9_pickmode_ctx_den_update( 1292 VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig, 1293 int ref_frame_cost[MAX_REF_FRAMES], 1294 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred, 1295 TX_SIZE best_tx_size, PREDICTION_MODE best_mode, 1296 MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter, 1297 uint8_t best_mode_skip_txfm) { 1298 ctx_den->zero_last_cost_orig = zero_last_cost_orig; 1299 ctx_den->ref_frame_cost = ref_frame_cost; 1300 ctx_den->frame_mv = frame_mv; 1301 ctx_den->reuse_inter_pred = reuse_inter_pred; 1302 ctx_den->best_tx_size = best_tx_size; 1303 ctx_den->best_mode = best_mode; 1304 ctx_den->best_ref_frame = best_ref_frame; 1305 ctx_den->best_pred_filter = best_pred_filter; 1306 ctx_den->best_mode_skip_txfm = best_mode_skip_txfm; 1307 } 1308 1309 static void recheck_zeromv_after_denoising( 1310 VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd, 1311 VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den, 1312 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize, 1313 int mi_row, int mi_col) { 1314 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on 1315 // denoised result. Only do this under noise conditions, and if rdcost of 1316 // ZEROMV onoriginal source is not significantly higher than rdcost of best 1317 // mode. 1318 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow && 1319 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) && 1320 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) || 1321 (ctx_den->best_ref_frame == GOLDEN_FRAME && 1322 cpi->svc.number_spatial_layers == 1 && 1323 decision == FILTER_ZEROMV_BLOCK))) { 1324 // Check if we should pick ZEROMV on denoised signal. 1325 int rate = 0; 1326 int64_t dist = 0; 1327 uint32_t var_y = UINT_MAX; 1328 uint32_t sse_y = UINT_MAX; 1329 RD_COST this_rdc; 1330 mi->mode = ZEROMV; 1331 mi->ref_frame[0] = LAST_FRAME; 1332 mi->ref_frame[1] = NONE; 1333 mi->mv[0].as_int = 0; 1334 mi->interp_filter = EIGHTTAP; 1335 xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0]; 1336 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); 1337 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y); 1338 this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] + 1339 cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]] 1340 [INTER_OFFSET(ZEROMV)]; 1341 this_rdc.dist = dist; 1342 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist); 1343 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source 1344 // is higher than best_ref mode (on original source). 1345 if (this_rdc.rdcost > best_rdc->rdcost) { 1346 this_rdc = *best_rdc; 1347 mi->mode = ctx_den->best_mode; 1348 mi->ref_frame[0] = ctx_den->best_ref_frame; 1349 mi->interp_filter = ctx_den->best_pred_filter; 1350 if (ctx_den->best_ref_frame == INTRA_FRAME) { 1351 mi->mv[0].as_int = INVALID_MV; 1352 mi->interp_filter = SWITCHABLE_FILTERS; 1353 } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) { 1354 mi->mv[0].as_int = 1355 ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame] 1356 .as_int; 1357 if (ctx_den->reuse_inter_pred) { 1358 xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0]; 1359 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); 1360 } 1361 } 1362 mi->tx_size = ctx_den->best_tx_size; 1363 x->skip_txfm[0] = ctx_den->best_mode_skip_txfm; 1364 } else { 1365 ctx_den->best_ref_frame = LAST_FRAME; 1366 *best_rdc = this_rdc; 1367 } 1368 } 1369 } 1370 #endif // CONFIG_VP9_TEMPORAL_DENOISING 1371 1372 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row, 1373 int mi_col, BLOCK_SIZE bsize) { 1374 const int i = (mi_row & 0x7) >> 1; 1375 const int j = (mi_col & 0x7) >> 1; 1376 int force_skip_low_temp_var = 0; 1377 // Set force_skip_low_temp_var based on the block size and block offset. 1378 if (bsize == BLOCK_64X64) { 1379 force_skip_low_temp_var = variance_low[0]; 1380 } else if (bsize == BLOCK_64X32) { 1381 if (!(mi_col & 0x7) && !(mi_row & 0x7)) { 1382 force_skip_low_temp_var = variance_low[1]; 1383 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) { 1384 force_skip_low_temp_var = variance_low[2]; 1385 } 1386 } else if (bsize == BLOCK_32X64) { 1387 if (!(mi_col & 0x7) && !(mi_row & 0x7)) { 1388 force_skip_low_temp_var = variance_low[3]; 1389 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) { 1390 force_skip_low_temp_var = variance_low[4]; 1391 } 1392 } else if (bsize == BLOCK_32X32) { 1393 if (!(mi_col & 0x7) && !(mi_row & 0x7)) { 1394 force_skip_low_temp_var = variance_low[5]; 1395 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) { 1396 force_skip_low_temp_var = variance_low[6]; 1397 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) { 1398 force_skip_low_temp_var = variance_low[7]; 1399 } else if ((mi_col & 0x7) && (mi_row & 0x7)) { 1400 force_skip_low_temp_var = variance_low[8]; 1401 } 1402 } else if (bsize == BLOCK_16X16) { 1403 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]]; 1404 } else if (bsize == BLOCK_32X16) { 1405 // The col shift index for the second 16x16 block. 1406 const int j2 = ((mi_col + 2) & 0x7) >> 1; 1407 // Only if each 16x16 block inside has low temporal variance. 1408 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] && 1409 variance_low[pos_shift_16x16[i][j2]]; 1410 } else if (bsize == BLOCK_16X32) { 1411 // The row shift index for the second 16x16 block. 1412 const int i2 = ((mi_row + 2) & 0x7) >> 1; 1413 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] && 1414 variance_low[pos_shift_16x16[i2][j]]; 1415 } 1416 return force_skip_low_temp_var; 1417 } 1418 1419 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data, 1420 int mi_row, int mi_col, RD_COST *rd_cost, 1421 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) { 1422 VP9_COMMON *const cm = &cpi->common; 1423 SPEED_FEATURES *const sf = &cpi->sf; 1424 const SVC *const svc = &cpi->svc; 1425 MACROBLOCKD *const xd = &x->e_mbd; 1426 MODE_INFO *const mi = xd->mi[0]; 1427 struct macroblockd_plane *const pd = &xd->plane[0]; 1428 PREDICTION_MODE best_mode = ZEROMV; 1429 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME; 1430 MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame; 1431 TX_SIZE best_tx_size = TX_SIZES; 1432 INTERP_FILTER best_pred_filter = EIGHTTAP; 1433 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; 1434 uint8_t mode_checked[MB_MODE_COUNT][MAX_REF_FRAMES]; 1435 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; 1436 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, 1437 VP9_ALT_FLAG }; 1438 RD_COST this_rdc, best_rdc; 1439 uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE; 1440 // var_y and