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 "entropymode.h" 18 #include "pickinter.h" 19 #include "findnearmv.h" 20 #include "encodemb.h" 21 #include "reconinter.h" 22 #include "reconintra.h" 23 #include "reconintra4x4.h" 24 #include "g_common.h" 25 #include "variance.h" 26 #include "mcomp.h" 27 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 36 extern int VP8_UVSSE(MACROBLOCK *x, const vp8_variance_rtcd_vtable_t *rtcd); 37 38 #ifdef SPEEDSTATS 39 extern unsigned int cnt_pm; 40 #endif 41 42 extern const MV_REFERENCE_FRAME vp8_ref_frame_order[MAX_MODES]; 43 extern const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES]; 44 45 46 extern unsigned int (*vp8_get16x16pred_error)(unsigned char *src_ptr, int src_stride, unsigned char *ref_ptr, int ref_stride); 47 extern unsigned int (*vp8_get4x4sse_cs)(unsigned char *src_ptr, int source_stride, unsigned char *ref_ptr, int recon_stride); 48 extern 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); 49 extern int vp8_cost_mv_ref(MB_PREDICTION_MODE m, const int near_mv_ref_ct[4]); 50 extern void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, MV *mv); 51 52 53 int 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 67 static 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 92 unsigned 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 } 123 124 125 unsigned 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 } 151 152 static 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 } 162 163 static 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 ENTROPY_CONTEXT *a, 172 ENTROPY_CONTEXT *l, 173 174 int *bestrate, 175 int *bestdistortion) 176 { 177 B_PREDICTION_MODE mode; 178 int best_rd = INT_MAX; // 1<<30 179 int rate; 180 int distortion; 181 unsigned int *mode_costs; 182 (void) l; 183 (void) a; 184 185 if (x->e_mbd.frame_type == KEY_FRAME) 186 { 187 mode_costs = x->bmode_costs[above][left]; 188 } 189 else 190 { 191 mode_costs = x->inter_bmode_costs; 192 } 193 194 for (mode = B_DC_PRED; mode <= B_HE_PRED /*B_HU_PRED*/; mode++) 195 { 196 int this_rd; 197 198 rate = mode_costs[mode]; 199 vp8_predict_intra4x4(b, mode, b->predictor); 200 distortion = get_prediction_error(be, b, &rtcd->variance); 201 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate, distortion); 202 203 if (this_rd < best_rd) 204 { 205 *bestrate = rate; 206 *bestdistortion = distortion; 207 best_rd = this_rd; 208 *best_mode = mode; 209 } 210 } 211 212 b->bmi.mode = (B_PREDICTION_MODE)(*best_mode); 213 vp8_encode_intra4x4block(rtcd, x, be, b, b->bmi.mode); 214 return best_rd; 215 } 216 217 218 int vp8_pick_intra4x4mby_modes(const VP8_ENCODER_RTCD *rtcd, MACROBLOCK *mb, int *Rate, int *best_dist) 219 { 220 MACROBLOCKD *const xd = &mb->e_mbd; 221 int i; 222 int cost = mb->mbmode_cost [xd->frame_type] [B_PRED]; 223 int error = RD_ESTIMATE(mb->rdmult, mb->rddiv, cost, 0); // Rd estimate for the cost of the block prediction mode 224 int distortion = 0; 225 ENTROPY_CONTEXT_PLANES t_above, t_left; 226 ENTROPY_CONTEXT *ta; 227 ENTROPY_CONTEXT *tl; 228 229 vpx_memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES)); 230 vpx_memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES)); 231 232 ta = (ENTROPY_CONTEXT *)&t_above; 233 tl = (ENTROPY_CONTEXT *)&t_left; 234 235 vp8_intra_prediction_down_copy(xd); 236 237 for (i = 0; i < 16; i++) 238 { 239 MODE_INFO *const mic = xd->mode_info_context; 240 const int mis = xd->mode_info_stride; 241 const B_PREDICTION_MODE A = vp8_above_bmi(mic, i, mis)->mode; 242 const B_PREDICTION_MODE L = vp8_left_bmi(mic, i)->mode; 243 B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode); 244 int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(d); 245 246 error += pick_intra4x4block(rtcd, 247 mb, mb->block + i, xd->block + i, &best_mode, A, L, 248 ta + vp8_block2above[i], 249 tl + vp8_block2left[i], &r, &d); 250 251 cost += r; 252 distortion += d; 253 254 mic->bmi[i].mode = xd->block[i].bmi.mode = best_mode; 255 256 // Break out case where we have already exceeded best so far value that was bassed in 257 if (distortion > *best_dist) 258 break; 259 } 260 261 for (i = 0; i < 16; i++) 262 xd->block[i].bmi.mv.as_int = 0; 263 264 *Rate = cost; 265 266 if (i == 16) 267 *best_dist = distortion; 268 else 269 *best_dist = INT_MAX; 270 271 272 return error; 273 } 274 275 int vp8_pick_intra_mbuv_mode(MACROBLOCK *mb) 276 { 277 278 MACROBLOCKD *x = &mb->e_mbd; 279 unsigned char *uabove_row = x->dst.u_buffer - x->dst.uv_stride; 280 unsigned char *vabove_row = x->dst.v_buffer - x->dst.uv_stride; 281 unsigned char *usrc_ptr = (mb->block[16].src + *mb->block[16].base_src); 282 unsigned char *vsrc_ptr = (mb->block[20].src + *mb->block[20].base_src); 283 int uvsrc_stride = mb->block[16].src_stride; 284 unsigned char uleft_col[8]; 285 unsigned char vleft_col[8]; 286 unsigned char utop_left = uabove_row[-1]; 287 unsigned char vtop_left = vabove_row[-1]; 288 int i, j; 289 int expected_udc; 290 int expected_vdc; 291 int shift; 292 int Uaverage = 0; 293 int Vaverage = 0; 294 int diff; 295 int pred_error[4] = {0, 0, 0, 0}, best_error = INT_MAX; 296 MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode); 297 298 299 for (i = 0; i < 8; i++) 300 { 301 uleft_col[i] = x->dst.u_buffer [i* x->dst.uv_stride -1]; 302 vleft_col[i] = x->dst.v_buffer [i* x->dst.uv_stride -1]; 303 } 304 305 if (!x->up_available && !x->left_available) 306 { 307 expected_udc = 128; 308 expected_vdc = 128; 309 } 310 else 311 { 312 shift = 2; 313 314 if (x->up_available) 315 { 316 317 for (i = 0; i < 8; i++) 318 { 319 Uaverage += uabove_row[i]; 320 Vaverage += vabove_row[i]; 321 } 322 323 shift ++; 324 325 } 326 327 if (x->left_available) 328 { 329 for (i = 0; i < 8; i++) 330 { 331 Uaverage += uleft_col[i]; 332 Vaverage += vleft_col[i]; 333 } 334 335 shift ++; 336 337 } 338 339 expected_udc = (Uaverage + (1 << (shift - 1))) >> shift; 340 expected_vdc = (Vaverage + (1 << (shift - 1))) >> shift; 341 } 342 343 344 for (i = 0; i < 8; i++) 345 { 346 for (j = 0; j < 8; j++) 347 { 348 349 int predu = uleft_col[i] + uabove_row[j] - utop_left; 350 int predv = vleft_col[i] + vabove_row[j] - vtop_left; 351 int u_p, v_p; 352 353 u_p = usrc_ptr[j]; 354 v_p = vsrc_ptr[j]; 355 356 if (predu < 0) 357 predu = 0; 358 359 if (predu > 255) 360 predu = 255; 361 362 if (predv < 0) 363 predv = 0; 364 365 if (predv > 255) 366 predv = 255; 367 368 369 diff = u_p - expected_udc; 370 pred_error[DC_PRED] += diff * diff; 371 diff = v_p - expected_vdc; 372 pred_error[DC_PRED] += diff * diff; 373 374 375 diff = u_p - uabove_row[j]; 376 pred_error[V_PRED] += diff * diff; 377 diff = v_p - vabove_row[j]; 378 pred_error[V_PRED] += diff * diff; 379 380 381 diff = u_p - uleft_col[i]; 382 pred_error[H_PRED] += diff * diff; 383 diff = v_p - vleft_col[i]; 384 pred_error[H_PRED] += diff * diff; 385 386 387 diff = u_p - predu; 388 