vp9_loopfilter.c 48.8 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
} 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;


John Koleszar's avatar
John Koleszar committed
225
static void lf_init_lut(loop_filter_info_n *lfi) {
226 227 228 229 230
  lfi->mode_lf_lut[DC_PRED] = 0;
  lfi->mode_lf_lut[D45_PRED] = 0;
  lfi->mode_lf_lut[D135_PRED] = 0;
  lfi->mode_lf_lut[D117_PRED] = 0;
  lfi->mode_lf_lut[D153_PRED] = 0;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
231
  lfi->mode_lf_lut[D207_PRED] = 0;
232 233 234 235 236 237 238 239
  lfi->mode_lf_lut[D63_PRED] = 0;
  lfi->mode_lf_lut[V_PRED] = 0;
  lfi->mode_lf_lut[H_PRED] = 0;
  lfi->mode_lf_lut[TM_PRED] = 0;
  lfi->mode_lf_lut[ZEROMV]  = 0;
  lfi->mode_lf_lut[NEARESTMV] = 1;
  lfi->mode_lf_lut[NEARMV] = 1;
  lfi->mode_lf_lut[NEWMV] = 1;
Attila Nagy's avatar
Attila Nagy committed
240 241
}

Frank Galligan's avatar
Frank Galligan committed
242
static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
243
  int lvl;
John Koleszar's avatar
John Koleszar committed
244

Dmitry Kovalev's avatar
Dmitry Kovalev committed
245 246 247 248
  // 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
249

John Koleszar's avatar
John Koleszar committed
250 251 252
    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
John Koleszar's avatar
John Koleszar committed
253
    }
John Koleszar's avatar
John Koleszar committed
254 255 256 257

    if (block_inside_limit < 1)
      block_inside_limit = 1;

258 259
    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
260 261
               SIMD_WIDTH);
  }
Attila Nagy's avatar
Attila Nagy committed
262
}
John Koleszar's avatar
John Koleszar committed
263

264
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
265
  loop_filter_info_n *lfi = &cm->lf_info;
266
  struct loopfilter *lf = &cm->lf;
267
  int lvl;
Attila Nagy's avatar
Attila Nagy committed
268

269
  // init limits for given sharpness
270 271
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
272

273
  // init LUT for lvl  and hev thr picking
John Koleszar's avatar
John Koleszar committed
274
  lf_init_lut(lfi);
Attila Nagy's avatar
Attila Nagy committed
275

276
  // init hev threshold const vectors
277 278
  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
279 280
}

Frank Galligan's avatar
Frank Galligan committed
281
void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
282
  int seg_id;
283 284 285
  // 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
286
  const int scale = 1 << (default_filt_lvl >> 5);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
287
  loop_filter_info_n *const lfi = &cm->lf_info;
288
  struct loopfilter *const lf = &cm->lf;
289
  const struct segmentation *const seg = &cm->seg;
John Koleszar's avatar
John Koleszar committed
290

Dmitry Kovalev's avatar
Dmitry Kovalev committed
291
  // update limits if sharpness has changed
292 293 294
  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
295 296
  }

297
  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
298
    int lvl_seg = default_filt_lvl;
299
    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
300 301
      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
302 303
                  ? data
                  : clamp(default_filt_lvl + data, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
304
    }
John Koleszar's avatar
John Koleszar committed
305

306
    if (!lf->mode_ref_delta_enabled) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
307 308
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
309
      vpx_memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
310 311 312 313 314 315 316 317 318 319 320
    } 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
321
      }
322
    }
John Koleszar's avatar
John Koleszar committed
323
  }
John Koleszar's avatar
John Koleszar committed
324 325
}

326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
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;
346 347
  unsigned int mask;

348 349 350 351 352
  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);
353