sse_y are saved to be used in skipping checking 1441 unsigned int var_y = UINT_MAX; 1442 unsigned int sse_y = UINT_MAX; 1443 const int intra_cost_penalty = 1444 vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q); 1445 int64_t inter_mode_thresh = 1446 RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0); 1447 const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize]; 1448 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2; 1449 int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES; 1450 const int *const rd_thresh_freq_fact = 1451 (cpi->sf.adaptive_rd_thresh_row_mt) 1452 ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx]) 1453 : tile_data->thresh_freq_fact[bsize]; 1454 1455 INTERP_FILTER filter_ref; 1456 const int bsl = mi_width_log2_lookup[bsize]; 1457 const int pred_filter_search = 1458 cm->interp_filter == SWITCHABLE 1459 ? (((mi_row + mi_col) >> bsl) + 1460 get_chessboard_index(cm->current_video_frame)) & 1461 0x1 1462 : 0; 1463 int const_motion[MAX_REF_FRAMES] = { 0 }; 1464 const int bh = num_4x4_blocks_high_lookup[bsize] << 2; 1465 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2; 1466 // For speed 6, the result of interp filter is reused later in actual encoding 1467 // process. 1468 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers. 1469 PRED_BUFFER tmp[4]; 1470 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]); 1471 #if CONFIG_VP9_HIGHBITDEPTH 1472 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]); 1473 #endif 1474 struct buf_2d orig_dst = pd->dst; 1475 PRED_BUFFER *best_pred = NULL; 1476 PRED_BUFFER *this_mode_pred = NULL; 1477 const int pixels_in_block = bh * bw; 1478 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready; 1479 int ref_frame_skip_mask = 0; 1480 int idx; 1481 int best_pred_sad = INT_MAX; 1482 int best_early_term = 0; 1483 int ref_frame_cost[MAX_REF_FRAMES]; 1484 int svc_force_zero_mode[3] = { 0 }; 1485 int perform_intra_pred = 1; 1486 int use_golden_nonzeromv = 1; 1487 int force_skip_low_temp_var = 0; 1488 int skip_ref_find_pred[4] = { 0 }; 1489 unsigned int sse_zeromv_normalized = UINT_MAX; 1490 unsigned int best_sse_sofar = UINT_MAX; 1491 unsigned int thresh_svc_skip_golden = 500; 1492 #if CONFIG_VP9_TEMPORAL_DENOISING 1493 VP9_PICKMODE_CTX_DEN ctx_den; 1494 int64_t zero_last_cost_orig = INT64_MAX; 1495 int denoise_svc_pickmode = 1; 1496 #endif 1497 INTERP_FILTER filter_gf_svc = EIGHTTAP; 1498 MV_REFERENCE_FRAME best_second_ref_frame = NONE; 1499 int comp_modes = 0; 1500 int num_inter_modes = (cpi->use_svc) ? RT_INTER_MODES_SVC : RT_INTER_MODES; 1501 int flag_svc_subpel = 0; 1502 int svc_mv_col = 0; 1503 int svc_mv_row = 0; 1504 1505 init_ref_frame_cost(cm, xd, ref_frame_cost); 1506 1507 memset(&mode_checked[0][0], 0, MB_MODE_COUNT * MAX_REF_FRAMES); 1508 1509 if (reuse_inter_pred) { 1510 int i; 1511 for (i = 0; i < 3; i++) { 1512 #if CONFIG_VP9_HIGHBITDEPTH 1513 if (cm->use_highbitdepth) 1514 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]); 1515 else 1516 tmp[i].data = &pred_buf[pixels_in_block * i]; 1517 #else 1518 tmp[i].data = &pred_buf[pixels_in_block * i]; 1519 #endif // CONFIG_VP9_HIGHBITDEPTH 1520 tmp[i].stride = bw; 1521 tmp[i].in_use = 0; 1522 } 1523 tmp[3].data = pd->dst.buf; 1524 tmp[3].stride = pd->dst.stride; 1525 tmp[3].in_use = 0; 1526 } 1527 1528 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; 1529 x->skip = 0; 1530 1531 // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign 1532 // filter_ref, we use a less strict condition on assigning filter_ref. 1533 // This is to reduce the probabily of entering the flow of not assigning 1534 // filter_ref and then skip filter search. 1535 if (xd->above_mi && is_inter_block(xd->above_mi)) 1536 filter_ref = xd->above_mi->interp_filter; 1537 else if (xd->left_mi && is_inter_block(xd->left_mi)) 1538 filter_ref = xd->left_mi->interp_filter; 1539 else 1540 filter_ref = cm->interp_filter; 1541 1542 // initialize mode decisions 1543 vp9_rd_cost_reset(&best_rdc); 1544 vp9_rd_cost_reset(rd_cost); 1545 mi->sb_type = bsize; 1546 mi->ref_frame[0] = NONE; 1547 mi->ref_frame[1] = NONE; 1548 1549 mi->tx_size = 1550 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]); 1551 1552 if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) { 1553 #if CONFIG_VP9_HIGHBITDEPTH 1554 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) 1555 x->source_variance = vp9_high_get_sby_perpixel_variance( 1556 cpi, &x->plane[0].src, bsize, xd->bd); 1557 else 1558 #endif // CONFIG_VP9_HIGHBITDEPTH 1559 x->source_variance = 1560 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize); 1561 } 1562 1563 #if CONFIG_VP9_TEMPORAL_DENOISING 1564 if (cpi->oxcf.noise_sensitivity > 0) { 1565 if (cpi->use_svc) { 1566 int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id, 1567 cpi->svc.temporal_layer_id, 1568 cpi->svc.number_temporal_layers); 1569 LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer]; 1570 denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame; 1571 } 1572 if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode) 1573 vp9_denoiser_reset_frame_stats(ctx); 1574 } 1575 #endif 1576 1577 if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc && 1578 !cpi->rc.alt_ref_gf_group && !cpi->rc.last_frame_is_src_altref) { 1579 usable_ref_frame = LAST_FRAME; 1580 } else { 1581 usable_ref_frame = GOLDEN_FRAME; 1582 } 1583 1584 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) { 1585 if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref) 1586 usable_ref_frame = ALTREF_FRAME; 1587 1588 if (cpi->rc.is_src_frame_alt_ref) { 1589 skip_ref_find_pred[LAST_FRAME] = 1; 1590 skip_ref_find_pred[GOLDEN_FRAME] = 1; 1591 } 1592 if (!cm->show_frame) { 1593 if (cpi->rc.frames_since_key == 1) { 1594 usable_ref_frame = LAST_FRAME; 1595 skip_ref_find_pred[GOLDEN_FRAME] = 1; 1596 skip_ref_find_pred[ALTREF_FRAME] = 1; 1597 } 1598 } 1599 } 1600 1601 // For svc mode, on spatial_layer_id > 0: if the reference has different scale 1602 // constrain the inter mode to only test zero motion. 1603 if (cpi->use_svc && svc->force_zero_mode_spatial_ref && 1604 cpi->svc.spatial_layer_id > 0) { 1605 if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) { 1606 struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf; 1607 if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1; 1608 } 1609 if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) { 1610 struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf; 1611 if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1; 1612 } 1613 } 1614 1615 if (cpi->sf.short_circuit_low_temp_var) { 1616 force_skip_low_temp_var = 1617 get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize); 1618 // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3, 1619 // skip golden reference. 