pred_error[TM_PRED] += diff * diff; 389 diff = v_p - predv; 390 pred_error[TM_PRED] += diff * diff; 391 392 393 } 394 395 usrc_ptr += uvsrc_stride; 396 vsrc_ptr += uvsrc_stride; 397 398 if (i == 3) 399 { 400 usrc_ptr = (mb->block[18].src + *mb->block[18].base_src); 401 vsrc_ptr = (mb->block[22].src + *mb->block[22].base_src); 402 } 403 404 405 406 } 407 408 409 for (i = DC_PRED; i <= TM_PRED; i++) 410 { 411 if (best_error > pred_error[i]) 412 { 413 best_error = pred_error[i]; 414 best_mode = (MB_PREDICTION_MODE)i; 415 } 416 } 417 418 419 mb->e_mbd.mode_info_context->mbmi.uv_mode = best_mode; 420 return best_error; 421 422 } 423 424 425 int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra) 426 { 427 BLOCK *b = &x->block[0]; 428 BLOCKD *d = &x->e_mbd.block[0]; 429 MACROBLOCKD *xd = &x->e_mbd; 430 B_MODE_INFO best_bmodes[16]; 431 MB_MODE_INFO best_mbmode; 432 PARTITION_INFO best_partition; 433 MV best_ref_mv1; 434 MV mode_mv[MB_MODE_COUNT]; 435 MB_PREDICTION_MODE this_mode; 436 int num00; 437 int i; 438 int mdcounts[4]; 439 int best_rd = INT_MAX; // 1 << 30; 440 int best_intra_rd = INT_MAX; 441 int mode_index; 442 int ref_frame_cost[MAX_REF_FRAMES]; 443 int rate; 444 int rate2; 445 int distortion2; 446 int bestsme; 447 //int all_rds[MAX_MODES]; // Experimental debug code. 448 int best_mode_index = 0; 449 int sse = INT_MAX; 450 451 MV nearest_mv[4]; 452 MV near_mv[4]; 453 MV best_ref_mv[4]; 454 int MDCounts[4][4]; 455 unsigned char *y_buffer[4]; 456 unsigned char *u_buffer[4]; 457 unsigned char *v_buffer[4]; 458 459 int skip_mode[4] = {0, 0, 0, 0}; 460 461 vpx_memset(mode_mv, 0, sizeof(mode_mv)); 462 vpx_memset(nearest_mv, 0, sizeof(nearest_mv)); 463 vpx_memset(near_mv, 0, sizeof(near_mv)); 464 vpx_memset(&best_mbmode, 0, sizeof(best_mbmode)); 465 466 467 // set up all the refframe dependent pointers. 468 if (cpi->ref_frame_flags & VP8_LAST_FLAG) 469 { 470 YV12_BUFFER_CONFIG *lst_yv12 = &cpi->common.yv12_fb[cpi->common.lst_fb_idx]; 471 472 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[LAST_FRAME], &near_mv[LAST_FRAME], 473 &best_ref_mv[LAST_FRAME], MDCounts[LAST_FRAME], LAST_FRAME, cpi->common.ref_frame_sign_bias); 474 475 y_buffer[LAST_FRAME] = lst_yv12->y_buffer + recon_yoffset; 476 u_buffer[LAST_FRAME] = lst_yv12->u_buffer + recon_uvoffset; 477 v_buffer[LAST_FRAME] = lst_yv12->v_buffer + recon_uvoffset; 478 } 479 else 480 skip_mode[LAST_FRAME] = 1; 481 482 if (cpi->ref_frame_flags & VP8_GOLD_FLAG) 483 { 484 YV12_BUFFER_CONFIG *gld_yv12 = &cpi->common.yv12_fb[cpi->common.gld_fb_idx]; 485 486 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[GOLDEN_FRAME], &near_mv[GOLDEN_FRAME], 487 &best_ref_mv[GOLDEN_FRAME], MDCounts[GOLDEN_FRAME], GOLDEN_FRAME, cpi->common.ref_frame_sign_bias); 488 489 y_buffer[GOLDEN_FRAME] = gld_yv12->y_buffer + recon_yoffset; 490 u_buffer[GOLDEN_FRAME] = gld_yv12->u_buffer + recon_uvoffset; 491 v_buffer[GOLDEN_FRAME] = gld_yv12->v_buffer + recon_uvoffset; 492 } 493 else 494 skip_mode[GOLDEN_FRAME] = 1; 495 496 if (cpi->ref_frame_flags & VP8_ALT_FLAG && cpi->source_alt_ref_active) 497 { 498 YV12_BUFFER_CONFIG *alt_yv12 = &cpi->common.yv12_fb[cpi->common.alt_fb_idx]; 499 500 vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context, &nearest_mv[ALTREF_FRAME], &near_mv[ALTREF_FRAME], 501 &best_ref_mv[ALTREF_FRAME], MDCounts[ALTREF_FRAME], ALTREF_FRAME, cpi->common.ref_frame_sign_bias); 502 503 y_buffer[ALTREF_FRAME] = alt_yv12->y_buffer + recon_yoffset; 504 u_buffer[ALTREF_FRAME] = alt_yv12->u_buffer + recon_uvoffset; 505 v_buffer[ALTREF_FRAME] = alt_yv12->v_buffer + recon_uvoffset; 506 } 507 else 508 skip_mode[ALTREF_FRAME] = 1; 509 510 cpi->mbs_tested_so_far++; // Count of the number of MBs tested so far this frame 511 512 *returnintra = best_intra_rd; 513 x->skip = 0; 514 515 ref_frame_cost[INTRA_FRAME] = vp8_cost_zero(cpi->prob_intra_coded); 516 517 // Special case treatment when GF and ARF are not sensible options for reference 518 if (cpi->ref_frame_flags == VP8_LAST_FLAG) 519 { 520 ref_frame_cost[LAST_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 521 + vp8_cost_zero(255); 522 ref_frame_cost[GOLDEN_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 523 + vp8_cost_one(255) 524 + vp8_cost_zero(128); 525 ref_frame_cost[ALTREF_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 526 + vp8_cost_one(255) 527 + vp8_cost_one(128); 528 } 529 else 530 { 531 ref_frame_cost[LAST_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 532 + vp8_cost_zero(cpi->prob_last_coded); 533 ref_frame_cost[GOLDEN_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 534 + vp8_cost_one(cpi->prob_last_coded) 535 + vp8_cost_zero(cpi->prob_gf_coded); 536 ref_frame_cost[ALTREF_FRAME] = vp8_cost_one(cpi->prob_intra_coded) 537 + vp8_cost_one(cpi->prob_last_coded) 538 + vp8_cost_one(cpi->prob_gf_coded); 539 } 540 541 542 543 best_rd = INT_MAX; 544 545 x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME; 546 547 // if we encode a new mv this is important 548 // find the best new motion vector 549 for (mode_index = 0; mode_index < MAX_MODES; mode_index++) 550 { 551 int frame_cost; 552 int this_rd = INT_MAX; 553 554 if (best_rd <= cpi->rd_threshes[mode_index]) 555 continue; 556 557 x->e_mbd.mode_info_context->mbmi.ref_frame = vp8_ref_frame_order[mode_index]; 558 559 if (skip_mode[x->e_mbd.mode_info_context->mbmi.ref_frame]) 560 continue; 561 562 // Check to see if the testing frequency for this mode is at its max 563 // If so then prevent it from being tested and increase the threshold for its testing 564 if (cpi->mode_test_hit_counts[mode_index] && (cpi->mode_check_freq[mode_index] > 1)) 565 { 566 //if ( (cpi->mbs_tested_so_far / cpi->mode_test_hit_counts[mode_index]) <= cpi->mode_check_freq[mode_index] ) 567 if (cpi->mbs_tested_so_far <= (cpi->mode_check_freq[mode_index] * cpi->mode_test_hit_counts[mode_index])) 568 { 569 // Increase the threshold for coding this mode to make it less likely to be chosen 570 cpi->rd_thresh_mult[mode_index] += 4; 571 572 if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT) 573 cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT; 574 575 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 576 577 continue; 578 } 579 } 580 581 // 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 582 cpi->mode_test_hit_counts[mode_index] ++; 583 584 rate2 = 0; 585 distortion2 = 0; 586 587 this_mode = vp8_mode_order[mode_index]; 588 589 // Experimental debug code. 590 //all_rds[mode_index] = -1; 591 592 x->e_mbd.mode_info_context->mbmi.mode = this_mode; 593 x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED; 594 595 // Work out the cost assosciated with selecting the reference frame 596 frame_cost = ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame]; 597 rate2 += frame_cost; 598 599 // everything but intra 600 if (x->e_mbd.mode_info_context->mbmi.ref_frame) 601 { 602 x->e_mbd.pre.y_buffer = y_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 603 x->e_mbd.pre.u_buffer = u_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 604 x->e_mbd.pre.v_buffer = v_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame]; 605 mode_mv[NEARESTMV] = nearest_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 606 mode_mv[NEARMV] = near_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 607 best_ref_mv1 = best_ref_mv[x->e_mbd.mode_info_context->mbmi.ref_frame]; 608 memcpy(mdcounts, MDCounts[x->e_mbd.mode_info_context->mbmi.ref_frame], sizeof(mdcounts)); 609 } 610 611 //Only consider ZEROMV/ALTREF_FRAME for alt ref frame. 612 if (cpi->is_src_frame_alt_ref) 613 { 614 if (this_mode != ZEROMV || x->e_mbd.mode_info_context->mbmi.ref_frame != ALTREF_FRAME) 615 continue; 616 } 617 618 switch (this_mode) 619 { 620 case B_PRED: 621 distortion2 = *returndistortion; // Best so far passed in as breakout value to vp8_pick_intra4x4mby_modes 622 vp8_pick_intra4x4mby_modes(IF_RTCD(&cpi->rtcd), x, &rate, &distortion2); 623 rate2 += rate; 624 distortion2 = VARIANCE_INVOKE(&cpi->rtcd.variance, get16x16prederror)(x->src.y_buffer, x->src.y_stride, x->e_mbd.predictor, 16, 0x7fffffff); 625 626 if (distortion2 == INT_MAX) 627 { 628 this_rd = INT_MAX; 629 } 630 else 631 { 632 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 633 634 if (this_rd < best_intra_rd) 635 { 636 best_intra_rd = this_rd; 637 *returnintra = best_intra_rd ; 638 } 639 } 640 641 break; 642 643 case SPLITMV: 644 645 // Split MV modes currently not supported when RD is nopt enabled. 646 break; 647 648 case DC_PRED: 649 case V_PRED: 650 case H_PRED: 651 case TM_PRED: 652 vp8_build_intra_predictors_mby_ptr(&x->e_mbd); 653 distortion2 = VARIANCE_INVOKE(&cpi->rtcd.variance, get16x16prederror)(x->src.y_buffer, x->src.y_stride, x->e_mbd.predictor, 16, 0x7fffffff); 654 rate2 += x->mbmode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.mode]; 655 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 656 657 if (this_rd < best_intra_rd) 658 { 659 best_intra_rd = this_rd; 660 *returnintra = best_intra_rd ; 661 } 662 663 break; 664 665 case NEWMV: 666 { 667 int thissme; 668 int step_param; 669 int further_steps; 670 int n = 0; 671 int sadpb = x->sadperbit16; 672 673 // Further step/diamond searches as necessary 674 if (cpi->Speed < 8) 675 { 676 step_param = cpi->sf.first_step + ((cpi->Speed > 5) ? 1 : 0); 677 further_steps = (cpi->sf.max_step_search_steps - 1) - step_param; 678 } 679 else 680 { 681 step_param = cpi->sf.first_step + 2; 682 further_steps = 0; 683 } 684 685 #if 0 686 687 // Initial step Search 688 bestsme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, cpi->mb.mvcost); 689 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 690 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 691 692 // Further step searches 693 while (n < further_steps) 694 { 695 n++; 696 697 if (num00) 698 num00--; 699 else 700 { 701 thissme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, x->mvcost); 702 703 if (thissme < bestsme) 704 { 705 bestsme = thissme; 706 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 707 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 708 } 709 else 710 { 711 d->bmi.mv.as_mv.row = mode_mv[NEWMV].row; 712 d->bmi.mv.as_mv.col = mode_mv[NEWMV].col; 713 } 714 } 715 } 716 717 #else 718 719 if (cpi->sf.search_method == HEX) 720 { 721 bestsme = vp8_hex_search(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, sadpb/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); 722 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 723 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 724 } 725 else 726 { 727 bestsme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb < 9 728 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 729 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 730 731 // Further step/diamond searches as necessary 732 n = 0; 733 //further_steps = (cpi->sf.max_step_search_steps - 1) - step_param; 734 735 n = num00; 736 num00 = 0; 737 738 while (n < further_steps) 739 { 740 n++; 741 742 if (num00) 743 num00--; 744 else 745 { 746 thissme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb = 9 747 748 if (thissme < bestsme) 749 { 750 bestsme = thissme; 751 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 752 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 753 } 754 else 755 { 756 d->bmi.mv.as_mv.row = mode_mv[NEWMV].row; 757 d->bmi.mv.as_mv.col = mode_mv[NEWMV].col; 758 } 759 } 760 } 761 } 762 763 #endif 764 } 765 766 if (bestsme < INT_MAX) 767 cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv.as_mv, &best_ref_mv1, x->errorperbit, &cpi->fn_ptr[BLOCK_16X16], cpi->mb.mvcost); 768 769 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row; 770 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col; 771 772 // mv cost; 773 rate2 += vp8_mv_bit_cost(&mode_mv[NEWMV], &best_ref_mv1, cpi->mb.mvcost, 128); 774 775 776 case NEARESTMV: 777 case NEARMV: 778 779 if (mode_mv[this_mode].row == 0 && mode_mv[this_mode].col == 0) 780 continue; 781 782 case ZEROMV: 783 784 // Trap vectors that reach beyond the UMV borders 785 // Note that ALL New MV, Nearest MV Near MV and Zero MV code drops through to this point 786 // because of the lack of break statements in the previous two cases. 787 if (((mode_mv[this_mode].row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].row >> 3) > x->mv_row_max) || 788 ((mode_mv[this_mode].col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].col >> 3) > x->mv_col_max)) 789 continue; 790 791 rate2 += vp8_cost_mv_ref(this_mode, mdcounts); 792 x->e_mbd.mode_info_context->mbmi.mode = this_mode; 793 x->e_mbd.mode_info_context->mbmi.mv.as_mv = mode_mv[this_mode]; 794 x->e_mbd.block[0].bmi.mode = this_mode; 795 x->e_mbd.block[0].bmi.mv.as_int = x->e_mbd.mode_info_context->mbmi.mv.as_int; 796 797 distortion2 = get_inter_mbpred_error(x, &cpi->fn_ptr[BLOCK_16X16], (unsigned int *)(&sse)); 798 799 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2); 800 801 if (cpi->active_map_enabled && x->active_ptr[0] == 0) 802 { 803 x->skip = 1; 804 } 805 else if (sse < x->encode_breakout) 806 { 807 // Check u and v to make sure skip is ok 808 int sse2 = 0; 809 810 sse2 = VP8_UVSSE(x, IF_RTCD(&cpi->rtcd.variance)); 811 812 if (sse2 * 2 < x->encode_breakout) 813 x->skip = 1; 814 else 815 x->skip = 0; 816 } 817 818 break; 819 default: 820 break; 821 } 822 823 // Experimental debug code. 824 //all_rds[mode_index] = this_rd; 825 826 if (this_rd < best_rd || x->skip) 827 { 828 // Note index of best mode 829 best_mode_index = mode_index; 830 831 *returnrate = rate2; 832 *returndistortion = distortion2; 833 best_rd = this_rd; 834 vpx_memcpy(&best_mbmode, &x->e_mbd.mode_info_context->mbmi, sizeof(MB_MODE_INFO)); 835 vpx_memcpy(&best_partition, x->partition_info, sizeof(PARTITION_INFO)); 