vp9_loopfilter.c 46.2 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
35
36
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
// 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;

204
205
206
207
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
208

Frank Galligan's avatar
Frank Galligan committed
209
static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
210
  int lvl;
John Koleszar's avatar
John Koleszar committed
211

Dmitry Kovalev's avatar
Dmitry Kovalev committed
212
213
214
215
  // 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
216

John Koleszar's avatar
John Koleszar committed
217
218
219
    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
John Koleszar's avatar
John Koleszar committed
220
    }
John Koleszar's avatar
John Koleszar committed
221
222
223
224

    if (block_inside_limit < 1)
      block_inside_limit = 1;

225
226
    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
227
228
               SIMD_WIDTH);
  }
Attila Nagy's avatar
Attila Nagy committed
229
}
John Koleszar's avatar
John Koleszar committed
230

Dmitry Kovalev's avatar
Dmitry Kovalev committed
231
232
233
234
235
236
static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
                                const MB_MODE_INFO *mbmi) {
  return lfi_n->lvl[mbmi->segment_id][mbmi->ref_frame[0]]
                   [mode_lf_lut[mbmi->mode]];
}

237
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
238
  loop_filter_info_n *lfi = &cm->lf_info;
239
  struct loopfilter *lf = &cm->lf;
240
  int lvl;
Attila Nagy's avatar
Attila Nagy committed
241

242
  // init limits for given sharpness
243
244
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
245

246
  // init hev threshold const vectors
247
248
  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
249
250
}

Frank Galligan's avatar
Frank Galligan committed
251
void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
252
  int seg_id;
253
254
255
  // 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
256
  const int scale = 1 << (default_filt_lvl >> 5);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
257
  loop_filter_info_n *const lfi = &cm->lf_info;
258
  struct loopfilter *const lf = &cm->lf;
259
  const struct segmentation *const seg = &cm->seg;
John Koleszar's avatar
John Koleszar committed
260

Dmitry Kovalev's avatar
Dmitry Kovalev committed
261
  // update limits if sharpness has changed
262
263
264
  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
265
266
  }

267
  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
268
    int lvl_seg = default_filt_lvl;
269
    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
270
      const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
271
272
273
      lvl_seg = clamp(seg->abs_delta == SEGMENT_ABSDATA ?
                      data : default_filt_lvl + data,
                      0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
274
    }
John Koleszar's avatar
John Koleszar committed
275

276
    if (!lf->mode_ref_delta_enabled) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
277
278
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
279
      vpx_memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
280
281
282
283
284
285
286
287
288
289
290
    } 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
291
      }
292
    }
John Koleszar's avatar
John Koleszar committed
293
  }
John Koleszar's avatar
John Koleszar committed
294
295
}

296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
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;
316
317
  unsigned int mask;

318
319
320
321
322
  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);
323

324
    // TODO(yunqingwang): count in loopfilter functions should be removed.
325
    if (mask & 1) {
326
327
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
328
329
          vp9_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                   lfi0->hev_thr);
330
        } else if (mask_16x16_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
331
332
          vp9_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
                              lfi0->hev_thr);
333
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
334
335
          vp9_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
                              lfi1->lim, lfi1->hev_thr);
336
        }
337
338
339
340
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
341
342
343
          vp9_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
344
        } else if (mask_8x8_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
345
346
          vp9_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
347
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
348
349
          vp9_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
350
351
352
353
354
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
355
356
357
          vp9_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
358
        } else if (mask_4x4_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
359
360
          vp9_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
361
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
362
363
          vp9_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
364
365
366
367
368
        }
      }

      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
369
370
371
          vp9_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
372
        } else if (mask_4x4_int_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
373
374
          vp9_lpf_vertical_4(s + 4, 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 + 4, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
378
        }
379
380
      }
    }
381

382
    s += 8;
Yunqing Wang's avatar
Yunqing Wang committed
383
    lfl += 1;
384
385
386
387
388
389
390
391
    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;
392
393
394
395
396
397
398
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
399
                                     unsigned int mask_4x4_int,
Yunqing Wang's avatar
Yunqing Wang committed
400
401
                                     const loop_filter_info_n *lfi_n,
                                     const uint8_t *lfl) {
402
  unsigned int mask;
403
  int count;
404

John Koleszar's avatar
John Koleszar committed
405
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
406
       mask; mask >>= count) {
Yunqing Wang's avatar
Yunqing Wang committed
407
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
408

Dmitry Kovalev's avatar
Dmitry Kovalev committed
409
    count = 1;
410
    if (mask & 1) {
411
      if (mask_16x16 & 1) {
412
        if ((mask_16x16 & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
413
414
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 2);
415
416
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
417
418
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 1);
419
        }
420
421
422
423
424
      } 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
425
426
427
          vp9_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
428
429

