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

231
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
232
  loop_filter_info_n *lfi = &cm->lf_info;
233
  struct loopfilter *lf = &cm->lf;
234
  int lvl;
Attila Nagy's avatar
Attila Nagy committed
235

236
  // init limits for given sharpness
237 238
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
239

240
  // init hev threshold const vectors
241 242
  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
243 244
}

Frank Galligan's avatar
Frank Galligan committed
245
void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
246
  int seg_id;
247 248 249
  // 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
250
  const int scale = 1 << (default_filt_lvl >> 5);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
251
  loop_filter_info_n *const lfi = &cm->lf_info;
252
  struct loopfilter *const lf = &cm->lf;
253
  const struct segmentation *const seg = &cm->seg;
John Koleszar's avatar
John Koleszar committed
254

Dmitry Kovalev's avatar
Dmitry Kovalev committed
255
  // update limits if sharpness has changed
256 257 258
  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
259 260
  }

261
  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
262
    int lvl_seg = default_filt_lvl;
263
    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
264
      const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
265 266 267
      lvl_seg = clamp(seg->abs_delta == SEGMENT_ABSDATA ?
                      data : default_filt_lvl + data,
                      0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
268
    }
John Koleszar's avatar
John Koleszar committed
269

270
    if (!lf->mode_ref_delta_enabled) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
271 272
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
273
      vpx_memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
274 275 276 277 278 279 280 281 282 283 284
    } 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
285
      }
286
    }
John Koleszar's avatar
John Koleszar committed
287
  }
John Koleszar's avatar
John Koleszar committed
288 289
}

290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
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;
310 311
  unsigned int mask;

312 313 314 315 316
  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);
317

318
    // TODO(yunqingwang): count in loopfilter functions should be removed.
319
    if (mask & 1) {
320 321
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
322 323
          vp9_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                   lfi0->hev_thr);
324
        } else if (mask_16x16_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
325 326
          vp9_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
                              lfi0->hev_thr);
327
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
328 329
          vp9_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
                              lfi1->lim, lfi1->hev_thr);
330
        }
331 332 333 334
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
335 336 337
          vp9_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
338
        } else if (mask_8x8_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
339 340
          vp9_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
341
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
342 343
          vp9_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
344 345 346 347 348
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
349 350 351
          vp9_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
352
        } else if (mask_4x4_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
353 354
          vp9_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
                             1);
355
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
356 357
          vp9_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
358 359 360 361 362
        }
      }

      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
363 364 365
          vp9_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
366
        } else if (mask_4x4_int_0 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
367 368
          vp9_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
                             lfi0->hev_thr, 1);
369
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
370 371
          vp9_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr, 1);
372
        }
373 374
      }
    }
375

376
    s += 8;
Yunqing Wang's avatar
Yunqing Wang committed
377
    lfl += 1;
378 379 380 381 382 383 384 385
    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;
386 387 388 389 390 391 392
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
393
                                     unsigned int mask_4x4_int,
Yunqing Wang's avatar
Yunqing Wang committed
394 395
                                     const loop_filter_info_n *lfi_n,
                                     const uint8_t *lfl) {
396
  unsigned int mask;
397
  int count;
398

John Koleszar's avatar
John Koleszar committed
399
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
400
       mask; mask >>= count) {
Yunqing Wang's avatar
Yunqing Wang committed
401
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
402

Dmitry Kovalev's avatar
Dmitry Kovalev committed
403
    count = 1;
404
    if (mask & 1) {
405
      if (mask_16x16 & 1) {
406
        if ((mask_16x16 & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
407 408
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 2);
409 410
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
411 412
          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
                                lfi->hev_thr, 1);
413
        }
414 415 416 417 418
      } 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
419 420 421
          vp9_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
422 423

          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
424 425 426
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
427
          } else {
428
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
429 430
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
431
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
432 433
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
434
          }
435 436
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
437
          vp9_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
438 439

          if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
440 441
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
442
        }
443 444 445 446
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
447

Jim Bankoski's avatar
Jim Bankoski committed
448 449 450
          vp9_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
451
          if ((mask_4x4_int & 3) == 3) {
Jim Bankoski's avatar
Jim Bankoski committed
452 453 454
            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
455 456
          } else {
            if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
457 458
              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr, 1);
459
            else if (mask_4x4_int & 2)
Jim Bankoski's avatar
Jim Bankoski committed
460 461
              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr, 1);
462 463 464
          }
          count = 2;
        } else {
Jim Bankoski's avatar
Jim Bankoski committed
465
          vp9_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
466