1620 if ((cpi->sf.short_circuit_low_temp_var == 1 || 1621 cpi->sf.short_circuit_low_temp_var == 3) && 1622 force_skip_low_temp_var) { 1623 usable_ref_frame = LAST_FRAME; 1624 } 1625 } 1626 1627 if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) && 1628 !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var)) 1629 use_golden_nonzeromv = 0; 1630 1631 if (cpi->oxcf.speed >= 8 && !cpi->use_svc && 1632 ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content || 1633 x->last_sb_high_content > 40 || cpi->rc.frames_since_golden > 120)) 1634 usable_ref_frame = LAST_FRAME; 1635 1636 // Compound prediction modes: (0,0) on LAST/GOLDEN and ARF. 1637 if (cm->reference_mode == REFERENCE_MODE_SELECT && 1638 cpi->sf.use_compound_nonrd_pickmode && usable_ref_frame == ALTREF_FRAME) 1639 comp_modes = 2; 1640 1641 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) { 1642 if (!skip_ref_find_pred[ref_frame]) { 1643 find_predictors(cpi, x, ref_frame, frame_mv, const_motion, 1644 &ref_frame_skip_mask, flag_list, tile_data, mi_row, 1645 mi_col, yv12_mb, bsize, force_skip_low_temp_var, 1646 comp_modes > 0); 1647 } 1648 } 1649 1650 if (cpi->use_svc || cpi->oxcf.speed <= 7 || bsize < BLOCK_32X32) 1651 x->sb_use_mv_part = 0; 1652 1653 // Set the flag_svc_subpel to 1 for SVC if the lower spatial layer used 1654 // an averaging filter for downsampling (phase = 8). If so, we will test 1655 // a nonzero motion mode on the spatial (goldeen) reference. 1656 // The nonzero motion is half pixel shifted to left and top (-4, -4). 1657 if (cpi->use_svc && cpi->svc.spatial_layer_id > 0 && 1658 svc_force_zero_mode[GOLDEN_FRAME - 1] && 1659 cpi->svc.downsample_filter_phase[cpi->svc.spatial_layer_id - 1] == 8) { 1660 svc_mv_col = -4; 1661 svc_mv_row = -4; 1662 flag_svc_subpel = 1; 1663 } 1664 1665 for (idx = 0; idx < num_inter_modes + comp_modes; ++idx) { 1666 int rate_mv = 0; 1667 int mode_rd_thresh; 1668 int mode_index; 1669 int i; 1670 int64_t this_sse; 1671 int is_skippable; 1672 int this_early_term = 0; 1673 int rd_computed = 0; 1674 int flag_preduv_computed[2] = { 0 }; 1675 int inter_mv_mode = 0; 1676 int skip_this_mv = 0; 1677 int comp_pred = 0; 1678 int force_gf_mv = 0; 1679 PREDICTION_MODE this_mode; 1680 second_ref_frame = NONE; 1681 1682 if (idx < num_inter_modes) { 1683 this_mode = ref_mode_set[idx].pred_mode; 1684 ref_frame = ref_mode_set[idx].ref_frame; 1685 1686 if (cpi->use_svc) { 1687 this_mode = ref_mode_set_svc[idx].pred_mode; 1688 ref_frame = ref_mode_set_svc[idx].ref_frame; 1689 } 1690 } else { 1691 // Add (0,0) compound modes. 1692 this_mode = ZEROMV; 1693 ref_frame = LAST_FRAME; 1694 if (idx == num_inter_modes + comp_modes - 1) ref_frame = GOLDEN_FRAME; 1695 second_ref_frame = ALTREF_FRAME; 1696 comp_pred = 1; 1697 } 1698 1699 if (ref_frame > usable_ref_frame) continue; 1700 if (skip_ref_find_pred[ref_frame]) continue; 1701 1702 if (flag_svc_subpel && ref_frame == GOLDEN_FRAME) { 1703 force_gf_mv = 1; 1704 // Only test mode if NEARESTMV/NEARMV is (svc_mv_col, svc_mv_row), 1705 // otherwise set NEWMV to (svc_mv_col, svc_mv_row). 1706 if (this_mode == NEWMV) { 1707 frame_mv[this_mode][ref_frame].as_mv.col = svc_mv_col; 1708 frame_mv[this_mode][ref_frame].as_mv.row = svc_mv_row; 1709 } else if (frame_mv[this_mode][ref_frame].as_mv.col != svc_mv_col || 1710 frame_mv[this_mode][ref_frame].as_mv.row != svc_mv_row) { 1711 continue; 1712 } 1713 } 1714 1715 if (comp_pred) { 1716 const struct segmentation *const seg = &cm->seg; 1717 if (!cpi->allow_comp_inter_inter) continue; 1718 // Skip compound inter modes if ARF is not available. 1719 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue; 1720 // Do not allow compound prediction if the segment level reference frame 1721 // feature is in use as in this case there can only be one reference. 1722 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) continue; 1723 } 1724 1725 // For SVC, skip the golden (spatial) reference search if sse of zeromv_last 1726 // is below threshold. 1727 if (cpi->use_svc && ref_frame == GOLDEN_FRAME && 1728 sse_zeromv_normalized < thresh_svc_skip_golden) 1729 continue; 1730 1731 if (sf->short_circuit_flat_blocks && x->source_variance == 0 && 1732 this_mode != NEARESTMV) { 1733 continue; 1734 } 1735 1736 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue; 1737 1738 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) { 1739 if (cpi->rc.is_src_frame_alt_ref && 1740 (ref_frame != ALTREF_FRAME || 1741 frame_mv[this_mode][ref_frame].as_int != 0)) 1742 continue; 1743 1744 if (!cm->show_frame && ref_frame == ALTREF_FRAME && 1745 frame_mv[this_mode][ref_frame].as_int != 0) 1746 continue; 1747 1748 if (cpi->rc.alt_ref_gf_group && cm->show_frame && 1749 cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) && 1750 ref_frame == GOLDEN_FRAME && 1751 frame_mv[this_mode][ref_frame].as_int != 0) 1752 continue; 1753 1754 if (cpi->rc.alt_ref_gf_group && cm->show_frame && 1755 cpi->rc.frames_since_golden > 0 && 1756 cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) && 1757 ref_frame == ALTREF_FRAME && 1758 frame_mv[this_mode][ref_frame].as_int != 0) 1759 continue; 1760 } 1761 1762 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue; 1763 1764 if (const_motion[ref_frame] && this_mode == NEARMV) continue; 1765 1766 // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var 1767 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped 1768 // later. 1769 if (!force_gf_mv && force_skip_low_temp_var && ref_frame == GOLDEN_FRAME && 1770 frame_mv[this_mode][ref_frame].as_int != 0) { 1771 continue; 1772 } 1773 1774 if (x->content_state_sb != kVeryHighSad && 1775 (cpi->sf.short_circuit_low_temp_var >= 2 || 1776 (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) && 1777 force_skip_low_temp_var && ref_frame == LAST_FRAME && 1778 this_mode == NEWMV) { 1779 continue; 1780 } 1781 1782 if (cpi->use_svc) { 1783 if (!force_gf_mv && svc_force_zero_mode[ref_frame - 1] && 1784 frame_mv[this_mode][ref_frame].as_int != 0) 1785 continue; 1786 } 1787 1788 if (sf->reference_masking && 1789 !(frame_mv[this_mode][ref_frame].as_int == 0 && 1790 ref_frame == LAST_FRAME)) { 1791 if (usable_ref_frame < ALTREF_FRAME) { 1792 if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) { 1793 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME; 1794 if ((cpi->ref_frame_flags & flag_list[i])) 1795 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1)) 1796 ref_frame_skip_mask |= (1 << ref_frame); 1797 } 1798 } else if (!cpi->rc.is_src_frame_alt_ref && 1799 !