aom_dsp/loopfilter.c

   1 /*
   2  * Copyright (c) 2016, Alliance for Open Media. All rights reserved
   3  *
   4  * This source code is subject to the terms of the BSD 2 Clause License and
   5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
   6  * was not distributed with this source code in the LICENSE file, you can
   7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
   8  * Media Patent License 1.0 was not distributed with this source code in the
   9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  10  */
  11
  12 #include <stdlib.h>
  13
  14 #include "config/aom_config.h"
  15 #include "config/aom_dsp_rtcd.h"
  16
  17 #include "aom_dsp/aom_dsp_common.h"
  18 #include "aom_ports/mem.h"
  19
  20 static INLINE int8_t signed_char_clamp(int t) {
  21   return (int8_t)clamp(t, -128, 127);
  22 }
  23
  24 #if CONFIG_AV1_HIGHBITDEPTH
  25 static INLINE int16_t signed_char_clamp_high(int t, int bd) {
  26   switch (bd) {
  27     case 10: return (int16_t)clamp(t, -128 * 4, 128 * 4 - 1);
  28     case 12: return (int16_t)clamp(t, -128 * 16, 128 * 16 - 1);
  29     case 8:
  30     default: return (int16_t)clamp(t, -128, 128 - 1);
  31   }
  32 }
  33 #endif
  34
  35 // should we apply any filter at all: 11111111 yes, 00000000 no
  36 static INLINE int8_t filter_mask2(uint8_t limit, uint8_t blimit, uint8_t p1,
  37                                   uint8_t p0, uint8_t q0, uint8_t q1) {
  38   int8_t mask = 0;
  39   mask |= (abs(p1 - p0) > limit) * -1;
  40   mask |= (abs(q1 - q0) > limit) * -1;
  41   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
  42   return ~mask;
  43 }
  44
  45 static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit, uint8_t p3,
  46                                  uint8_t p2, uint8_t p1, uint8_t p0, uint8_t q0,
  47                                  uint8_t q1, uint8_t q2, uint8_t q3) {
  48   int8_t mask = 0;
  49   mask |= (abs(p3 - p2) > limit) * -1;
  50   mask |= (abs(p2 - p1) > limit) * -1;
  51   mask |= (abs(p1 - p0) > limit) * -1;
  52   mask |= (abs(q1 - q0) > limit) * -1;
  53   mask |= (abs(q2 - q1) > limit) * -1;
  54   mask |= (abs(q3 - q2) > limit) * -1;
  55   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
  56   return ~mask;
  57 }
  58
  59 static INLINE int8_t filter_mask3_chroma(uint8_t limit, uint8_t blimit,
  60                                          uint8_t p2, uint8_t p1, uint8_t p0,
  61                                          uint8_t q0, uint8_t q1, uint8_t q2) {
  62   int8_t mask = 0;
  63   mask |= (abs(p2 - p1) > limit) * -1;
  64   mask |= (abs(p1 - p0) > limit) * -1;
  65   mask |= (abs(q1 - q0) > limit) * -1;
  66   mask |= (abs(q2 - q1) > limit) * -1;
  67   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
  68   return ~mask;
  69 }
  70
  71 static INLINE int8_t flat_mask3_chroma(uint8_t thresh, uint8_t p2, uint8_t p1,
  72                                        uint8_t p0, uint8_t q0, uint8_t q1,
  73                                        uint8_t q2) {
  74   int8_t mask = 0;
  75   mask |= (abs(p1 - p0) > thresh) * -1;
  76   mask |= (abs(q1 - q0) > thresh) * -1;
  77   mask |= (abs(p2 - p0) > thresh) * -1;
  78   mask |= (abs(q2 - q0) > thresh) * -1;
  79   return ~mask;
  80 }
  81
  82 static INLINE int8_t flat_mask4(uint8_t thresh, uint8_t p3, uint8_t p2,
  83                                 uint8_t p1, uint8_t p0, uint8_t q0, uint8_t q1,
  84                                 uint8_t q2, uint8_t q3) {
  85   int8_t mask = 0;
  86   mask |= (abs(p1 - p0) > thresh) * -1;
  87   mask |= (abs(q1 - q0) > thresh) * -1;
  88   mask |= (abs(p2 - p0) > thresh) * -1;
  89   mask |= (abs(q2 - q0) > thresh) * -1;
  90   mask |= (abs(p3 - p0) > thresh) * -1;
  91   mask |= (abs(q3 - q0) > thresh) * -1;
  92   return ~mask;
  93 }
  94
  95 // is there high edge variance internal edge: 11111111 yes, 00000000 no
  96 static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
  97                               uint8_t q0, uint8_t q1) {
  98   int8_t hev = 0;
  99   hev |= (abs(p1 - p0) > thresh) * -1;
 100   hev |= (abs(q1 - q0) > thresh) * -1;
 101   return hev;
 102 }
 103
 104 static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
 105                            uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
 106   int8_t filter1, filter2;
 107
 108   const int8_t ps1 = (int8_t)(*op1 ^ 0x80);
 109   const int8_t ps0 = (int8_t)(*op0 ^ 0x80);
 110   const int8_t qs0 = (int8_t)(*oq0 ^ 0x80);
 111   const int8_t qs1 = (int8_t)(*oq1 ^ 0x80);
 112   const int8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
 113
 114   // add outer taps if we have high edge variance
 115   int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
 116
 117   // inner taps
 118   filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
 119
 120   // save bottom 3 bits so that we round one side +4 and the other +3
 121   // if it equals 4 we'll set to adjust by -1 to account for the fact
 122   // we'd round 3 the other way
 123   filter1 = signed_char_clamp(filter + 4) >> 3;
 124   filter2 = signed_char_clamp(filter + 3) >> 3;
 125
 126   *oq0 = (uint8_t)(signed_char_clamp(qs0 - filter1) ^ 0x80);
 127   *op0 = (uint8_t)(signed_char_clamp(ps0 + filter2) ^ 0x80);
 128
 129   // outer tap adjustments
 130   filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
 131
 132   *oq1 = (uint8_t)(signed_char_clamp(qs1 - filter) ^ 0x80);
 133   *op1 = (uint8_t)(signed_char_clamp(ps1 + filter) ^ 0x80);
 134 }
 135
 136 void aom_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
 137                             const uint8_t *blimit, const uint8_t *limit,
 138                             const uint8_t *thresh) {
 139   int i;
 140   int count = 4;
 141
 142   // loop filter designed to work using chars so that we can make maximum use
 143   // of 8 bit simd instructions.
