1/* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 12#include <limits.h> 13#include "vpx_ports/config.h" 14#include "onyx_int.h" 15#include "modecosts.h" 16#include "encodeintra.h" 17#include "vp8/common/entropymode.h" 18#include "pickinter.h" 19#include "vp8/common/findnearmv.h" 20#include "encodemb.h" 21#include "vp8/common/reconinter.h" 22#include "vp8/common/reconintra.h" 23#include "vp8/common/reconintra4x4.h" 24#include "vp8/common/g_common.h" 25#include "variance.h" 26#include "mcomp.h" 27#include "rdopt.h" 28#include "vpx_mem/vpx_mem.h" 29 30#if CONFIG_RUNTIME_CPU_DETECT 31#define IF_RTCD(x) (x) 32#else 33#define IF_RTCD(x) NULL 34#endif 35 36extern int VP8_UVSSE(MACROBLOCK *x, const vp8_variance_rtcd_vtable_t *rtcd); 37 38#ifdef SPEEDSTATS 39extern unsigned int cnt_pm; 40#endif 41 42extern const MV_REFERENCE_FRAME vp8_ref_frame_order[MAX_MODES]; 43extern const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES]; 44 45 46extern unsigned int (*vp8_get16x16pred_error)(unsigned char *src_ptr, int src_stride, unsigned char *ref_ptr, int ref_stride); 47extern unsigned int (*vp8_get4x4sse_cs)(unsigned char *src_ptr, int source_stride, unsigned char *ref_ptr, int recon_stride); 48extern int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x, MV *best_ref_mv, int best_rd, int *, int *, int *, int, int *mvcost[2], int, int fullpixel); 49extern int vp8_cost_mv_ref(MB_PREDICTION_MODE m, const int near_mv_ref_ct[4]); 50extern void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, MV *mv); 51 52 53int vp8_skip_fractional_mv_step(MACROBLOCK *mb, BLOCK *b, BLOCKD *d, MV *bestmv, MV *ref_mv, int error_per_bit, const vp8_variance_fn_ptr_t *vfp, int *mvcost[2]) 54{ 55 (void) b; 56 (void) d; 57 (void) ref_mv; 58 (void) error_per_bit; 59 (void) vfp; 60 (void) mvcost; 61 bestmv->row <<= 3; 62 bestmv->col <<= 3; 63 return 0; 64} 65 66 67static int get_inter_mbpred_error(MACROBLOCK *mb, const vp8_variance_fn_ptr_t *vfp, unsigned int *sse) 68{ 69 70 BLOCK *b = &mb->block[0]; 71 BLOCKD *d = &mb->e_mbd.block[0]; 72 unsigned char *what = (*(b->base_src) + b->src); 73 int what_stride = b->src_stride; 74 unsigned char *in_what = *(d->base_pre) + d->pre ; 75 int in_what_stride = d->pre_stride; 76 int xoffset = d->bmi.mv.as_mv.col & 7; 77 int yoffset = d->bmi.mv.as_mv.row & 7; 78 79 in_what += (d->bmi.mv.as_mv.row >> 3) * d->pre_stride + (d->bmi.mv.as_mv.col >> 3); 80 81 if (xoffset | yoffset) 82 { 83 return vfp->svf(in_what, in_what_stride, xoffset, yoffset, what, what_stride, sse); 84 } 85 else 86 { 87 return vfp->vf(what, what_stride, in_what, in_what_stride, sse); 88 } 89 90} 91 92unsigned int vp8_get16x16pred_error_c 93( 94 const unsigned char *src_ptr, 95 int src_stride, 96 const unsigned char *ref_ptr, 97 int ref_stride, 98 int max_sad 99) 100{ 101 unsigned pred_error = 0; 102 int i, j; 103 int sum = 0; 104 105 for (i = 0; i < 16; i++) 106 { 107 int diff; 108 109 for (j = 0; j < 16; j++) 110 { 111 diff = src_ptr[j] - ref_ptr[j]; 112 sum += diff; 113 pred_error += diff * diff; 114 } 115 116 src_ptr += src_stride; 117 ref_ptr += ref_stride; 118 } 119 120 pred_error -= sum * sum / 256; 121 return pred_error; 122} 123 124 125unsigned int vp8_get4x4sse_cs_c 126( 127 const unsigned char *src_ptr, 128 int source_stride, 129 const unsigned char *ref_ptr, 130 int recon_stride, 131 int max_sad 132) 133{ 134 int distortion = 0; 135 int r, c; 136 137 for (r = 0; r < 4; r++) 138 { 139 for (c = 0; c < 4; c++) 140 { 141 int diff = src_ptr[c] - ref_ptr[c]; 