vp9_loopfilter.c 47.2 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5 6
 *  that can be found in the LICENSE file in the root of the source
 *  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.
John Koleszar's avatar
John Koleszar committed
9 10
 */

11
#include "./vpx_config.h"
12 13
#include "vp9/common/vp9_loopfilter.h"
#include "vp9/common/vp9_onyxc_int.h"
14
#include "vp9/common/vp9_reconinter.h"
Attila Nagy's avatar
Attila Nagy committed
15
#include "vpx_mem/vpx_mem.h"
John Koleszar's avatar
John Koleszar committed
16

17
#include "vp9/common/vp9_seg_common.h"
18

Jim Bankoski's avatar
Jim Bankoski committed
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
// This structure holds bit masks for all 8x8 blocks in a 64x64 region.
// Each 1 bit represents a position in which we want to apply the loop filter.
// Left_ entries refer to whether we apply a filter on the border to the
// left of the block.   Above_ entries refer to whether or not to apply a
// filter on the above border.   Int_ entries refer to whether or not to
// apply borders on the 4x4 edges within the 8x8 block that each bit
// represents.
// Since each transform is accompanied by a potentially different type of
// loop filter there is a different entry in the array for each transform size.
typedef struct {
  uint64_t left_y[TX_SIZES];
  uint64_t above_y[TX_SIZES];
  uint64_t int_4x4_y;
  uint16_t left_uv[TX_SIZES];
  uint16_t above_uv[TX_SIZES];
  uint16_t int_4x4_uv;
35 36
  uint8_t lfl_y[64];
  uint8_t lfl_uv[16];
Jim Bankoski's avatar
Jim Bankoski committed
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
} LOOP_FILTER_MASK;

// 64 bit masks for left transform size.  Each 1 represents a position where
// we should apply a loop filter across the left border of an 8x8 block
// boundary.
//
// In the case of TX_16X16->  ( in low order byte first we end up with
// a mask that looks like this
//
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//
// A loopfilter should be applied to every other 8x8 horizontally.
static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
    0xffffffffffffffff,  // TX_4X4
    0xffffffffffffffff,  // TX_8x8
    0x5555555555555555,  // TX_16x16
    0x1111111111111111,  // TX_32x32
};

// 64 bit masks for above transform size.  Each 1 represents a position where
// we should apply a loop filter across the top border of an 8x8 block
// boundary.
//
// In the case of TX_32x32 ->  ( in low order byte first we end up with
// a mask that looks like this
//
//    11111111
//    00000000
//    00000000
//    00000000
//    11111111
//    00000000
//    00000000
//    00000000
//
// A loopfilter should be applied to every other 4 the row vertically.
static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
    0xffffffffffffffff,  // TX_4X4
    0xffffffffffffffff,  // TX_8x8
    0x00ff00ff00ff00ff,  // TX_16x16
    0x000000ff000000ff,  // TX_32x32
};

// 64 bit masks for prediction sizes (left).  Each 1 represents a position
// where left border of an 8x8 block.  These are aligned to the right most
// appropriate bit,  and then shifted into place.
//
// In the case of TX_16x32 ->  ( low order byte first ) we end up with
// a mask that looks like this :
//
//  10000000
//  10000000
//  10000000
//  10000000
//  00000000
//  00000000
//  00000000
//  00000000
static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
    0x0000000000000001,  // BLOCK_4X4,
    0x0000000000000001,  // BLOCK_4X8,
    0x0000000000000001,  // BLOCK_8X4,
    0x0000000000000001,  // BLOCK_8X8,
    0x0000000000000101,  // BLOCK_8X16,
    0x0000000000000001,  // BLOCK_16X8,
    0x0000000000000101,  // BLOCK_16X16,
    0x0000000001010101,  // BLOCK_16X32,
    0x0000000000000101,  // BLOCK_32X16,
    0x0000000001010101,  // BLOCK_32X32,
    0x0101010101010101,  // BLOCK_32X64,
    0x0000000001010101,  // BLOCK_64X32,
    0x0101010101010101,  // BLOCK_64X64
};

