1 /* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include <math.h> 12 #include <limits.h> 13 14 #include "vp9/common/vp9_alloccommon.h" 15 #include "vp9/common/vp9_onyxc_int.h" 16 #include "vp9/common/vp9_quant_common.h" 17 #include "vp9/common/vp9_reconinter.h" 18 #include "vp9/common/vp9_systemdependent.h" 19 #include "vp9/encoder/vp9_extend.h" 20 #include "vp9/encoder/vp9_firstpass.h" 21 #include "vp9/encoder/vp9_mcomp.h" 22 #include "vp9/encoder/vp9_onyx_int.h" 23 #include "vp9/encoder/vp9_quantize.h" 24 #include "vp9/encoder/vp9_ratectrl.h" 25 #include "vp9/encoder/vp9_segmentation.h" 26 #include "vpx_mem/vpx_mem.h" 27 #include "vpx_ports/vpx_timer.h" 28 #include "vpx_scale/vpx_scale.h" 29 30 #define ALT_REF_MC_ENABLED 1 // dis/enable MC in AltRef filtering 31 32 static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd, 33 uint8_t *y_mb_ptr, 34 uint8_t *u_mb_ptr, 35 uint8_t *v_mb_ptr, 36 int stride, 37 int uv_block_size, 38 int mv_row, 39 int mv_col, 40 uint8_t *pred, 41 struct scale_factors *scale, 42 int x, int y) { 43 const int which_mv = 0; 44 const MV mv = { mv_row, mv_col }; 45 const InterpKernel *const kernel = 46 vp9_get_interp_kernel(xd->mi[0]->mbmi.interp_filter); 47 48 enum mv_precision mv_precision_uv; 49 int uv_stride; 50 if (uv_block_size == 8) { 51 uv_stride = (stride + 1) >> 1; 52 mv_precision_uv = MV_PRECISION_Q4; 53 } else { 54 uv_stride = stride; 55 mv_precision_uv = MV_PRECISION_Q3; 56 } 57 58 vp9_build_inter_predictor(y_mb_ptr, stride, 59 &pred[0], 16, 60 &mv, 61 scale, 62 16, 16, 63 which_mv, 64 kernel, MV_PRECISION_Q3, x, y); 65 66 vp9_build_inter_predictor(u_mb_ptr, uv_stride, 67 &pred[256], uv_block_size, 68 &mv, 69 scale, 70 uv_block_size, uv_block_size, 71 which_mv, 72 kernel, mv_precision_uv, x, y); 73 74 vp9_build_inter_predictor(v_mb_ptr, uv_stride, 75 &pred[512], uv_block_size, 76 &mv, 77 scale, 78 uv_block_size, uv_block_size, 79 which_mv, 80 kernel, mv_precision_uv, x, y); 81 } 82 83 void vp9_temporal_filter_apply_c(uint8_t *frame1, 84 unsigned int stride, 85 uint8_t *frame2, 86 unsigned int block_size, 87 int strength, 88 int filter_weight, 89 unsigned int *accumulator, 90 uint16_t *count) { 91 unsigned int i, j, k; 92 int modifier; 93 int byte = 0; 94 95 for (i = 0, k = 0; i < block_size; i++) { 96 for (j = 0; j < block_size; j++, k++) { 97 int src_byte = frame1[byte]; 98 int pixel_value = *frame2++; 99 100 modifier = src_byte - pixel_value; 101 // This is an integer approximation of: 102 // float coeff = (3.0 * modifer * modifier) / pow(2, strength); 103 // modifier = (int)roundf(coeff > 16 ? 0 : 16-coeff); 104 modifier *= modifier; 105 modifier *= 3; 106 modifier += 1 << (strength - 1); 107 modifier >>= strength; 108 109 if (modifier > 16) 110 modifier = 16; 111 112 modifier = 16 - modifier; 113 modifier *= filter_weight; 114 115 count[k] += modifier; 116 accumulator[k] += modifier * pixel_value; 117 118 byte++; 119 } 120 121 byte += stride - block_size; 122 } 123 } 124 125 #if ALT_REF_MC_ENABLED 126 127 static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi, 128 uint8_t *arf_frame_buf, 129 uint8_t *frame_ptr_buf, 130 int stride) { 131 MACROBLOCK *x = &cpi->mb; 132 MACROBLOCKD* const xd = &x->e_mbd; 133 int step_param; 134 int sadpb = x->sadperbit16; 135 int