 144   for (i = 0; i < count; ++i) {
 145     const uint8_t p1 = s[-2 * p], p0 = s[-p];
 146     const uint8_t q0 = s[0 * p], q1 = s[1 * p];
 147     const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1);
 148     filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
 149     ++s;
 150   }
 151 }
 152
 153 void aom_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
 154                                  const uint8_t *limit0, const uint8_t *thresh0,
 155                                  const uint8_t *blimit1, const uint8_t *limit1,
 156                                  const uint8_t *thresh1) {
 157   aom_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0);
 158   aom_lpf_horizontal_4_c(s + 4, p, blimit1, limit1, thresh1);
 159 }
 160
 161 void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
 162                           const uint8_t *limit, const uint8_t *thresh) {
 163   int i;
 164   int count = 4;
 165
 166   // loop filter designed to work using chars so that we can make maximum use
 167   // of 8 bit simd instructions.
 168   for (i = 0; i < count; ++i) {
 169     const uint8_t p1 = s[-2], p0 = s[-1];
 170     const uint8_t q0 = s[0], q1 = s[1];
 171     const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1);
 172     filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
 173     s += pitch;
 174   }
 175 }
 176
 177 void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
 178                                const uint8_t *limit0, const uint8_t *thresh0,
 179                                const uint8_t *blimit1, const uint8_t *limit1,
 180                                const uint8_t *thresh1) {
 181   aom_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0);
 182   aom_lpf_vertical_4_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1);
 183 }
 184
 185 static INLINE void filter6(int8_t mask, uint8_t thresh, int8_t flat,
 186                            uint8_t *op2, uint8_t *op1, uint8_t *op0,
 187                            uint8_t *oq0, uint8_t *oq1, uint8_t *oq2) {
 188   if (flat && mask) {
 189     const uint8_t p2 = *op2, p1 = *op1, p0 = *op0;
 190     const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2;
 191
 192     // 5-tap filter [1, 2, 2, 2, 1]
 193     *op1 = ROUND_POWER_OF_TWO(p2 * 3 + p1 * 2 + p0 * 2 + q0, 3);
 194     *op0 = ROUND_POWER_OF_TWO(p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1, 3);
 195     *oq0 = ROUND_POWER_OF_TWO(p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2, 3);
 196     *oq1 = ROUND_POWER_OF_TWO(p0 + q0 * 2 + q1 * 2 + q2 * 3, 3);
 197   } else {
 198     filter4(mask, thresh, op1, op0, oq0, oq1);
 199   }
 200 }
 201
 202 static INLINE void filter8(int8_t mask, uint8_t thresh, int8_t flat,
 203                            uint8_t *op3, uint8_t *op2, uint8_t *op1,
 204                            uint8_t *op0, uint8_t *oq0, uint8_t *oq1,
 205                            uint8_t *oq2, uint8_t *oq3) {
 206   if (flat && mask) {
 207     const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
 208     const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
 209
 210     // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
 211     *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
 212     *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
 213     *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
 214     *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
 215     *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
 216     *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
 217   } else {
 218     filter4(mask, thresh, op1, op0, oq0, oq1);
 219   }
 220 }
 221
 222 void aom_lpf_horizontal_6_c(uint8_t *s, int p, const uint8_t *blimit,
 223                             const uint8_t *limit, const uint8_t *thresh) {
 224   int i;
 225   int count = 4;
 226
 227   // loop filter designed to work using chars so that we can make maximum use
 228   // of 8 bit simd instructions.
