src/third_party/libvpx/source/libvpx/vp8/encoder/denoising.c

   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 * 60;
  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 {
  63     unsigned char *running_avg_y_start = running_avg_y;
  64     unsigned char *sig_start = sig;
  65     int sum_diff_thresh;
  66     int r, c;
  67     int sum_diff = 0;
  68     int adj_val[3] = {3, 4, 6};
  69     int shift_inc1 = 0;
  70     int shift_inc2 = 1;
  71     int col_sum[16] = {0, 0, 0, 0,
  72                        0, 0, 0, 0,
  73                        0, 0, 0, 0,
  74                        0, 0, 0, 0};
  75     /* If motion_magnitude is small, making the denoiser more aggressive by
  76      * increasing the adjustment for each level. Add another increment for
  77      * blocks that are labeled for increase denoising. */
  78     if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
  79     {
  80       if (increase_denoising) {
  81         shift_inc1 = 1;
  82         shift_inc2 = 2;
  83       }
  84       adj_val[0] += shift_inc2;
  85       adj_val[1] += shift_inc2;
  86       adj_val[2] += shift_inc2;
  87     }
  88
  89     for (r = 0; r < 16; ++r)
  90     {
  91         for (c = 0; c < 16; ++c)
  92         {
  93             int diff = 0;
  94             int adjustment = 0;
  95             int absdiff = 0;
  96
  97             diff = mc_running_avg_y[c] - sig[c];
  98             absdiff = abs(diff);
  99
 100             // When |diff| <= |3 + shift_inc1|, use pixel value from
 101             // last denoised raw.
 102             if (absdiff <= 3 + shift_inc1)
 103             {
 104                 running_avg_y[c] = mc_running_avg_y[c];
 105                 col_sum[c] += diff;
 106             }
 107             else
 108             {
 109                 if (absdiff >= 4 + shift_inc1 && absdiff <= 7)
 110                     adjustment = adj_val[0];
 111                 else if (absdiff >= 8 && absdiff <= 15)
 112                     adjustment = adj_val[1];
 113                 else
 114                     adjustment = adj_val[2];
 115
 116                 if (diff > 0)
 117                 {
 118                     if ((sig[c] + adjustment) > 255)
 119                         running_avg_y[c] = 255;
 120                     else
 121                         running_avg_y[c] = sig[c] + adjustment;
 122
 123                     col_sum[c] += adjustment;
 124                 }
 125                 else
 126                 {
 127                     if ((sig[c] - adjustment) < 0)
 128                         running_avg_y[c] = 0;
 129                     else
 130                         running_avg_y[c] = sig[c] - adjustment;
 131
 132                     col_sum[c] -= adjustment;
 133                 }
 134             }
 135         }
 136
 137         /* Update pointers for next iteration. */
 138         sig += sig_stride;
 139         mc_running_avg_y += mc_avg_y_stride;
 140         running_avg_y += avg_y_stride;
 141     }
 142
 143     for (c = 0; c < 16; ++c) {
 144       // Below we clip the value in the same way which SSE code use.
 145       // When adopting aggressive denoiser, the adj_val for each pixel
 146       // could be at most 8 (this is current max adjustment of the map).
 147       // In SSE code, we calculate the sum of adj_val for
 148       // the columns, so the sum could be upto 128(16 rows). However,
 149       // the range of the value is -128 ~ 127 in SSE code, that's why
 150       // we do this change in C code.
 151       // We don't do this for UV denoiser, since there are only 8 rows,
 152       // and max adjustments <= 8, so the sum of the columns will not
 153       // exceed 64.
 154       if (col_sum[c] >= 128) {
 155         col_sum[c] = 127;
 156       }
 157       sum_diff += col_sum[c];
 158     }
 159
 160     sum_diff_thresh= SUM_DIFF_THRESHOLD;
 161     if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
 162     if (abs(sum_diff) > sum_diff_thresh) {
 163       // Before returning to copy the block (i.e., apply no denoising), check
 164       // if we can still apply some (weaker) temporal filtering to this block,
 165       // that would otherwise not be denoised at all. Simplest is to apply
 166       // an additional adjustment to running_avg_y to bring it closer to sig.