142 distortion += diff * diff; 143 } 144 145 src_ptr += source_stride; 146 ref_ptr += recon_stride; 147 } 148 149 return distortion; 150} 151 152static int get_prediction_error(BLOCK *be, BLOCKD *b, const vp8_variance_rtcd_vtable_t *rtcd) 153{ 154 unsigned char *sptr; 155 unsigned char *dptr; 156 sptr = (*(be->base_src) + be->src); 157 dptr = b->predictor; 158 159 return VARIANCE_INVOKE(rtcd, get4x4sse_cs)(sptr, be->src_stride, dptr, 16, 0x7fffffff); 160 161} 162 163static int pick_intra4x4block( 164 const VP8_ENCODER_RTCD *rtcd, 165 MACROBLOCK *x, 166 BLOCK *be, 167 BLOCKD *b, 168 B_PREDICTION_MODE *best_mode, 169 B_PREDICTION_MODE above, 170 B_PREDICTION_MODE left, 171 172 int *bestrate, 173 int *bestdistortion) 174{ 175 B_PREDICTION_MODE mode; 176 int best_rd = INT_MAX; // 1<<30 177 int rate; 178 int distortion; 179 unsigned int *mode_costs; 180 181 if (x->e_mbd.frame_type == KEY_FRAME) 182 { 183 mode_costs = x->bmode_costs[above][left]; 184 } 185 else 186 { 187 mode_costs = x->inter_bmode_costs; 188 } 189 190 for (mode = B_DC_PRED; mode <= B_HE_PRED /*B_HU_PRED*/; mode++) 191 { 192 int this_rd; 193 194 rate = mode_costs[mode]; 195 vp8_predict_intra4x4(b, mode, b->predictor); 196 distortion = get_prediction_error(be, b, &rtcd->variance); 197 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate, distortion); 198 199 if (this_rd < best_rd) 200 { 201 *bestrate = rate; 202 *bestdistortion = distortion; 203 best_rd = this_rd; 204 *best_mode = mode; 205 } 206 } 207 208 b->bmi.mode = (B_PREDICTION_MODE)(*best_mode); 209 vp8_encode_intra4x4block(rtcd, x, be, b, b->bmi.mode); 210 211 return best_rd; 212} 213 214 215int vp8_pick_intra4x4mby_modes(const VP8_ENCODER_RTCD *rtcd, MACROBLOCK *mb, int *Rate, int *best_dist) 216{ 217 MACROBLOCKD *const xd = &mb->e_mbd; 218 int i; 219 int cost = mb->mbmode_cost [xd->frame_type] [B_PRED]; 220 int error; 221 int distortion = 0; 222 223 vp8_intra_prediction_down_copy(xd); 224 225 for (i = 0; i < 16; i++) 226 { 227 MODE_INFO *const mic = xd->mode_info_context; 228 const int mis = xd->mode_info_stride; 229 const B_PREDICTION_MODE A = vp8_above_bmi(mic, i, mis)->mode; 230 const B_PREDICTION_MODE L = vp8_left_bmi(mic, i)->mode; 231 B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode); 232 int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(d); 233 234 pick_intra4x4block(rtcd, mb, mb->block + i, xd->block + i, 235 &best_mode, A, L, &r, &d); 236 237 cost += r; 238 distortion += d; 239 240 mic->bmi[i].mode = xd->block[i].bmi.mode = best_mode; 241 242 // Break out case where we have already exceeded best so far value that was bassed in 243 if (distortion > *best_dist) 244 break; 245 } 246 247 for (i = 0; i < 16; i++) 248 xd->block[i].bmi.mv.as_int = 0; 249 250 *Rate = cost; 251 252 if (i == 16) 253 { 254 *best_dist = distortion; 255 error = RD_ESTIMATE(mb->rdmult, mb->rddiv, cost, distortion); 256 } 257 else 258 { 259 *best_dist = INT_MAX; 260 error = INT_MAX; 261 } 262 263 return error; 264} 265 266int vp8_pick_intra_mbuv_mode(MACROBLOCK *mb) 267{ 268 269 MACROBLOCKD *x = &mb->e_mbd; 270 unsigned char *uabove_row = x->dst.u_buffer - x->dst.uv_stride; 271 unsigned char *vabove_row = x->dst.v_buffer - x->dst.uv_stride; 272 unsigned char *usrc_ptr = (mb->block[16].src + *mb->block[16].base_src); 273 unsigned char *vsrc_ptr = (mb->block[20].src + *mb->block[20].base_src); 274 int uvsrc_stride = mb->block[16].src_stride; 275 unsigned char uleft_col[8]; 276 unsigned char vleft_col[8]; 277 unsigned char