 229   for (i = 0; i < count; ++i) {
 230     const uint8_t p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
 231     const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p];
 232
 233     const int8_t mask =
 234         filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2);
 235     const int8_t flat = flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2);
 236     filter6(mask, *thresh, flat, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p,
 237             s + 2 * p);
 238     ++s;
 239   }
 240 }
 241
 242 void aom_lpf_horizontal_6_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
 243                                  const uint8_t *limit0, const uint8_t *thresh0,
 244                                  const uint8_t *blimit1, const uint8_t *limit1,
 245                                  const uint8_t *thresh1) {
 246   aom_lpf_horizontal_6_c(s, p, blimit0, limit0, thresh0);
 247   aom_lpf_horizontal_6_c(s + 4, p, blimit1, limit1, thresh1);
 248 }
 249
 250 void aom_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
 251                             const uint8_t *limit, const uint8_t *thresh) {
 252   int i;
 253   int count = 4;
 254
 255   // loop filter designed to work using chars so that we can make maximum use
 256   // of 8 bit simd instructions.
 257   for (i = 0; i < count; ++i) {
 258     const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
 259     const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
 260
 261     const int8_t mask =
 262         filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
 263     const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
 264     filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s,
 265             s + 1 * p, s + 2 * p, s + 3 * p);
 266     ++s;
 267   }
 268 }
 269
 270 void aom_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
 271                                  const uint8_t *limit0, const uint8_t *thresh0,
 272                                  const uint8_t *blimit1, const uint8_t *limit1,
 273                                  const uint8_t *thresh1) {
 274   aom_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0);
 275   aom_lpf_horizontal_8_c(s + 4, p, blimit1, limit1, thresh1);
 276 }
 277
 278 void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit,
 279                           const uint8_t *limit, const uint8_t *thresh) {
 280   int i;
 281   int count = 4;
 282
 283   for (i = 0; i < count; ++i) {
 284     const uint8_t p2 = s[-3], p1 = s[-2], p0 = s[-1];
 285     const uint8_t q0 = s[0], q1 = s[1], q2 = s[2];
 286     const int8_t mask =
 287         filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2);
 288     const int8_t flat = flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2);
 289     filter6(mask, *thresh, flat, s - 3, s - 2, s - 1, s, s + 1, s + 2);
 290     s += pitch;
 291   }
 292 }
 293
 294 void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
 295                                const uint8_t *limit0, const uint8_t *thresh0,
 296                                const uint8_t *blimit1, const uint8_t *limit1,
 297                                const uint8_t *thresh1) {
 298   aom_lpf_vertical_6_c(s, pitch, blimit0, limit0, thresh0);
 299   aom_lpf_vertical_6_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1);
 300 }
 301
 302 void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
 303                           const uint8_t *limit, const uint8_t *thresh) {
 304   int i;
 305   int count = 4;
 306
 307   for (i = 0; i < count; ++i) {
 308     const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
 309     const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
 310     const int8_t mask =
 311         filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
 312     const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
 313     filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1, s + 2,
 314             s + 3);
 315     s += pitch;
 316   }
 317 }
 318
 319 void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
 320                                const uint8_t *limit0, const uint8_t *thresh0,
 321                                const uint8_t *blimit1, const uint8_t *limit1,
 322                                const uint8_t *thresh1) {
 323   aom_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0);
 324   aom_lpf_vertical_8_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1);
 325 }
 326
 327 static INLINE void filter14(int8_t mask, uint8_t thresh, int8_t flat,
 328                             int8_t flat2, uint8_t *op6, uint8_t *op5,
 329                             uint8_t *op4, uint8_t *op3, uint8_t *op2,
 330                             uint8_t *op1, uint8_t *op0, uint8_t *oq0,
 331                             uint8_t *oq1, uint8_t *oq2, uint8_t *oq3,
 332                             uint8_t *oq4, uint8_t *oq5, uint8_t *oq6) {
 333   if (flat2 && flat && mask) {
 334     const uint8_t p6 = *op6, p5 = *op5, p4 = *op4, p3 = *op3, p2 = *op2,
 335                   p1 = *op1, p0 = *op0;
 336     const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, q4 = *oq4,
 337                   q5 = *oq5, q6 = *oq6;
 338
 339     // 13-tap filter [1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1]
 340     *op5 = ROUND_POWER_OF_TWO(p6 * 7 + p5 * 2 + p4 * 2 + p3 + p2 + p1 + p0 + q0,
 341                               4);
 342     *op4 = ROUND_POWER_OF_TWO(