(frame_mv[this_mode][ref_frame].as_int == 0 && 1800 ref_frame == ALTREF_FRAME)) { 1801 int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME; 1802 int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME; 1803 if (((cpi->ref_frame_flags & flag_list[ref1]) && 1804 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) || 1805 ((cpi->ref_frame_flags & flag_list[ref2]) && 1806 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1)))) 1807 ref_frame_skip_mask |= (1 << ref_frame); 1808 } 1809 } 1810 if (ref_frame_skip_mask & (1 << ref_frame)) continue; 1811 1812 // Select prediction reference frames. 1813 for (i = 0; i < MAX_MB_PLANE; i++) { 1814 xd->plane[i].pre[0] = yv12_mb[ref_frame][i]; 1815 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i]; 1816 } 1817 1818 mi->ref_frame[0] = ref_frame; 1819 mi->ref_frame[1] = second_ref_frame; 1820 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame); 1821 1822 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)]; 1823 mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1 1824 : rd_threshes[mode_index]; 1825 1826 // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding 1827 // speed with little/no subjective quality loss. 1828 if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME && 1829 cpi->rc.frames_since_golden > 4) 1830 mode_rd_thresh = mode_rd_thresh << 3; 1831 1832 if ((cpi->sf.adaptive_rd_thresh_row_mt && 1833 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh, 1834 &rd_thresh_freq_fact[mode_index])) || 1835 (!cpi->sf.adaptive_rd_thresh_row_mt && 1836 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh, 1837 &rd_thresh_freq_fact[mode_index]))) 1838 continue; 1839 1840 if (this_mode == NEWMV && !force_gf_mv) { 1841 if (ref_frame > LAST_FRAME && !cpi->use_svc && 1842 cpi->oxcf.rc_mode == VPX_CBR) { 1843 int tmp_sad; 1844 uint32_t dis; 1845 int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX }; 1846 1847 if (bsize < BLOCK_16X16) continue; 1848 1849 tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col); 1850 1851 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue; 1852 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad) 1853 continue; 1854 1855 frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int; 1856 rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv, 1857 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv, 1858 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 1859 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3; 1860 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3; 1861 1862 cpi->find_fractional_mv_step( 1863 x, &frame_mv[NEWMV][ref_frame].as_mv, 1864 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv, 1865 cpi->common.allow_high_precision_mv, x->errorperbit, 1866 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop, 1867 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list), 1868 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0, 1869 0); 1870 } else if (svc->use_base_mv && svc->spatial_layer_id) { 1871 if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) { 1872 const int pre_stride = xd->plane[0].pre[0].stride; 1873 unsigned int base_mv_sse = UINT_MAX; 1874 int scale = (cpi->rc.avg_frame_low_motion > 60) ? 2 : 4; 1875 const uint8_t *const pre_buf = 1876 xd->plane[0].pre[0].buf + 1877 (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride + 1878 (frame_mv[NEWMV][ref_frame].as_mv.col >> 3); 1879 cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride, 1880 pre_buf, pre_stride, &base_mv_sse); 1881 1882 // Exit NEWMV search if base_mv is (0,0) && bsize < BLOCK_16x16, 1883 // for SVC encoding. 1884 if (cpi->use_svc && cpi->svc.use_base_mv && bsize < BLOCK_16X16 && 1885 frame_mv[NEWMV][ref_frame].as_mv.row == 0 && 1886 frame_mv[NEWMV][ref_frame].as_mv.col == 0) 1887 continue; 1888 1889 // Exit NEWMV search if base_mv_sse is large. 1890 if (sf->base_mv_aggressive && base_mv_sse > (best_sse_sofar << scale)) 1891 continue; 1892 if (base_mv_sse < (best_sse_sofar << 1)) { 1893 // Base layer mv is good. 1894 // Exit NEWMV search if the base_mv is (0, 0) and sse is low, since 1895 // (0, 0) mode is already tested. 1896 unsigned int base_mv_sse_normalized = 1897 base_mv_sse >> 1898 (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]); 1899 if (sf->base_mv_aggressive && base_mv_sse <= best_sse_sofar && 1900 base_mv_sse_normalized < 400 && 1901 frame_mv[NEWMV][ref_frame].as_mv.row == 0 && 1902 frame_mv[NEWMV][ref_frame].as_mv.col == 0) 1903 continue; 1904 if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col, 1905 &frame_mv[NEWMV][ref_frame], &rate_mv, 1906 best_rdc.rdcost, 1)) { 1907 continue; 1908 } 1909 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col, 1910 &frame_mv[NEWMV][ref_frame], 1911 &rate_mv, best_rdc.rdcost, 0)) { 1912 continue; 1913 } 1914 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col, 1915 &frame_mv[NEWMV][ref_frame], 1916 &rate_mv, best_rdc.rdcost, 0)) { 1917 continue; 1918 } 1919 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col, 1920 &frame_mv[NEWMV][ref_frame], &rate_mv, 1921 best_rdc.rdcost, 0)) { 1922 continue; 1923 } 1924 } 1925 1926 // TODO(jianj): Skipping the testing of (duplicate) non-zero motion vector 1927 // causes some regression, leave it for duplicate zero-mv for now, until 1928 // regression issue is resolved. 1929 for (inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV; inter_mv_mode++) { 1930 if (inter_mv_mode == this_mode || comp_pred) continue; 1931 if (mode_checked[inter_mv_mode][ref_frame] && 1932 frame_mv[this_mode][ref_frame].as_int == 1933 frame_mv[inter_mv_mode][ref_frame].as_int && 1934 frame_mv[inter_mv_mode][ref_frame].as_int == 0) { 1935 skip_this_mv = 1; 1936 break; 1937 } 1938 } 1939 1940 if (skip_this_mv) continue; 1941 1942 // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no 1943 // need to compute best_pred_sad which is only used to skip golden NEWMV. 