utop_left = uabove_row[-1]; 278 unsigned char vtop_left = vabove_row[-1]; 279 int i, j; 280 int expected_udc; 281 int expected_vdc; 282 int shift; 283 int Uaverage = 0; 284 int Vaverage = 0; 285 int diff; 286 int pred_error[4] = {0, 0, 0, 0}, best_error = INT_MAX; 287 MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode); 288 289 290 for (i = 0; i < 8; i++) 291 { 292 uleft_col[i] = x->dst.u_buffer [i* x->dst.uv_stride -1]; 293 vleft_col[i] = x->dst.v_buffer [i* x->dst.uv_stride -1]; 294 } 295 296 if (!x->up_available && !x->left_available) 297 { 298 expected_udc = 128; 299 expected_vdc = 128; 300 } 301 else 302 { 303 shift = 2; 304 305 if (x->up_available) 306 { 307 308 for (i = 0; i < 8; i++) 309 { 310 Uaverage += uabove_row[i]; 311 Vaverage += vabove_row[i]; 312 } 313 314 shift ++; 315 316 } 317 318 if (x->left_available) 319 { 320 for (i = 0; i < 8; i++) 321 { 322 Uaverage += uleft_col[i]; 323 Vaverage += vleft_col[i]; 324 } 325 326 shift ++; 327 328 } 329 330 expected_udc = (Uaverage + (1 << (shift - 1))) >> shift; 331 expected_vdc = (Vaverage + (1 << (shift - 1))) >> shift; 332 } 333 334 335 for (i = 0; i < 8; i++) 336 { 337 for (j = 0; j < 8; j++) 338 { 339 340 int predu = uleft_col[i] + uabove_row[j] - utop_left; 341 int predv = vleft_col[i] + vabove_row[j] - vtop_left; 342 int u_p, v_p; 343 344 u_p = usrc_ptr[j]; 345 v_p = vsrc_ptr[j]; 346 347 if (predu < 0) 348 predu = 0; 349 350 if (predu > 255) 351 predu = 255; 352 353 if (predv < 0) 354 predv = 0; 355 356 if (predv > 255) 357 predv = 255; 358 359 360 diff = u_p - expected_udc; 361 pred_error[DC_PRED] += diff * diff; 362 diff = v_p - expected_vdc; 363 pred_error[DC_PRED] += diff * diff; 364 365 366 diff = u_p - uabove_row[j]; 367 pred_error[V_PRED] += diff * diff; 368 diff = v_p - vabove_row[j]; 369 pred_error[V_PRED] += diff * diff; 370 371 372 diff = u_p - uleft_col[i]; 373 pred_error[H_PRED] += diff * diff; 374 diff = v_p - vleft_col[i]; 375 pred_error[H_PRED] += diff * diff; 376 377 378 diff = u_p - predu; 379 pred_error[TM_PRED] += diff * diff; 380 diff = v_p - predv; 381 pred_error[TM_PRED] += diff * diff; 382 383 384 } 385 386 usrc_ptr += uvsrc_stride; 387 vsrc_ptr += uvsrc_stride; 388 389 if (i == 3) 390 { 391 usrc_ptr = (mb->block[18].src + *mb->block[18].base_src); 392 vsrc_ptr = (mb->block[22].src + *mb->block[22].base_src); 393 } 394 395 396 397 } 398 399 400 for (i = DC_PRED; i <= TM_PRED; i++) 401 { 402 if (best_error > pred_error[i]) 403 { 404 best_error = pred_error[i]; 405 best_mode = (MB_PREDICTION_MODE)i; 406 } 407 } 408 409 410 mb->e_mbd.mode_info_context->mbmi.uv_mode = best_mode; 411 return best_error; 412 413} 414 415int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra) 416{ 417 BLOCK *b = &x->block[0]; 418 BLOCKD *d = &x->e_mbd.block[0]; 419 MACROBLOCKD *xd = &x->e_mbd; 420 B_MODE_INFO best_bmodes[16]; 421 MB_MODE_INFO best_mbmode; 422 PARTITION_INFO best_partition; 423 MV best_ref_mv; 424 MV mode_mv[MB_MODE_COUNT]; 425 MB_PREDICTION_MODE this_mode; 426 int num00; 427 int i; 428 int mdcounts[4]; 429 int best_rd = INT_MAX; // 1 << 30; 430 int best_intra_rd = INT_MAX; 431 int mode_index; 432 int ref_frame_cost[MAX_REF_FRAMES]; 433 int rate; 434 int rate2; 435 int distortion2; 436 int bestsme; 437 //int all_rds[MAX_MODES]; // Experimental debug code. 438 int best_mode_index = 0; 439 int sse = INT_MAX; 440 441 MV mvp; 442 int near_sadidx[8] = {0, 1, 2, 3, 4, 5, 6, 7}; 443 int saddone=0; 444 int sr=0; //search range got from mv_pred(). It uses step_param levels. (0-7) 445 446 MV nearest_mv[4]; 447 MV near_mv[4]; 448 MV frame_best_ref_mv[4]; 449 int MDCounts[4][4]; 450 unsigned char *y_buffer[4]; 451 unsigned char *u_buffer[4]; 452 unsigned char *v_buffer[4]; 453 454 int skip_mode[4] = {0, 0, 0, 0}; 455 456 vpx_memset(mode_mv, 0, sizeof(mode_mv)); 457 vpx_memset(nearest_mv, 0, sizeof(nearest_mv)); 458 vpx_memset(near_mv, 0, sizeof(near_mv)); 459 vpx_memset(&best_mbmode, 0, sizeof(best_mbmode)); 460 461 462 // set up all the refframe dependent pointers. 