354
    // TODO(yunqingwang): count in loopfilter functions should be removed.
355
    if (mask & 1) {
356 357 358 359 360 361 362 363 364 365 366
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
          vp9_mb_lpf_vertical_edge_w_16(s, pitch, lfi0->mblim, lfi0->lim,
                                     lfi0->hev_thr);
        } else if (mask_16x16_0 & 1) {
          vp9_mb_lpf_vertical_edge_w(s, pitch, lfi0->mblim, lfi0->lim,
                                     lfi0->hev_thr);
        } else {
          vp9_mb_lpf_vertical_edge_w(s + 8 *pitch, pitch, lfi1->mblim,
                                     lfi1->lim, lfi1->hev_thr);
        }
367 368 369 370
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
371 372 373
          vp9_mbloop_filter_vertical_edge_16(s, pitch, lfi0->mblim, lfi0->lim,
                                          lfi0->hev_thr, lfi1->mblim,
                                          lfi1->lim, lfi1->hev_thr);
374 375 376 377 378 379 380 381 382 383 384
        } else if (mask_8x8_0 & 1) {
          vp9_mbloop_filter_vertical_edge(s, pitch, lfi0->mblim, lfi0->lim,
                                          lfi0->hev_thr, 1);
        } else {
          vp9_mbloop_filter_vertical_edge(s + 8 *pitch, pitch, lfi1->mblim,
                                          lfi1->lim, lfi1->hev_thr, 1);
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
385 386 387
          vp9_loop_filter_vertical_edge_16(s, pitch, lfi0->mblim, lfi0->lim,
                                        lfi0->hev_thr, lfi1->mblim,
                                        lfi1->lim, lfi1->hev_thr);
388 389 390 391 392 393 394 395 396 397 398
        } else if (mask_4x4_0 & 1) {
          vp9_loop_filter_vertical_edge(s, pitch, lfi0->mblim, lfi0->lim,
                                        lfi0->hev_thr, 1);
        } else {
          vp9_loop_filter_vertical_edge(s + 8 *pitch, pitch, lfi1->mblim,
                                        lfi1->lim, lfi1->hev_thr, 1);
        }
      }

      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
399 400 401
          vp9_loop_filter_vertical_edge_16(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                        lfi0->hev_thr, lfi1->mblim,
                                        lfi1->lim, lfi1->hev_thr);
402 403 404 405 406 407 408
        } else if (mask_4x4_int_0 & 1) {
          vp9_loop_filter_vertical_edge(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                        lfi0->hev_thr, 1);
        } else {
          vp9_loop_filter_vertical_edge(s + 8 *pitch + 4, pitch, lfi1->mblim,
                                        lfi1->lim, lfi1->hev_thr, 1);
        }
409 410
      }
    }
411

412
    s += 8;
Yunqing Wang's avatar
Yunqing Wang committed
413
    lfl += 1;
414 415 416 417 418 419 420 421
    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;
422 423 424 425 426 427 428
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
429
                                     unsigned int mask_4x4_int,
Yunqing Wang's avatar
Yunqing Wang committed
430 431
                                     const loop_filter_info_n *lfi_n,
                                     const uint8_t *lfl) {
432
  unsigned int mask;
433
  int count;
434

John Koleszar's avatar
John Koleszar committed
435
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
436
       mask; mask >>= count) {
Yunqing Wang's avatar
Yunqing Wang committed
437
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
438

Dmitry Kovalev's avatar
Dmitry Kovalev committed
439
    count = 1;
440
    if (mask & 1) {
441
      if (mask_16x16 & 1) {
442 443 444 445 446 447 448 449
        if ((mask_16x16 & 3) == 3) {
          vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim,
                                       lfi->hev_thr, 2);
          count = 2;
        } else {
          vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim,
                                       lfi->hev_thr, 1);
        }
450 451 452 453 454
      } else if (mask_8x8 & 1) {
        if ((mask_8x8 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

455 456 457 458
          vp9_mbloop_filter_horizontal_edge_16(s, pitch, lfi->mblim,
                                               lfi->lim, lfi->hev_thr,
                                               lfin->mblim, lfin->lim,
                                               lfin->hev_thr);
459 460

          if ((mask_4x4_int & 3) == 3) {
461 462 463 464
            vp9_loop_filter_horizontal_edge_16(s + 4 * pitch, pitch, lfi->mblim,
                                               lfi->lim, lfi->hev_thr,
                                               lfin->mblim, lfin->lim,
                                               lfin->hev_thr);
465
          } else {
466 467 468 469 470 471 472
            if (mask_4x4_int & 1)
              vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                              lfi->lim, lfi->hev_thr, 1);
            else if (mask_4x4_int & 2)
              vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch,
                                              lfin->mblim, lfin->lim,
                                              lfin->hev_thr, 1);
473
          }
474 475
          count = 2;
        } else {
476 477
          vp9_mbloop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim,
                                            lfi->hev_thr, 1);
478 479 480 481

          if (mask_4x4_int & 1)
            vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                            lfi->lim, lfi->hev_thr, 1);
482
        }
483 484 485 486
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
487

