Home | History | Annotate | Download | only in encoder
      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_uv,
    217                              int mc_avg_uv_stride,
    218                              unsigned char *running_avg_uv, int avg_uv_stride,
    219                              unsigned char *sig, int sig_stride,
    220                              unsigned int motion_magnitude,
    221                              int increase_denoising) {
    222   unsigned char *running_avg_uv_start = running_avg_uv;
    223   unsigned char *sig_start = sig;
    224   int sum_diff_thresh;
    225   int r, c;
    226   int sum_diff = 0;
    227   int sum_block = 0;
    228   int adj_val[3] = { 3, 4, 6 };
    229   int shift_inc1 = 0;
    230   int shift_inc2 = 1;
    231   /* If motion_magnitude is small, making the denoiser more aggressive by
    232    * increasing the adjustment for each level. Add another increment for
    233    * blocks that are labeled for increase denoising. */
    234   if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
    235     if (increase_denoising) {
    236       shift_inc1 = 1;
    237       shift_inc2 = 2;
    238     }
    239     adj_val[0] += shift_inc2;
    240     adj_val[1] += shift_inc2;
    241     adj_val[2] += shift_inc2;
    242   }
    243 
    244   // Avoid denoising color signal if its close to average level.
    245   for (r = 0; r < 8; ++r) {
    246     for (c = 0; c < 8; ++c) {
    247       sum_block += sig[c];
    248     }
    249     sig += sig_stride;
    250   }
    251   if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
    252     return COPY_BLOCK;
    253   }
    254 
    255   sig -= sig_stride * 8;
    256   for (r = 0; r < 8; ++r) {
    257     for (c = 0; c < 8; ++c) {
    258       int diff = 0;
    259       int adjustment = 0;
    260       int absdiff = 0;
    261 
    262       diff = mc_running_avg_uv[c] - sig[c];
    263       absdiff = abs(diff);
    264 
    265       // When |diff| <= |3 + shift_inc1|, use pixel value from
    266       // last denoised raw.
    267       if (absdiff <= 3 + shift_inc1) {
    268         running_avg_uv[c] = mc_running_avg_uv[c];
    269         sum_diff += diff;
    270       } else {
    271         if (absdiff >= 4 && absdiff <= 7) {
    272           adjustment = adj_val[0];
    273         } else if (absdiff >= 8 && absdiff <= 15) {
    274           adjustment = adj_val[1];
    275         } else {
    276           adjustment = adj_val[2];
    277         }
    278         if (diff > 0) {
    279           if ((sig[c] + adjustment) > 255) {
    280             running_avg_uv[c] = 255;
    281           } else {
    282             running_avg_uv[c] = sig[c] + adjustment;
    283           }
    284           sum_diff += adjustment;
    285         } else {
    286           if ((sig[c] - adjustment) < 0) {
    287             running_avg_uv[c] = 0;
    288           } else {
    289             running_avg_uv[c] = sig[c] - adjustment;
    290           }
    291           sum_diff -= adjustment;
    292         }
    293       }
    294     }
    295     /* Update pointers for next iteration. */
    296     sig += sig_stride;
    297     mc_running_avg_uv += mc_avg_uv_stride;
    298     running_avg_uv += avg_uv_stride;
    299   }
    300 
    301   sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
    302   if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
    303   if (abs(sum_diff) > sum_diff_thresh) {
    304     // Before returning to copy the block (i.e., apply no denoising), check
    305     // if we can still apply some (weaker) temporal filtering to this block,
    306     // that would otherwise not be denoised at all. Simplest is to apply
    307     // an additional adjustment to running_avg_y to bring it closer to sig.
    308     // The adjustment is capped by a maximum delta, and chosen such that
    309     // in most cases the resulting sum_diff will be within the
    310     // accceptable range given by sum_diff_thresh.
    311 
    312     // The delta is set by the excess of absolute pixel diff over threshold.
    313     int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
    314     // Only apply the adjustment for max delta up to 3.
    315     if (delta < 4) {
    316       sig -= sig_stride * 8;
    317       mc_running_avg_uv -= mc_avg_uv_stride * 8;
    318       running_avg_uv -= avg_uv_stride * 8;
    319       for (r = 0; r < 8; ++r) {
    320         for (c = 0; c < 8; ++c) {
    321           int diff = mc_running_avg_uv[c] - sig[c];
    322           int adjustment = abs(diff);
    323           if (adjustment > delta) adjustment = delta;
    324           if (diff > 0) {
    325             // Bring denoised signal down.
