1 /* 2 * Copyright (c) 2012 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 <limits.h> 12 13 #include "denoising.h" 14 15 #include "vp8/common/reconinter.h" 16 #include "vpx/vpx_integer.h" 17 #include "vpx_mem/vpx_mem.h" 18 #include "vp8_rtcd.h" 19 20 static const unsigned int NOISE_MOTION_THRESHOLD = 25 * 25; 21 /* SSE_DIFF_THRESHOLD is selected as ~95% confidence assuming 22 * var(noise) ~= 100. 23 */ 24 static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20; 25 static const unsigned int SSE_THRESHOLD = 16 * 16 * 40; 26 static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 80; 27 28 /* 29 * The filter function was modified to reduce the computational complexity. 30 * Step 1: 31 * Instead of applying tap coefficients for each pixel, we calculated the 32 * pixel adjustments vs. pixel diff value ahead of time. 33 * adjustment = filtered_value - current_raw 34 * = (filter_coefficient * diff + 128) >> 8 35 * where 36 * filter_coefficient = (255 << 8) / (256 + ((absdiff * 330) >> 3)); 37 * filter_coefficient += filter_coefficient / 38 * (3 + motion_magnitude_adjustment); 39 * filter_coefficient is clamped to 0 ~ 255. 40 * 41 * Step 2: 42 * The adjustment vs. diff curve becomes flat very quick when diff increases. 43 * This allowed us to use only several levels to approximate the curve without 44 * changing the filtering algorithm too much. 45 * The adjustments were further corrected by checking the motion magnitude. 46 * The levels used are: 47 * diff adjustment w/o motion correction adjustment w/ motion correction 48 * [-255, -16] -6 -7 49 * [-15, -8] -4 -5 50 * [-7, -4] -3 -4 51 * [-3, 3] diff diff 52 * [4, 7] 3 4 53 * [8, 15] 4 5 54 * [16, 255] 6 7 55 */ 56 57 int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, 58 unsigned char *running_avg_y, int avg_y_stride, 59 unsigned char *sig, int sig_stride, 60 unsigned int motion_magnitude, 61 int increase_denoising) { 62 unsigned char *running_avg_y_start = running_avg_y; 63 unsigned char *sig_start = sig; 64 int sum_diff_thresh; 65 int r, c; 66 int sum_diff = 0; 67 int adj_val[3] = { 3, 4, 6 }; 68 int shift_inc1 = 0; 69 int shift_inc2 = 1; 70 int col_sum[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }; 71 /* If motion_magnitude is small, making the denoiser more aggressive by 72 * increasing the adjustment for each level. Add another increment for 73 * blocks that are labeled for increase denoising. */ 74 if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) { 75 if (increase_denoising) { 76 shift_inc1 = 1; 77 shift_inc2 = 2; 78 } 79 adj_val[0] += shift_inc2; 80 adj_val[1] += shift_inc2; 81 adj_val[2] += shift_inc2; 82 } 83 84 for (r = 0; r < 16; ++r) { 85 for (c = 0; c < 16; ++c) { 86 int diff = 0; 87 int adjustment = 0; 88 int absdiff = 0; 89 90 diff = mc_running_avg_y[c] - sig[c]; 91 absdiff = abs(diff); 92 93 // When |diff| <= |3 + shift_inc1|, use pixel value from 94 // last denoised raw. 95 if (absdiff <= 3 + shift_inc1) { 96 running_avg_y[c] = mc_running_avg_y[c]; 97 col_sum[c] += diff; 98 } else { 99 if (absdiff >= 4 + shift_inc1 && absdiff <= 7) { 100 adjustment = adj_val[0]; 101 } else if (absdiff >= 8 && absdiff <= 15) { 102 adjustment = adj_val[1]; 103 } else { 104 adjustment = adj_val[2]; 105 } 106 107 if (diff > 0) { 108 if ((sig[c] + adjustment) > 255) { 109 running_avg_y[c] = 255; 110 } else { 111 running_avg_y[c] = sig[c] + adjustment; 112 } 113 114 col_sum[c] += adjustment; 115 } else { 116 if ((sig[c] - adjustment) < 0) { 117 running_avg_y[c] = 0; 118 } else { 119 running_avg_y[c] = sig[c] - adjustment; 120 } 121 122 col_sum[c] -= adjustment; 123 } 124 } 125 } 126 127 /* Update pointers for next iteration. */ 128 sig += sig_stride; 129 mc_running_avg_y += mc_avg_y_stride; 130 running_avg_y += avg_y_stride; 131 } 132 133 for (c = 0; c < 16; ++c) { 134 // Below we clip the value in the same way which SSE code use. 135 // When adopting aggressive denoiser, the adj_val for each pixel 136 // could be at most 8 (this is current max adjustment of the map). 