488 489 490 491
          vp9_loop_filter_horizontal_edge_16(s, pitch, lfi->mblim,
                                             lfi->lim, lfi->hev_thr,
                                             lfin->mblim, lfin->lim,
                                             lfin->hev_thr);
492
          if ((mask_4x4_int & 3) == 3) {
493 494 495 496
            vp9_loop_filter_horizontal_edge_16(s + 4 * pitch, pitch, lfi->mblim,
                                               lfi->lim, lfi->hev_thr,
                                               lfin->mblim, lfin->lim,
                                               lfin->hev_thr);
497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
          } else {
            if (mask_4x4_int & 1)
              vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                              lfi->lim, lfi->hev_thr, 1);
            else if (mask_4x4_int & 2)
              vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch,
                                              lfin->mblim, lfin->lim,
                                              lfin->hev_thr, 1);
          }
          count = 2;
        } else {
        vp9_loop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim,
                                        lfi->hev_thr, 1);

        if (mask_4x4_int & 1)
          vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                          lfi->lim, lfi->hev_thr, 1);
        }
      } else if (mask_4x4_int & 1) {
516 517
        vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                        lfi->lim, lfi->hev_thr, 1);
518
      }
519
    }
520
    s += 8 * count;
Yunqing Wang's avatar
Yunqing Wang committed
521
    lfl += count;
522 523 524 525
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
526 527 528
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
// 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];
  const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];
  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
554 555 556 557
  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
558
  // If filter level is 0 we don't loop filter.
559
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
560
    return;
561 562 563 564 565 566 567
  } 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
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632

  // 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];
  const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];
  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
633 634 635 636
  int i;
  int w = num_8x8_blocks_wide_lookup[block_size];
  int h = num_8x8_blocks_high_lookup[block_size];

637
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
638
    return;
639 640 641 642 643 644 645
  } 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
646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667

  *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,
668
                       MODE_INFO **mi_8x8, const int mode_info_stride,
Jim Bankoski's avatar
Jim Bankoski committed
669 670 671
                       LOOP_FILTER_MASK *lfm) {
  int idx_32, idx_16, idx_8;
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
672 673
  MODE_INFO **mip = mi_8x8;
  MODE_INFO **mip2 = mi_8x8;
Jim Bankoski's avatar
Jim Bankoski committed
674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704

  // 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.
705
  switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
706
    case BLOCK_64X64:
707
      build_masks(lfi_n, mip[0] , 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
708 709
      break;
    case BLOCK_64X32:
710
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
711
      mip2 = mip + mode_info_stride * 4;
712 713
      if (4 >= max_rows)
        break;
714
      build_masks(lfi_n, mip2[0], 32, 8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
715 716
      break;
    case BLOCK_32X64:
717
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
718
      mip2 = mip + 4;
719 720
      if (4 >= max_cols)
        break;
721
      build_masks(lfi_n, mip2[0], 4, 2, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
722 723 724 725 726 727 728 729 730
      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;
731
        switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
732
          case BLOCK_32X32:
733
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
734 735
            break;
          case BLOCK_32X16:
736
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
737 738
            if (mi_32_row_offset + 2 >= max_rows)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
739
            mip2 = mip + mode_info_stride * 2;
740
            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
741 742
            break;
          case BLOCK_16X32:
743
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
744 745
            if (mi_32_col_offset + 2 >= max_cols)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
746
            mip2 = mip + 2;
747
            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
748 749 750 751 752 753 754 755 756 757 758 759 760
            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;

761
              switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
762
                case BLOCK_16X16:
763
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
764 765
                  break;
                case BLOCK_16X8:
766
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
767 768
                  if (mi_16_row_offset + 1 >= max_rows)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
769
                  mip2 = mip + mode_info_stride;
770
                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
771 772
                  break;
                case BLOCK_8X16:
773
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
774 775
                  if (mi_16_col_offset +1 >= max_cols)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
776
                  mip2 = mip + 1;
777
                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
778 779 780 781 782
                  break;
                default: {
                  const int shift_y = shift_32_y[idx_32] +
                                      shift_16_y[idx_16] +
                                      shift_8_y[0];
783
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
784 785 786 787 788 789 790 791 792 793 794 795 796
                  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;
797
                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896
                  }
                  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
897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914

  // 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
915
}
Yunqing Wang's avatar
Yunqing Wang committed
916

917
#if CONFIG_NON420
918 919 920 921 922 923 924 925 926 927
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];
  const int mode = lfi_n->mode_lf_lut[mbmi->mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];

  return filter_level;
}

928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964
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) {
        vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr);
      } else if (mask_8x8 & 1) {
        vp9_mbloop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim,
                                        lfi->hev_thr, 1);
      } else if (mask_4x4 & 1) {
        vp9_loop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim,
                                      lfi->hev_thr, 1);
      }
    }
    if (mask_4x4_int & 1)
      vp9_loop_filter_vertical_edge(s + 4, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, 1);
    s += 8;
    lfl += 1;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
    mask_4x4_int >>= 1;
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
965 966
static void filter_block_plane_non420(VP9_COMMON *cm,
                                      struct macroblockd_plane *plane,
967
                                      MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
968
                                      int mi_row, int mi_col) {
969 970
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
971 972
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
973 974
  const int row_step_stride = cm->mode_info_stride * row_step;
  struct buf_2d *const dst = &plane->dst;
975
  uint8_t* const dst0 = dst->buf;
976 977 978 979
  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};
980
  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
981
  int r, c;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
982