 343         p6 * 5 + p5 * 2 + p4 * 2 + p3 * 2 + p2 + p1 + p0 + q0 + q1, 4);
 344     *op3 = ROUND_POWER_OF_TWO(
 345         p6 * 4 + p5 + p4 * 2 + p3 * 2 + p2 * 2 + p1 + p0 + q0 + q1 + q2, 4);
 346     *op2 = ROUND_POWER_OF_TWO(
 347         p6 * 3 + p5 + p4 + p3 * 2 + p2 * 2 + p1 * 2 + p0 + q0 + q1 + q2 + q3,
 348         4);
 349     *op1 = ROUND_POWER_OF_TWO(p6 * 2 + p5 + p4 + p3 + p2 * 2 + p1 * 2 + p0 * 2 +
 350                                   q0 + q1 + q2 + q3 + q4,
 351                               4);
 352     *op0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 * 2 +
 353                                   q0 * 2 + q1 + q2 + q3 + q4 + q5,
 354                               4);
 355     *oq0 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 * 2 +
 356                                   q1 * 2 + q2 + q3 + q4 + q5 + q6,
 357                               4);
 358     *oq1 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 * 2 +
 359                                   q2 * 2 + q3 + q4 + q5 + q6 * 2,
 360                               4);
 361     *oq2 = ROUND_POWER_OF_TWO(
 362         p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 * 2 + q3 * 2 + q4 + q5 + q6 * 3,
 363         4);
 364     *oq3 = ROUND_POWER_OF_TWO(
 365         p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 * 2 + q4 * 2 + q5 + q6 * 4, 4);
 366     *oq4 = ROUND_POWER_OF_TWO(
 367         p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 * 2 + q5 * 2 + q6 * 5, 4);
 368     *oq5 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 2 + q6 * 7,
 369                               4);
 370   } else {
 371     filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
 372   }
 373 }
 374
 375 static void mb_lpf_horizontal_edge_w(uint8_t *s, int p, const uint8_t *blimit,
 376                                      const uint8_t *limit,
 377                                      const uint8_t *thresh, int count) {
 378   int i;
 379   int step = 4;
 380
 381   // loop filter designed to work using chars so that we can make maximum use
 382   // of 8 bit simd instructions.
 383   for (i = 0; i < step * count; ++i) {
 384     const uint8_t p6 = s[-7 * p], p5 = s[-6 * p], p4 = s[-5 * p],
 385                   p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
 386     const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p],
 387                   q4 = s[4 * p], q5 = s[5 * p], q6 = s[6 * p];
 388     const int8_t mask =
 389         filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
 390     const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
 391     const int8_t flat2 = flat_mask4(1, p6, p5, p4, p0, q0, q4, q5, q6);
 392
 393     filter14(mask, *thresh, flat, flat2, s - 7 * p, s - 6 * p, s - 5 * p,
 394              s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p,
 395              s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p, s + 6 * p);
 396     ++s;
 397   }
 398 }
 399
 400 void aom_lpf_horizontal_14_c(uint8_t *s, int p, const uint8_t *blimit,
 401                              const uint8_t *limit, const uint8_t *thresh) {
 402   mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1);
 403 }
 404
 405 void aom_lpf_horizontal_14_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
 406                                   const uint8_t *limit0, const uint8_t *thresh0,
 407                                   const uint8_t *blimit1, const uint8_t *limit1,
 408                                   const uint8_t *thresh1) {
 409   mb_lpf_horizontal_edge_w(s, p, blimit0, limit0, thresh0, 1);
 410   mb_lpf_horizontal_edge_w(s + 4, p, blimit1, limit1, thresh1, 1);
 411 }
 412
 413 static void mb_lpf_vertical_edge_w(uint8_t *s, int p, const uint8_t *blimit,
 414                                    const uint8_t *limit, const uint8_t *thresh,
 415                                    int count) {
 416   int i;
 417
 418   for (i = 0; i < count; ++i) {
 419     const uint8_t p6 = s[-7], p5 = s[-6], p4 = s[-5], p3 = s[-4], p2 = s[-3],
 420                   p1 = s[-2], p0 = s[-1];
 421     const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3], q4 = s[4],
 422                   q5 = s[5], q6 = s[6];
 423     const int8_t mask =
 424         filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
 425     const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
 426     const int8_t flat2 = flat_mask4(1, p6, p5, p4, p0, q0, q4, q5, q6);
 427
 428     filter14(mask, *thresh, flat, flat2, s - 7, s - 6, s - 5, s - 4, s - 3,
 429              s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6);
 430     s += p;
 431   }
 432 }
 433
 434 void aom_lpf_vertical_14_c(uint8_t *s, int p, const uint8_t *blimit,
 435                            const uint8_t *limit, const uint8_t *thresh) {
 436   mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4);
 437 }
 438
 439 void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
 440                                 const uint8_t *limit0, const uint8_t *thresh0,
 441                                 const uint8_t *blimit1, const uint8_t *limit1,
 442                                 const uint8_t *thresh1) {
 443   mb_lpf_vertical_edge_w(s, pitch, blimit0, limit0, thresh0, 4);
 444   mb_lpf_vertical_edge_w(s + 4 * pitch, pitch, blimit1, limit1, thresh1, 4);
 445 }
 446
 447 #if CONFIG_AV1_HIGHBITDEPTH
 448 // Should we apply any filter at all: 11111111 yes, 00000000 no ?