    326             if (running_avg_uv[c] - adjustment < 0) {
    327               running_avg_uv[c] = 0;
    328             } else {
    329               running_avg_uv[c] = running_avg_uv[c] - adjustment;
    330             }
    331             sum_diff -= adjustment;
    332           } else if (diff < 0) {
    333             // Bring denoised signal up.
    334             if (running_avg_uv[c] + adjustment > 255) {
    335               running_avg_uv[c] = 255;
    336             } else {
    337               running_avg_uv[c] = running_avg_uv[c] + adjustment;
    338             }
    339             sum_diff += adjustment;
    340           }
    341         }
    342         // TODO(marpan): Check here if abs(sum_diff) has gone below the
    343         // threshold sum_diff_thresh, and if so, we can exit the row loop.
    344         sig += sig_stride;
    345         mc_running_avg_uv += mc_avg_uv_stride;
    346         running_avg_uv += avg_uv_stride;
    347       }
    348       if (abs(sum_diff) > sum_diff_thresh) return COPY_BLOCK;
    349     } else {
    350       return COPY_BLOCK;
    351     }
    352   }
    353 
    354   vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start, sig_stride);
    355   return FILTER_BLOCK;
    356 }
    357 
    358 void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) {
    359   assert(mode > 0);  // Denoiser is allocated only if mode > 0.
    360   if (mode == 1) {
    361     denoiser->denoiser_mode = kDenoiserOnYOnly;
    362   } else if (mode == 2) {
    363     denoiser->denoiser_mode = kDenoiserOnYUV;
    364   } else if (mode == 3) {
    365     denoiser->denoiser_mode = kDenoiserOnYUVAggressive;
    366   } else {
    367     denoiser->denoiser_mode = kDenoiserOnYUV;
    368   }
    369   if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) {
    370     denoiser->denoise_pars.scale_sse_thresh = 1;
    371     denoiser->denoise_pars.scale_motion_thresh = 8;
    372     denoiser->denoise_pars.scale_increase_filter = 0;
    373     denoiser->denoise_pars.denoise_mv_bias = 95;
    374     denoiser->denoise_pars.pickmode_mv_bias = 100;
    375     denoiser->denoise_pars.qp_thresh = 0;
    376     denoiser->denoise_pars.consec_zerolast = UINT_MAX;
    377     denoiser->denoise_pars.spatial_blur = 0;
    378   } else {
    379     denoiser->denoise_pars.scale_sse_thresh = 2;
    380     denoiser->denoise_pars.scale_motion_thresh = 16;
    381     denoiser->denoise_pars.scale_increase_filter = 1;
    382     denoiser->denoise_pars.denoise_mv_bias = 60;
    383     denoiser->denoise_pars.pickmode_mv_bias = 75;
    384     denoiser->denoise_pars.qp_thresh = 80;
    385     denoiser->denoise_pars.consec_zerolast = 15;
    386     denoiser->denoise_pars.spatial_blur = 0;
    387   }
    388 }
    389 
    390 int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
    391                           int num_mb_rows, int num_mb_cols, int mode) {
    392   int i;
    393   assert(denoiser);
    394   denoiser->num_mb_cols = num_mb_cols;
    395 
    396   for (i = 0; i < MAX_REF_FRAMES; ++i) {
    397     denoiser->yv12_running_avg[i].flags = 0;
    398 
    399     if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width,
    400                                     height, VP8BORDERINPIXELS) < 0) {
    401       vp8_denoiser_free(denoiser);
    402       return 1;
    403     }
    404     memset(denoiser->yv12_running_avg[i].buffer_alloc, 0,
    405            denoiser->yv12_running_avg[i].frame_size);
    406   }
    407   denoiser->yv12_mc_running_avg.flags = 0;
    408 
    409   if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width,
    410                                   height, VP8BORDERINPIXELS) < 0) {
    411     vp8_denoiser_free(denoiser);
    412     return 1;
    413   }
    414 
    415   memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0,
    416          denoiser->yv12_mc_running_avg.frame_size);
    417 
    418   if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width, height,
    419                                   VP8BORDERINPIXELS) < 0) {
    420     vp8_denoiser_free(denoiser);
    421     return 1;
    422   }
    423   memset(denoiser->yv12_last_source.buffer_alloc, 0,
    424          denoiser->yv12_last_source.frame_size);
    425 
    426   denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1);
    427   if (!denoiser->denoise_state) {
    428     vp8_denoiser_free(denoiser);
    429     return 1;
    430   }
    431   memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols));
    432   vp8_denoiser_set_parameters(denoiser, mode);
    433   denoiser->nmse_source_diff = 0;
    434   denoiser->nmse_source_diff_count = 0;
    435   denoiser->qp_avg = 0;
    436   // QP threshold below which we can go up to aggressive mode.