983
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
984 985 986 987 988 989
    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
990
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
991
      const MODE_INFO *mi = mi_8x8[c];
992
      const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
993 994
      const int skip_this = mi[0].mbmi.skip_coeff
                            && is_inter_block(&mi[0].mbmi);
995
      // left edge of current unit is block/partition edge -> no skip
996 997
      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
998
      const int skip_this_c = skip_this && !block_edge_left;
999
      // top edge of current unit is block/partition edge -> no skip
1000 1001
      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
1002
      const int skip_this_r = skip_this && !block_edge_above;
1003
      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
1004 1005
                            ? get_uv_tx_size(&mi[0].mbmi)
                            : mi[0].mbmi.tx_size;
1006
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
1007
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
1008 1009

      // Filter level can vary per MI
1010 1011
      if (!(lfl[(r << 3) + (c >> ss_x)] =
          build_lfi(&cm->lf_info, &mi[0].mbmi)))
1012 1013 1014 1015
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027
        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);
        }
1028
      } else if (tx_size == TX_16X16) {
1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
        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);
        }
1041 1042 1043
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
1044
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
1045 1046 1047 1048 1049 1050
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
1051
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
1052 1053 1054 1055 1056
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

1057 1058
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
1059 1060 1061 1062 1063 1064 1065 1066 1067
      }
    }

    // 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
1068
                            mask_4x4_int[r],
1069
                            &cm->lf_info, &lfl[r << 3]);
1070
    dst->buf += 8 * dst->stride;
1071
    mi_8x8 += row_step_stride;
1072 1073 1074 1075
  }

  // Now do horizontal pass
  dst->buf = dst0;
1076
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1077 1078 1079
    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];

1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093
    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];
    }

1094
    filter_selectively_horiz(dst->buf, dst->stride,
1095 1096 1097 1098
                             mask_16x16_r,
                             mask_8x8_r,
                             mask_4x4_r,
                             mask_4x4_int_r,
1099
                             &cm->lf_info, &lfl[r << 3]);
1100 1101 1102
    dst->buf += 8 * dst->stride;
  }
}
1103
#endif
1104

Jim Bankoski's avatar
Jim Bankoski committed
1105 1106
static void filter_block_plane(VP9_COMMON *const cm,
                               struct macroblockd_plane *const plane,
1107
                               int mi_row,
Jim Bankoski's avatar
Jim Bankoski committed
1108 1109 1110 1111 1112
                               LOOP_FILTER_MASK *lfm) {
  struct buf_2d *const dst = &plane->dst;
  uint8_t* const dst0 = dst->buf;
  int r, c;

Yunqing Wang's avatar
Yunqing Wang committed
1113 1114 1115 1116 1117
  if (!plane->plane_type) {
    uint64_t mask_16x16 = lfm->left_y[TX_16X16];
    uint64_t mask_8x8 = lfm->left_y[TX_8X8];
    uint64_t mask_4x4 = lfm->left_y[TX_4X4];
    uint64_t mask_4x4_int = lfm->int_4x4_y;
Jim Bankoski's avatar
Jim Bankoski committed
1118

1119 1120 1121 1122 1123 1124
    // Vertical pass: do 2 rows at one time
    for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
      unsigned int mask_16x16_l = mask_16x16 & 0xffff;
      unsigned int mask_8x8_l = mask_8x8 & 0xffff;
      unsigned int mask_4x4_l = mask_4x4 & 0xffff;
      unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
1125

Jim Bankoski's avatar
Jim Bankoski committed
1126
      // Disable filtering on the leftmost column
1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139
      filter_selectively_vert_row2(plane->plane_type,
                                   dst->buf, dst->stride,
                                   mask_16x16_l,
                                   mask_8x8_l,
                                   mask_4x4_l,
                                   mask_4x4_int_l,
                                   &cm->lf_info, &lfm->lfl_y[r << 3]);

      dst->buf += 16 * dst->stride;
      mask_16x16 >>= 16;
      mask_8x8 >>= 16;
      mask_4x4 >>= 16;
      mask_4x4_int >>= 16;
Yunqing Wang's avatar
Yunqing Wang committed
1140 1141 1142 1143 1144 1145 1146
    }

    // Horizontal pass
    dst->buf = dst0;
    mask_16x16 = lfm->above_y[TX_16X16];
    mask_8x8 = lfm->above_y[TX_8X8];
    mask_4x4 = lfm->above_y[TX_4X4];
1147
    mask_4x4_int = lfm->int_4x4_y;
Yunqing Wang's avatar
Yunqing Wang committed
1148 1149

    for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163
      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 & 0xff;
        mask_8x8_r = mask_8x8 & 0xff;
        mask_4x4_r = mask_4x4 & 0xff;
      }