463 if (cpi->ref_frame_flags & VP8_LAST_FLAG) 464 { 465 YV12_BUFFER_CONFIG *lst_yv12 = &cpi->common.yv12_fb[cpi->common.lst_fb_idx]; 466 467 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[LAST_FRAME], &near_mv[LAST_FRAME], 468 &frame_best_ref_mv[LAST_FRAME], MDCounts[LAST_FRAME], LAST_FRAME, cpi->common.ref_frame_sign_bias); 469 470 y_buffer[LAST_FRAME] = lst_yv12->y_buffer + recon_yoffset; 471 u_buffer[LAST_FRAME] = lst_yv12->u_buffer + recon_uvoffset; 472 v_buffer[LAST_FRAME] = lst_yv12->v_buffer + recon_uvoffset; 473 } 474 else 475 skip_mode[LAST_FRAME] = 1; 476 477 if (cpi->ref_frame_flags & VP8_GOLD_FLAG) 478 { 479 YV12_BUFFER_CONFIG *gld_yv12 = &cpi->common.yv12_fb[cpi->common.gld_fb_idx]; 480 481 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[GOLDEN_FRAME], &near_mv[GOLDEN_FRAME], 482 &frame_best_ref_mv[GOLDEN_FRAME], MDCounts[GOLDEN_FRAME], GOLDEN_FRAME, cpi->common.ref_frame_sign_bias); 483 484 y_buffer[GOLDEN_FRAME] = gld_yv12->y_buffer + recon_yoffset; 485 u_buffer[GOLDEN_FRAME] = gld_yv12->u_buffer + recon_uvoffset; 486 v_buffer[GOLDEN_FRAME] = gld_yv12->v_buffer + recon_uvoffset; 487 } 488 else 489 skip_mode[GOLDEN_FRAME] = 1; 490 491 if (cpi->ref_frame_flags & VP8_ALT_FLAG && cpi->source_alt_ref_active) 492 { 493 YV12_BUFFER_CONFIG *alt_yv12 = &cpi->common.yv12_fb[cpi->common.alt_fb_idx]; 494 495 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[ALTREF_FRAME], &near_mv[ALTREF_FRAME], 496 &frame_best_ref_mv[ALTREF_FRAME], MDCounts[ALTREF_FRAME], ALTREF_FRAME, cpi->common.ref_frame_sign_bias); 497 498 y_buffer[ALTREF_FRAME] = alt_yv12->y_buffer + recon_yoffset; 499 u_buffer[ALTREF_FRAME] = alt_yv12->u_buffer + recon_uvoffset; 500 v_buffer[ALTREF_FRAME] = alt_yv12->v_buffer + recon_uvoffset; 501 } 502 else 503 skip_mode[ALTREF_FRAME] = 1; 504 505 cpi->mbs_tested_so_far++; // Count of the number of MBs tested so far this frame 506 507 *returnintra = best_intra_rd; 508 x->skip = 0; 509 510 ref_frame_cost[INTRA_FRAME] = vp8_cost_zero(cpi->prob_intra_coded); 511 512 // Special case treatment when GF and ARF are not sensible options for reference 513 if (cpi->ref_frame_flags == VP8_LAST_FLAG) 514 { 515 ref_frame_cost[LAST_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 516 + vp8_cost_zero(255); 517 ref_frame_cost[GOLDEN_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 518 + vp8_cost_one(255) 519 + vp8_cost_zero(128); 520 ref_frame_cost[ALTREF_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 521 + vp8_cost_one(255) 522 + vp8_cost_one(128); 523 } 524 else 525 { 526 ref_frame_cost[LAST_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 527 + vp8_cost_zero(cpi->prob_last_coded); 528 ref_frame_cost[GOLDEN_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 529 + vp8_cost_one(cpi->prob_last_coded) 530 + vp8_cost_zero(cpi->prob_gf_coded); 531 ref_frame_cost[ALTREF_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 532 + vp8_cost_one(cpi->prob_last_coded) 533 + vp8_cost_one(cpi->prob_gf_coded); 534 } 535 