bestsme = INT_MAX; 136 137 MV best_ref_mv1 = {0, 0}; 138 MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */ 139 MV *ref_mv = &x->e_mbd.mi[0]->bmi[0].as_mv[0].as_mv; 140 141 // Save input state 142 struct buf_2d src = x->plane[0].src; 143 struct buf_2d pre = xd->plane[0].pre[0]; 144 145 best_ref_mv1_full.col = best_ref_mv1.col >> 3; 146 best_ref_mv1_full.row = best_ref_mv1.row >> 3; 147 148 // Setup frame pointers 149 x->plane[0].src.buf = arf_frame_buf; 150 x->plane[0].src.stride = stride; 151 xd->plane[0].pre[0].buf = frame_ptr_buf; 152 xd->plane[0].pre[0].stride = stride; 153 154 // Further step/diamond searches as necessary 155 if (cpi->speed < 8) 156 step_param = cpi->sf.reduce_first_step_size + ((cpi->speed > 5) ? 1 : 0); 157 else 158 step_param = cpi->sf.reduce_first_step_size + 2; 159 step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2)); 160 161 /*cpi->sf.search_method == HEX*/ 162 // Ignore mv costing by sending NULL pointer instead of cost arrays 163 vp9_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1, 164 &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1, ref_mv); 165 166 // Try sub-pixel MC? 167 // if (bestsme > error_thresh && bestsme < INT_MAX) 168 { 169 int distortion; 170 unsigned int sse; 171 // Ignore mv costing by sending NULL pointer instead of cost array 172 bestsme = cpi->find_fractional_mv_step(x, ref_mv, 173 &best_ref_mv1, 174 cpi->common.allow_high_precision_mv, 175 x->errorperbit, 176 &cpi->fn_ptr[BLOCK_16X16], 177 0, cpi->sf.subpel_iters_per_step, 178 NULL, NULL, 179 &distortion, &sse); 180 } 181 182 // Restore input state 183 x->plane[0].src = src; 184 xd->plane[0].pre[0] = pre; 185 186 return bestsme; 187 } 188 #endif 189 190 static void temporal_filter_iterate_c(VP9_COMP *cpi, 191 int frame_count, 192 int alt_ref_index, 193 int strength, 194 struct scale_factors *scale) { 195 int byte; 196 int frame; 197 int mb_col, mb_row; 198 unsigned int filter_weight; 199 int mb_cols = cpi->common.mb_cols; 200 int mb_rows = cpi->common.mb_rows; 201 int mb_y_offset = 0; 202 int mb_uv_offset = 0; 203 DECLARE_ALIGNED_ARRAY(16, unsigned int, accumulator, 16 * 16 * 3); 204 DECLARE_ALIGNED_ARRAY(16, uint16_t, count, 16 * 16 * 3); 205 MACROBLOCKD *mbd = &cpi->mb.e_mbd; 206 YV12_BUFFER_CONFIG *f = cpi->frames[alt_ref_index]; 207 uint8_t *dst1, *dst2; 208 DECLARE_ALIGNED_ARRAY(16, uint8_t, predictor, 16 * 16 * 3); 209 const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y; 210 211 // Save input state 212 uint8_t* input_buffer[MAX_MB_PLANE]; 213 int i; 214 215 // TODO(aconverse): Add 4:2:2 support 216 assert(mbd->plane[1].subsampling_x == mbd->plane[1].subsampling_y); 217 218 for (i = 0; i < MAX_MB_PLANE; i++) 219 input_buffer[i] = mbd->plane[i].pre[0].buf; 220 221 for (mb_row = 0; mb_row < mb_rows; mb_row++) { 222 #if ALT_REF_MC_ENABLED 223 // Source frames are extended to 16 pixels. This is different than 224 // L/A/G reference frames that have a border of 32 (VP9ENCBORDERINPIXELS) 225 // A 6/8 tap filter is used for motion search. This requires 2 pixels 226 // before and 3 pixels after. So the largest Y mv on a border would 227 // then be 16 - VP9_INTERP_EXTEND. The UV blocks are half the size of the 228 // Y and therefore only extended by 8. The largest mv that a UV block 229 // can support is 8 - VP9_INTERP_EXTEND. A UV mv is half of a Y mv. 230 // (16 - VP9_INTERP_EXTEND) >> 1 which is greater than 231 // 8 - VP9_INTERP_EXTEND. 