 449 static INLINE int8_t highbd_filter_mask2(uint8_t limit, uint8_t blimit,
 450                                          uint16_t p1, uint16_t p0, uint16_t q0,
 451                                          uint16_t q1, int bd) {
 452   int8_t mask = 0;
 453   int16_t limit16 = (uint16_t)limit << (bd - 8);
 454   int16_t blimit16 = (uint16_t)blimit << (bd - 8);
 455   mask |= (abs(p1 - p0) > limit16) * -1;
 456   mask |= (abs(q1 - q0) > limit16) * -1;
 457   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
 458   return ~mask;
 459 }
 460
 461 // Should we apply any filter at all: 11111111 yes, 00000000 no ?
 462 static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
 463                                         uint16_t p3, uint16_t p2, uint16_t p1,
 464                                         uint16_t p0, uint16_t q0, uint16_t q1,
 465                                         uint16_t q2, uint16_t q3, int bd) {
 466   int8_t mask = 0;
 467   int16_t limit16 = (uint16_t)limit << (bd - 8);
 468   int16_t blimit16 = (uint16_t)blimit << (bd - 8);
 469   mask |= (abs(p3 - p2) > limit16) * -1;
 470   mask |= (abs(p2 - p1) > limit16) * -1;
 471   mask |= (abs(p1 - p0) > limit16) * -1;
 472   mask |= (abs(q1 - q0) > limit16) * -1;
 473   mask |= (abs(q2 - q1) > limit16) * -1;
 474   mask |= (abs(q3 - q2) > limit16) * -1;
 475   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
 476   return ~mask;
 477 }
 478
 479 static INLINE int8_t highbd_filter_mask3_chroma(uint8_t limit, uint8_t blimit,
 480                                                 uint16_t p2, uint16_t p1,
 481                                                 uint16_t p0, uint16_t q0,
 482                                                 uint16_t q1, uint16_t q2,
 483                                                 int bd) {
 484   int8_t mask = 0;
 485   int16_t limit16 = (uint16_t)limit << (bd - 8);
 486   int16_t blimit16 = (uint16_t)blimit << (bd - 8);
 487   mask |= (abs(p2 - p1) > limit16) * -1;
 488   mask |= (abs(p1 - p0) > limit16) * -1;
 489   mask |= (abs(q1 - q0) > limit16) * -1;
 490   mask |= (abs(q2 - q1) > limit16) * -1;
 491   mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
 492   return ~mask;
 493 }
 494
 495 static INLINE int8_t highbd_flat_mask3_chroma(uint8_t thresh, uint16_t p2,
 496                                               uint16_t p1, uint16_t p0,
 497                                               uint16_t q0, uint16_t q1,
 498                                               uint16_t q2, int bd) {
 499   int8_t mask = 0;
 500   int16_t thresh16 = (uint16_t)thresh << (bd - 8);
 501   mask |= (abs(p1 - p0) > thresh16) * -1;
 502   mask |= (abs(q1 - q0) > thresh16) * -1;
 503   mask |= (abs(p2 - p0) > thresh16) * -1;
 504   mask |= (abs(q2 - q0) > thresh16) * -1;
 505   return ~mask;
 506 }
 507
 508 static INLINE int8_t highbd_flat_mask4(uint8_t thresh, uint16_t p3, uint16_t p2,
 509                                        uint16_t p1, uint16_t p0, uint16_t q0,
 510                                        uint16_t q1, uint16_t q2, uint16_t q3,
 511                                        int bd) {
 512   int8_t mask = 0;
 513   int16_t thresh16 = (uint16_t)thresh << (bd - 8);
 514   mask |= (abs(p1 - p0) > thresh16) * -1;
 515   mask |= (abs(q1 - q0) > thresh16) * -1;
 516   mask |= (abs(p2 - p0) > thresh16) * -1;
 517   mask |= (abs(q2 - q0) > thresh16) * -1;
 518   mask |= (abs(p3 - p0) > thresh16) * -1;
 519   mask |= (abs(q3 - q0) > thresh16) * -1;
 520   return ~mask;
 521 }
 522
 523 // Is there high edge variance internal edge:
 524 // 11111111_11111111 yes, 00000000_00000000 no ?
 525 static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
 526                                       uint16_t q0, uint16_t q1, int bd) {
 527   int16_t hev = 0;
 528   int16_t thresh16 = (uint16_t)thresh << (bd - 8);
 529   hev |= (abs(p1 - p0) > thresh16) * -1;
 530   hev |= (abs(q1 - q0) > thresh16) * -1;
 531   return hev;
 532 }
 533
 534 static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
 535                                   uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
 536                                   int bd) {
 537   int16_t filter1, filter2;
 538   // ^0x80 equivalent to subtracting 0x80 from the values to turn them
 539   // into -128 to +127 instead of 0 to 255.
 540   int shift = bd - 8;
 541   const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
 542   const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
 543   const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
 544   const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
 545   const int16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
 546
 547   // Add outer taps if we have high edge variance.
 548   int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
 549
 550   // Inner taps.
 551   filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
 552
 553   // Save bottom 3 bits so that we round one side +4 and the other +3
 554   // if it equals 4 we'll set to adjust by -1 to account for the fact
 555   // we'd round 3 the other way.
 556   filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
 557   filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
 558
 559   *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
 560   *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
 561
 562   // Outer tap adjustments.