536 x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME; 537 538 // if we encode a new mv this is important 539 // find the best new motion vector 540 for (mode_index = 0; mode_index < MAX_MODES; mode_index++) 541 { 542 int frame_cost; 543 int this_rd = INT_MAX; 544 545 if (best_rd <= cpi->rd_threshes[mode_index]) 546 continue; 547 548 x->e_mbd.mode_info_context->mbmi.ref_frame = vp8_ref_frame_order[mode_index]; 549 550 if (skip_mode[x->e_mbd.mode_info_context->mbmi.ref_frame]) 551 continue; 552 553 // Check to see if the testing frequency for this mode is at its max 554 // If so then prevent it from being tested and increase the threshold for its testing 555 if (cpi->mode_test_hit_counts[mode_index] && (cpi->mode_check_freq[mode_index] > 1)) 556 { 557 //if ( (cpi->mbs_tested_so_far / cpi->mode_test_hit_counts[mode_index]) <= cpi->mode_check_freq[mode_index] ) 558 if (cpi->mbs_tested_so_far <= (cpi->mode_check_freq[mode_index] * cpi->mode_test_hit_counts[mode_index])) 559 { 560 // Increase the threshold for coding this mode to make it less likely to be chosen 561 cpi->rd_thresh_mult[mode_index] += 4; 562 563 if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT) 564 cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT; 565 566 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 567 568 continue; 569 } 570 } 571 572 // We have now reached the point where we are going to test the current mode so increment the counter for the number of times it has been tested 573 cpi->mode_test_hit_counts[mode_index] ++; 574 575 rate2 = 0; 576 distortion2 = 0; 577 578 this_mode = vp8_mode_order[mode_index]; 579 580 // Experimental debug code. 581 //all_rds[mode_index] = -1; 582 583 x->e_mbd.mode_info_context->mbmi.mode = this_mode; 584 x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED; 585 586 // Work out the cost assosciated with selecting the reference frame 587 frame_cost = ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame]; 588 rate2 += frame_cost; 589 590 // everything but intra 591 if (x->e_mbd.mode_info_context->mbmi.ref_frame) 592 { 593 x->e_mbd.pre.y_buffer = y_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 594 x->e_mbd.pre.u_buffer = u_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 595 x->e_mbd.pre.v_buffer = v_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 596 mode_mv[NEARESTMV] = nearest_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 597 mode_mv[NEARMV] = near_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 598 best_ref_mv = frame_best_ref_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 599 memcpy(mdcounts, MDCounts[x->e_mbd.mode_info_context->mbmi.ref_frame], sizeof(mdcounts)); 600 } 601 602 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame, 603 // unless ARNR filtering is enabled in which case we want 604 // an unfiltered alternative 605 if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) 606 { 607 if (this_mode != ZEROMV || x->e_mbd.mode_info_context->mbmi.ref_frame != ALTREF_FRAME) 608 continue; 609 } 610 611 if(cpi->sf.improved_mv_pred && x->e_mbd.mode_info_context->mbmi.mode == NEWMV) 612 { 613 if(!saddone) 614 { 615 vp8_cal_sad(cpi,xd,x, recon_yoffset ,&near_sadidx[0] ); 616 saddone = 1; 617 } 618 619 vp8_mv_pred(cpi, &x->e_mbd, x->e_mbd.mode_info_context, &mvp, 620 x->e_mbd.mode_info_context->mbmi.ref_frame, cpi->common.ref_frame_sign_bias, &sr, &near_sadidx[0]); 621 622 /* adjust mvp to make sure it is within MV range */ 623 if(mvp.row > best_ref_mv.row + MAX_FULL_PEL_VAL) 