// 64 bit mask to shift and set for each prediction size.
static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
    0x0000000000000001,  // BLOCK_4X4
    0x0000000000000001,  // BLOCK_4X8
    0x0000000000000001,  // BLOCK_8X4
    0x0000000000000001,  // BLOCK_8X8
    0x0000000000000001,  // BLOCK_8X16,
    0x0000000000000003,  // BLOCK_16X8
    0x0000000000000003,  // BLOCK_16X16
    0x0000000000000003,  // BLOCK_16X32,
    0x000000000000000f,  // BLOCK_32X16,
    0x000000000000000f,  // BLOCK_32X32,
    0x000000000000000f,  // BLOCK_32X64,
    0x00000000000000ff,  // BLOCK_64X32,
    0x00000000000000ff,  // BLOCK_64X64
};
// 64 bit mask to shift and set for each prediction size.  A bit is set for
// each 8x8 block that would be in the left most block of the given block
// size in the 64x64 block.
static const uint64_t size_mask[BLOCK_SIZES] = {
    0x0000000000000001,  // BLOCK_4X4
    0x0000000000000001,  // BLOCK_4X8
    0x0000000000000001,  // BLOCK_8X4
    0x0000000000000001,  // BLOCK_8X8
    0x0000000000000101,  // BLOCK_8X16,
    0x0000000000000003,  // BLOCK_16X8
    0x0000000000000303,  // BLOCK_16X16
    0x0000000003030303,  // BLOCK_16X32,
    0x0000000000000f0f,  // BLOCK_32X16,
    0x000000000f0f0f0f,  // BLOCK_32X32,
    0x0f0f0f0f0f0f0f0f,  // BLOCK_32X64,
    0x00000000ffffffff,  // BLOCK_64X32,
    0xffffffffffffffff,  // BLOCK_64X64
};

// These are used for masking the left and above borders.
static const uint64_t left_border =  0x1111111111111111;
static const uint64_t above_border = 0x000000ff000000ff;

// 16 bit masks for uv transform sizes.
static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
    0xffff,  // TX_4X4
    0xffff,  // TX_8x8
    0x5555,  // TX_16x16
    0x1111,  // TX_32x32
};

static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
    0xffff,  // TX_4X4
    0xffff,  // TX_8x8
    0x0f0f,  // TX_16x16
    0x000f,  // TX_32x32
};

// 16 bit left mask to shift and set for each uv prediction size.
static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
    0x0001,  // BLOCK_4X4,
    0x0001,  // BLOCK_4X8,
    0x0001,  // BLOCK_8X4,
    0x0001,  // BLOCK_8X8,
    0x0001,  // BLOCK_8X16,
    0x0001,  // BLOCK_16X8,
    0x0001,  // BLOCK_16X16,
    0x0011,  // BLOCK_16X32,
    0x0001,  // BLOCK_32X16,
    0x0011,  // BLOCK_32X32,
    0x1111,  // BLOCK_32X64
    0x0011,  // BLOCK_64X32,
    0x1111,  // BLOCK_64X64
};
// 16 bit above mask to shift and set for uv each prediction size.
static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
    0x0001,  // BLOCK_4X4
    0x0001,  // BLOCK_4X8
    0x0001,  // BLOCK_8X4
    0x0001,  // BLOCK_8X8
    0x0001,  // BLOCK_8X16,
    0x0001,  // BLOCK_16X8
    0x0001,  // BLOCK_16X16
    0x0001,  // BLOCK_16X32,
    0x0003,  // BLOCK_32X16,
    0x0003,  // BLOCK_32X32,
    0x0003,  // BLOCK_32X64,
    0x000f,  // BLOCK_64X32,
    0x000f,  // BLOCK_64X64
};

// 64 bit mask to shift and set for each uv prediction size
static const uint16_t size_mask_uv[BLOCK_SIZES] = {
    0x0001,  // BLOCK_4X4
    0x0001,  // BLOCK_4X8
    0x0001,  // BLOCK_8X4
    0x0001,  // BLOCK_8X8
    0x0001,  // BLOCK_8X16,
    0x0001,  // BLOCK_16X8
    0x0001,  // BLOCK_16X16
    0x0011,  // BLOCK_16X32,
    0x0003,  // BLOCK_32X16,
    0x0033,  // BLOCK_32X32,
    0x3333,  // BLOCK_32X64,
    0x00ff,  // BLOCK_64X32,
    0xffff,  // BLOCK_64X64
};
static const uint16_t left_border_uv =  0x1111;
static const uint16_t above_border_uv = 0x000f;

224 225 226 227
static const int mode_lf_lut[MB_MODE_COUNT] = {
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
  1, 1, 0, 1                     // INTER_MODES (ZEROMV == 0)
};
Attila Nagy's avatar
Attila Nagy committed
228