 563   filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
 564
 565   *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
 566   *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
 567 }
 568
 569 void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
 570                                    const uint8_t *blimit, const uint8_t *limit,
 571                                    const uint8_t *thresh, int bd) {
 572   int i;
 573   int count = 4;
 574
 575   // loop filter designed to work using chars so that we can make maximum use
 576   // of 8 bit simd instructions.
 577   for (i = 0; i < count; ++i) {
 578     const uint16_t p1 = s[-2 * p];
 579     const uint16_t p0 = s[-p];
 580     const uint16_t q0 = s[0 * p];
 581     const uint16_t q1 = s[1 * p];
 582     const int8_t mask =
 583         highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd);
 584     highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
 585     ++s;
 586   }
 587 }
 588
 589 void aom_highbd_lpf_horizontal_4_dual_c(
 590     uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
 591     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 592     const uint8_t *thresh1, int bd) {
 593   aom_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, bd);
 594   aom_highbd_lpf_horizontal_4_c(s + 4, p, blimit1, limit1, thresh1, bd);
 595 }
 596
 597 void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
 598                                  const uint8_t *limit, const uint8_t *thresh,
 599                                  int bd) {
 600   int i;
 601   int count = 4;
 602
 603   // loop filter designed to work using chars so that we can make maximum use
 604   // of 8 bit simd instructions.
 605   for (i = 0; i < count; ++i) {
 606     const uint16_t p1 = s[-2], p0 = s[-1];
 607     const uint16_t q0 = s[0], q1 = s[1];
 608     const int8_t mask =
 609         highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd);
 610     highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
 611     s += pitch;
 612   }
 613 }
 614
 615 void aom_highbd_lpf_vertical_4_dual_c(
 616     uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
 617     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 618     const uint8_t *thresh1, int bd) {
 619   aom_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, bd);
 620   aom_highbd_lpf_vertical_4_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1,
 621                               bd);
 622 }
 623
 624 static INLINE void highbd_filter6(int8_t mask, uint8_t thresh, int8_t flat,
 625                                   uint16_t *op2, uint16_t *op1, uint16_t *op0,
 626                                   uint16_t *oq0, uint16_t *oq1, uint16_t *oq2,
 627                                   int bd) {
 628   if (flat && mask) {
 629     const uint16_t p2 = *op2, p1 = *op1, p0 = *op0;
 630     const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2;
 631
 632     // 5-tap filter [1, 2, 2, 2, 1]
 633     *op1 = ROUND_POWER_OF_TWO(p2 * 3 + p1 * 2 + p0 * 2 + q0, 3);
 634     *op0 = ROUND_POWER_OF_TWO(p2 + p1 * 2 + p0 * 2 + q0 * 2 + q1, 3);
 635     *oq0 = ROUND_POWER_OF_TWO(p1 + p0 * 2 + q0 * 2 + q1 * 2 + q2, 3);
 636     *oq1 = ROUND_POWER_OF_TWO(p0 + q0 * 2 + q1 * 2 + q2 * 3, 3);
 637   } else {
 638     highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd);
 639   }
 640 }
 641
 642 static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, int8_t flat,
 643                                   uint16_t *op3, uint16_t *op2, uint16_t *op1,
 644                                   uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
 645                                   uint16_t *oq2, uint16_t *oq3, int bd) {
 646   if (flat && mask) {
 647     const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
 648     const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
 649
 650     // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
 651     *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
 652     *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
 653     *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
 654     *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
 655     *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
 656     *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
 657   } else {
 658     highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd);
 659   }
 660 }
 661
 662 void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
 663                                    const uint8_t *limit, const uint8_t *thresh,
 664                                    int bd) {
 665   int i;
 666   int count = 4;
 667
 668   // loop filter designed to work using chars so that we can make maximum use
 669   // of 8 bit simd instructions.
 670   for (i = 0; i < count; ++i) {
 671     const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
 672     const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
 673
 674     const int8_t mask =
 675         highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 676     const int8_t flat =
 677         highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 678     highbd_filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p,
 679                    s - 1 * p, s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
 680     ++s;
 681   }
 682 }
 683
 684 void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int p, const uint8_t *blimit,
 685                                    const uint8_t *limit, const uint8_t *thresh,
 686                                    int bd) {
 687   int i;
 688   int count = 4;
 689
 690   // loop filter designed to work using chars so that we can make maximum use
 691   // of 8 bit simd instructions.