 167       // The adjustment is capped by a maximum delta, and chosen such that
 168       // in most cases the resulting sum_diff will be within the
 169       // accceptable range given by sum_diff_thresh.
 170
 171       // The delta is set by the excess of absolute pixel diff over threshold.
 172       int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
 173       // Only apply the adjustment for max delta up to 3.
 174       if (delta < 4) {
 175         sig -= sig_stride * 16;
 176         mc_running_avg_y -= mc_avg_y_stride * 16;
 177         running_avg_y -= avg_y_stride * 16;
 178         for (r = 0; r < 16; ++r) {
 179           for (c = 0; c < 16; ++c) {
 180             int diff = mc_running_avg_y[c] - sig[c];
 181             int adjustment = abs(diff);
 182             if (adjustment > delta)
 183               adjustment = delta;
 184             if (diff > 0) {
 185               // Bring denoised signal down.
 186               if (running_avg_y[c] - adjustment < 0)
 187                 running_avg_y[c] = 0;
 188               else
 189                 running_avg_y[c] = running_avg_y[c] - adjustment;
 190               col_sum[c] -= adjustment;
 191             } else if (diff < 0) {
 192               // Bring denoised signal up.
 193               if (running_avg_y[c] + adjustment > 255)
 194                 running_avg_y[c] = 255;
 195               else
 196                 running_avg_y[c] = running_avg_y[c] + adjustment;
 197               col_sum[c] += adjustment;
 198             }
 199           }
 200           // TODO(marpan): Check here if abs(sum_diff) has gone below the
 201           // threshold sum_diff_thresh, and if so, we can exit the row loop.
 202           sig += sig_stride;
 203           mc_running_avg_y += mc_avg_y_stride;
 204           running_avg_y += avg_y_stride;
 205         }
 206
 207         sum_diff = 0;
 208         for (c = 0; c < 16; ++c) {
 209           if (col_sum[c] >= 128) {
 210             col_sum[c] = 127;
 211           }
 212           sum_diff += col_sum[c];
 213         }
 214
 215         if (abs(sum_diff) > sum_diff_thresh)
 216           return COPY_BLOCK;
 217       } else {
 218         return COPY_BLOCK;
 219       }
 220     }
 221
 222     vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
 223     return FILTER_BLOCK;
 224 }
 225
 226 int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg_uv,
 227                              int mc_avg_uv_stride,
 228                              unsigned char *running_avg_uv,
 229                              int avg_uv_stride,
 230                              unsigned char *sig,
 231                              int sig_stride,
 232                              unsigned int motion_magnitude,
 233                              int increase_denoising) {
 234     unsigned char *running_avg_uv_start = running_avg_uv;
 235     unsigned char *sig_start = sig;
 236     int sum_diff_thresh;
 237     int r, c;
 238     int sum_diff = 0;
 239     int sum_block = 0;
 240     int adj_val[3] = {3, 4, 6};
 241     int shift_inc1 = 0;
 242     int shift_inc2 = 1;
 243     /* If motion_magnitude is small, making the denoiser more aggressive by
 244      * increasing the adjustment for each level. Add another increment for
 245      * blocks that are labeled for increase denoising. */
 246     if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
 247       if (increase_denoising) {
 248         shift_inc1 = 1;
 249         shift_inc2 = 2;
 250       }
 251       adj_val[0] += shift_inc2;
 252       adj_val[1] += shift_inc2;
 253       adj_val[2] += shift_inc2;
 254     }
 255
 256     // Avoid denoising color signal if its close to average level.