137 // In SSE code, we calculate the sum of adj_val for 138 // the columns, so the sum could be upto 128(16 rows). However, 139 // the range of the value is -128 ~ 127 in SSE code, that's why 140 // we do this change in C code. 141 // We don't do this for UV denoiser, since there are only 8 rows, 142 // and max adjustments <= 8, so the sum of the columns will not 143 // exceed 64. 144 if (col_sum[c] >= 128) { 145 col_sum[c] = 127; 146 } 147 sum_diff += col_sum[c]; 148 } 149 150 sum_diff_thresh = SUM_DIFF_THRESHOLD; 151 if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH; 152 if (abs(sum_diff) > sum_diff_thresh) { 153 // Before returning to copy the block (i.e., apply no denoising), check 154 // if we can still apply some (weaker) temporal filtering to this block, 155 // that would otherwise not be denoised at all. Simplest is to apply 156 // an additional adjustment to running_avg_y to bring it closer to sig. 157 // The adjustment is capped by a maximum delta, and chosen such that 158 // in most cases the resulting sum_diff will be within the 159 // accceptable range given by sum_diff_thresh. 160 161 // The delta is set by the excess of absolute pixel diff over threshold. 162 int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; 163 // Only apply the adjustment for max delta up to 3. 164 if (delta < 4) { 165 sig -= sig_stride * 16; 166 mc_running_avg_y -= mc_avg_y_stride * 16; 167 running_avg_y -= avg_y_stride * 16; 168 for (r = 0; r < 16; ++r) { 169 for (c = 0; c < 16; ++c) { 170 int diff = mc_running_avg_y[c] - sig[c]; 171 int adjustment = abs(diff); 172 if (adjustment > delta) adjustment = delta; 173 if (diff > 0) { 174 // Bring denoised signal down. 175 if (running_avg_y[c] - adjustment < 0) { 176 running_avg_y[c] = 0; 177 } else { 178 running_avg_y[c] = running_avg_y[c] - adjustment; 179 } 180 col_sum[c] -= adjustment; 181 } else if (diff < 0) { 182 // Bring denoised signal up. 183 if (running_avg_y[c] + adjustment > 255) { 184 running_avg_y[c] = 255; 185 } else { 186 running_avg_y[c] = running_avg_y[c] + adjustment; 187 } 188 col_sum[c] += adjustment; 189 } 190 } 191 // TODO(marpan): Check here if abs(sum_diff) has gone below the 192 // threshold sum_diff_thresh, and if so, we can exit the row loop. 193 sig += sig_stride; 194 mc_running_avg_y += mc_avg_y_stride; 195 running_avg_y += avg_y_stride; 196 } 197 198 sum_diff = 0; 199 for (c = 0; c < 16; ++c) { 200 if (col_sum[c] >= 128) { 201 col_sum[c] = 127; 202 } 203 sum_diff += col_sum[c]; 204 } 205 206 if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; 207 } else { 208 return COPY_BLOCK; 209 } 210 } 211 212 vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride); 213 return FILTER_BLOCK; 214 } 215 216 int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, 217 unsigned char *running_avg, int avg_stride, 218 unsigned char *sig, int sig_stride, 219 unsigned int motion_magnitude, 220 int increase_denoising) { 221 unsigned char *running_avg_start = running_avg; 222 unsigned char *sig_start = sig; 223 int sum_diff_thresh; 224 int r, c; 225 int sum_diff = 0; 226 int sum_block = 0; 227 int adj_val[3] = { 3, 4, 6 }; 228 int shift_inc1 = 0; 229 int shift_inc2 = 1; 230 /* If motion_magnitude is small, making the denoiser more aggressive by 231 * increasing the adjustment for each level. Add another increment for 232 * blocks that are labeled for increase denoising. */ 233 if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) { 234 if (increase_denoising) { 235 shift_inc1 = 1; 236 shift_inc2 = 2; 237 } 238 adj_val[0] += shift_inc2; 239 adj_val[1] += shift_inc2; 240 adj_val[2] += shift_inc2; 241 } 242 243 // Avoid denoising color signal if its close to average level. 