Jim Bankoski's avatar
Jim Bankoski committed
467 468 469
          if (mask_4x4_int & 1)
            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr, 1);
470 471
        }
      } else if (mask_4x4_int & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
472 473
        vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                             lfi->hev_thr, 1);
474
      }
475
    }
476
    s += 8 * count;
Yunqing Wang's avatar
Yunqing Wang committed
477
    lfl += count;
478 479 480 481
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
482 483 484
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
485 486 487 488 489 490 491 492 493 494 495 496 497 498
// 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);
499
  const int skip = mi->mbmi.skip;
Jim Bankoski's avatar
Jim Bankoski committed
500 501
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
502
  const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
Jim Bankoski's avatar
Jim Bankoski committed
503 504 505 506 507 508
  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
509 510 511 512
  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
513
  // If filter level is 0 we don't loop filter.
514
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
515
    return;
516 517 518 519 520 521 522
  } 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
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579

  // 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;
580
  const int skip = mi->mbmi.skip;
Jim Bankoski's avatar
Jim Bankoski committed
581 582
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
583
  const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
Jim Bankoski's avatar
Jim Bankoski committed
584 585 586
  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
587 588 589 590
  int i;
  int w = num_8x8_blocks_wide_lookup[block_size];
  int h = num_8x8_blocks_high_lookup[block_size];

591
  if (!filter_level) {
Jim Bankoski's avatar
Jim Bankoski committed
592
    return;
593 594 595 596 597 598 599
  } 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
600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620

  *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.
621 622 623
void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
                    MODE_INFO **mi_8x8, const int mode_info_stride,
                    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_8x8;
  MODE_INFO **mip2 = mi_8x8;
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
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];
875
  return lfi_n->lvl[seg][ref][mode_lf_lut[mbmi->mode]];
876 877
}

878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
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
893
        vp9_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
894
      } else if (mask_8x8 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
895
        vp9_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
896
      } else if (mask_4x4 & 1) {
Jim Bankoski's avatar
Jim Bankoski committed
897
        vp9_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
898 899 900
      }
    }
    if (mask_4x4_int & 1)
Jim Bankoski's avatar
Jim Bankoski committed
901
      vp9_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
902 903 904 905 906 907 908 909 910
    s += 8;
    lfl += 1;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
    mask_4x4_int >>= 1;
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
911 912
static void filter_block_plane_non420(VP9_COMMON *cm,
                                      struct macroblockd_plane *plane,
913
                                      MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
914
                                      int mi_row, int mi_col) {
915 916
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
917 918
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
919 920
  const int row_step_stride = cm->mode_info_stride * row_step;
  struct buf_2d *const dst = &plane->dst;
921
  uint8_t* const dst0 = dst->buf;
922 923 924 925
  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};
926
  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
927
  int r, c;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
928

929
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
930 931 932 933 934 935
    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
936
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
937
      const MODE_INFO *mi = mi_8x8[c];
938
      const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
939
      const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
940
      // left edge of current unit is block/partition edge -> no skip
941 942
      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
943
      const int skip_this_c = skip_this && !block_edge_left;
944
      // top edge of current unit is block/partition edge -> no skip
945 946
      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
947
      const int skip_this_r = skip_this && !block_edge_above;
948
      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
949 950
                            ? get_uv_tx_size(&mi[0].mbmi)
                            : mi[0].mbmi.tx_size;
951
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
952
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
953 954

      // Filter level can vary per MI
955 956
      if (!(lfl[(r << 3) + (c >> ss_x)] =
          build_lfi(&cm->lf_info, &mi[0].mbmi)))
957 958 959 960
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
961 962 963 964 965 966 967 968 969 970 971 972
        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);
        }
973
      } else if (tx_size == TX_16X16) {
974 975 976 977 978 979 980 981 982 983 984 985
        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);
        }
986 987 988
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
989
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
990 991 992 993 994 995
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
996
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
997 998 999 1000 1001
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

1002 1003
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
1004 1005 1006 1007 1008 1009 1010 1011 1012
      }
    }

    // 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
1013
                            mask_4x4_int[r],
1014
                            &cm->lf_info, &lfl[r << 3]);
1015
    dst->buf += 8 * dst->stride;
1016
    mi_8x8 += row_step_stride;
1017 1018 1019 1020
  }

  // Now do horizontal pass
  dst->buf = dst0;
1021
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
1022 1023 1024
    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];

1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
    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];
    }

1039
    filter_selectively_horiz(dst->buf, dst->stride,
1040 1041 1042 1043
                             mask_16x16_r,
                             mask_8x8_r,
                             mask_4x4_r,
                             mask_4x4_int_r,
1044
                             &cm->lf_info, &lfl[r << 3]);
1045 1046 1047 1048
    dst->buf += 8 * dst->stride;
  }
}

1049 1050 1051 1052
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
1053 1054 1055 1056
  struct buf_2d *const dst = &plane->dst;
  uint8_t* const dst0 = dst->buf;
  int r, c;

Yunqing Wang's avatar
Yunqing Wang committed
1057 1058 1059 1060 1061
  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