1944 if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME && 1945 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) { 1946 const int pre_stride = xd->plane[0].pre[0].stride; 1947 const uint8_t *const pre_buf = 1948 xd->plane[0].pre[0].buf + 1949 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride + 1950 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3); 1951 best_pred_sad = cpi->fn_ptr[bsize].sdf( 1952 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride); 1953 x->pred_mv_sad[LAST_FRAME] = best_pred_sad; 1954 } 1955 1956 if (this_mode != NEARESTMV && !comp_pred && 1957 frame_mv[this_mode][ref_frame].as_int == 1958 frame_mv[NEARESTMV][ref_frame].as_int) 1959 continue; 1960 1961 mi->mode = this_mode; 1962 mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int; 1963 mi->mv[1].as_int = 0; 1964 1965 // Search for the best prediction filter type, when the resulting 1966 // motion vector is at sub-pixel accuracy level for luma component, i.e., 1967 // the last three bits are all zeros. 1968 if (reuse_inter_pred) { 1969 if (!this_mode_pred) { 1970 this_mode_pred = &tmp[3]; 1971 } else { 1972 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)]; 1973 pd->dst.buf = this_mode_pred->data; 1974 pd->dst.stride = bw; 1975 } 1976 } 1977 1978 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) && 1979 pred_filter_search && 1980 (ref_frame == LAST_FRAME || 1981 (ref_frame == GOLDEN_FRAME && !force_gf_mv && 1982 (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) && 1983 (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) { 1984 int pf_rate[3]; 1985 int64_t pf_dist[3]; 1986 int curr_rate[3]; 1987 unsigned int pf_var[3]; 1988 unsigned int pf_sse[3]; 1989 TX_SIZE pf_tx_size[3]; 1990 int64_t best_cost = INT64_MAX; 1991 INTERP_FILTER best_filter = SWITCHABLE, filter; 1992 PRED_BUFFER *current_pred = this_mode_pred; 1993 rd_computed = 1; 1994 1995 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) { 1996 int64_t cost; 1997 mi->interp_filter = filter; 1998 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); 1999 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter], 2000 &pf_var[filter], &pf_sse[filter]); 2001 curr_rate[filter] = pf_rate[filter]; 2002 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd); 2003 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]); 2004 pf_tx_size[filter] = mi->tx_size; 2005 if (cost < best_cost) { 2006 best_filter = filter; 2007 best_cost = cost; 2008 skip_txfm = x->skip_txfm[0]; 2009 2010 if (reuse_inter_pred) { 2011 if (this_mode_pred != current_pred) { 2012 free_pred_buffer(this_mode_pred); 2013 this_mode_pred = current_pred; 2014 } 2015 current_pred = &tmp[get_pred_buffer(tmp, 3)]; 2016 pd->dst.buf = current_pred->data; 2017 pd->dst.stride = bw; 2018 } 2019 } 2020 } 2021 2022 if (reuse_inter_pred && this_mode_pred != current_pred) 2023 free_pred_buffer(current_pred); 2024 2025 mi->interp_filter = best_filter; 2026 mi->tx_size = pf_tx_size[best_filter]; 2027 this_rdc.rate = curr_rate[best_filter]; 2028 this_rdc.dist = pf_dist[best_filter]; 2029 var_y = pf_var[best_filter]; 2030 sse_y = pf_sse[best_filter]; 2031 x->skip_txfm[0] = skip_txfm; 2032 if (reuse_inter_pred) { 2033 pd->dst.buf = this_mode_pred->data; 2034 pd->dst.stride = this_mode_pred->stride; 2035 } 2036 } else { 2037 // For low motion content use x->sb_is_skin in addition to VeryHighSad 2038 // for setting large_block. 2039 const int large_block = 2040 (x->content_state_sb == kVeryHighSad || 2041 (x->sb_is_skin && cpi->rc.avg_frame_low_motion > 70) || 2042 cpi->oxcf.speed < 7) 2043 ? bsize > BLOCK_32X32 2044 : bsize >= BLOCK_32X32; 2045 mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref; 2046 2047 if (cpi->use_svc && ref_frame == GOLDEN_FRAME && 2048 svc_force_zero_mode[ref_frame - 1]) 2049 mi->interp_filter = filter_gf_svc; 2050 2051 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); 2052 2053 // For large partition blocks, extra testing is done. 2054 if (cpi->oxcf.rc_mode == VPX_CBR && large_block && 2055 !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) && 2056 cm->base_qindex) { 2057 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate, 2058 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col, 2059 &this_early_term, flag_preduv_computed); 2060 } else { 2061 rd_computed = 1; 2062 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist, 2063 &var_y, &sse_y); 2064 } 2065 // Save normalized sse (between current and last frame) for (0, 0) motion. 2066 if (cpi->use_svc && ref_frame == LAST_FRAME && 2067 frame_mv[this_mode][ref_frame].as_int == 0) { 2068 sse_zeromv_normalized = 2069 sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]); 2070 } 2071 if (sse_y < best_sse_sofar) best_sse_sofar = sse_y; 2072 } 2073 2074 if (!this_early_term) { 2075 this_sse = (int64_t)sse_y; 2076 block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize, 2077 VPXMIN(mi->tx_size, TX_16X16), rd_computed); 2078 2079 x->skip_txfm[0] = is_skippable; 2080 if (is_skippable) { 2081 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2082 } else { 2083 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) < 2084 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) { 2085 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0); 2086 } else { 2087 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2088 this_rdc.dist = this_sse; 2089 x->skip_txfm[0] = SKIP_TXFM_AC_DC; 2090 } 2091 } 2092 2093 if (cm->interp_filter == SWITCHABLE) { 2094 if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) 2095 this_rdc.rate += vp9_get_switchable_rate(cpi, xd); 2096 } 2097 } else { 2098 this_rdc.rate += cm->interp_filter == SWITCHABLE 2099 ? vp9_get_switchable_rate(cpi, xd) 2100 : 0; 2101 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1); 2102 } 2103 2104 if (!this_early_term && 2105 (x->color_sensitivity[0] || x->color_sensitivity[1])) { 2106 RD_COST rdc_uv; 2107 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]); 2108 if (x->color_sensitivity[0] && !flag_preduv_computed[0]) { 2109 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1); 2110 flag_preduv_computed[0] = 1; 2111 } 2112 if (x->color_sensitivity[1] && !flag_preduv_computed[1]) { 2113 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2); 2114 flag_preduv_computed[1] = 1; 2115 } 2116 model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2); 2117 this_rdc.rate += rdc_uv.rate; 2118 this_rdc.dist += rdc_uv.dist; 2119 } 2120 2121 this_rdc.rate += rate_mv; 2122 this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]] 2123 [INTER_OFFSET(this_mode)]; 2124 // TODO(marpan): Add costing for compound mode. 2125 this_rdc.rate += ref_frame_cost[ref_frame]; 2126 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist); 2127 2128 // Bias against NEWMV that is very different from its neighbors, and bias 2129 // to small motion-lastref for noisy input. 