Frank Galligan's avatar
Frank Galligan committed
229
static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
230
  int lvl;
John Koleszar's avatar
John Koleszar committed
231

Dmitry Kovalev's avatar
Dmitry Kovalev committed
232 233 234 235
  // For each possible value for the loop filter fill out limits
  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
    // Set loop filter paramaeters that control sharpness.
    int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
John Koleszar's avatar
John Koleszar committed
236

John Koleszar's avatar
John Koleszar committed
237 238 239
    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
John Koleszar's avatar
John Koleszar committed
240
    }
John Koleszar's avatar
John Koleszar committed
241 242 243 244

    if (block_inside_limit < 1)
      block_inside_limit = 1;

245 246
    vpx_memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
    vpx_memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
John Koleszar's avatar
John Koleszar committed
247 248
               SIMD_WIDTH);
  }
Attila Nagy's avatar
Attila Nagy committed
249
}
John Koleszar's avatar
John Koleszar committed
250

251
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
252
  loop_filter_info_n *lfi = &cm->lf_info;
253
  struct loopfilter *lf = &cm->lf;
254
  int lvl;
Attila Nagy's avatar
Attila Nagy committed
255

256
  // init limits for given sharpness
257 258
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
259

260
  // init hev threshold const vectors
261 262
  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
    vpx_memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
John Koleszar's avatar
John Koleszar committed
263 264
}

Frank Galligan's avatar
Frank Galligan committed
265
void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
266
  int seg_id;
267 268 269
  // n_shift is the a multiplier for lf_deltas
  // the multiplier is 1 for when filter_lvl is between 0 and 31;
  // 2 when filter_lvl is between 32 and 63
270
  const int scale = 1 << (default_filt_lvl >> 5);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
271
  loop_filter_info_n *const lfi = &cm->lf_info;
272
  struct loopfilter *const lf = &cm->lf;
273
  const struct segmentation *const seg = &cm->seg;
John Koleszar's avatar
John Koleszar committed
274

Dmitry Kovalev's avatar
Dmitry Kovalev committed
275
  // update limits if sharpness has changed
276 277 278
  if (lf->last_sharpness_level != lf->sharpness_level) {
    update_sharpness(lfi, lf->sharpness_level);
    lf->last_sharpness_level = lf->sharpness_level;
John Koleszar's avatar
John Koleszar committed
279 280
  }

281
  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
282
    int lvl_seg = default_filt_lvl;
283
    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
284 285
      const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
      lvl_seg = seg->abs_delta == SEGMENT_ABSDATA
Dmitry Kovalev's avatar
Dmitry Kovalev committed
286 287
                  ? data
                  : clamp(default_filt_lvl + data, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
288
    }
John Koleszar's avatar
John Koleszar committed
289

290
    if (!lf->mode_ref_delta_enabled) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
291 292
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
293
      vpx_memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
294 295 296 297 298 299 300 301 302 303 304
    } else {
      int ref, mode;
      const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
      lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);

      for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) {
        for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
          const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale
                                        + lf->mode_deltas[mode] * scale;
          lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
        }
John Koleszar's avatar
John Koleszar committed
305
      }
306
    }
John Koleszar's avatar
John Koleszar committed
307
  }
John Koleszar's avatar
John Koleszar committed
308 309
}

310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
static void filter_selectively_vert_row2(PLANE_TYPE plane_type,
                                         uint8_t *s, int pitch,
                                         unsigned int mask_16x16_l,
                                         unsigned int mask_8x8_l,
                                         unsigned int mask_4x4_l,
                                         unsigned int mask_4x4_int_l,
                                         const loop_filter_info_n *lfi_n,
                                         const uint8_t *lfl) {
  const int mask_shift = plane_type ? 4 : 8;
  const int mask_cutoff = plane_type ? 0xf : 0xff;
  const int lfl_forward = plane_type ? 4 : 8;

  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
330 331
  unsigned int mask;

332 333 334 335 336
  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
      mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
      mask; mask >>= 1) {
    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
337

338
    // TODO(yunqingwang): count in loopfilter functions should be removed.
339
    if (mask & 1) {
340 341
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
342 343
          vp9_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                   lfi0->hev_thr);
344
        } else if (mask_16x16_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
345 346
          vp9_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
                              lfi0->hev_thr);
347
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
348 349
          vp9_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
                              lfi1->lim, lfi1->hev_thr);
350
        }
351 352 353 354
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
355 356 357
          vp9_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
358
        } else if (mask_8x8_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
359 360
          vp9_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
361
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
362 363
          vp9_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
364 365 366 367 368
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
369 370 371
          vp9_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
372
        } else if (mask_4x4_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
373 374
          vp9_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
375
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
376 377
          vp9_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
378 379 380 381 382
        }
      }