 257     for (r = 0; r < 8; ++r) {
 258       for (c = 0; c < 8; ++c) {
 259         sum_block += sig[c];
 260       }
 261       sig += sig_stride;
 262     }
 263     if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
 264       return COPY_BLOCK;
 265     }
 266
 267     sig -= sig_stride * 8;
 268     for (r = 0; r < 8; ++r) {
 269       for (c = 0; c < 8; ++c) {
 270         int diff = 0;
 271         int adjustment = 0;
 272         int absdiff = 0;
 273
 274         diff = mc_running_avg_uv[c] - sig[c];
 275         absdiff = abs(diff);
 276
 277         // When |diff| <= |3 + shift_inc1|, use pixel value from
 278         // last denoised raw.
 279         if (absdiff <= 3 + shift_inc1) {
 280           running_avg_uv[c] = mc_running_avg_uv[c];
 281           sum_diff += diff;
 282         } else {
 283           if (absdiff >= 4 && absdiff <= 7)
 284             adjustment = adj_val[0];
 285           else if (absdiff >= 8 && absdiff <= 15)
 286             adjustment = adj_val[1];
 287           else
 288             adjustment = adj_val[2];
 289           if (diff > 0) {
 290             if ((sig[c] + adjustment) > 255)
 291               running_avg_uv[c] = 255;
 292             else
 293               running_avg_uv[c] = sig[c] + adjustment;
 294             sum_diff += adjustment;
 295           } else {
 296             if ((sig[c] - adjustment) < 0)
 297               running_avg_uv[c] = 0;
 298             else
 299               running_avg_uv[c] = sig[c] - adjustment;
 300             sum_diff -= adjustment;
 301           }
 302         }
 303       }
 304       /* Update pointers for next iteration. */
 305       sig += sig_stride;
 306       mc_running_avg_uv += mc_avg_uv_stride;
 307       running_avg_uv += avg_uv_stride;
 308     }
 309
 310     sum_diff_thresh= SUM_DIFF_THRESHOLD_UV;
 311     if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
 312     if (abs(sum_diff) > sum_diff_thresh) {
 313       // Before returning to copy the block (i.e., apply no denoising), check
 314       // if we can still apply some (weaker) temporal filtering to this block,
 315       // that would otherwise not be denoised at all. Simplest is to apply
 316       // an additional adjustment to running_avg_y to bring it closer to sig.
 317       // The adjustment is capped by a maximum delta, and chosen such that
 318       // in most cases the resulting sum_diff will be within the
 319       // accceptable range given by sum_diff_thresh.
 320
 321       // The delta is set by the excess of absolute pixel diff over threshold.
 322       int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
 323       // Only apply the adjustment for max delta up to 3.
 324       if (delta < 4) {
 325         sig -= sig_stride * 8;
 326         mc_running_avg_uv -= mc_avg_uv_stride * 8;
 327         running_avg_uv -= avg_uv_stride * 8;
 328         for (r = 0; r < 8; ++r) {
 329           for (c = 0; c < 8; ++c) {
 330             int diff = mc_running_avg_uv[c] - sig[c];
 331             int adjustment = abs(diff);
 332             if (adjustment > delta)
 333               adjustment = delta;
 334             if (diff > 0) {
 335               // Bring denoised signal down.
 336               if (running_avg_uv[c] - adjustment < 0)
 337                 running_avg_uv[c] = 0;
 338               else
 339                 running_avg_uv[c] = running_avg_uv[c] - adjustment;
 340               sum_diff -= adjustment;
 341             } else if (diff < 0) {
 342               // Bring denoised signal up.
 343               if (running_avg_uv[c] + adjustment > 255)
 344                 running_avg_uv[c] = 255;
 345               else
 346                 running_avg_uv[c] = running_avg_uv[c] + adjustment;
 347               sum_diff += adjustment;
 348             }
 349           }
 350           // TODO(marpan): Check here if abs(sum_diff) has gone below the
 351           // threshold sum_diff_thresh, and if so, we can exit the row loop.