    437   denoiser->qp_threshold_up = 80;
    438   // QP threshold above which we can go back down to normal mode.
    439   // For now keep this second threshold high, so not used currently.
    440   denoiser->qp_threshold_down = 128;
    441   // Bitrate thresholds and noise metric (nmse) thresholds for switching to
    442   // aggressive mode.
    443   // TODO(marpan): Adjust thresholds, including effect on resolution.
    444   denoiser->bitrate_threshold = 400000;  // (bits/sec).
    445   denoiser->threshold_aggressive_mode = 80;
    446   if (width * height > 1280 * 720) {
    447     denoiser->bitrate_threshold = 3000000;
    448     denoiser->threshold_aggressive_mode = 200;
    449   } else if (width * height > 960 * 540) {
    450     denoiser->bitrate_threshold = 1200000;
    451     denoiser->threshold_aggressive_mode = 120;
    452   } else if (width * height > 640 * 480) {
    453     denoiser->bitrate_threshold = 600000;
    454     denoiser->threshold_aggressive_mode = 100;
    455   }
    456   return 0;
    457 }
    458 
    459 void vp8_denoiser_free(VP8_DENOISER *denoiser) {
    460   int i;
    461   assert(denoiser);
    462 
    463   for (i = 0; i < MAX_REF_FRAMES; ++i) {
    464     vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]);
    465   }
    466   vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg);
    467   vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source);
    468   vpx_free(denoiser->denoise_state);
    469 }
    470 
    471 void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser, MACROBLOCK *x,
    472                              unsigned int best_sse, unsigned int zero_mv_sse,
    473                              int recon_yoffset, int recon_uvoffset,
    474                              loop_filter_info_n *lfi_n, int mb_row, int mb_col,
    475                              int block_index, int consec_zero_last)
    476 
    477 {
    478   int mv_row;
    479   int mv_col;
    480   unsigned int motion_threshold;
    481   unsigned int motion_magnitude2;
    482   unsigned int sse_thresh;
    483   int sse_diff_thresh = 0;
    484   // Spatial loop filter: only applied selectively based on
    485   // temporal filter state of block relative to top/left neighbors.
    486   int apply_spatial_loop_filter = 1;
    487   MV_REFERENCE_FRAME frame = x->best_reference_frame;
    488   MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
    489 
    490   enum vp8_denoiser_decision decision = FILTER_BLOCK;
    491   enum vp8_denoiser_decision decision_u = COPY_BLOCK;
    492   enum vp8_denoiser_decision decision_v = COPY_BLOCK;
    493 
    494   if (zero_frame) {
    495     YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame];
    496     YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg;
    497     YV12_BUFFER_CONFIG saved_pre, saved_dst;
    498     MB_MODE_INFO saved_mbmi;
    499     MACROBLOCKD *filter_xd = &x->e_mbd;
    500     MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi;
    501     int sse_diff = 0;
    502     // Bias on zero motion vector sse.
    503     const int zero_bias = denoiser->denoise_pars.denoise_mv_bias;
    504     zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100);
    505     sse_diff = (int)zero_mv_sse - (int)best_sse;
    506 
    507     saved_mbmi = *mbmi;
    508 
    509     /* Use the best MV for the compensation. */
    510     mbmi->ref_frame = x->best_reference_frame;
    511     mbmi->mode = x->best_sse_inter_mode;
    512     mbmi->mv = x->best_sse_mv;
    513     mbmi->need_to_clamp_mvs = x->need_to_clamp_best_mvs;
    514     mv_col = x->best_sse_mv.as_mv.col;
    515     mv_row = x->best_sse_mv.as_mv.row;
    516     // Bias to zero_mv if small amount of motion.
    517     // Note sse_diff_thresh is intialized to zero, so this ensures
    518     // we will always choose zero_mv for denoising if
    519     // zero_mv_see <= best_sse (i.e., sse_diff <= 0).