 692   for (i = 0; i < count; ++i) {
 693     const uint16_t p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
 694     const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p];
 695
 696     const int8_t mask =
 697         highbd_filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2, bd);
 698     const int8_t flat = highbd_flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2, bd);
 699     highbd_filter6(mask, *thresh, flat, s - 3 * p, s - 2 * p, s - 1 * p, s,
 700                    s + 1 * p, s + 2 * p, bd);
 701     ++s;
 702   }
 703 }
 704
 705 void aom_highbd_lpf_horizontal_6_dual_c(
 706     uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
 707     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 708     const uint8_t *thresh1, int bd) {
 709   aom_highbd_lpf_horizontal_6_c(s, p, blimit0, limit0, thresh0, bd);
 710   aom_highbd_lpf_horizontal_6_c(s + 4, p, blimit1, limit1, thresh1, bd);
 711 }
 712
 713 void aom_highbd_lpf_horizontal_8_dual_c(
 714     uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
 715     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 716     const uint8_t *thresh1, int bd) {
 717   aom_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, bd);
 718   aom_highbd_lpf_horizontal_8_c(s + 4, p, blimit1, limit1, thresh1, bd);
 719 }
 720
 721 void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit,
 722                                  const uint8_t *limit, const uint8_t *thresh,
 723                                  int bd) {
 724   int i;
 725   int count = 4;
 726
 727   for (i = 0; i < count; ++i) {
 728     const uint16_t p2 = s[-3], p1 = s[-2], p0 = s[-1];
 729     const uint16_t q0 = s[0], q1 = s[1], q2 = s[2];
 730     const int8_t mask =
 731         highbd_filter_mask3_chroma(*limit, *blimit, p2, p1, p0, q0, q1, q2, bd);
 732     const int8_t flat = highbd_flat_mask3_chroma(1, p2, p1, p0, q0, q1, q2, bd);
 733     highbd_filter6(mask, *thresh, flat, s - 3, s - 2, s - 1, s, s + 1, s + 2,
 734                    bd);
 735     s += pitch;
 736   }
 737 }
 738
 739 void aom_highbd_lpf_vertical_6_dual_c(
 740     uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
 741     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 742     const uint8_t *thresh1, int bd) {
 743   aom_highbd_lpf_vertical_6_c(s, pitch, blimit0, limit0, thresh0, bd);
 744   aom_highbd_lpf_vertical_6_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1,
 745                               bd);
 746 }
 747
 748 void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
 749                                  const uint8_t *limit, const uint8_t *thresh,
 750                                  int bd) {
 751   int i;
 752   int count = 4;
 753
 754   for (i = 0; i < count; ++i) {
 755     const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
 756     const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
 757     const int8_t mask =
 758         highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 759     const int8_t flat =
 760         highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 761     highbd_filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1,
 762                    s + 2, s + 3, bd);
 763     s += pitch;
 764   }
 765 }
 766
 767 void aom_highbd_lpf_vertical_8_dual_c(
 768     uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
 769     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 770     const uint8_t *thresh1, int bd) {
 771   aom_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, bd);
 772   aom_highbd_lpf_vertical_8_c(s + 4 * pitch, pitch, blimit1, limit1, thresh1,
 773                               bd);
 774 }
 775
 776 static INLINE void highbd_filter14(int8_t mask, uint8_t thresh, int8_t flat,
 777                                    int8_t flat2, uint16_t *op6, uint16_t *op5,
 778                                    uint16_t *op4, uint16_t *op3, uint16_t *op2,
 779                                    uint16_t *op1, uint16_t *op0, uint16_t *oq0,
 780                                    uint16_t *oq1, uint16_t *oq2, uint16_t *oq3,
 781                                    uint16_t *oq4, uint16_t *oq5, uint16_t *oq6,
 782                                    int bd) {
 783   if (flat2 && flat && mask) {
 784     const uint16_t p6 = *op6;
 785     const uint16_t p5 = *op5;
 786     const uint16_t p4 = *op4;
 787     const uint16_t p3 = *op3;
 788     const uint16_t p2 = *op2;
 789     const uint16_t p1 = *op1;
 790     const uint16_t p0 = *op0;
 791     const uint16_t q0 = *oq0;
 792     const uint16_t q1 = *oq1;
 793     const uint16_t q2 = *oq2;
 794     const uint16_t q3 = *oq3;
 795     const uint16_t q4 = *oq4;
 796     const uint16_t q5 = *oq5;
 797     const uint16_t q6 = *oq6;
 798
 799     // 13-tap filter [1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1]
 800     *op5 = ROUND_POWER_OF_TWO(p6 * 7 + p5 * 2 + p4 * 2 + p3 + p2 + p1 + p0 + q0,
 801                               4);
 802     *op4 = ROUND_POWER_OF_TWO(
 803         p6 * 5 + p5 * 2 + p4 * 2 + p3 * 2 + p2 + p1 + p0 + q0 + q1, 4);
 804     *op3 = ROUND_POWER_OF_TWO(
 805         p6 * 4 + p5 + p4 * 2 + p3 * 2 + p2 * 2 + p1 + p0 + q0 + q1 + q2, 4);
 806     *op2 = ROUND_POWER_OF_TWO(
 807         p6 * 3 + p5 + p4 + p3 * 2 + p2 * 2 + p1 * 2 + p0 + q0 + q1 + q2 + q3,
 808         4);
 809     *op1 = ROUND_POWER_OF_TWO(p6 * 2 + p5 + p4 + p3 + p2 * 2 + p1 * 2 + p0 * 2 +
 810                                   q0 + q1 + q2 + q3 + q4,
 811                               4);
 812     *op0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 * 2 +
 813                                   q0 * 2 + q1 + q2 + q3 + q4 + q5,
 814                               4);
 815     *oq0 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 * 2 +
 816                                   q1 * 2 + q2 + q3 + q4 + q5 + q6,
 817                               4);
 818     *oq1 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 * 2 +
 819                                   q2 * 2 + q3 + q4 + q5 + q6 * 2,
 820                               4);
 821     *oq2 = ROUND_POWER_OF_TWO(
 822         p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 * 2 + q3 * 2 + q4 + q5 + q6 * 3,
 823         4);
 824     *oq3 = ROUND_POWER_OF_TWO(
 825         p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 * 2 + q4 * 2 + q5 + q6 * 4, 4);
 826     *oq4 = ROUND_POWER_OF_TWO(
 827         p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 * 2 + q5 * 2 + q6 * 5, 4);
 828     *oq5 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 2 + q6 * 7,
 829                               4);
 830   } else {
 831     highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
 832                    bd);
 833   }
 834 }
 835
 836 static void highbd_mb_lpf_horizontal_edge_w(uint16_t *s, int p,
 837                                             const uint8_t *blimit,
 838                                             const uint8_t *limit,
 839                                             const uint8_t *thresh, int count,
 840                                             int bd) {
 841   int i;
 842   int step = 4;
 843
 844   // loop filter designed to work using chars so that we can make maximum use
 845   // of 8 bit simd instructions.