          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
430
431
432
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
433
          } else {
434
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
435
436
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
437
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
438
439
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
440
          }
441
442
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
443
          vp9_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
444
445

          if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
446
447
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
448
        }
449
450
451
452
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
453

Jim Bankoski's avatar
Jim Bankoski committed
454
455
456
          vp9_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
457
          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
458
459
460
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
461
462
          } else {
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
463
464
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
465
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
466
467
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
468
469
470
          }
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
471
          vp9_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
472

Jim Bankoski's avatar
Jim Bankoski committed
473
474
475
          if (mask_4x4_int & 1)
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
476
477
        }
      } else if (mask_4x4_int & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
478
479
        vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                             lfi->hev_thr, 1);
480
      }
481
    }
482
    s += 8 * count;
Yunqing Wang's avatar
Yunqing Wang committed
483
    lfl += count;
484
485
486
487
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
488
489
490
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
491
492
493
494
495
496
497
498
499
500
501
// 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) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
502
503
504
  const MB_MODE_INFO *mbmi = &mi->mbmi;
  const BLOCK_SIZE block_size = mbmi->sb_type;
  const TX_SIZE tx_size_y = mbmi->tx_size;
505
  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
506
507
508
509
510
511
512
  const int filter_level = get_filter_level(lfi_n, mbmi);
  uint64_t *const left_y = &lfm->left_y[tx_size_y];
  uint64_t *const above_y = &lfm->above_y[tx_size_y];
  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
  uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
Yunqing Wang's avatar
Yunqing Wang committed
513
514
  int i;

Jim Bankoski's avatar
Jim Bankoski committed
515
  // If filter level is 0 we don't loop filter.
516
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
517
    return;
518
  } else {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
519
520
    const int w = num_8x8_blocks_wide_lookup[block_size];
    const int h = num_8x8_blocks_high_lookup[block_size];
521
522
523
524
525
526
    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
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546

  // 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.
Dmitry Kovalev's avatar
Dmitry Kovalev committed
547
  if (mbmi->skip && is_inter_block(mbmi))
Jim Bankoski's avatar
Jim Bankoski committed
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
    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.
Dmitry Kovalev's avatar
Dmitry Kovalev committed
568
  if (tx_size_y == TX_4X4)
Jim Bankoski's avatar
Jim Bankoski committed
569
    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
570
571

  if (tx_size_uv == TX_4X4)
Jim Bankoski's avatar
Jim Bankoski committed
572
573
574
575
576
577
578
579
580
    *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) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
581
582
583
584
585
586
587
  const MB_MODE_INFO *mbmi = &mi->mbmi;
  const BLOCK_SIZE block_size = mbmi->sb_type;
  const TX_SIZE tx_size_y = mbmi->tx_size;
  const int filter_level = get_filter_level(lfi_n, mbmi);
  uint64_t *const left_y = &lfm->left_y[tx_size_y];
  uint64_t *const above_y = &lfm->above_y[tx_size_y];
  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
Yunqing Wang's avatar
Yunqing Wang committed
588
589
  int i;

590
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
591
    return;
592
  } else {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
593
594
    const int w = num_8x8_blocks_wide_lookup[block_size];
    const int h = num_8x8_blocks_high_lookup[block_size];
595
596
597
598
599
600
    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
601
602
603
604

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

Dmitry Kovalev's avatar
Dmitry Kovalev committed
605
  if (mbmi->skip && is_inter_block(mbmi))
Jim Bankoski's avatar
Jim Bankoski committed
606
607
608
609
610
611
612
613
    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;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
614
  if (tx_size_y == TX_4X4)
Jim Bankoski's avatar
Jim Bankoski committed
615
616
617
618
619
620
    *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.
621
void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
622
                    MODE_INFO **mi, const int mode_info_stride,
623
                    LOOP_FILTER_MASK *lfm) {
Jim Bankoski's avatar
Jim Bankoski committed
624
625
  int idx_32, idx_16, idx_8;
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
626
627
  MODE_INFO **mip = mi;
  MODE_INFO **mip2 = mi;
Jim Bankoski's avatar
Jim Bankoski committed
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658