    520     if ((unsigned int)(mv_row * mv_row + mv_col * mv_col) <=
    521         NOISE_MOTION_THRESHOLD) {
    522       sse_diff_thresh = (int)SSE_DIFF_THRESHOLD;
    523     }
    524 
    525     if (frame == INTRA_FRAME || sse_diff <= sse_diff_thresh) {
    526       /*
    527        * Handle intra blocks as referring to last frame with zero motion
    528        * and let the absolute pixel difference affect the filter factor.
    529        * Also consider small amount of motion as being random walk due
    530        * to noise, if it doesn't mean that we get a much bigger error.
    531        * Note that any changes to the mode info only affects the
    532        * denoising.
    533        */
    534       x->denoise_zeromv = 1;
    535       mbmi->ref_frame = x->best_zeromv_reference_frame;
    536 
    537       src = &denoiser->yv12_running_avg[zero_frame];
    538 
    539       mbmi->mode = ZEROMV;
    540       mbmi->mv.as_int = 0;
    541       x->best_sse_inter_mode = ZEROMV;
    542       x->best_sse_mv.as_int = 0;
    543       best_sse = zero_mv_sse;
    544     }
    545 
    546     mv_row = x->best_sse_mv.as_mv.row;
    547     mv_col = x->best_sse_mv.as_mv.col;
    548     motion_magnitude2 = mv_row * mv_row + mv_col * mv_col;
    549     motion_threshold =
    550         denoiser->denoise_pars.scale_motion_thresh * NOISE_MOTION_THRESHOLD;
    551 
    552     if (motion_magnitude2 <
    553         denoiser->denoise_pars.scale_increase_filter * NOISE_MOTION_THRESHOLD) {
    554       x->increase_denoising = 1;
    555     }
    556 
    557     sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD;
    558     if (x->increase_denoising) {
    559       sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD_HIGH;
    560     }
    561 
    562     if (best_sse > sse_thresh || motion_magnitude2 > motion_threshold) {
    563       decision = COPY_BLOCK;
    564     }
    565 
    566     // If block is considered skin, don't denoise if the block
    567     // (1) is selected as non-zero motion for current frame, or
    568     // (2) has not been selected as ZERO_LAST mode at least x past frames
    569     // in a row.
    570     // TODO(marpan): Parameter "x" should be varied with framerate.
    571     // In particualar, should be reduced for layers (base layer/LAST).
    572     if (x->is_skin && (consec_zero_last < 2 || motion_magnitude2 > 0)) {
    573       decision = COPY_BLOCK;
    574     }
    575 
    576     if (decision == FILTER_BLOCK) {
    577       saved_pre = filter_xd->pre;
    578       saved_dst = filter_xd->dst;
    579 
    580       /* Compensate the running average. */
    581       filter_xd->pre.y_buffer = src->y_buffer + recon_yoffset;
    582       filter_xd->pre.u_buffer = src->u_buffer + recon_uvoffset;
    583       filter_xd->pre.v_buffer = src->v_buffer + recon_uvoffset;
    584       /* Write the compensated running average to the destination buffer. */
    585       filter_xd->dst.y_buffer = dst->y_buffer + recon_yoffset;
    586       filter_xd->dst.u_buffer = dst->u_buffer + recon_uvoffset;
    587       filter_xd->dst.v_buffer = dst->v_buffer + recon_uvoffset;
    588 
    589       if (!x->skip) {
    590         vp8_build_inter_predictors_mb(filter_xd);
    591       } else {
    592         vp8_build_inter16x16_predictors_mb(
    593             filter_xd, filter_xd->dst.y_buffer, filter_xd->dst.u_buffer,
    594             filter_xd->dst.v_buffer, filter_xd->dst.y_stride,
    595             filter_xd->dst.uv_stride);
    596       }
    597       filter_xd->pre = saved_pre;
    598       filter_xd->dst = saved_dst;
    599       *mbmi = saved_mbmi;
    600     }
    601   } else {
    602     // zero_frame should always be 1 for real-time mode, as the
    603     // ZEROMV mode is always checked, so we should never go into this branch.
    604     // If case ZEROMV is not checked, then we will force no denoise (COPY).
    605     decision = COPY_BLOCK;
    606   }
    607 
    608   if (decision == FILTER_BLOCK) {
    609     unsigned char *mc_running_avg_y =
    610         denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset;
    611     int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride;
    612     unsigned char *running_avg_y =
    613         denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset;
    614     int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
    615 
    616     /* Filter. */
    617     decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
    618                                    running_avg_y, avg_y_stride, x->thismb, 16,
    619                                    motion_magnitude2, x->increase_denoising);
    620     denoiser->denoise_state[block_index] =
    621         motion_magnitude2 > 0 ? kFilterNonZeroMV : kFilterZeroMV;
    622     // Only denoise UV for zero motion, and if y channel was denoised.
