vp9_loopfilter.c 38.7 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
 */

Attila Nagy's avatar
Attila Nagy committed
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

19 20 21 22 23 24
struct loop_filter_info {
  const uint8_t *mblim;
  const uint8_t *lim;
  const uint8_t *hev_thr;
};

Jim Bankoski's avatar
Jim Bankoski committed
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
// This structure holds bit masks for all 8x8 blocks in a 64x64 region.
// Each 1 bit represents a position in which we want to apply the loop filter.
// Left_ entries refer to whether we apply a filter on the border to the
// left of the block.   Above_ entries refer to whether or not to apply a
// filter on the above border.   Int_ entries refer to whether or not to
// apply borders on the 4x4 edges within the 8x8 block that each bit
// represents.
// Since each transform is accompanied by a potentially different type of
// loop filter there is a different entry in the array for each transform size.
typedef struct {
  uint64_t left_y[TX_SIZES];
  uint64_t above_y[TX_SIZES];
  uint64_t int_4x4_y;
  uint16_t left_uv[TX_SIZES];
  uint16_t above_uv[TX_SIZES];
  uint16_t int_4x4_uv;
} LOOP_FILTER_MASK;

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

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

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

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

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

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

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

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

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


John Koleszar's avatar
John Koleszar committed
229
static void lf_init_lut(loop_filter_info_n *lfi) {
230 231 232 233 234
  lfi->mode_lf_lut[DC_PRED] = 0;
  lfi->mode_lf_lut[D45_PRED] = 0;
  lfi->mode_lf_lut[D135_PRED] = 0;
  lfi->mode_lf_lut[D117_PRED] = 0;
  lfi->mode_lf_lut[D153_PRED] = 0;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
235
  lfi->mode_lf_lut[D207_PRED] = 0;
236 237 238 239 240 241 242 243
  lfi->mode_lf_lut[D63_PRED] = 0;
  lfi->mode_lf_lut[V_PRED] = 0;
  lfi->mode_lf_lut[H_PRED] = 0;
  lfi->mode_lf_lut[TM_PRED] = 0;
  lfi->mode_lf_lut[ZEROMV]  = 0;
  lfi->mode_lf_lut[NEARESTMV] = 1;
  lfi->mode_lf_lut[NEARMV] = 1;
  lfi->mode_lf_lut[NEWMV] = 1;
Attila Nagy's avatar
Attila Nagy committed
244 245
}

Frank Galligan's avatar
Frank Galligan committed
246
static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
247
  int lvl;
John Koleszar's avatar
John Koleszar committed
248

Dmitry Kovalev's avatar
Dmitry Kovalev committed
249 250 251 252
  // 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
253

John Koleszar's avatar
John Koleszar committed
254 255 256
    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
John Koleszar's avatar
John Koleszar committed
257
    }
John Koleszar's avatar
John Koleszar committed
258 259 260 261

    if (block_inside_limit < 1)
      block_inside_limit = 1;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
262 263
    vpx_memset(lfi->lim[lvl], block_inside_limit, SIMD_WIDTH);
    vpx_memset(lfi->mblim[lvl], (2 * (lvl + 2) + block_inside_limit),
John Koleszar's avatar
John Koleszar committed
264 265
               SIMD_WIDTH);
  }
Attila Nagy's avatar
Attila Nagy committed
266
}
John Koleszar's avatar
John Koleszar committed
267

268
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
269
  loop_filter_info_n *lfi = &cm->lf_info;
270
  struct loopfilter *lf = &cm->lf;
John Koleszar's avatar
John Koleszar committed
271
  int i;
Attila Nagy's avatar
Attila Nagy committed
272

273
  // init limits for given sharpness
274 275
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
276

277
  // init LUT for lvl  and hev thr picking
John Koleszar's avatar
John Koleszar committed
278
  lf_init_lut(lfi);
Attila Nagy's avatar
Attila Nagy committed
279

280 281
  // init hev threshold const vectors
  for (i = 0; i < 4; i++)
John Koleszar's avatar
John Koleszar committed
282
    vpx_memset(lfi->hev_thr[i], i, SIMD_WIDTH);
John Koleszar's avatar
John Koleszar committed
283 284
}

Frank Galligan's avatar
Frank Galligan committed
285
void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
286
  int seg_id;
287 288 289
  // 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
Dmitry Kovalev's avatar
Dmitry Kovalev committed
290 291
  const int n_shift = default_filt_lvl >> 5;
  loop_filter_info_n *const lfi = &cm->lf_info;
292
  struct loopfilter *const lf = &cm->lf;
293
  struct segmentation *const seg = &cm->seg;
John Koleszar's avatar
John Koleszar committed
294