232 // To keep the mv in play for both Y and UV planes the max that it 233 // can be on a border is therefore 16 - (2*VP9_INTERP_EXTEND+1). 234 cpi->mb.mv_row_min = -((mb_row * 16) + (17 - 2 * VP9_INTERP_EXTEND)); 235 cpi->mb.mv_row_max = ((cpi->common.mb_rows - 1 - mb_row) * 16) 236 + (17 - 2 * VP9_INTERP_EXTEND); 237 #endif 238 239 for (mb_col = 0; mb_col < mb_cols; mb_col++) { 240 int i, j, k; 241 int stride; 242 243 vpx_memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0])); 244 vpx_memset(count, 0, 16 * 16 * 3 * sizeof(count[0])); 245 246 #if ALT_REF_MC_ENABLED 247 cpi->mb.mv_col_min = -((mb_col * 16) + (17 - 2 * VP9_INTERP_EXTEND)); 248 cpi->mb.mv_col_max = ((cpi->common.mb_cols - 1 - mb_col) * 16) 249 + (17 - 2 * VP9_INTERP_EXTEND); 250 #endif 251 252 for (frame = 0; frame < frame_count; frame++) { 253 if (cpi->frames[frame] == NULL) 254 continue; 255 256 mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0; 257 mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0; 258 259 if (frame == alt_ref_index) { 260 filter_weight = 2; 261 } else { 262 int err = 0; 263 #if ALT_REF_MC_ENABLED 264 #define THRESH_LOW 10000 265 #define THRESH_HIGH 20000 266 267 // Find best match in this frame by MC 268 err = temporal_filter_find_matching_mb_c 269 (cpi, 270 cpi->frames[alt_ref_index]->y_buffer + mb_y_offset, 271 cpi->frames[frame]->y_buffer + mb_y_offset, 272 cpi->frames[frame]->y_stride); 273 #endif 274 // Assign higher weight to matching MB if it's error 275 // score is lower. If not applying MC default behavior 276 // is to weight all MBs equal. 277 filter_weight = err < THRESH_LOW 278 ? 2 : err < THRESH_HIGH ? 1 : 0; 279 } 280 281 if (filter_weight != 0) { 282 // Construct the predictors 283 temporal_filter_predictors_mb_c 284 (mbd, 285 cpi->frames[frame]->y_buffer + mb_y_offset, 286 cpi->frames[frame]->u_buffer + mb_uv_offset, 287 cpi->frames[frame]->v_buffer + mb_uv_offset, 288 cpi->frames[frame]->y_stride, 289 mb_uv_height, 290 mbd->mi[0]->bmi[0].as_mv[0].as_mv.row, 291 mbd->mi[0]->bmi[0].as_mv[0].as_mv.col, 292 predictor, scale, 293 mb_col * 16, mb_row * 16); 294 295 // Apply the filter (YUV) 296 vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride, 297 predictor, 16, strength, filter_weight, 298 accumulator, count); 299 300 vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride, 301 predictor + 256, mb_uv_height, strength, 302 filter_weight, accumulator + 256, 303 count + 256); 304 305 vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride, 306 predictor + 512, mb_uv_height, strength, 307 filter_weight, accumulator + 512, 308 count + 512); 309 } 310 } 311 312 // Normalize filter output to produce AltRef frame 313 dst1 = cpi->alt_ref_buffer.y_buffer; 314 stride = cpi->alt_ref_buffer.y_stride; 315 byte = mb_y_offset; 316 for (i = 0, k = 0; i < 16; i++) { 317 for (j = 0; j < 16; j++, k++) { 318 unsigned int pval = accumulator[k] + (count[k] >> 1); 319 pval *= cpi->fixed_divide[count[k]]; 320 pval >>= 19; 321 322 dst1[byte] = (uint8_t)pval; 323 324 // move to next pixel 325 byte++; 326 } 327 328 byte += stride - 16; 329 } 330 331 dst1 = cpi->alt_ref_buffer.u_buffer; 332 dst2 = cpi->alt_ref_buffer.v_buffer; 333 stride = cpi->alt_ref_buffer.uv_stride; 334 byte = mb_uv_offset; 335 for (i = 0, k = 256; i < mb_uv_height; i++) { 336 for (j = 0; j < mb_uv_height; j++, k++) { 337 int m = k + 256; 338 339 // U 340 unsigned int pval = accumulator[k] + (count[k] >> 1); 341 pval *= cpi->fixed_divide[count[k]]; 342 pval >>= 19; 343 dst1[byte] = (uint8_t)pval; 344 345 // V 346 pval = accumulator[m] + (count[m] >> 1); 347 pval *= cpi->fixed_divide[count[m]]; 348 pval >>= 19; 349 dst2[byte] = (uint8_t)pval; 350 351 // move to next pixel 352 byte++; 353 } 354 355 byte += stride - mb_uv_height; 356 } 357 358 mb_y_offset += 16; 359 mb_uv_offset += mb_uv_height; 360 } 361 362 mb_y_offset += 16 * (f->y_stride - mb_cols); 363 mb_uv_offset += mb_uv_height * (f->uv_stride - mb_cols); 364 } 365 366 // Restore input state 367 for (i = 0; i < MAX_MB_PLANE; i++) 368 mbd->plane[i].pre[0].buf = input_buffer[i]; 369 } 370 371 void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance) { 372 VP9_COMMON *const cm = &cpi->common; 373 374 int frame = 0; 375 376 int frames_to_blur_backward = 0; 377 int frames_to_blur_forward = 0; 378 int frames_to_blur = 0; 379 int start_frame = 0; 380 381 int strength = cpi->active_arnr_strength; 382 int blur_type = cpi->oxcf.arnr_type; 383 int max_frames = cpi->active_arnr_frames; 384 385 const int num_frames_backward = distance; 386 const int num_frames_forward = vp9_lookahead_depth(cpi->lookahead) 387 - (num_frames_backward + 1); 388 struct scale_factors sf; 389 390 switch (blur_type) { 391 case 1: 392 // Backward Blur 393 frames_to_blur_backward = num_frames_backward; 394 395 if (frames_to_blur_backward >= max_frames) 396 frames_to_blur_backward = max_frames - 1; 397 398 frames_to_blur = frames_to_blur_backward + 1; 399 break; 400 401 case 2: 402 // Forward Blur 403 frames_to_blur_forward = num_frames_forward; 404 405 if (frames_to_blur_forward >= max_frames) 406 frames_to_blur_forward = max_frames - 1; 407 408 frames_to_blur = frames_to_blur_forward + 1; 409 break; 410 411 case 3: 412 default: 413 // Center Blur 414 frames_to_blur_forward = num_frames_forward; 415 frames_to_blur_backward = num_frames_backward; 416 417 if (frames_to_blur_forward > frames_to_blur_backward) 418 frames_to_blur_forward = frames_to_blur_backward; 419 420 if (frames_to_blur_backward > frames_to_blur_forward) 421 frames_to_blur_backward = frames_to_blur_forward; 422 423 // When max_frames is even we have 1 more frame backward than forward 424 if (frames_to_blur_forward > (max_frames - 1) / 2) 425 frames_to_blur_forward = ((max_frames - 1) / 2); 426 427 if (frames_to_blur_backward > (max_frames / 2)) 428 frames_to_blur_backward = (max_frames / 2); 429 430 frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1; 431 break; 432 } 433 434 start_frame = distance + frames_to_blur_forward; 435 436 #ifdef DEBUGFWG 437 // DEBUG FWG 438 printf( 439 "max:%d FBCK:%d FFWD:%d ftb:%d ftbbck:%d ftbfwd:%d sei:%d lasei:%d " 440 "start:%d", 441 max_frames, num_frames_backward, num_frames_forward, frames_to_blur, 442 frames_to_blur_backward, frames_to_blur_forward, cpi->source_encode_index, 443 cpi->last_alt_ref_sei, start_frame); 444 #endif 445 446 // Setup scaling factors. Scaling