 352           sig += sig_stride;
 353           mc_running_avg_uv += mc_avg_uv_stride;
 354           running_avg_uv += avg_uv_stride;
 355         }
 356         if (abs(sum_diff) > sum_diff_thresh)
 357           return COPY_BLOCK;
 358       } else {
 359         return COPY_BLOCK;
 360       }
 361     }
 362
 363     vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start,
 364                     sig_stride);
 365     return FILTER_BLOCK;
 366 }
 367
 368 void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) {
 369   assert(mode > 0);  // Denoiser is allocated only if mode > 0.
 370   if (mode == 1) {
 371     denoiser->denoiser_mode = kDenoiserOnYOnly;
 372   } else if (mode == 2) {
 373     denoiser->denoiser_mode = kDenoiserOnYUV;
 374   } else if (mode == 3) {
 375     denoiser->denoiser_mode = kDenoiserOnYUVAggressive;
 376   } else {
 377     denoiser->denoiser_mode = kDenoiserOnAdaptive;
 378   }
 379   if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) {
 380     denoiser->denoise_pars.scale_sse_thresh = 1;
 381     denoiser->denoise_pars.scale_motion_thresh = 8;
 382     denoiser->denoise_pars.scale_increase_filter = 0;
 383     denoiser->denoise_pars.denoise_mv_bias = 95;
 384     denoiser->denoise_pars.pickmode_mv_bias = 100;
 385     denoiser->denoise_pars.qp_thresh = 0;
 386     denoiser->denoise_pars.consec_zerolast = UINT_MAX;
 387     denoiser->denoise_pars.spatial_blur = 0;
 388   } else {
 389     denoiser->denoise_pars.scale_sse_thresh = 2;
 390     denoiser->denoise_pars.scale_motion_thresh = 16;
 391     denoiser->denoise_pars.scale_increase_filter = 1;
 392     denoiser->denoise_pars.denoise_mv_bias = 60;
 393     denoiser->denoise_pars.pickmode_mv_bias = 60;
 394     denoiser->denoise_pars.qp_thresh = 100;
 395     denoiser->denoise_pars.consec_zerolast = 10;
 396     denoiser->denoise_pars.spatial_blur = 20;
 397   }
 398 }
 399
 400 int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
 401                           int num_mb_rows, int num_mb_cols, int mode)
 402 {
 403     int i;
 404     assert(denoiser);
 405     denoiser->num_mb_cols = num_mb_cols;
 406
 407     for (i = 0; i < MAX_REF_FRAMES; i++)
 408     {
 409         denoiser->yv12_running_avg[i].flags = 0;
 410
 411         if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width,
 412                                         height, VP8BORDERINPIXELS)
 413             < 0)
 414         {
 415             vp8_denoiser_free(denoiser);
 416             return 1;
 417         }
 418         vpx_memset(denoiser->yv12_running_avg[i].buffer_alloc, 0,
 419                    denoiser->yv12_running_avg[i].frame_size);
 420
 421     }
 422     denoiser->yv12_mc_running_avg.flags = 0;
 423
 424     if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width,
 425                                    height, VP8BORDERINPIXELS) < 0)
 426     {
 427         vp8_denoiser_free(denoiser);
 428         return 1;
 429     }
 430
 431     vpx_memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0,
 432                denoiser->yv12_mc_running_avg.frame_size);
 433
 434     if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width,
 435                                     height, VP8BORDERINPIXELS) < 0) {
 436       vp8_denoiser_free(denoiser);
 437       return 1;
 438     }
 439     vpx_memset(denoiser->yv12_last_source.buffer_alloc, 0,
 440                denoiser->yv12_last_source.frame_size);
 441
 442     denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1);
 443     vpx_memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols));
 444     vp8_denoiser_set_parameters(denoiser, mode);
 445     denoiser->nmse_source_diff = 0;
 446     denoiser->nmse_source_diff_count = 0;
 447     denoiser->qp_avg = 0;
 448     // QP threshold below which we can go up to aggressive mode.
 449     denoiser->qp_threshold_up = 80;
 450     // QP threshold above which we can go back down to normal mode.
 451     // For now keep this second threshold high, so not used currently.