  // 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.
659
  switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
660
    case BLOCK_64X64:
661
      build_masks(lfi_n, mip[0] , 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
662
663
      break;
    case BLOCK_64X32:
664
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
665
      mip2 = mip + mode_info_stride * 4;
666
667
      if (4 >= max_rows)
        break;
668
      build_masks(lfi_n, mip2[0], 32, 8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
669
670
      break;
    case BLOCK_32X64:
671
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
672
      mip2 = mip + 4;
673
674
      if (4 >= max_cols)
        break;
675
      build_masks(lfi_n, mip2[0], 4, 2, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
676
677
678
679
680
681
682
683
684
      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;
685
        switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
686
          case BLOCK_32X32:
687
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
688
689
            break;
          case BLOCK_32X16:
690
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
691
692
            if (mi_32_row_offset + 2 >= max_rows)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
693
            mip2 = mip + mode_info_stride * 2;
694
            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
695
696
            break;
          case BLOCK_16X32:
697
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
698
699
            if (mi_32_col_offset + 2 >= max_cols)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
700
            mip2 = mip + 2;
701
            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
702
703
704
705
706
707
708
709
710
711
712
713
714
            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;

715
              switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
716
                case BLOCK_16X16:
717
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
718
719
                  break;
                case BLOCK_16X8:
720
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
721
722
                  if (mi_16_row_offset + 1 >= max_rows)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
723
                  mip2 = mip + mode_info_stride;
724
                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
725
726
                  break;
                case BLOCK_8X16:
727
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
728
729
                  if (mi_16_col_offset +1 >= max_cols)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
730
                  mip2 = mip + 1;
731
                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
732
733
734
735
736
                  break;
                default: {
                  const int shift_y = shift_32_y[idx_32] +
                                      shift_16_y[idx_16] +
                                      shift_8_y[0];
737
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
738
739
740
741
742
743
744
745
746
747
748
749
750
                  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;
751
                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
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
                  }
                  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
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868

  // 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
869
}
Yunqing Wang's avatar
Yunqing Wang committed
870

871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
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
886
        vp9_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
887
      } else if (mask_8x8 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
888
        vp9_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
889
      } else if (mask_4x4 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
890
        vp9_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
891
892
893
      }
    }
    if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
894
      vp9_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
895
896
897
898
899
900
901
902
903
    s += 8;
    lfl += 1;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
    mask_4x4_int >>= 1;
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
904
905
static void filter_block_plane_non420(VP9_COMMON *cm,
                                      struct macroblockd_plane *plane,
906
                                      MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
907
                                      int mi_row, int mi_col) {
908
909
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
910
911
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
912
  const int row_step_stride = cm->mi_stride * row_step;
913
  struct buf_2d *const dst = &plane->dst;
914
  uint8_t* const dst0 = dst->buf;
915
916
917
918
  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};
919
  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
920
  int r, c;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
921

922
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
923
924
925
926
927
928
    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
929
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
930
      const MODE_INFO *mi = mi_8x8[c];
931
      const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
932
      const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
933
      // left edge of current unit is block/partition edge -> no skip
934
935
      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
936
      const int skip_this_c = skip_this && !block_edge_left;
937
      // top edge of current unit is block/partition edge -> no skip
938
939
      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
940
      const int skip_this_r = skip_this && !block_edge_above;
941
      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
942
                            ? get_uv_tx_size(&mi[0].mbmi, plane)
943
                            : mi[0].mbmi.tx_size;
944
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
945
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
946
947

      // Filter level can vary per MI
948
      if (!(lfl[(r << 3) + (c >> ss_x)] =
Dmitry Kovalev's avatar
Dmitry Kovalev committed
949
            get_filter_level(&cm->lf_info, &mi[0].mbmi)))
950
951
952
953
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
954
955
956
957
958
959
960
961
962
963
964
965
        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);
        }
966
      } else if (tx_size == TX_16X16) {
967
968
969
970
971
972
973
974
975
976
977
978
        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);
        }
979
980
981
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
982
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
983
984
985
986
987
988
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
989
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
990
991
992
993
994
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

995
996
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
997
998
999
1000
1001
1002
1003
1004
1005
      }
    }

    // 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
1006
                            mask_4x4_int[r],
1007
                            &cm->lf_info, &lfl[r << 3]);
1008
    dst->buf += 8 * dst->stride;
1009
    mi_8x8 += row_step_stride;
1010
1011
1012
1013
  }

  // Now do horizontal pass
  dst->buf = dst0;
1014
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1015
1016
1017
    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];

1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
    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];
    }

1032
    filter_selectively_horiz(dst->buf, dst->stride,
1033
1034
1035
1036
                             mask_16x16_r,
                             mask_8x8_r,
                             mask_4x4_r,
                             mask_4x4_int_r,
1037
                             &cm->lf_info, &lfl[r << 3]);
1038
1039
1040
1041
    dst->buf += 8 * dst->stride;
  }
}