      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
383 384 385
          vp9_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
386
        } else if (mask_4x4_int_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
387 388
          vp9_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
                             lfi0->hev_thr, 1);
389
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
390 391
          vp9_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
392
        }
393 394
      }
    }
395

396
    s += 8;
Yunqing Wang's avatar
Yunqing Wang committed
397
    lfl += 1;
398 399 400 401 402 403 404 405
    mask_16x16_0 >>= 1;
    mask_8x8_0 >>= 1;
    mask_4x4_0 >>= 1;
    mask_4x4_int_0 >>= 1;
    mask_16x16_1 >>= 1;
    mask_8x8_1 >>= 1;
    mask_4x4_1 >>= 1;
    mask_4x4_int_1 >>= 1;
406 407 408 409 410 411 412
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
413
                                     unsigned int mask_4x4_int,
Yunqing Wang's avatar
Yunqing Wang committed
414 415
                                     const loop_filter_info_n *lfi_n,
                                     const uint8_t *lfl) {
416
  unsigned int mask;
417
  int count;
418

John Koleszar's avatar
John Koleszar committed
419
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
420
       mask; mask >>= count) {
Yunqing Wang's avatar
Yunqing Wang committed
421
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
422

Dmitry Kovalev's avatar
Dmitry Kovalev committed
423
    count = 1;
424
    if (mask & 1) {
425
      if (mask_16x16 & 1) {
426
        if ((mask_16x16 & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
427 428
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 2);
429 430
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
431 432
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 1);
433
        }
434 435 436 437 438
      } else if (mask_8x8 & 1) {
        if ((mask_8x8 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

Jim Bankoski's avatar
Jim Bankoski committed
439 440 441
          vp9_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
442 443

          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
444 445 446
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
447
          } else {
448
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
449 450
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
451
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
452 453
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
454
          }
455 456
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
457
          vp9_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
458 459

          if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
460 461
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
462
        }
463 464 465 466
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
467

Jim Bankoski's avatar
Jim Bankoski committed
468 469 470
          vp9_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
471
          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
472 473 474
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
475 476
          } else {
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
477 478
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
479
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
480 481
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
482 483 484
          }
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
485
          vp9_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
486

Jim Bankoski's avatar
Jim Bankoski committed
487 488 489
          if (mask_4x4_int & 1)
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
490 491
        }
      } else if (mask_4x4_int & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
492 493
        vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                             lfi->hev_thr, 1);
494
      }
495
    }
496
    s += 8 * count;
Yunqing Wang's avatar
Yunqing Wang committed
497
    lfl += count;
498 499 500 501
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
502 503 504
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521
// This function ors into the current lfm structure, where to do loop
// filters for the specific mi we are looking at.   It uses information
// including the block_size_type (32x16, 32x32, etc),  the transform size,
// whether there were any coefficients encoded, and the loop filter strength
// block we are currently looking at. Shift is used to position the
// 1's we produce.
// TODO(JBB) Need another function for different resolution color..
static void build_masks(const loop_filter_info_n *const lfi_n,
                        const MODE_INFO *mi, const int shift_y,
                        const int shift_uv,
                        LOOP_FILTER_MASK *lfm) {
  const BLOCK_SIZE block_size = mi->mbmi.sb_type;
  const TX_SIZE tx_size_y = mi->mbmi.tx_size;
  const TX_SIZE tx_size_uv = get_uv_tx_size(&mi->mbmi);
  const int skip = mi->mbmi.skip_coeff;
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
522
  const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
Jim Bankoski's avatar
Jim Bankoski committed
523 524 525 526 527 528
  uint64_t *left_y = &lfm->left_y[tx_size_y];
  uint64_t *above_y = &lfm->above_y[tx_size_y];
  uint64_t *int_4x4_y = &lfm->int_4x4_y;
  uint16_t *left_uv = &lfm->left_uv[tx_size_uv];
  uint16_t *above_uv = &lfm->above_uv[tx_size_uv];
  uint16_t *int_4x4_uv = &lfm->int_4x4_uv;
Yunqing Wang's avatar
Yunqing Wang committed
529 530 531 532
  int i;
  int w = num_8x8_blocks_wide_lookup[block_size];
  int h = num_8x8_blocks_high_lookup[block_size];