Dmitry Kovalev's avatar
Dmitry Kovalev committed
295
  // update limits if sharpness has changed
296 297 298
  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
299 300
  }

301
  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
302
    int lvl_seg = default_filt_lvl, ref, mode, intra_lvl;
John Koleszar's avatar
John Koleszar committed
303 304

    // Set the baseline filter values for each segment
305
    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
306 307
      const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
      lvl_seg = seg->abs_delta == SEGMENT_ABSDATA
Dmitry Kovalev's avatar
Dmitry Kovalev committed
308 309
                  ? data
                  : clamp(default_filt_lvl + data, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
310
    }
John Koleszar's avatar
John Koleszar committed
311

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

Yaowu Xu's avatar
Yaowu Xu committed
319
    intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * (1 << n_shift);
320
    lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
321

Dmitry Kovalev's avatar
Dmitry Kovalev committed
322 323
    for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref)
      for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
Yaowu Xu's avatar
Yaowu Xu committed
324 325
        const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * (1 << n_shift)
                                      + lf->mode_deltas[mode] * (1 << n_shift);
326
        lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
327 328
      }
  }
John Koleszar's avatar
John Koleszar committed
329 330
}

Frank Galligan's avatar
Frank Galligan committed
331 332 333
static int build_lfi(const loop_filter_info_n *lfi_n,
                     const MB_MODE_INFO *mbmi,
                     struct loop_filter_info *lfi) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
334 335 336 337 338 339
  const int seg = mbmi->segment_id;
  const int ref = mbmi->ref_frame[0];
  const int mode = lfi_n->mode_lf_lut[mbmi->mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];

  if (filter_level > 0) {
340 341
    lfi->mblim = lfi_n->mblim[filter_level];
    lfi->lim = lfi_n->lim[filter_level];
Dmitry Kovalev's avatar
Dmitry Kovalev committed
342
    lfi->hev_thr = lfi_n->hev_thr[filter_level >> 4];
343
    return 1;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
344 345
  } else {
    return 0;
346 347 348 349 350 351 352
  }
}

static void filter_selectively_vert(uint8_t *s, int pitch,
                                    unsigned int mask_16x16,
                                    unsigned int mask_8x8,
                                    unsigned int mask_4x4,
353
                                    unsigned int mask_4x4_int,
354 355 356
                                    const struct loop_filter_info *lfi) {
  unsigned int mask;

John Koleszar's avatar
John Koleszar committed
357 358
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
       mask; mask >>= 1) {
359 360 361
    if (mask & 1) {
      if (mask_16x16 & 1) {
        vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim,
362
                                   lfi->hev_thr);
363 364
        assert(!(mask_8x8 & 1));
        assert(!(mask_4x4 & 1));
365
        assert(!(mask_4x4_int & 1));
366 367 368 369 370 371 372 373 374 375 376 377
      } else if (mask_8x8 & 1) {
        vp9_mbloop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim,
                                        lfi->hev_thr, 1);
        assert(!(mask_16x16 & 1));
        assert(!(mask_4x4 & 1));
      } else if (mask_4x4 & 1) {
        vp9_loop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim,
                                      lfi->hev_thr, 1);
        assert(!(mask_16x16 & 1));
        assert(!(mask_8x8 & 1));
      }
    }
378 379 380
    if (mask_4x4_int & 1)
      vp9_loop_filter_vertical_edge(s + 4, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, 1);
381 382 383 384 385
    s += 8;
    lfi++;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
386
    mask_4x4_int >>= 1;
387 388 389 390 391 392 393
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
394
                                     unsigned int mask_4x4_int,
395 396 397
                                     int only_4x4_1,
                                     const struct loop_filter_info *lfi) {
  unsigned int mask;
398
  int count;
399

John Koleszar's avatar
John Koleszar committed
400
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
401
       mask; mask >>= count) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
402
    count = 1;
403 404 405
    if (mask & 1) {
      if (!only_4x4_1) {
        if (mask_16x16 & 1) {
406 407 408 409 410 411 412 413
          if ((mask_16x16 & 3) == 3) {
            vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim,
                                         lfi->hev_thr, 2);
            count = 2;
          } else {
            vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim,
                                         lfi->hev_thr, 1);
          }
414 415
          assert(!(mask_8x8 & 1));
          assert(!(mask_4x4 & 1));
416
          assert(!(mask_4x4_int & 1));
417 418 419 420 421 422 423 424 425 426 427 428 429
        } else if (mask_8x8 & 1) {
          vp9_mbloop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim,
                                            lfi->hev_thr, 1);
          assert(!(mask_16x16 & 1));
          assert(!(mask_4x4 & 1));
        } else if (mask_4x4 & 1) {
          vp9_loop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim,
                                          lfi->hev_thr, 1);
          assert(!(mask_16x16 & 1));
          assert(!(mask_8x8 & 1));
        }
      }