244 for (r = 0; r < 8; ++r) { 245 for (c = 0; c < 8; ++c) { 246 sum_block += sig[c]; 247 } 248 sig += sig_stride; 249 } 250 if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) { 251 return COPY_BLOCK; 252 } 253 254 sig -= sig_stride * 8; 255 for (r = 0; r < 8; ++r) { 256 for (c = 0; c < 8; ++c) { 257 int diff = 0; 258 int adjustment = 0; 259 int absdiff = 0; 260 261 diff = mc_running_avg[c] - sig[c]; 262 absdiff = abs(diff); 263 264 // When |diff| <= |3 + shift_inc1|, use pixel value from 265 // last denoised raw. 266 if (absdiff <= 3 + shift_inc1) { 267 running_avg[c] = mc_running_avg[c]; 268 sum_diff += diff; 269 } else { 270 if (absdiff >= 4 && absdiff <= 7) { 271 adjustment = adj_val[0]; 272 } else if (absdiff >= 8 && absdiff <= 15) { 273 adjustment = adj_val[1]; 274 } else { 275 adjustment = adj_val[2]; 276 } 277 if (diff > 0) { 278 if ((sig[c] + adjustment) > 255) { 279 running_avg[c] = 255; 280 } else { 281 running_avg[c] = sig[c] + adjustment; 282 } 283 sum_diff += adjustment; 284 } else { 285 if ((sig[c] - adjustment) < 0) { 286 running_avg[c] = 0; 287 } else { 288 running_avg[c] = sig[c] - adjustment; 289 } 290 sum_diff -= adjustment; 291 } 292 } 293 } 294 /* Update pointers for next iteration. */ 295 sig += sig_stride; 296 mc_running_avg += mc_avg_stride; 297 running_avg += avg_stride; 298 } 299 300 sum_diff_thresh = SUM_DIFF_THRESHOLD_UV; 301 if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV; 302 if (abs(sum_diff) > sum_diff_thresh) { 303 // Before returning to copy the block (i.e., apply no denoising), check 304 // if we can still apply some (weaker) temporal filtering to this block, 305 // that would otherwise not be denoised at all. Simplest is to apply 306 // an additional adjustment to running_avg_y to bring it closer to sig. 307 // The adjustment is capped by a maximum delta, and chosen such that 308 // in most cases the resulting sum_diff will be within the 309 // accceptable range given by sum_diff_thresh. 310 311 // The delta is set by the excess of absolute pixel diff over threshold. 312 int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; 313 // Only apply the adjustment for max delta up to 3. 314 if (delta < 4) { 315 sig -= sig_stride * 8; 316 mc_running_avg -= mc_avg_stride * 8; 317 running_avg -= avg_stride * 8; 318 for (r = 0; r < 8; ++r) { 319 for (c = 0; c < 8; ++c) { 320 int diff = mc_running_avg[c] - sig[c]; 321 int adjustment = abs(diff); 322 if (adjustment > delta) adjustment = delta; 323 if (diff > 0) { 324 // Bring denoised signal down. 325 if (running_avg[c] - adjustment < 0) { 326 running_avg[c] = 0; 327 } else { 328 running_avg[c] = running_avg[c] - adjustment; 329 } 330 sum_diff -= adjustment; 331 } else if (diff < 0) { 332 // Bring denoised signal up. 333 if (running_avg[c] + adjustment > 255) { 334 running_avg[c] = 255; 335 } else { 336 running_avg[c] = running_avg[c] + adjustment; 337 } 338 sum_diff += adjustment; 339 } 340 } 341 // TODO(marpan): Check here if abs(sum_diff) has gone below the 342 // threshold sum_diff_thresh, and if so, we can exit the row loop. 343 sig += sig_stride; 344 mc_running_avg += mc_avg_stride; 345 running_avg += avg_stride; 346 } 347 if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK; 348 } else { 349 return COPY_BLOCK; 350 } 351 } 352 353 vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride); 354 return FILTER_BLOCK; 355 } 356 357 void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) { 358 assert(mode > 0); // Denoiser is allocated only if mode > 0. 