Jim Bankoski's avatar
Jim Bankoski committed
533
  // If filter level is 0 we don't loop filter.
534
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
535
    return;
536 537 538 539 540 541 542
  } else {
    int index = shift_y;
    for (i = 0; i < h; i++) {
      vpx_memset(&lfm->lfl_y[index], filter_level, w);
      index += 8;
    }
  }
Jim Bankoski's avatar
Jim Bankoski committed
543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602

  // These set 1 in the current block size for the block size edges.
  // For instance if the block size is 32x16,   we'll set :
  //    above =   1111
  //              0000
  //    and
  //    left  =   1000
  //          =   1000
  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
  //        1,  not 8...
  //
  // U and v set things on a 16 bit scale.
  //
  *above_y |= above_prediction_mask[block_size] << shift_y;
  *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
  *left_y |= left_prediction_mask[block_size] << shift_y;
  *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;

  // If the block has no coefficients and is not intra we skip applying
  // the loop filter on block edges.
  if (skip && ref > INTRA_FRAME)
    return;

  // Here we are adding a mask for the transform size.  The transform
  // size mask is set to be correct for a 64x64 prediction block size. We
  // mask to match the size of the block we are working on and then shift it
  // into place..
  *above_y |= (size_mask[block_size] &
               above_64x64_txform_mask[tx_size_y]) << shift_y;
  *above_uv |= (size_mask_uv[block_size] &
                above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;

  *left_y |= (size_mask[block_size] &
              left_64x64_txform_mask[tx_size_y]) << shift_y;
  *left_uv |= (size_mask_uv[block_size] &
               left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;

  // Here we are trying to determine what to do with the internal 4x4 block
  // boundaries.  These differ from the 4x4 boundaries on the outside edge of
  // an 8x8 in that the internal ones can be skipped and don't depend on
  // the prediction block size.
  if (tx_size_y == TX_4X4) {
    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y;
  }
  if (tx_size_uv == TX_4X4) {
    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
  }
}

// This function does the same thing as the one above with the exception that
// it only affects the y masks.   It exists because for blocks < 16x16 in size,
// we only update u and v masks on the first block.
static void build_y_mask(const loop_filter_info_n *const lfi_n,
                         const MODE_INFO *mi, const int shift_y,
                         LOOP_FILTER_MASK *lfm) {
  const BLOCK_SIZE block_size = mi->mbmi.sb_type;
  const TX_SIZE tx_size_y = mi->mbmi.tx_size;
  const int skip = mi->mbmi.skip_coeff;
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
603
  const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
Jim Bankoski's avatar
Jim Bankoski committed
604 605 606
  uint64_t *left_y = &lfm->left_y[tx_size_y];
  uint64_t *above_y = &lfm->above_y[tx_size_y];
  uint64_t *int_4x4_y = &lfm->int_4x4_y;
Yunqing Wang's avatar
Yunqing Wang committed
607 608 609 610
  int i;
  int w = num_8x8_blocks_wide_lookup[block_size];
  int h = num_8x8_blocks_high_lookup[block_size];

611
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
612
    return;
613 614 615 616 617 618 619
  } else {
    int index = shift_y;
    for (i = 0; i < h; i++) {
      vpx_memset(&lfm->lfl_y[index], filter_level, w);
      index += 8;
    }
  }
Jim Bankoski's avatar
Jim Bankoski committed
620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641

  *above_y |= above_prediction_mask[block_size] << shift_y;
  *left_y |= left_prediction_mask[block_size] << shift_y;

  if (skip && ref > INTRA_FRAME)
    return;

  *above_y |= (size_mask[block_size] &
               above_64x64_txform_mask[tx_size_y]) << shift_y;

  *left_y |= (size_mask[block_size] &
              left_64x64_txform_mask[tx_size_y]) << shift_y;

  if (tx_size_y == TX_4X4) {
    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y;
  }
}

// This function sets up the bit masks for the entire 64x64 region represented
// by mi_row, mi_col.
// TODO(JBB): This function only works for yv12.
static void setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
642
                       MODE_INFO **mi_8x8, const int mode_info_stride,
Jim Bankoski's avatar
Jim Bankoski committed
643 644 645
                       LOOP_FILTER_MASK *lfm) {
  int idx_32, idx_16, idx_8;
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
646 647
  MODE_INFO **mip = mi_8x8;
  MODE_INFO **mip2 = mi_8x8;
Jim Bankoski's avatar
Jim Bankoski committed
648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678