430
      if (mask_4x4_int & 1)
431 432 433
        vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                        lfi->lim, lfi->hev_thr, 1);
    }
434 435 436 437 438 439
    s += 8 * count;
    lfi += count;
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
440 441 442
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561
// This function ors into the current lfm structure, where to do loop
// filters for the specific mi we are looking at.   It uses information
// including the block_size_type (32x16, 32x32, etc),  the transform size,
// whether there were any coefficients encoded, and the loop filter strength
// block we are currently looking at. Shift is used to position the
// 1's we produce.
// TODO(JBB) Need another function for different resolution color..
static void build_masks(const loop_filter_info_n *const lfi_n,
                        const MODE_INFO *mi, const int shift_y,
                        const int shift_uv,
                        LOOP_FILTER_MASK *lfm) {
  const BLOCK_SIZE block_size = mi->mbmi.sb_type;
  const TX_SIZE tx_size_y = mi->mbmi.tx_size;
  const TX_SIZE tx_size_uv = get_uv_tx_size(&mi->mbmi);
  const int skip = mi->mbmi.skip_coeff;
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
  const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];
  uint64_t *left_y = &lfm->left_y[tx_size_y];
  uint64_t *above_y = &lfm->above_y[tx_size_y];
  uint64_t *int_4x4_y = &lfm->int_4x4_y;
  uint16_t *left_uv = &lfm->left_uv[tx_size_uv];
  uint16_t *above_uv = &lfm->above_uv[tx_size_uv];
  uint16_t *int_4x4_uv = &lfm->int_4x4_uv;

  // If filter level is 0 we don't loop filter.
  if (!filter_level)
    return;

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

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

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

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

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

// This function does the same thing as the one above with the exception that
// it only affects the y masks.   It exists because for blocks < 16x16 in size,
// we only update u and v masks on the first block.
static void build_y_mask(const loop_filter_info_n *const lfi_n,
                         const MODE_INFO *mi, const int shift_y,
                         LOOP_FILTER_MASK *lfm) {
  const BLOCK_SIZE block_size = mi->mbmi.sb_type;
  const TX_SIZE tx_size_y = mi->mbmi.tx_size;
  const int skip = mi->mbmi.skip_coeff;
  const int seg = mi->mbmi.segment_id;
  const int ref = mi->mbmi.ref_frame[0];
  const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode];
  const int filter_level = lfi_n->lvl[seg][ref][mode];
  uint64_t *left_y = &lfm->left_y[tx_size_y];
  uint64_t *above_y = &lfm->above_y[tx_size_y];
  uint64_t *int_4x4_y = &lfm->int_4x4_y;

  if (!filter_level)
    return;

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

  if (skip && ref > INTRA_FRAME)
    return;

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

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

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

// This function sets up the bit masks for the entire 64x64 region represented
// by mi_row, mi_col.
// TODO(JBB): This function only works for yv12.
static void setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
562
                       MODE_INFO **mi_8x8, const int mode_info_stride,
Jim Bankoski's avatar
Jim Bankoski committed
563 564 565
                       LOOP_FILTER_MASK *lfm) {
  int idx_32, idx_16, idx_8;
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
566 567
  MODE_INFO **mip = mi_8x8;
  MODE_INFO **mip2 = mi_8x8;
Jim Bankoski's avatar
Jim Bankoski committed
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598

  // 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.
599
  switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
600
    case BLOCK_64X64:
601
      build_masks(lfi_n, mip[0] , 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
602 603
      break;
    case BLOCK_64X32:
604
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
605
      mip2 = mip + mode_info_stride * 4;
606 607
      if (4 >= max_rows)
        break;
608
      build_masks(lfi_n, mip2[0], 32, 8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
609 610
      break;
    case BLOCK_32X64:
611
      build_masks(lfi_n, mip[0], 0, 0, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
612
      mip2 = mip + 4;
613 614
      if (4 >= max_cols)
        break;
615
      build_masks(lfi_n, mip2[0], 4, 2, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
616 617 618 619 620 621 622 623 624
      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;
625
        switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
626
          case BLOCK_32X32:
627
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
628 629
            break;
          case BLOCK_32X16:
630
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
631 632
            if (mi_32_row_offset + 2 >= max_rows)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
633
            mip2 = mip + mode_info_stride * 2;
634
            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
635 636
            break;
          case BLOCK_16X32:
637
            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
638 639
            if (mi_32_col_offset + 2 >= max_cols)
              continue;
Jim Bankoski's avatar
Jim Bankoski committed
640
            mip2 = mip + 2;
641
            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
642 643 644 645 646 647 648 649 650 651 652 653 654
            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;