359 if (mode == 1) { 360 denoiser->denoiser_mode = kDenoiserOnYOnly; 361 } else if (mode == 2) { 362 denoiser->denoiser_mode = kDenoiserOnYUV; 363 } else if (mode == 3) { 364 denoiser->denoiser_mode = kDenoiserOnYUVAggressive; 365 } else { 366 denoiser->denoiser_mode = kDenoiserOnYUV; 367 } 368 if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) { 369 denoiser->denoise_pars.scale_sse_thresh = 1; 370 denoiser->denoise_pars.scale_motion_thresh = 8; 371 denoiser->denoise_pars.scale_increase_filter = 0; 372 denoiser->denoise_pars.denoise_mv_bias = 95; 373 denoiser->denoise_pars.pickmode_mv_bias = 100; 374 denoiser->denoise_pars.qp_thresh = 0; 375 denoiser->denoise_pars.consec_zerolast = UINT_MAX; 376 denoiser->denoise_pars.spatial_blur = 0; 377 } else { 378 denoiser->denoise_pars.scale_sse_thresh = 2; 379 denoiser->denoise_pars.scale_motion_thresh = 16; 380 denoiser->denoise_pars.scale_increase_filter = 1; 381 denoiser->denoise_pars.denoise_mv_bias = 60; 382 denoiser->denoise_pars.pickmode_mv_bias = 75; 383 denoiser->denoise_pars.qp_thresh = 80; 384 denoiser->denoise_pars.consec_zerolast = 15; 385 denoiser->denoise_pars.spatial_blur = 0; 386 } 387 } 388 389 int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height, 390 int num_mb_rows, int num_mb_cols, int mode) { 391 int i; 392 assert(denoiser); 393 denoiser->num_mb_cols = num_mb_cols; 394 395 for (i = 0; i < MAX_REF_FRAMES; ++i) { 396 denoiser->yv12_running_avg[i].flags = 0; 397 398 if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width, 399 height, VP8BORDERINPIXELS) < 0) { 400 vp8_denoiser_free(denoiser); 401 return 1; 402 } 403 memset(denoiser->yv12_running_avg[i].buffer_alloc, 0, 404 denoiser->yv12_running_avg[i].frame_size); 405 } 406 denoiser->yv12_mc_running_avg.flags = 0; 407 408 if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width, 409 height, VP8BORDERINPIXELS) < 0) { 410 vp8_denoiser_free(denoiser); 411 return 1; 412 } 413 414 memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0, 415 denoiser->yv12_mc_running_avg.frame_size); 416 417 if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width, height, 418 VP8BORDERINPIXELS) < 0) { 419 vp8_denoiser_free(denoiser); 420 return 1; 421 } 422 memset(denoiser->yv12_last_source.buffer_alloc, 0, 423 denoiser->yv12_last_source.frame_size); 424 425 denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1); 426 if (!denoiser->denoise_state) { 427 vp8_denoiser_free(denoiser); 428 return 1; 429 } 430 memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols)); 431 vp8_denoiser_set_parameters(denoiser, mode); 432 denoiser->nmse_source_diff = 0; 433 denoiser->nmse_source_diff_count = 0; 434 denoiser->qp_avg = 0; 435 // QP threshold below which we can go up to aggressive mode. 436 denoiser->qp_threshold_up = 80; 437 // QP threshold above which we can go back down to normal mode. 438 // For now keep this second threshold high, so not used currently. 439 denoiser->qp_threshold_down = 128; 440 // Bitrate thresholds and noise metric (nmse) thresholds for switching to 441 // aggressive mode. 442 // TODO(marpan): Adjust thresholds, including effect on resolution. 443 denoiser->bitrate_threshold = 400000; // (bits/sec). 