836 837 if (this_mode == B_PRED || this_mode == SPLITMV) 838 for (i = 0; i < 16; i++) 839 { 840 vpx_memcpy(&best_bmodes[i], &x->e_mbd.block[i].bmi, sizeof(B_MODE_INFO)); 841 } 842 else 843 { 844 best_bmodes[0].mv = x->e_mbd.block[0].bmi.mv; 845 } 846 847 // Testing this mode gave rise to an improvement in best error score. Lower threshold a bit for next time 848 cpi->rd_thresh_mult[mode_index] = (cpi->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ? cpi->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT; 849 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 850 } 851 852 // If the mode did not help improve the best error case then raise the threshold for testing that mode next time around. 853 else 854 { 855 cpi->rd_thresh_mult[mode_index] += 4; 856 857 if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT) 858 cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT; 859 860 cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index]; 861 } 862 863 if (x->skip) 864 break; 865 } 866 867 // Reduce the activation RD thresholds for the best choice mode 868 if ((cpi->rd_baseline_thresh[best_mode_index] > 0) && (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2))) 869 { 870 int best_adjustment = (cpi->rd_thresh_mult[best_mode_index] >> 3); 871 872 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; 873 cpi->rd_threshes[best_mode_index] = (cpi->rd_baseline_thresh[best_mode_index] >> 7) * cpi->rd_thresh_mult[best_mode_index]; 874 } 875 876 // Keep a record of best mode index for use in next loop 877 cpi->last_best_mode_index = best_mode_index; 878 879 if (best_mbmode.mode <= B_PRED) 880 { 881 x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME; 882 vp8_pick_intra_mbuv_mode(x); 883 best_mbmode.uv_mode = x->e_mbd.mode_info_context->mbmi.uv_mode; 884 } 885 886 887 { 888 int this_rdbin = (*returndistortion >> 7); 889 890 if (this_rdbin >= 1024) 891 { 892 this_rdbin = 1023; 893 } 894 895 cpi->error_bins[this_rdbin] ++; 896 } 897 898 899 if (cpi->is_src_frame_alt_ref && (best_mbmode.mode != ZEROMV || best_mbmode.ref_frame != ALTREF_FRAME)) 900 { 901 best_mbmode.mode = ZEROMV; 902 best_mbmode.ref_frame = ALTREF_FRAME; 903 best_mbmode.mv.as_int = 0; 904 best_mbmode.uv_mode = 0; 905 best_mbmode.mb_skip_coeff = (cpi->common.mb_no_coeff_skip) ? 1 : 0; 906 best_mbmode.partitioning = 0; 907 best_mbmode.dc_diff = 0; 908 909 vpx_memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode, sizeof(MB_MODE_INFO)); 910 vpx_memcpy(x->partition_info, &best_partition, sizeof(PARTITION_INFO)); 911 912 for (i = 0; i < 16; i++) 913 { 914 vpx_memset(&x->e_mbd.block[i].bmi, 0, sizeof(B_MODE_INFO)); 915 } 916 917 x->e_mbd.mode_info_context->mbmi.mv.as_int = 0; 918 919 return best_rd; 920 } 921 922 923 // macroblock modes 924 vpx_memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode, sizeof(MB_MODE_INFO)); 925 vpx_memcpy(x->partition_info, &best_partition, sizeof(PARTITION_INFO)); 926 927 if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED || x->e_mbd.mode_info_context->mbmi.mode == SPLITMV) 928 for (i = 0; i < 16; i++) 929 { 930 vpx_memcpy(&x->e_mbd.block[i].bmi, &best_bmodes[i], sizeof(B_MODE_INFO)); 931 932 } 933 else 934 { 935 vp8_set_mbmode_and_mvs(x, x->e_mbd.mode_info_context->mbmi.mode, &best_bmodes[0].mv.as_mv); 936 } 937 938 x->e_mbd.mode_info_context->mbmi.mv.as_mv = x->e_mbd.block[15].bmi.mv.as_mv; 939 940 return best_rd; 941 } 942