1042
1043
1044
1045
void vp9_filter_block_plane(VP9_COMMON *const cm,
                            struct macroblockd_plane *const plane,
                            int mi_row,
                            LOOP_FILTER_MASK *lfm) {
Jim Bankoski's avatar
Jim Bankoski committed
1046
1047
1048
1049
  struct buf_2d *const dst = &plane->dst;
  uint8_t* const dst0 = dst->buf;
  int r, c;

Yunqing Wang's avatar
Yunqing Wang committed
1050
1051
1052
1053
1054
  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
1055

1056
1057
1058
1059
1060
1061
    // 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;
1062

Jim Bankoski's avatar
Jim Bankoski committed
1063
      // Disable filtering on the leftmost column
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
      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
1077
1078
1079
1080
1081
1082
1083
    }

    // 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];
1084
    mask_4x4_int = lfm->int_4x4_y;
Yunqing Wang's avatar
Yunqing Wang committed
1085
1086

    for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
      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;
      }

Yunqing Wang's avatar
Yunqing Wang committed
1101
      filter_selectively_horiz(dst->buf, dst->stride,
1102
1103
1104
                               mask_16x16_r,
                               mask_8x8_r,
                               mask_4x4_r,
1105
                               mask_4x4_int & 0xff,
1106
                               &cm->lf_info, &lfm->lfl_y[r << 3]);
Yunqing Wang's avatar
Yunqing Wang committed
1107
1108
1109
1110
1111

      dst->buf += 8 * dst->stride;
      mask_16x16 >>= 8;
      mask_8x8 >>= 8;
      mask_4x4 >>= 8;
1112
      mask_4x4_int >>= 8;
Yunqing Wang's avatar
Yunqing Wang committed
1113
1114
1115
1116
1117
1118
1119
    }
  } else {
    uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
    uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
    uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
    uint16_t mask_4x4_int = lfm->int_4x4_uv;

1120
1121
    // Vertical pass: do 2 rows at one time
    for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
Yunqing Wang's avatar
Yunqing Wang committed
1122
      if (plane->plane_type == 1) {
1123
        for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
1124
          lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
1125
1126
1127
          lfm->lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) +
                                                       (c << 1)];
        }
Yunqing Wang's avatar
Yunqing Wang committed
1128
1129
      }

1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
      {
        unsigned int mask_16x16_l = mask_16x16 & 0xff;
        unsigned int mask_8x8_l = mask_8x8 & 0xff;
        unsigned int mask_4x4_l = mask_4x4 & 0xff;
        unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;

        // Disable filtering on the leftmost column
        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_uv[r << 1]);

        dst->buf += 16 * dst->stride;
        mask_16x16 >>= 8;
        mask_8x8 >>= 8;
        mask_4x4 >>= 8;
        mask_4x4_int >>= 8;
      }
Jim Bankoski's avatar
Jim Bankoski committed
1151
1152
    }

Yunqing Wang's avatar
Yunqing Wang committed
1153
1154
1155
1156
1157
    // Horizontal pass
    dst->buf = dst0;
    mask_16x16 = lfm->above_uv[TX_16X16];
    mask_8x8 = lfm->above_uv[TX_8X8];
    mask_4x4 = lfm->above_uv[TX_4X4];
1158
    mask_4x4_int = lfm->int_4x4_uv;
Yunqing Wang's avatar
Yunqing Wang committed
1159
1160
1161
1162

    for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
      const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
      const unsigned int mask_4x4_int_r = skip_border_4x4_r ?
1163
          0 : (mask_4x4_int & 0xf);
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
      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 & 0xf;
        mask_8x8_r = mask_8x8 & 0xf;
        mask_4x4_r = mask_4x4 & 0xf;
      }
Jim Bankoski's avatar
Jim Bankoski committed
1177
1178

      filter_selectively_horiz(dst->buf, dst->stride,
1179
1180
1181
                               mask_16x16_r,
                               mask_8x8_r,
                               mask_4x4_r,
Jim Bankoski's avatar
Jim Bankoski committed
1182
                               mask_4x4_int_r,
1183
                               &cm->lf_info, &lfm->lfl_uv[r << 1]);
Yunqing Wang's avatar
Yunqing Wang committed
1184
1185
1186
1187
1188

      dst->buf += 8 * dst->stride;
      mask_16x16 >>= 4;
      mask_8x8 >>= 4;
      mask_4x4 >>= 4;
1189
      mask_4x4_int >>= 4;
Jim Bankoski's avatar
Jim Bankoski committed
1190
1191
1192
1193
    }
  }
}

1194
void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
1195
1196
                          VP9_COMMON *cm,
                          struct macroblockd_plane planes[MAX_MB_PLANE],
1197
                          int start, int stop, int y_only) {
1198
  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
1199
1200