444 denoiser->threshold_aggressive_mode = 80; 445 if (width * height > 1280 * 720) { 446 denoiser->bitrate_threshold = 3000000; 447 denoiser->threshold_aggressive_mode = 200; 448 } else if (width * height > 960 * 540) { 449 denoiser->bitrate_threshold = 1200000; 450 denoiser->threshold_aggressive_mode = 120; 451 } else if (width * height > 640 * 480) { 452 denoiser->bitrate_threshold = 600000; 453 denoiser->threshold_aggressive_mode = 100; 454 } 455 return 0; 456 } 457 458 void vp8_denoiser_free(VP8_DENOISER *denoiser) { 459 int i; 460 assert(denoiser); 461 462 for (i = 0; i < MAX_REF_FRAMES; ++i) { 463 vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]); 464 } 465 vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg); 466 vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source); 467 vpx_free(denoiser->denoise_state); 468 } 469 470 void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser, MACROBLOCK *x, 471 unsigned int best_sse, unsigned int zero_mv_sse, 472 int recon_yoffset, int recon_uvoffset, 473 loop_filter_info_n *lfi_n, int mb_row, int mb_col, 474 int block_index, int consec_zero_last) 475 476 { 477 int mv_row; 478 int mv_col; 479 unsigned int motion_threshold; 480 unsigned int motion_magnitude2; 481 unsigned int sse_thresh; 482 int sse_diff_thresh = 0; 483 // Spatial loop filter: only applied selectively based on 484 // temporal filter state of block relative to top/left neighbors. 485 int apply_spatial_loop_filter = 1; 486 MV_REFERENCE_FRAME frame = x->best_reference_frame; 487 MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame; 488 489 enum vp8_denoiser_decision decision = FILTER_BLOCK; 490 enum vp8_denoiser_decision decision_u = COPY_BLOCK; 491 enum vp8_denoiser_decision decision_v = COPY_BLOCK; 492 493 if (zero_frame) { 494 YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame]; 495 YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg; 496 YV12_BUFFER_CONFIG saved_pre, saved_dst; 497 MB_MODE_INFO saved_mbmi; 498 MACROBLOCKD *filter_xd = &x->e_mbd; 499 MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi; 500 int sse_diff = 0; 501 // Bias on zero motion vector sse. 502 const int zero_bias = denoiser->denoise_pars.denoise_mv_bias; 503 zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100); 504 sse_diff = (int)zero_mv_sse - (int)best_sse; 505 506 saved_mbmi = *mbmi; 507 508 /* Use the best MV for the compensation. */ 509 mbmi->ref_frame = x->best_reference_frame; 510 mbmi->mode = x->best_sse_inter_mode; 511 mbmi->mv = x->best_sse_mv; 512 mbmi->need_to_clamp_mvs = x->need_to_clamp_best_mvs; 513 mv_col = x->best_sse_mv.as_mv.col; 514 mv_row = x->best_sse_mv.as_mv.row; 515 // Bias to zero_mv if small amount of motion. 516 // Note sse_diff_thresh is intialized to zero, so this ensures 517 // we will always choose zero_mv for denoising if 518 // zero_mv_see <= best_sse (i.e., sse_diff <= 0). 519 if ((unsigned int)(mv_row * mv_row + mv_col * mv_col) <= 520 NOISE_MOTION_THRESHOLD) { 521 sse_diff_thresh = (int)SSE_DIFF_THRESHOLD; 522 } 523 524 if (frame == INTRA_FRAME || sse_diff <= sse_diff_thresh) { 525 /* 526 * Handle intra blocks as referring to last frame with zero motion 527 * and let the absolute pixel difference affect the filter factor. 528 * Also consider small amount of motion as being random walk due 529 * to noise, if it doesn't mean that we get a much bigger error. 530 * Note that any changes to the mode info only affects the 531 * denoising. 