2130 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 && 2131 cpi->oxcf.content != VP9E_CONTENT_SCREEN) { 2132 vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize, 2133 frame_mv[this_mode][ref_frame].as_mv.row, 2134 frame_mv[this_mode][ref_frame].as_mv.col, 2135 ref_frame == LAST_FRAME, x->lowvar_highsumdiff, 2136 x->sb_is_skin); 2137 } 2138 2139 // Skipping checking: test to see if this block can be reconstructed by 2140 // prediction only. 2141 if (cpi->allow_encode_breakout) { 2142 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode, 2143 var_y, sse_y, yv12_mb, &this_rdc.rate, 2144 &this_rdc.dist, flag_preduv_computed); 2145 if (x->skip) { 2146 this_rdc.rate += rate_mv; 2147 this_rdc.rdcost = 2148 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist); 2149 } 2150 } 2151 2152 #if CONFIG_VP9_TEMPORAL_DENOISING 2153 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode && 2154 cpi->denoiser.denoising_level > kDenLowLow) { 2155 vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx); 2156 // Keep track of zero_last cost. 2157 if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0) 2158 zero_last_cost_orig = this_rdc.rdcost; 2159 } 2160 #else 2161 (void)ctx; 2162 #endif 2163 2164 mode_checked[this_mode][ref_frame] = 1; 2165 2166 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) { 2167 best_rdc = this_rdc; 2168 best_mode = this_mode; 2169 best_pred_filter = mi->interp_filter; 2170 best_tx_size = mi->tx_size; 2171 best_ref_frame = ref_frame; 2172 best_mode_skip_txfm = x->skip_txfm[0]; 2173 best_early_term = this_early_term; 2174 best_second_ref_frame = second_ref_frame; 2175 2176 if (reuse_inter_pred) { 2177 free_pred_buffer(best_pred); 2178 best_pred = this_mode_pred; 2179 } 2180 } else { 2181 if (reuse_inter_pred) free_pred_buffer(this_mode_pred); 2182 } 2183 2184 if (x->skip) break; 2185 2186 // If early termination flag is 1 and at least 2 modes are checked, 2187 // the mode search is terminated. 2188 if (best_early_term && idx > 0) { 2189 x->skip = 1; 2190 break; 2191 } 2192 } 2193 2194 mi->mode = best_mode; 2195 mi->interp_filter = best_pred_filter; 2196 mi->tx_size = best_tx_size; 2197 mi->ref_frame[0] = best_ref_frame; 2198 mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int; 2199 xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int; 2200 x->skip_txfm[0] = best_mode_skip_txfm; 2201 mi->ref_frame[1] = best_second_ref_frame; 2202 2203 // For spatial enhancemanent layer: perform intra prediction only if base 2204 // layer is chosen as the reference. Always perform intra prediction if 2205 // LAST is the only reference or is_key_frame is set. 2206 if (cpi->svc.spatial_layer_id) { 2207 perform_intra_pred = 2208 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame || 2209 !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) || 2210 (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame && 2211 svc_force_zero_mode[best_ref_frame - 1]); 2212 inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh; 2213 } 2214 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR && 2215 cpi->rc.is_src_frame_alt_ref) 2216 perform_intra_pred = 0; 2217 // Perform intra prediction search, if the best SAD is above a certain 2218 // threshold. 2219 if (best_rdc.rdcost == INT64_MAX || 2220 ((!force_skip_low_temp_var || bsize < BLOCK_32X32 || 2221 x->content_state_sb == kVeryHighSad) && 2222 perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh && 2223 bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad && 2224 !x->lowvar_highsumdiff)) { 2225 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 }; 2226 int i; 2227 TX_SIZE best_intra_tx_size = TX_SIZES; 2228 TX_SIZE intra_tx_size = 2229 VPXMIN(max_txsize_lookup[bsize], 2230 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); 2231 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16) 2232 intra_tx_size = TX_16X16; 2233 2234 if (reuse_inter_pred && best_pred != NULL) { 2235 if (best_pred->data == orig_dst.buf) { 2236 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)]; 2237 #if CONFIG_VP9_HIGHBITDEPTH 2238 if (cm->use_highbitdepth) 2239 vpx_highbd_convolve_copy( 2240 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride, 2241 CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride, 2242 NULL, 0, 0, 0, 0, bw, bh, xd->bd); 2243 else 2244 vpx_convolve_copy(best_pred->data, best_pred->stride, 2245 this_mode_pred->data, this_mode_pred->stride, NULL, 2246 0, 0, 0, 0, bw, bh); 2247 #else 2248 vpx_convolve_copy(best_pred->data, best_pred->stride, 2249 this_mode_pred->data, this_mode_pred->stride, NULL, 0, 2250 0, 0, 0, bw, bh); 2251 #endif // CONFIG_VP9_HIGHBITDEPTH 2252 best_pred = this_mode_pred; 2253 } 2254 } 2255 pd->dst = orig_dst; 2256 2257 for (i = 0; i < 4; ++i) { 2258 const PREDICTION_MODE this_mode = intra_mode_list[i]; 2259 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)]; 2260 int mode_rd_thresh = rd_threshes[mode_index]; 2261 if (sf->short_circuit_flat_blocks && x->source_variance == 0 && 2262 this_mode != DC_PRED) { 2263 continue; 2264 } 2265 2266 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize])) 2267 continue; 2268 2269 if ((cpi->sf.adaptive_rd_thresh_row_mt && 2270 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh, 2271 &rd_thresh_freq_fact[mode_index])) || 2272 (!cpi->sf.adaptive_rd_thresh_row_mt && 2273 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh, 2274 &rd_thresh_freq_fact[mode_index]))) 2275 continue; 2276 2277 mi->mode = this_mode; 2278 mi->ref_frame[0] = INTRA_FRAME; 2279 this_rdc.dist = this_rdc.rate = 0; 2280 args.mode = this_mode; 2281 args.skippable = 1; 2282 args.rdc = &this_rdc; 2283 mi->tx_size = intra_tx_size; 2284 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra, 2285 &args); 2286 // Check skip cost here since skippable is not set for for uv, this 2287 // mirrors the behavior used by inter 2288 if (args.skippable) { 2289 x->skip_txfm[0] = SKIP_TXFM_AC_DC; 2290 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1); 2291 } else { 2292 x->skip_txfm[0] = SKIP_TXFM_NONE; 2293 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0); 2294 } 2295 // Inter and intra RD will mismatch in scale for non-screen content. 