  // These are offsets to the next mi in the 64x64 block. It is what gets
  // added to the mi ptr as we go through each loop.  It helps us to avoids
  // setting up special row and column counters for each index.  The last step
  // brings us out back to the starting position.
  const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
                           -(mode_info_stride << 2) - 4};
  const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
                           -(mode_info_stride << 1) - 2};
  const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};

  // Following variables represent shifts to position the current block
  // mask over the appropriate block.   A shift of 36 to the left will move
  // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
  // 4 rows to the appropriate spot.
  const int shift_32_y[] = {0, 4, 32, 36};
  const int shift_16_y[] = {0, 2, 16, 18};
  const int shift_8_y[] = {0, 1, 8, 9};
  const int shift_32_uv[] = {0, 2, 8, 10};
  const int shift_16_uv[] = {0, 1, 4, 5};
  int i;
  const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
                        cm->mi_rows - mi_row : MI_BLOCK_SIZE);
  const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
                        cm->mi_cols - mi_col : MI_BLOCK_SIZE);

  vp9_zero(*lfm);

  // TODO(jimbankoski): Try moving most of the following code into decode
  // loop and storing lfm in the mbmi structure so that we don't have to go
  // through the recursive loop structure multiple times.
679
  switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
680
    case BLOCK_64X64:
681
      build_masks(lfi_n, mip[0] , 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
682 683
      break;
    case BLOCK_64X32:
684
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
685
      mip2 = mip + mode_info_stride * 4;
686 687
      if (4 >= max_rows)
        break;
688
      build_masks(lfi_n, mip2[0], 32, 8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
689 690
      break;
    case BLOCK_32X64:
691
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
692
      mip2 = mip + 4;
693 694
      if (4 >= max_cols)
        break;
695
      build_masks(lfi_n, mip2[0], 4, 2, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
696 697 698 699 700 701 702 703 704
      break;
    default:
      for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
        const int shift_y = shift_32_y[idx_32];
        const int shift_uv = shift_32_uv[idx_32];
        const int mi_32_col_offset = ((idx_32 & 1) << 2);
        const int mi_32_row_offset = ((idx_32 >> 1) << 2);
        if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
          continue;
705
        switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
706
          case BLOCK_32X32:
707
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
708 709
            break;
          case BLOCK_32X16:
710
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
711 712
            if (mi_32_row_offset + 2 >= max_rows)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
713
            mip2 = mip + mode_info_stride * 2;
714
            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
715 716
            break;
          case BLOCK_16X32:
717
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
718 719
            if (mi_32_col_offset + 2 >= max_cols)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
720
            mip2 = mip + 2;
721
            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
722 723 724 725 726 727 728 729 730 731 732 733 734
            break;
          default:
            for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
              const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
              const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
              const int mi_16_col_offset = mi_32_col_offset +
                  ((idx_16 & 1) << 1);
              const int mi_16_row_offset = mi_32_row_offset +
                  ((idx_16 >> 1) << 1);

              if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
                continue;

735
              switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
736
                case BLOCK_16X16:
737
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
738 739
                  break;
                case BLOCK_16X8:
740
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
741 742
                  if (mi_16_row_offset + 1 >= max_rows)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
743
                  mip2 = mip + mode_info_stride;
744
                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
745 746
                  break;
                case BLOCK_8X16:
747
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
748 749
                  if (mi_16_col_offset +1 >= max_cols)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
750
                  mip2 = mip + 1;
751
                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
752 753 754 755 756
                  break;
                default: {
                  const int shift_y = shift_32_y[idx_32] +
                                      shift_16_y[idx_16] +
                                      shift_8_y[0];
757
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
758 759 760 761 762 763 764 765 766 767 768 769 770
                  mip += offset[0];
                  for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
                    const int shift_y = shift_32_y[idx_32] +
                                        shift_16_y[idx_16] +
                                        shift_8_y[idx_8];
                    const int mi_8_col_offset = mi_16_col_offset +
                        ((idx_8 & 1));
                    const int mi_8_row_offset = mi_16_row_offset +
                        ((idx_8 >> 1));

                    if (mi_8_col_offset >= max_cols ||
                        mi_8_row_offset >= max_rows)
                      continue;
771
                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870
                  }
                  break;
                }
              }
            }
            break;
        }
      }
      break;
  }
  // The largest loopfilter we have is 16x16 so we use the 16x16 mask
  // for 32x32 transforms also also.
  lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
  lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
  lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
  lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];