532 */ 533 x->denoise_zeromv = 1; 534 mbmi->ref_frame = x->best_zeromv_reference_frame; 535 536 src = &denoiser->yv12_running_avg[zero_frame]; 537 538 mbmi->mode = ZEROMV; 539 mbmi->mv.as_int = 0; 540 x->best_sse_inter_mode = ZEROMV; 541 x->best_sse_mv.as_int = 0; 542 best_sse = zero_mv_sse; 543 } 544 545 mv_row = x->best_sse_mv.as_mv.row; 546 mv_col = x->best_sse_mv.as_mv.col; 547 motion_magnitude2 = mv_row * mv_row + mv_col * mv_col; 548 motion_threshold = 549 denoiser->denoise_pars.scale_motion_thresh * NOISE_MOTION_THRESHOLD; 550 551 if (motion_magnitude2 < 552 denoiser->denoise_pars.scale_increase_filter * NOISE_MOTION_THRESHOLD) { 553 x->increase_denoising = 1; 554 } 555 556 sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD; 557 if (x->increase_denoising) { 558 sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD_HIGH; 559 } 560 561 if (best_sse > sse_thresh || motion_magnitude2 > motion_threshold) { 562 decision = COPY_BLOCK; 563 } 564 565 // If block is considered skin, don't denoise if the block 566 // (1) is selected as non-zero motion for current frame, or 567 // (2) has not been selected as ZERO_LAST mode at least x past frames 568 // in a row. 569 // TODO(marpan): Parameter "x" should be varied with framerate. 570 // In particualar, should be reduced for layers (base layer/LAST). 571 if (x->is_skin && (consec_zero_last < 2 || motion_magnitude2 > 0)) { 572 decision = COPY_BLOCK; 573 } 574 575 if (decision == FILTER_BLOCK) { 576 saved_pre = filter_xd->pre; 577 saved_dst = filter_xd->dst; 578 579 /* Compensate the running average. */ 580 filter_xd->pre.y_buffer = src->y_buffer + recon_yoffset; 581 filter_xd->pre.u_buffer = src->u_buffer + recon_uvoffset; 582 filter_xd->pre.v_buffer = src->v_buffer + recon_uvoffset; 583 /* Write the compensated running average to the destination buffer. */ 584 filter_xd->dst.y_buffer = dst->y_buffer + recon_yoffset; 585 filter_xd->dst.u_buffer = dst->u_buffer + recon_uvoffset; 586 filter_xd->dst.v_buffer = dst->v_buffer + recon_uvoffset; 587 588 if (!x->skip) { 589 vp8_build_inter_predictors_mb(filter_xd); 590 } else { 591 vp8_build_inter16x16_predictors_mb( 592 filter_xd, filter_xd->dst.y_buffer, filter_xd->dst.u_buffer, 593 filter_xd->dst.v_buffer, filter_xd->dst.y_stride, 594 filter_xd->dst.uv_stride); 595 } 596 filter_xd->pre = saved_pre; 597 filter_xd->dst = saved_dst; 598 *mbmi = saved_mbmi; 599 } 600 } else { 601 // zero_frame should always be 1 for real-time mode, as the 602 // ZEROMV mode is always checked, so we should never go into this branch. 603 // If case ZEROMV is not checked, then we will force no denoise (COPY). 604 decision = COPY_BLOCK; 605 } 606 607 if (decision == FILTER_BLOCK) { 608 unsigned char *mc_running_avg_y = 609 denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset; 610 int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride; 611 unsigned char *running_avg_y = 612 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset; 613 int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride; 614 615 /* Filter. */ 616 decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride, 617 running_avg_y, avg_y_stride, x->thismb, 16, 618 motion_magnitude2, x->increase_denoising); 619 denoiser->denoise_state[block_index] = 620 motion_magnitude2 > 0 ? kFilterNonZeroMV : kFilterZeroMV; 621 // Only denoise UV for zero motion, and if y channel was denoised. 622 if (denoiser->denoiser_mode != kDenoiserOnYOnly && motion_magnitude2 == 0 && 623 decision == FILTER_BLOCK) { 624 unsigned char *mc_running_avg_u = 625 denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset; 626 unsigned char *running_avg_u = 627 denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset; 628 unsigned char *mc_running_avg_v = 629 denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset; 630 unsigned char *running_avg_v = 631 denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset; 632 