624 mvp.row = best_ref_mv.row + MAX_FULL_PEL_VAL; 625 else if(mvp.row < best_ref_mv.row - MAX_FULL_PEL_VAL) 626 mvp.row = best_ref_mv.row - MAX_FULL_PEL_VAL; 627 if(mvp.col > best_ref_mv.col + MAX_FULL_PEL_VAL) 628 mvp.col = best_ref_mv.col + MAX_FULL_PEL_VAL; 629 else if(mvp.col < best_ref_mv.col - MAX_FULL_PEL_VAL) 630 mvp.col = best_ref_mv.col - MAX_FULL_PEL_VAL; 631 } 632 633 switch (this_mode) 634 { 635 case B_PRED: 636 distortion2 = *returndistortion; // Best so far passed in as breakout value to vp8_pick_intra4x4mby_modes 637 vp8_pick_intra4x4mby_modes(IF_RTCD(&cpi->rtcd), x, &rate, &distortion2); 638 rate2 += rate; 639 distortion2 = VARIANCE_INVOKE(&cpi->rtcd.variance, get16x16prederror)(x->src.y_buffer, x->src.y_stride, x->e_mbd.predictor, 16, 0x7fffffff); 640 641 if (distortion2 == INT_MAX) 642 { 643 this_rd = INT_MAX; 644 } 645 else 646 { 647 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 648 649 if (this_rd < best_intra_rd) 650 { 651 best_intra_rd = this_rd; 652 *returnintra = best_intra_rd ; 653 } 654 } 655 656 break; 657 658 case SPLITMV: 659 660 // Split MV modes currently not supported when RD is nopt enabled. 661 break; 662 663 case DC_PRED: 664 case V_PRED: 665 case H_PRED: 666 case TM_PRED: 667 RECON_INVOKE(&cpi->common.rtcd.recon, build_intra_predictors_mby) 668 (&x->e_mbd); 669 distortion2 = VARIANCE_INVOKE(&cpi->rtcd.variance, get16x16prederror)(x->src.y_buffer, x->src.y_stride, x->e_mbd.predictor, 16, 0x7fffffff); 670 rate2 += x->mbmode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.mode]; 671 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 672 673 if (this_rd < best_intra_rd) 674 { 675 best_intra_rd = this_rd; 676 *returnintra = best_intra_rd ; 677 } 678 679 break; 680 681 case NEWMV: 682 { 683 int thissme; 684 int step_param; 685 int further_steps; 686 int n = 0; 687 int sadpb = x->sadperbit16; 688 689 int col_min; 690 int col_max; 691 int row_min; 692 int row_max; 693 694 int tmp_col_min = x->mv_col_min; 695 int tmp_col_max = x->mv_col_max; 696 int tmp_row_min = x->mv_row_min; 697 int tmp_row_max = x->mv_row_max; 698 699 int speed_adjust = (cpi->Speed > 5) ? ((cpi->Speed >= 8)? 3 : 2) : 1; 700 701 // Further step/diamond searches as necessary 702 step_param = cpi->sf.first_step + speed_adjust; 703 704 if(cpi->sf.improved_mv_pred) 705 { 706 sr += speed_adjust; 707 //adjust search range according to sr from mv prediction 708 if(sr > step_param) 709 step_param = sr; 710 711 col_min = (best_ref_mv.col - MAX_FULL_PEL_VAL) >>3; 712 col_max = (best_ref_mv.col + MAX_FULL_PEL_VAL) >>3; 713 row_min = (best_ref_mv.row - MAX_FULL_PEL_VAL) >>3; 714 row_max = (best_ref_mv.row + MAX_FULL_PEL_VAL) >>3; 715 716 // Get intersection of UMV window and valid MV window to reduce # of checks in diamond search. 717 if (x->mv_col_min < col_min ) 718 x->mv_col_min = col_min; 719 if (x->mv_col_max > col_max ) 720 x->mv_col_max = col_max; 721 if (x->mv_row_min < row_min ) 722 x->mv_row_min = row_min; 723 if (x->mv_row_max > row_max ) 724 x->mv_row_max = row_max; 725 }else 726 { 727 mvp.row = best_ref_mv.row; 728 mvp.col = best_ref_mv.col; 729 } 730 731 further_steps = (cpi->Speed >= 8)? 