 452     denoiser->qp_threshold_down = 128;
 453     // Bitrate thresholds and noise metric (nmse) thresholds for switching to
 454     // aggressive mode.
 455     // TODO(marpan): Adjust thresholds, including effect on resolution.
 456     denoiser->bitrate_threshold = 200000;  // (bits/sec).
 457     denoiser->threshold_aggressive_mode = 35;
 458     if (width * height > 640 * 480) {
 459       denoiser->bitrate_threshold = 500000;
 460       denoiser->threshold_aggressive_mode = 100;
 461     } else if (width * height > 960 * 540) {
 462       denoiser->bitrate_threshold = 800000;
 463       denoiser->threshold_aggressive_mode = 150;
 464     } else if (width * height > 1280 * 720) {
 465       denoiser->bitrate_threshold = 2000000;
 466       denoiser->threshold_aggressive_mode = 1400;
 467     }
 468     return 0;
 469 }
 470
 471
 472 void vp8_denoiser_free(VP8_DENOISER *denoiser)
 473 {
 474     int i;
 475     assert(denoiser);
 476
 477     for (i = 0; i < MAX_REF_FRAMES ; i++)
 478     {
 479         vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]);
 480     }
 481     vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg);
 482     vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source);
 483     vpx_free(denoiser->denoise_state);
 484 }
 485
 486
 487 void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
 488                              MACROBLOCK *x,
 489                              unsigned int best_sse,
 490                              unsigned int zero_mv_sse,
 491                              int recon_yoffset,
 492                              int recon_uvoffset,
 493                              loop_filter_info_n *lfi_n,
 494                              int mb_row,
 495                              int mb_col,
 496                              int block_index)
 497
 498 {
 499     int mv_row;
 500     int mv_col;
 501     unsigned int motion_threshold;
 502     unsigned int motion_magnitude2;
 503     unsigned int sse_thresh;
 504     int sse_diff_thresh = 0;
 505     // Spatial loop filter: only applied selectively based on
 506     // temporal filter state of block relative to top/left neighbors.
 507     int apply_spatial_loop_filter = 1;
 508     MV_REFERENCE_FRAME frame = x->best_reference_frame;
 509     MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
 510
 511     enum vp8_denoiser_decision decision = FILTER_BLOCK;
 512     enum vp8_denoiser_decision decision_u = COPY_BLOCK;
 513     enum vp8_denoiser_decision decision_v = COPY_BLOCK;
 514
 515     if (zero_frame)
 516     {
 517         YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame];
 518         YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg;
 519         YV12_BUFFER_CONFIG saved_pre,saved_dst;
 520         MB_MODE_INFO saved_mbmi;
 521         MACROBLOCKD *filter_xd = &x->e_mbd;
 522         MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi;
 523         int sse_diff = 0;
 524         // Bias on zero motion vector sse.
 525         const int zero_bias = denoiser->denoise_pars.denoise_mv_bias;
 526         zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100);
 527         sse_diff = zero_mv_sse - best_sse;
 528
 529         saved_mbmi = *mbmi;
 530
 531         /* Use the best MV for the compensation. */
 532         mbmi->ref_frame = x->best_reference_frame;
 533         mbmi->mode = x->best_sse_inter_mode;
 534         mbmi->mv = x->best_sse_mv;
 535         mbmi->need_to_clamp_mvs = x->need_to_clamp_best_mvs;
 536         mv_col = x->best_sse_mv.as_mv.col;
 537         mv_row = x->best_sse_mv.as_mv.row;
 538         // Bias to zero_mv if small amount of motion.
 539         // Note sse_diff_thresh is intialized to zero, so this ensures
 540         // we will always choose zero_mv for denoising if
 541         // zero_mv_see <= best_sse (i.e., sse_diff <= 0).
 542         if ((unsigned int)(mv_row * mv_row + mv_col * mv_col)
 543             <= NOISE_MOTION_THRESHOLD)
 544             sse_diff_thresh = (int)SSE_DIFF_THRESHOLD;
 545
 546         if (frame == INTRA_FRAME ||
 547             sse_diff <= sse_diff_thresh)
 548         {
 549             /*
 550              * Handle intra blocks as referring to last frame with zero motion
 551              * and let the absolute pixel difference affect the filter factor.