  // We do at least 8 tap filter on every 32x32 even if the transform size
  // is 4x4.  So if the 4x4 is set on a border pixel add it to the 8x8 and
  // remove it from the 4x4.
  lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
  lfm->left_y[TX_4X4] &= ~left_border;
  lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
  lfm->above_y[TX_4X4] &= ~above_border;
  lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
  lfm->left_uv[TX_4X4] &= ~left_border_uv;
  lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
  lfm->above_uv[TX_4X4] &= ~above_border_uv;

  // We do some special edge handling.
  if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
    const uint64_t rows = cm->mi_rows - mi_row;

    // Each pixel inside the border gets a 1,
    const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
    const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);

    // Remove values completely outside our border.
    for (i = 0; i < TX_32X32; i++) {
      lfm->left_y[i] &= mask_y;
      lfm->above_y[i] &= mask_y;
      lfm->left_uv[i] &= mask_uv;
      lfm->above_uv[i] &= mask_uv;
    }
    lfm->int_4x4_y &= mask_y;
    lfm->int_4x4_uv &= mask_uv;

    // We don't apply a wide loop filter on the last uv block row.  If set
    // apply the shorter one instead.
    if (rows == 1) {
      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
      lfm->above_uv[TX_16X16] = 0;
    }
    if (rows == 5) {
      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
      lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
    }
  }

  if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
    const uint64_t columns = cm->mi_cols - mi_col;

    // Each pixel inside the border gets a 1, the multiply copies the border
    // to where we need it.
    const uint64_t mask_y  = (((1 << columns) - 1)) * 0x0101010101010101;
    const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;

    // Internal edges are not applied on the last column of the image so
    // we mask 1 more for the internal edges
    const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;

    // Remove the bits outside the image edge.
    for (i = 0; i < TX_32X32; i++) {
      lfm->left_y[i] &= mask_y;
      lfm->above_y[i] &= mask_y;
      lfm->left_uv[i] &= mask_uv;
      lfm->above_uv[i] &= mask_uv;
    }
    lfm->int_4x4_y &= mask_y;
    lfm->int_4x4_uv &= mask_uv_int;

    // We don't apply a wide loop filter on the last uv column.  If set
    // apply the shorter one instead.
    if (columns == 1) {
      lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
      lfm->left_uv[TX_16X16] = 0;
    }
    if (columns == 5) {
      lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
      lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
    }
  }
  // We don't a loop filter on the first column in the image.  Mask that out.
  if (mi_col == 0) {
    for (i = 0; i < TX_32X32; i++) {
      lfm->left_y[i] &= 0xfefefefefefefefe;
      lfm->left_uv[i] &= 0xeeee;
    }
  }
Jim Bankoski's avatar
Jim Bankoski committed
871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888

  // Assert if we try to apply 2 different loop filters at the same position.
  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
  assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
  assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
  assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
  assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
  assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
  assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
  assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
  assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
  assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
  assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
Jim Bankoski's avatar
Jim Bankoski committed
889
}
Yunqing Wang's avatar
Yunqing Wang committed
890

891
#if CONFIG_NON420
892 893 894 895
static uint8_t build_lfi(const loop_filter_info_n *lfi_n,
                     const MB_MODE_INFO *mbmi) {
  const int seg = mbmi->segment_id;
  const int ref = mbmi->ref_frame[0];
896
  return lfi_n->lvl[seg][ref][mode_lf_lut[mbmi->mode]];
897 898
}

899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
static void filter_selectively_vert(uint8_t *s, int pitch,
                                    unsigned int mask_16x16,
                                    unsigned int mask_8x8,
                                    unsigned int mask_4x4,
                                    unsigned int mask_4x4_int,
                                    const loop_filter_info_n *lfi_n,
                                    const uint8_t *lfl) {
  unsigned int mask;