2296 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) { 2297 if (x->color_sensitivity[0]) 2298 vp9_foreach_transformed_block_in_plane(xd, bsize, 1, 2299 estimate_block_intra, &args); 2300 if (x->color_sensitivity[1]) 2301 vp9_foreach_transformed_block_in_plane(xd, bsize, 2, 2302 estimate_block_intra, &args); 2303 } 2304 this_rdc.rate += cpi->mbmode_cost[this_mode]; 2305 this_rdc.rate += ref_frame_cost[INTRA_FRAME]; 2306 this_rdc.rate += intra_cost_penalty; 2307 this_rdc.rdcost = 2308 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist); 2309 2310 if (this_rdc.rdcost < best_rdc.rdcost) { 2311 best_rdc = this_rdc; 2312 best_mode = this_mode; 2313 best_intra_tx_size = mi->tx_size; 2314 best_ref_frame = INTRA_FRAME; 2315 best_second_ref_frame = NONE; 2316 mi->uv_mode = this_mode; 2317 mi->mv[0].as_int = INVALID_MV; 2318 mi->mv[1].as_int = INVALID_MV; 2319 best_mode_skip_txfm = x->skip_txfm[0]; 2320 } 2321 } 2322 2323 // Reset mb_mode_info to the best inter mode. 2324 if (best_ref_frame != INTRA_FRAME) { 2325 mi->tx_size = best_tx_size; 2326 } else { 2327 mi->tx_size = best_intra_tx_size; 2328 } 2329 } 2330 2331 pd->dst = orig_dst; 2332 mi->mode = best_mode; 2333 mi->ref_frame[0] = best_ref_frame; 2334 mi->ref_frame[1] = best_second_ref_frame; 2335 x->skip_txfm[0] = best_mode_skip_txfm; 2336 2337 if (!is_inter_block(mi)) { 2338 mi->interp_filter = SWITCHABLE_FILTERS; 2339 } 2340 2341 if (reuse_inter_pred && best_pred != NULL) { 2342 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) { 2343 #if CONFIG_VP9_HIGHBITDEPTH 2344 if (cm->use_highbitdepth) 2345 vpx_highbd_convolve_copy( 2346 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride, 2347 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, 0, 0, 0, 2348 bw, bh, xd->bd); 2349 else 2350 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf, 2351 pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh); 2352 #else 2353 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf, 2354 pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh); 2355 #endif // CONFIG_VP9_HIGHBITDEPTH 2356 } 2357 } 2358 2359 #if CONFIG_VP9_TEMPORAL_DENOISING 2360 if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 && 2361 denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow && 2362 cpi->denoiser.reset == 0) { 2363 VP9_DENOISER_DECISION decision = COPY_BLOCK; 2364 ctx->sb_skip_denoising = 0; 2365 // TODO(marpan): There is an issue with denoising when the 2366 // superblock partitioning scheme is based on the pickmode. 2367 // Remove this condition when the issue is resolved. 2368 if (x->sb_pickmode_part) ctx->sb_skip_denoising = 1; 2369 vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost, 2370 frame_mv, reuse_inter_pred, best_tx_size, 2371 best_mode, best_ref_frame, best_pred_filter, 2372 best_mode_skip_txfm); 2373 vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision); 2374 recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb, 2375 &best_rdc, bsize, mi_row, mi_col); 2376 best_ref_frame = ctx_den.best_ref_frame; 2377 } 2378 #endif 2379 2380 if (best_ref_frame == ALTREF_FRAME || best_second_ref_frame == ALTREF_FRAME) 2381 x->arf_frame_usage++; 2382 else if (best_ref_frame != INTRA_FRAME) 2383 x->lastgolden_frame_usage++; 2384 2385 if (cpi->sf.adaptive_rd_thresh) { 2386 THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)]; 2387 2388 if (best_ref_frame == INTRA_FRAME) { 2389 // Only consider the modes that are included in the intra_mode_list. 2390 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE); 2391 int i; 2392 2393 // TODO(yunqingwang): Check intra mode mask and only update freq_fact 2394 // for those valid modes. 2395 for (i = 0; i < intra_modes; i++) { 2396 if (cpi->sf.adaptive_rd_thresh_row_mt) 2397 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance, 2398 thresh_freq_fact_idx, INTRA_FRAME, 2399 best_mode_idx, intra_mode_list[i]); 2400 else 2401 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize, 2402 INTRA_FRAME, best_mode_idx, 2403 intra_mode_list[i]); 2404 } 2405 } else { 2406 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) { 2407 PREDICTION_MODE this_mode; 2408 if (best_ref_frame != ref_frame) continue; 2409 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) { 2410 if (cpi->sf.adaptive_rd_thresh_row_mt) 2411 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance, 2412 thresh_freq_fact_idx, ref_frame, 2413 best_mode_idx, this_mode); 2414 else 2415 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize, 2416 ref_frame, best_mode_idx, this_mode); 2417 } 2418 } 2419 } 2420 } 2421 2422 *rd_cost = best_rdc; 2423 } 2424 2425 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row, 2426 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize, 2427 PICK_MODE_CONTEXT *ctx) { 2428 VP9_COMMON *const cm = &cpi->common; 2429 SPEED_FEATURES *const sf = &cpi->sf; 2430 MACROBLOCKD *const xd = &x->e_mbd; 2431 MODE_INFO *const mi = xd->mi[0]; 2432 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext; 2433 const struct segmentation *const seg = &cm->seg; 2434 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE; 2435 MV_REFERENCE_FRAME best_ref_frame = NONE; 2436 unsigned char segment_id = mi->segment_id; 2437 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; 2438 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, 2439 VP9_ALT_FLAG }; 2440 int64_t best_rd = INT64_MAX; 2441 b_mode_info bsi[MAX_REF_FRAMES][4]; 2442 int ref_frame_skip_mask = 0; 2443 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize]; 2444 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize]; 2445 int idx, idy; 2446 2447 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; 2448 ctx->pred_pixel_ready = 0; 2449 2450 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) { 2451 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame); 2452 int_mv dummy_mv[2]; 2453 x->pred_mv_sad[ref_frame] = INT_MAX; 2454 2455 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) { 2456 int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame]; 2457 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf; 2458 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, 2459 sf); 2460 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col, 2461 mbmi_ext->mode_context); 2462 2463 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates, 2464 &dummy_mv[0], &dummy_mv[1]); 2465 } else { 2466 ref_frame_skip_mask |= (1 << ref_frame); 2467 } 2468 } 2469 2470 mi->sb_type = bsize; 2471 mi->tx_size = TX_4X4; 2472 mi->uv_mode = DC_PRED; 2473 mi->ref_frame[0] = LAST_FRAME; 2474 mi->ref_frame[1] = NONE; 2475 mi->interp_filter = 2476 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter; 2477 2478 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) { 2479 int64_t this_rd = 0; 2480 int plane; 2481 2482 if (ref_frame_skip_mask & (1 << ref_frame)) continue; 2483 2484 #if CONFIG_BETTER_HW_COMPATIBILITY 2485 if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME && 2486 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf)) 2487 continue; 2488 #endif 2489 2490 // TODO(jingning, agrange): Scaling reference frame not supported for 2491 // sub8x8 blocks. Is this supported now? 2492 if (ref_frame > INTRA_FRAME && 2493 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf)) 2494 continue; 2495 2496 // If the segment reference frame feature is enabled.... 