 846   for (i = 0; i < step * count; ++i) {
 847     const uint16_t p3 = s[-4 * p];
 848     const uint16_t p2 = s[-3 * p];
 849     const uint16_t p1 = s[-2 * p];
 850     const uint16_t p0 = s[-p];
 851     const uint16_t q0 = s[0 * p];
 852     const uint16_t q1 = s[1 * p];
 853     const uint16_t q2 = s[2 * p];
 854     const uint16_t q3 = s[3 * p];
 855     const int8_t mask =
 856         highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 857     const int8_t flat =
 858         highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 859
 860     const int8_t flat2 =
 861         highbd_flat_mask4(1, s[-7 * p], s[-6 * p], s[-5 * p], p0, q0, s[4 * p],
 862                           s[5 * p], s[6 * p], bd);
 863
 864     highbd_filter14(mask, *thresh, flat, flat2, s - 7 * p, s - 6 * p, s - 5 * p,
 865                     s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p,
 866                     s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p, s + 6 * p, bd);
 867     ++s;
 868   }
 869 }
 870
 871 void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch,
 872                                     const uint8_t *blimit, const uint8_t *limit,
 873                                     const uint8_t *thresh, int bd) {
 874   highbd_mb_lpf_horizontal_edge_w(s, pitch, blimit, limit, thresh, 1, bd);
 875 }
 876
 877 void aom_highbd_lpf_horizontal_14_dual_c(
 878     uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
 879     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 880     const uint8_t *thresh1, int bd) {
 881   highbd_mb_lpf_horizontal_edge_w(s, p, blimit0, limit0, thresh0, 1, bd);
 882   highbd_mb_lpf_horizontal_edge_w(s + 4, p, blimit1, limit1, thresh1, 1, bd);
 883 }
 884
 885 static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
 886                                           const uint8_t *blimit,
 887                                           const uint8_t *limit,
 888                                           const uint8_t *thresh, int count,
 889                                           int bd) {
 890   int i;
 891
 892   for (i = 0; i < count; ++i) {
 893     const uint16_t p3 = s[-4];
 894     const uint16_t p2 = s[-3];
 895     const uint16_t p1 = s[-2];
 896     const uint16_t p0 = s[-1];
 897     const uint16_t q0 = s[0];
 898     const uint16_t q1 = s[1];
 899     const uint16_t q2 = s[2];
 900     const uint16_t q3 = s[3];
 901     const int8_t mask =
 902         highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 903     const int8_t flat =
 904         highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
 905     const int8_t flat2 =
 906         highbd_flat_mask4(1, s[-7], s[-6], s[-5], p0, q0, s[4], s[5], s[6], bd);
 907
 908     highbd_filter14(mask, *thresh, flat, flat2, s - 7, s - 6, s - 5, s - 4,
 909                     s - 3, s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5,
 910                     s + 6, bd);
 911     s += p;
 912   }
 913 }
 914
 915 void aom_highbd_lpf_vertical_14_c(uint16_t *s, int p, const uint8_t *blimit,
 916                                   const uint8_t *limit, const uint8_t *thresh,
 917                                   int bd) {
 918   highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4, bd);
 919 }
 920
 921 void aom_highbd_lpf_vertical_14_dual_c(
 922     uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
 923     const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
 924     const uint8_t *thresh1, int bd) {
 925   highbd_mb_lpf_vertical_edge_w(s, pitch, blimit0, limit0, thresh0, 4, bd);
 926   highbd_mb_lpf_vertical_edge_w(s + 4 * pitch, pitch, blimit1, limit1, thresh1,
 927                                 4, bd);
 928 }
 929 #endif  // CONFIG_AV1_HIGHBITDEPTH