int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride; 633 int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride; 634 int signal_stride = x->block[16].src_stride; 635 decision_u = vp8_denoiser_filter_uv( 636 mc_running_avg_u, mc_avg_uv_stride, running_avg_u, avg_uv_stride, 637 x->block[16].src + *x->block[16].base_src, signal_stride, 638 motion_magnitude2, 0); 639 decision_v = vp8_denoiser_filter_uv( 640 mc_running_avg_v, mc_avg_uv_stride, running_avg_v, avg_uv_stride, 641 x->block[20].src + *x->block[20].base_src, signal_stride, 642 motion_magnitude2, 0); 643 } 644 } 645 if (decision == COPY_BLOCK) { 646 /* No filtering of this block; it differs too much from the predictor, 647 * or the motion vector magnitude is considered too big. 648 */ 649 x->denoise_zeromv = 0; 650 vp8_copy_mem16x16( 651 x->thismb, 16, 652 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset, 653 denoiser->yv12_running_avg[INTRA_FRAME].y_stride); 654 denoiser->denoise_state[block_index] = kNoFilter; 655 } 656 if (denoiser->denoiser_mode != kDenoiserOnYOnly) { 657 if (decision_u == COPY_BLOCK) { 658 vp8_copy_mem8x8( 659 x->block[16].src + *x->block[16].base_src, x->block[16].src_stride, 660 denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset, 661 denoiser->yv12_running_avg[INTRA_FRAME].uv_stride); 662 } 663 if (decision_v == COPY_BLOCK) { 664 vp8_copy_mem8x8( 665 x->block[20].src + *x->block[20].base_src, x->block[16].src_stride, 666 denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset, 667 denoiser->yv12_running_avg[INTRA_FRAME].uv_stride); 668 } 669 } 670 // Option to selectively deblock the denoised signal, for y channel only. 671 if (apply_spatial_loop_filter) { 672 loop_filter_info lfi; 673 int apply_filter_col = 0; 674 int apply_filter_row = 0; 675 int apply_filter = 0; 676 int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride; 677 int uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride; 678 679 // Fix filter level to some nominal value for now. 680 int filter_level = 48; 681 682 int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level]; 683 lfi.mblim = lfi_n->mblim[filter_level]; 684 lfi.blim = lfi_n->blim[filter_level]; 685 lfi.lim = lfi_n->lim[filter_level]; 686 lfi.hev_thr = lfi_n->hev_thr[hev_index]; 687 688 // Apply filter if there is a difference in the denoiser filter state 689 // between the current and left/top block, or if non-zero motion vector 690 // is used for the motion-compensated filtering. 691 if (mb_col > 0) { 692 apply_filter_col = 693 !((denoiser->denoise_state[block_index] == 694 denoiser->denoise_state[block_index - 1]) && 695 denoiser->denoise_state[block_index] != kFilterNonZeroMV); 696 if (apply_filter_col) { 697 // Filter left vertical edge. 698 apply_filter = 1; 699 vp8_loop_filter_mbv( 700 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset, 701 NULL, NULL, y_stride, uv_stride, &lfi); 702 } 703 } 704 if (mb_row > 0) { 705 apply_filter_row = 706 !((denoiser->denoise_state[block_index] == 707 denoiser->denoise_state[block_index - denoiser->num_mb_cols]) && 708 denoiser->denoise_state[block_index] != kFilterNonZeroMV); 709 if (apply_filter_row) { 710 // Filter top horizontal edge. 711 apply_filter = 1; 712 vp8_loop_filter_mbh( 713 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset, 714 NULL, NULL, y_stride, uv_stride, &lfi); 715 } 716 } 717 if (apply_filter) { 718 // Update the signal block |x|. Pixel changes are only to top and/or 719 // left boundary pixels: can we avoid full block copy here. 720 vp8_copy_mem16x16( 721 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset, 722 y_stride, x->thismb, 16); 723 } 724 } 725 } 726