2497 // then do nothing if the current ref frame is not allowed.. 2498 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) && 2499 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) 2500 continue; 2501 2502 mi->ref_frame[0] = ref_frame; 2503 x->skip = 0; 2504 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame); 2505 2506 // Select prediction reference frames. 2507 for (plane = 0; plane < MAX_MB_PLANE; plane++) 2508 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane]; 2509 2510 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) { 2511 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) { 2512 int_mv b_mv[MB_MODE_COUNT]; 2513 int64_t b_best_rd = INT64_MAX; 2514 const int i = idy * 2 + idx; 2515 PREDICTION_MODE this_mode; 2516 RD_COST this_rdc; 2517 unsigned int var_y, sse_y; 2518 2519 struct macroblock_plane *p = &x->plane[0]; 2520 struct macroblockd_plane *pd = &xd->plane[0]; 2521 2522 const struct buf_2d orig_src = p->src; 2523 const struct buf_2d orig_dst = pd->dst; 2524 struct buf_2d orig_pre[2]; 2525 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre)); 2526 2527 // set buffer pointers for sub8x8 motion search. 2528 p->src.buf = 2529 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)]; 2530 pd->dst.buf = 2531 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)]; 2532 pd->pre[0].buf = 2533 &pd->pre[0] 2534 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)]; 2535 2536 b_mv[ZEROMV].as_int = 0; 2537 b_mv[NEWMV].as_int = INVALID_MV; 2538 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col, 2539 &b_mv[NEARESTMV], &b_mv[NEARMV], 2540 mbmi_ext->mode_context); 2541 2542 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) { 2543 int b_rate = 0; 2544 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int; 2545 2546 if (this_mode == NEWMV) { 2547 const int step_param = cpi->sf.mv.fullpel_search_step_param; 2548 MV mvp_full; 2549 MV tmp_mv; 2550 int cost_list[5]; 2551 const MvLimits tmp_mv_limits = x->mv_limits; 2552 uint32_t dummy_dist; 2553 2554 if (i == 0) { 2555 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3; 2556 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3; 2557 } else { 2558 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3; 2559 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3; 2560 } 2561 2562 vp9_set_mv_search_range(&x->mv_limits, 2563 &mbmi_ext->ref_mvs[ref_frame][0].as_mv); 2564 2565 vp9_full_pixel_search( 2566 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, 2567 x->sadperbit4, cond_cost_list(cpi, cost_list), 2568 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0); 2569 2570 x->mv_limits = tmp_mv_limits; 2571 2572 // calculate the bit cost on motion vector 2573 mvp_full.row = tmp_mv.row * 8; 2574 mvp_full.col = tmp_mv.col * 8; 2575 2576 b_rate += vp9_mv_bit_cost( 2577 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv, 2578 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); 2579 2580 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]] 2581 [INTER_OFFSET(NEWMV)]; 2582 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue; 2583 2584 cpi->find_fractional_mv_step( 2585 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv, 2586 cpi->common.allow_high_precision_mv, x->errorperbit, 2587 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop, 2588 cpi->sf.mv.subpel_iters_per_step, 2589 cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost, 2590 &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0); 2591 2592 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv; 2593 } else { 2594 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]] 2595 [INTER_OFFSET(this_mode)]; 2596 } 2597 2598 #if CONFIG_VP9_HIGHBITDEPTH 2599 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { 2600 vp9_highbd_build_inter_predictor( 2601 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride, 2602 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, 2603 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf, 2604 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0, 2605 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3, 2606 mi_col * MI_SIZE + 4 * (i & 0x01), 2607 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd); 2608 } else { 2609 #endif 2610 vp9_build_inter_predictor( 2611 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride, 2612 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf, 2613 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0, 2614 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3, 2615 mi_col * MI_SIZE + 4 * (i & 0x01), 2616 mi_row * MI_SIZE + 4 * (i >> 1)); 2617 2618 #if CONFIG_VP9_HIGHBITDEPTH 2619 } 2620 #endif 2621 2622 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist, 2623 &var_y, &sse_y); 2624 2625 this_rdc.rate += b_rate; 2626 this_rdc.rdcost = 2627 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist); 2628 if (this_rdc.rdcost < b_best_rd) { 2629 b_best_rd = this_rdc.rdcost; 2630 bsi[ref_frame][i].as_mode = this_mode; 2631 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv; 2632 } 2633 } // mode search 2634 2635 // restore source and prediction buffer pointers. 2636 p->src = orig_src; 2637 pd->pre[0] = orig_pre[0]; 2638 pd->dst = orig_dst; 2639 this_rd += b_best_rd; 2640 2641 xd->mi[0]->bmi[i] = bsi[ref_frame][i]; 2642 if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i]; 2643 if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i]; 2644 } 2645 } // loop through sub8x8 blocks 2646 2647 if (this_rd < best_rd) { 2648 best_rd = this_rd; 2649 best_ref_frame = ref_frame; 2650 } 2651 } // reference frames 2652 2653 mi->tx_size = TX_4X4; 2654 mi->ref_frame[0] = best_ref_frame; 2655 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) { 2656 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) { 2657 const int block = idy * 2 + idx; 2658 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block]; 2659 if (num_4x4_blocks_wide > 1) 2660 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block]; 2661 if (num_4x4_blocks_high > 1) 2662 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block]; 2663 } 2664 } 2665 mi->mode = xd->mi[0]->bmi[3].as_mode; 2666 ctx->mic = *(xd->mi[0]); 2667 ctx->mbmi_ext = *x->mbmi_ext; 2668 ctx->skip_txfm[0] = SKIP_TXFM_NONE; 2669 ctx->skip = 0; 2670 // Dummy assignment for speed -5. No effect in speed -6. 2671 rd_cost->rdcost = best_rd; 2672 } 2673