    623     if (denoiser->denoiser_mode != kDenoiserOnYOnly && motion_magnitude2 == 0 &&
    624         decision == FILTER_BLOCK) {
    625       unsigned char *mc_running_avg_u =
    626           denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
    627       unsigned char *running_avg_u =
    628           denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
    629       unsigned char *mc_running_avg_v =
    630           denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
    631       unsigned char *running_avg_v =
    632           denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
    633       int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
    634       int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
    635       int signal_stride = x->block[16].src_stride;
    636       decision_u = vp8_denoiser_filter_uv(
    637           mc_running_avg_u, mc_avg_uv_stride, running_avg_u, avg_uv_stride,
    638           x->block[16].src + *x->block[16].base_src, signal_stride,
    639           motion_magnitude2, 0);
    640       decision_v = vp8_denoiser_filter_uv(
    641           mc_running_avg_v, mc_avg_uv_stride, running_avg_v, avg_uv_stride,
    642           x->block[20].src + *x->block[20].base_src, signal_stride,
    643           motion_magnitude2, 0);
    644     }
    645   }
    646   if (decision == COPY_BLOCK) {
    647     /* No filtering of this block; it differs too much from the predictor,
    648      * or the motion vector magnitude is considered too big.
    649      */
    650     x->denoise_zeromv = 0;
    651     vp8_copy_mem16x16(
    652         x->thismb, 16,
    653         denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
    654         denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
    655     denoiser->denoise_state[block_index] = kNoFilter;
    656   }
    657   if (denoiser->denoiser_mode != kDenoiserOnYOnly) {
    658     if (decision_u == COPY_BLOCK) {
    659       vp8_copy_mem8x8(
    660           x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
    661           denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
    662           denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
    663     }
    664     if (decision_v == COPY_BLOCK) {
    665       vp8_copy_mem8x8(
    666           x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
    667           denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
    668           denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
    669     }
    670   }
    671   // Option to selectively deblock the denoised signal, for y channel only.
    672   if (apply_spatial_loop_filter) {
    673     loop_filter_info lfi;
    674     int apply_filter_col = 0;
    675     int apply_filter_row = 0;
    676     int apply_filter = 0;
    677     int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
    678     int uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
    679 
    680     // Fix filter level to some nominal value for now.
    681     int filter_level = 48;
    682 
    683     int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level];
    684     lfi.mblim = lfi_n->mblim[filter_level];
    685     lfi.blim = lfi_n->blim[filter_level];
    686     lfi.lim = lfi_n->lim[filter_level];
    687     lfi.hev_thr = lfi_n->hev_thr[hev_index];
    688 
    689     // Apply filter if there is a difference in the denoiser filter state
    690     // between the current and left/top block, or if non-zero motion vector
    691     // is used for the motion-compensated filtering.
    692     if (mb_col > 0) {
    693       apply_filter_col =
    694           !((denoiser->denoise_state[block_index] ==
    695              denoiser->denoise_state[block_index - 1]) &&
    696             denoiser->denoise_state[block_index] != kFilterNonZeroMV);
    697       if (apply_filter_col) {
    698         // Filter left vertical edge.
    699         apply_filter = 1;
    700         vp8_loop_filter_mbv(
    701             denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
    702             NULL, NULL, y_stride, uv_stride, &lfi);
    703       }
    704     }
    705     if (mb_row > 0) {
    706       apply_filter_row =
    707           !((denoiser->denoise_state[block_index] ==
    708              denoiser->denoise_state[block_index - denoiser->num_mb_cols]) &&
    709             denoiser->denoise_state[block_index] != kFilterNonZeroMV);
    710       if (apply_filter_row) {
    711         // Filter top horizontal edge.
    712         apply_filter = 1;
    713         vp8_loop_filter_mbh(
    714             denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
    715             NULL, NULL, y_stride, uv_stride, &lfi);
    716       }
    717     }
    718     if (apply_filter) {
    719       // Update the signal block |x|. Pixel changes are only to top and/or
    720       // left boundary pixels: can we avoid full block copy here.
    721       vp8_copy_mem16x16(
    722           denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
    723           y_stride, x->thismb, 16);
    724     }
    725   }
    726 }
    727