 552              * Also consider small amount of motion as being random walk due
 553              * to noise, if it doesn't mean that we get a much bigger error.
 554              * Note that any changes to the mode info only affects the
 555              * denoising.
 556              */
 557             mbmi->ref_frame =
 558                     x->best_zeromv_reference_frame;
 559
 560             src = &denoiser->yv12_running_avg[zero_frame];
 561
 562             mbmi->mode = ZEROMV;
 563             mbmi->mv.as_int = 0;
 564             x->best_sse_inter_mode = ZEROMV;
 565             x->best_sse_mv.as_int = 0;
 566             best_sse = zero_mv_sse;
 567         }
 568
 569         saved_pre = filter_xd->pre;
 570         saved_dst = filter_xd->dst;
 571
 572         /* Compensate the running average. */
 573         filter_xd->pre.y_buffer = src->y_buffer + recon_yoffset;
 574         filter_xd->pre.u_buffer = src->u_buffer + recon_uvoffset;
 575         filter_xd->pre.v_buffer = src->v_buffer + recon_uvoffset;
 576         /* Write the compensated running average to the destination buffer. */
 577         filter_xd->dst.y_buffer = dst->y_buffer + recon_yoffset;
 578         filter_xd->dst.u_buffer = dst->u_buffer + recon_uvoffset;
 579         filter_xd->dst.v_buffer = dst->v_buffer + recon_uvoffset;
 580
 581         if (!x->skip)
 582         {
 583             vp8_build_inter_predictors_mb(filter_xd);
 584         }
 585         else
 586         {
 587             vp8_build_inter16x16_predictors_mb(filter_xd,
 588                                                filter_xd->dst.y_buffer,
 589                                                filter_xd->dst.u_buffer,
 590                                                filter_xd->dst.v_buffer,
 591                                                filter_xd->dst.y_stride,
 592                                                filter_xd->dst.uv_stride);
 593         }
 594         filter_xd->pre = saved_pre;
 595         filter_xd->dst = saved_dst;
 596         *mbmi = saved_mbmi;
 597
 598     }
 599
 600     mv_row = x->best_sse_mv.as_mv.row;
 601     mv_col = x->best_sse_mv.as_mv.col;
 602     motion_magnitude2 = mv_row * mv_row + mv_col * mv_col;
 603     motion_threshold = denoiser->denoise_pars.scale_motion_thresh *
 604         NOISE_MOTION_THRESHOLD;
 605
 606     if (motion_magnitude2 <
 607         denoiser->denoise_pars.scale_increase_filter * NOISE_MOTION_THRESHOLD)
 608       x->increase_denoising = 1;
 609
 610     sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD;
 611     if (x->increase_denoising)
 612       sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD_HIGH;
 613
 614     if (best_sse > sse_thresh || motion_magnitude2 > motion_threshold)
 615       decision = COPY_BLOCK;
 616
 617     if (decision == FILTER_BLOCK)
 618     {
 619         unsigned char *mc_running_avg_y =
 620             denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset;
 621         int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride;
 622         unsigned char *running_avg_y =
 623             denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset;
 624         int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
 625
 626         /* Filter. */
 627         decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
 628                                        running_avg_y, avg_y_stride,
 629                                        x->thismb, 16, motion_magnitude2,
 630                                        x->increase_denoising);
 631         denoiser->denoise_state[block_index] = motion_magnitude2 > 0 ?
 632             kFilterNonZeroMV : kFilterZeroMV;
 633         // Only denoise UV for zero motion, and if y channel was denoised.