on each of the arnr frames is not supported 447 vp9_setup_scale_factors_for_frame(&sf, 448 get_frame_new_buffer(cm)->y_crop_width, 449 get_frame_new_buffer(cm)->y_crop_height, 450 cm->width, cm->height); 451 452 // Setup frame pointers, NULL indicates frame not included in filter 453 vp9_zero(cpi->frames); 454 for (frame = 0; frame < frames_to_blur; frame++) { 455 int which_buffer = start_frame - frame; 456 struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 457 which_buffer); 458 cpi->frames[frames_to_blur - 1 - frame] = &buf->img; 459 } 460 461 temporal_filter_iterate_c(cpi, frames_to_blur, frames_to_blur_backward, 462 strength, &sf); 463 } 464 465 void vp9_configure_arnr_filter(VP9_COMP *cpi, 466 const unsigned int frames_to_arnr, 467 const int group_boost) { 468 int half_gf_int; 469 int frames_after_arf; 470 int frames_bwd = cpi->oxcf.arnr_max_frames - 1; 471 int frames_fwd = cpi->oxcf.arnr_max_frames - 1; 472 int q; 473 474 // Define the arnr filter width for this group of frames. We only 475 // filter frames that lie within a distance of half the GF interval 476 // from the ARF frame. We also have to trap cases where the filter 477 // extends beyond the end of the lookahead buffer. 478 // Note: frames_to_arnr parameter is the offset of the arnr 479 // frame from the current frame. 480 half_gf_int = cpi->rc.baseline_gf_interval >> 1; 481 frames_after_arf = vp9_lookahead_depth(cpi->lookahead) 482 - frames_to_arnr - 1; 483 484 switch (cpi->oxcf.arnr_type) { 485 case 1: // Backward filter 486 frames_fwd = 0; 487 if (frames_bwd > half_gf_int) 488 frames_bwd = half_gf_int; 489 break; 490 491 case 2: // Forward filter 492 if (frames_fwd > half_gf_int) 493 frames_fwd = half_gf_int; 494 if (frames_fwd > frames_after_arf) 495 frames_fwd = frames_after_arf; 496 frames_bwd = 0; 497 break; 498 499 case 3: // Centered filter 500 default: 501 frames_fwd >>= 1; 502 if (frames_fwd > frames_after_arf) 503 frames_fwd = frames_after_arf; 504 if (frames_fwd > half_gf_int) 505 frames_fwd = half_gf_int; 506 507 frames_bwd = frames_fwd; 508 509 // For even length filter there is one more frame backward 510 // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff. 511 if (frames_bwd < half_gf_int) 512 frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1; 513 break; 514 } 515 516 cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd; 517 518 // Adjust the strength based on active max q 519 if (cpi->common.current_video_frame > 1) 520 q = ((int)vp9_convert_qindex_to_q( 521 cpi->rc.avg_frame_qindex[INTER_FRAME])); 522 else 523 q = ((int)vp9_convert_qindex_to_q( 524 cpi->rc.avg_frame_qindex[KEY_FRAME])); 525 if (q > 16) { 526 cpi->active_arnr_strength = cpi->oxcf.arnr_strength; 527 } else { 528 cpi->active_arnr_strength = cpi->oxcf.arnr_strength - ((16 - q) / 2); 529 if (cpi->active_arnr_strength < 0) 530 cpi->active_arnr_strength = 0; 531 } 532 533 // Adjust number of frames in filter and strength based on gf boost level. 534 if (cpi->active_arnr_frames > (group_boost / 150)) { 535 cpi->active_arnr_frames = (group_boost / 150); 536 cpi->active_arnr_frames += !(cpi->active_arnr_frames & 1); 537 } 538 if (cpi->active_arnr_strength > (group_boost / 300)) { 539 cpi->active_arnr_strength = (group_boost / 300); 540 } 541 } 542