  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
       mask; mask >>= 1) {
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;

    if (mask & 1) {
      if (mask_16x16 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
914
        vp9_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
915
      } else if (mask_8x8 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
916
        vp9_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
917
      } else if (mask_4x4 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
918
        vp9_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
919 920 921
      }
    }
    if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
922
      vp9_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
923 924 925 926 927 928 929 930 931
    s += 8;
    lfl += 1;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
    mask_4x4_int >>= 1;
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
932 933
static void filter_block_plane_non420(VP9_COMMON *cm,
                                      struct macroblockd_plane *plane,
934
                                      MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
935
                                      int mi_row, int mi_col) {
936 937
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
938 939
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
940 941
  const int row_step_stride = cm->mode_info_stride * row_step;
  struct buf_2d *const dst = &plane->dst;
942
  uint8_t* const dst0 = dst->buf;
943 944 945 946
  unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
  unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
  unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
  unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
947
  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
948
  int r, c;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
949

950
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
951 952 953 954 955 956
    unsigned int mask_16x16_c = 0;
    unsigned int mask_8x8_c = 0;
    unsigned int mask_4x4_c = 0;
    unsigned int border_mask;

    // Determine the vertical edges that need filtering
957
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
958
      const MODE_INFO *mi = mi_8x8[c];
959
      const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
960 961
      const int skip_this = mi[0].mbmi.skip_coeff
                            && is_inter_block(&mi[0].mbmi);
962
      // left edge of current unit is block/partition edge -> no skip
963 964
      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
965
      const int skip_this_c = skip_this && !block_edge_left;
966
      // top edge of current unit is block/partition edge -> no skip
967 968
      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
969
      const int skip_this_r = skip_this && !block_edge_above;
970
      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
971 972
                            ? get_uv_tx_size(&mi[0].mbmi)
                            : mi[0].mbmi.tx_size;
973
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
974
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
975 976

      // Filter level can vary per MI
977 978
      if (!(lfl[(r << 3) + (c >> ss_x)] =
          build_lfi(&cm->lf_info, &mi[0].mbmi)))
979 980 981 982
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
983 984 985 986 987 988 989 990 991 992 993 994
        if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
          if (!skip_border_4x4_c)
            mask_16x16_c |= 1 << (c >> ss_x);
          else
            mask_8x8_c |= 1 << (c >> ss_x);
        }
        if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
          if (!skip_border_4x4_r)
            mask_16x16[r] |= 1 << (c >> ss_x);
          else
            mask_8x8[r] |= 1 << (c >> ss_x);
        }
995
      } else if (tx_size == TX_16X16) {
996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
        if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
          if (!skip_border_4x4_c)
            mask_16x16_c |= 1 << (c >> ss_x);
          else
            mask_8x8_c |= 1 << (c >> ss_x);
        }
        if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
          if (!skip_border_4x4_r)
            mask_16x16[r] |= 1 << (c >> ss_x);
          else
            mask_8x8[r] |= 1 << (c >> ss_x);
        }
1008 1009 1010
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
1011
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
1012 1013 1014 1015 1016 1017
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
1018
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
1019 1020 1021 1022 1023
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

1024 1025
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
1026 1027 1028 1029 1030 1031 1032 1033 1034
      }
    }

    // Disable filtering on the leftmost column
    border_mask = ~(mi_col == 0);
    filter_selectively_vert(dst->buf, dst->stride,
                            mask_16x16_c & border_mask,
                            mask_8x8_c & border_mask,
                            mask_4x4_c & border_mask,
Yunqing Wang's avatar
Yunqing Wang committed
1035
                            mask_4x4_int[r],
1036
                            &cm->lf_info, &lfl[r << 3]);
1037
    dst->buf += 8 * dst->stride;
1038
    mi_8x8 += row_step_stride;
1039 1040 1041 1042
  }

  // Now do horizontal pass
  dst->buf = dst0;
1043
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1044 1045 1046
    const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
    const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];

1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060
    unsigned int mask_16x16_r;
    unsigned int mask_8x8_r;
    unsigned int mask_4x4_r;

    if (mi_row + r == 0) {
      mask_16x16_r = 0;
      mask_8x8_r = 0;
      mask_4x4_r = 0;
    } else {
      mask_16x16_r = mask_16x16[r];
      mask_8x8_r = mask_8x8[r];
      mask_4x4_r = mask_4x4[r];
    }

1061
    filter_selectively_horiz(dst->buf, dst->stride,
1062 1063 1064 1065
                             mask_16x16_r,
                             mask_8x8_r,
                             mask_4x4_r,
                             mask_4x4_int_r,
1066
                             &cm->lf_info, &lfl[r << 3]);
1067 1068 1069
    dst->buf += 8 * dst->stride;
  }
}
1070
#endif
1071

Jim Bankoski's avatar
Jim Bankoski committed
1072 <