0: (cpi->sf.max_step_search_steps - 1 - step_param); 732 733 if (cpi->sf.search_method == HEX) 734 { 735 bestsme = vp8_hex_search(x, b, d, &mvp, &d->bmi.mv.as_mv, step_param, sadpb/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv); 736 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 737 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 738 } 739 else 740 { 741 bestsme = cpi->diamond_search_sad(x, b, d, &mvp, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv); //sadpb < 9 742 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 743 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 744 745 // Further step/diamond searches as necessary 746 n = 0; 747 //further_steps = (cpi->sf.max_step_search_steps - 1) - step_param; 748 749 n = num00; 750 num00 = 0; 751 752 while (n < further_steps) 753 { 754 n++; 755 756 if (num00) 757 num00--; 758 else 759 { 760 thissme = cpi->diamond_search_sad(x, b, d, &mvp, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv); //sadpb = 9 761 762 if (thissme < bestsme) 763 { 764 bestsme = thissme; 765 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 766 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 767 } 768 else 769 { 770 d->bmi.mv.as_mv.row = mode_mv[NEWMV].row; 771 d->bmi.mv.as_mv.col = mode_mv[NEWMV].col; 772 } 773 } 774 } 775 } 776 777 if(cpi->sf.improved_mv_pred) 778 { 779 x->mv_col_min = tmp_col_min; 780 x->mv_col_max = tmp_col_max; 781 x->mv_row_min = tmp_row_min; 782 x->mv_row_max = tmp_row_max; 783 } 784 785 if (bestsme < INT_MAX) 786 cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv.as_mv, &best_ref_mv, x->errorperbit, &cpi->fn_ptr[BLOCK_16X16], cpi->mb.mvcost); 787 788 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 789 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 790 791 // mv cost; 792 rate2 += vp8_mv_bit_cost(&mode_mv[NEWMV], &best_ref_mv, cpi->mb.mvcost, 128); 793 } 794 795 case NEARESTMV: 796 case NEARMV: 797 798 if (mode_mv[this_mode].row == 0 && mode_mv[this_mode].col == 0) 799 continue; 800 801 case ZEROMV: 802 803 // Trap vectors that reach beyond the UMV borders 804 // Note that ALL New MV, Nearest MV Near MV and Zero MV code drops through to this point 805 // because of the lack of break statements in the previous two cases. 806 if (((mode_mv[this_mode].row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].row >> 3) > x->mv_row_max) || 807 ((mode_mv[this_mode].col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].col >> 3) > x->mv_col_max)) 808 continue; 809 810 rate2 += vp8_cost_mv_ref(this_mode, mdcounts); 811 x->e_mbd.mode_info_context->mbmi.mode = this_mode; 812 x->e_mbd.mode_info_context->mbmi.mv.as_mv = mode_mv[this_mode]; 813 x->e_mbd.block[0].bmi.mode = this_mode; 814 x->e_mbd.block[0].bmi.mv.as_int = x->e_mbd.mode_info_context->mbmi.mv.as_int; 815 816 distortion2 = get_inter_mbpred_error(x, &cpi->fn_ptr[BLOCK_16X16], (unsigned int *)(&sse)); 817 818 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 819 820 if (cpi->active_map_enabled && x->active_ptr[0] == 0) 821 { 822 x->skip = 1; 823 } 824 else if (sse < x->encode_breakout) 825 { 826 // Check u and v to make sure skip is ok 827 int sse2 = 0; 828 829 sse2 = VP8_UVSSE(x, IF_RTCD(&cpi->rtcd.variance)); 830 831 if (sse2 * 2 < x->encode_breakout) 832 x->skip = 1; 833 else 834 x->skip = 0; 835 } 836 837 break; 838 default: 839 break; 840 } 841 842 // Experimental debug code. 843 //all_rds[mode_index] = this_rd; 844 845 if (this_rd < best_rd || x->skip) 846 { 847 // Note index of best mode 848 best_mode_index = mode_index; 849 850 *returnrate = rate2; 851 *returndistortion = distortion2; 852 best_rd = this_rd; 853 vpx_memcpy(&best_mbmode, &x->e_mbd.mode_info_context->mbmi, sizeof(MB_MODE_INFO)); 854 vpx_memcpy(&best_partition, x->partition_info, sizeof(PARTITION_INFO)); 855 856 if (this_mode == B_PRED || this_mode == SPLITMV) 857 for (i = 0; i < 16; i++) 858 { 859 vpx_memcpy(&best_bmodes[i], &x->e_mbd.block[i].bmi, sizeof(B_MODE_INFO)); 860 } 861 else 862 { 863 best_bmodes[0].mv = x->e_mbd.block[0].bmi.mv; 864 } 865 866 // Testing this mode gave rise to an improvement in best error score. Lower threshold a bit for next time 867 cpi->rd_thresh_mult[mode_index] = (cpi->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ? cpi->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT; 868 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 869 } 870 871 // If the mode did not help improve the best error case then raise the threshold for testing that mode next time around. 872 else 873 { 874 cpi->rd_thresh_mult[mode_index] += 4; 875 876 if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT) 877 cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT; 878 879 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 880 } 881 882 if (x->skip) 883 break; 884 } 885 886 // Reduce the activation RD thresholds for the best choice mode 887 if ((cpi->rd_baseline_thresh[best_mode_index] > 0) && (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2))) 888 { 889 int best_adjustment = (cpi->rd_thresh_mult[best_mode_index] >> 3); 890 891 cpi->rd_thresh_mult[best_mode_index] = (cpi->rd_thresh_mult[best_mode_index] >= (MIN_THRESHMULT + best_adjustment)) ? cpi->rd_thresh_mult[best_mode_index] - best_adjustment : MIN_THRESHMULT; 892 cpi->rd_threshes[best_mode_index] = (cpi->rd_baseline_thresh[best_mode_index] >> 7) * cpi->rd_thresh_mult[best_mode_index]; 893 } 894 895 // Keep a record of best mode index for use in next loop 896 cpi->last_best_mode_index = best_mode_index; 897 898 if (best_mbmode.mode <= B_PRED) 899 { 900 x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME; 901 vp8_pick_intra_mbuv_mode(x); 902 best_mbmode.uv_mode = x->e_mbd.mode_info_context->mbmi.uv_mode; 903 } 904 905 906 { 907 int this_rdbin = (*returndistortion >> 7); 908 909 if (this_rdbin >= 1024) 910 { 911 this_rdbin = 1023; 912 } 913 914 cpi->error_bins[this_rdbin] ++; 915 } 916 917 918 if (cpi->is_src_frame_alt_ref && (best_mbmode.mode != ZEROMV || best_mbmode.ref_frame != ALTREF_FRAME)) 919 { 920 best_mbmode.mode = ZEROMV; 921 best_mbmode.ref_frame = ALTREF_FRAME; 922 best_mbmode.mv.as_int = 0; 923 best_mbmode.uv_mode = 0; 924 best_mbmode.mb_skip_coeff = (cpi->common.mb_no_coeff_skip) ? 1 : 0; 925 best_mbmode.partitioning = 0; 926 best_mbmode.dc_diff = 0; 927 928 vpx_memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode, sizeof(MB_MODE_INFO)); 929 vpx_memcpy(x->partition_info, &best_partition, sizeof(PARTITION_INFO)); 930 931 for (i = 0; i < 16; i++) 932 { 933 vpx_memset(&x->e_mbd.block[i].bmi, 0, sizeof(B_MODE_INFO)); 934 } 935 936 x->e_mbd.mode_info_context->mbmi.mv.as_int = 0; 937 938 return best_rd; 939 } 940 941 942 // macroblock modes 943 vpx_memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode, sizeof(MB_MODE_INFO)); 944 vpx_memcpy(x->partition_info, &best_partition, sizeof(PARTITION_INFO)); 945 946 if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED || x->e_mbd.mode_info_context->mbmi.mode == SPLITMV) 947 for (i = 0; i < 16; i++) 948 { 949 vpx_memcpy(&x->e_mbd.block[i].bmi, &best_bmodes[i], sizeof(B_MODE_INFO)); 950 951 } 952 else 953 { 954 vp8_set_mbmode_and_mvs(x, x->e_mbd.mode_info_context->mbmi.mode, &best_bmodes[0].mv.as_mv); 955 } 956 957 x->e_mbd.mode_info_context->mbmi.mv.as_mv = x->e_mbd.block[15].bmi.mv.as_mv; 958 959 return best_rd; 960} 961