 634         if (denoiser->denoiser_mode != kDenoiserOnYOnly &&
 635             motion_magnitude2 == 0 &&
 636             decision == FILTER_BLOCK) {
 637           unsigned char *mc_running_avg_u =
 638               denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
 639           unsigned char *running_avg_u =
 640               denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
 641           unsigned char *mc_running_avg_v =
 642               denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
 643           unsigned char *running_avg_v =
 644               denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
 645           int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
 646           int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
 647           int signal_stride = x->block[16].src_stride;
 648           decision_u =
 649               vp8_denoiser_filter_uv(mc_running_avg_u, mc_avg_uv_stride,
 650                                       running_avg_u, avg_uv_stride,
 651                                       x->block[16].src + *x->block[16].base_src,
 652                                       signal_stride, motion_magnitude2, 0);
 653           decision_v =
 654               vp8_denoiser_filter_uv(mc_running_avg_v, mc_avg_uv_stride,
 655                                       running_avg_v, avg_uv_stride,
 656                                       x->block[20].src + *x->block[20].base_src,
 657                                       signal_stride, motion_magnitude2, 0);
 658         }
 659     }
 660     if (decision == COPY_BLOCK)
 661     {
 662         /* No filtering of this block; it differs too much from the predictor,
 663          * or the motion vector magnitude is considered too big.
 664          */
 665         vp8_copy_mem16x16(
 666                 x->thismb, 16,
 667                 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
 668                 denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
 669         denoiser->denoise_state[block_index] = kNoFilter;
 670     }
 671     if (denoiser->denoiser_mode != kDenoiserOnYOnly) {
 672       if (decision_u == COPY_BLOCK) {
 673         vp8_copy_mem8x8(
 674             x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
 675             denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
 676             denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
 677       }
 678       if (decision_v == COPY_BLOCK) {
 679         vp8_copy_mem8x8(
 680             x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
 681             denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
 682             denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
 683       }
 684     }
 685     // Option to selectively deblock the denoised signal, for y channel only.
 686     if (apply_spatial_loop_filter) {
 687       loop_filter_info lfi;
 688       int apply_filter_col = 0;
 689       int apply_filter_row = 0;
 690       int apply_filter = 0;
 691       int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
 692       int uv_stride =denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
 693
 694       // Fix filter level to some nominal value for now.
 695       int filter_level = 32;
 696
 697       int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level];
 698       lfi.mblim = lfi_n->mblim[filter_level];
 699       lfi.blim = lfi_n->blim[filter_level];
 700       lfi.lim = lfi_n->lim[filter_level];
 701       lfi.hev_thr = lfi_n->hev_thr[hev_index];
 702
 703       // Apply filter if there is a difference in the denoiser filter state
 704       // between the current and left/top block, or if non-zero motion vector
 705       // is used for the motion-compensated filtering.
 706       if (mb_col > 0) {
 707         apply_filter_col = !((denoiser->denoise_state[block_index] ==
 708             denoiser->denoise_state[block_index - 1]) &&
 709             denoiser->denoise_state[block_index] != kFilterNonZeroMV);
 710         if (apply_filter_col) {
 711           // Filter left vertical edge.
 712           apply_filter = 1;
 713           vp8_loop_filter_mbv(
 714               denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
 715               NULL, NULL, y_stride, uv_stride, &lfi);
 716         }
 717       }
 718       if (mb_row > 0) {
 719         apply_filter_row = !((denoiser->denoise_state[block_index] ==
 720             denoiser->denoise_state[block_index - denoiser->num_mb_cols]) &&
 721             denoiser->denoise_state[block_index] != kFilterNonZeroMV);
 722         if (apply_filter_row) {
 723           // Filter top horizontal edge.
 724           apply_filter = 1;
 725           vp8_loop_filter_mbh(
 726               denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
 727               NULL, NULL, y_stride, uv_stride, &lfi);
 728         }
 729       }
 730       if (apply_filter) {
 731         // Update the signal block |x|. Pixel changes are only to top and/or
 732         // left boundary pixels: can we avoid full block copy here.
 733         vp8_copy_mem16x16(
 734             denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
 735             y_stride, x->thismb, 16);
 736       }
 737     }
 738 }