655
              switch (mip[0]->mbmi.sb_type) {
Jim Bankoski's avatar
Jim Bankoski committed
656
                case BLOCK_16X16:
657
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
658 659
                  break;
                case BLOCK_16X8:
660
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
661 662
                  if (mi_16_row_offset + 1 >= max_rows)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
663
                  mip2 = mip + mode_info_stride;
664
                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
665 666
                  break;
                case BLOCK_8X16:
667
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
668 669
                  if (mi_16_col_offset +1 >= max_cols)
                    continue;
Jim Bankoski's avatar
Jim Bankoski committed
670
                  mip2 = mip + 1;
671
                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
672 673 674 675 676
                  break;
                default: {
                  const int shift_y = shift_32_y[idx_32] +
                                      shift_16_y[idx_16] +
                                      shift_8_y[0];
677
                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
678 679 680 681 682 683 684 685 686 687 688 689 690
                  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;
691
                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
Jim Bankoski's avatar
Jim Bankoski committed
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791
                  }
                  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;
    }
  }
}
792
#if CONFIG_NON420
Jim Bankoski's avatar
Jim Bankoski committed
793 794
static void filter_block_plane_non420(VP9_COMMON *cm,
                                      struct macroblockd_plane *plane,
795
                                      MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
796
                                      int mi_row, int mi_col) {
797 798
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
799 800
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
801 802
  const int row_step_stride = cm->mode_info_stride * row_step;
  struct buf_2d *const dst = &plane->dst;
803
  uint8_t* const dst0 = dst->buf;
804 805 806 807 808
  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};
  struct loop_filter_info lfi[MI_BLOCK_SIZE][MI_BLOCK_SIZE];
809
  int r, c;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
810

811
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
812 813 814 815 816 817
    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
818
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
819 820 821
      const MODE_INFO *mi = mi_8x8[c];
      const int skip_this = mi[0].mbmi.skip_coeff
                            && is_inter_block(&mi[0].mbmi);
822
      // left edge of current unit is block/partition edge -> no skip
823 824
      const int block_edge_left = b_width_log2(mi[0].mbmi.sb_type) ?
          !(c & ((1 << (b_width_log2(mi[0].mbmi.sb_type)-1)) - 1)) : 1;
825
      const int skip_this_c = skip_this && !block_edge_left;
826
      // top edge of current unit is block/partition edge -> no skip
827 828
      const int block_edge_above = b_height_log2(mi[0].mbmi.sb_type) ?
          !(r & ((1 << (b_height_log2(mi[0].mbmi.sb_type)-1)) - 1)) : 1;
829
      const int skip_this_r = skip_this && !block_edge_above;
830
      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
831 832
                            ? get_uv_tx_size(&mi[0].mbmi)
                            : mi[0].mbmi.tx_size;
833
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
834
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
835 836

      // Filter level can vary per MI
837
      if (!build_lfi(&cm->lf_info, &mi[0].mbmi, lfi[r] + (c >> ss_x)))
838 839 840 841
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
842 843 844 845 846 847 848 849 850 851 852 853
        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);
        }
854
      } else if (tx_size == TX_16X16) {
855 856 857 858 859 860 861 862 863 864 865 866
        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);
        }
867 868 869
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
870
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
871 872 873 874 875 876
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
877
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
878 879 880 881 882
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

883 884
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
885 886 887 888 889 890 891 892 893
      }
    }

    // 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,
894
                            mask_4x4_int[r], lfi[r]);
895
    dst->buf += 8 * dst->stride;
896
    mi_8x8 += row_step_stride;
897 898 899 900
  }

  // Now do horizontal pass
  dst->buf = dst0;
901
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
902 903 904
    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];

905 906 907 908
    filter_selectively_horiz(dst->buf, dst->stride,
                             mask_16x16[r],
                             mask_8x8[r],
                             mask_4x4[r],
909
                             mask_4x4_int_r, mi_row + r == 0, lfi[r]);
910 911 912
    dst->buf += 8 * dst->stride;
  }
}
913
#endif
914

Jim Bankoski's avatar
Jim Bankoski committed
915 916
static void filter_block_plane(VP9_COMMON *const cm,
                               struct macroblockd_plane *const plane,
917
                               MODE_INFO **mi_8x8,
Jim Bankoski's avatar
Jim Bankoski committed
918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939
                               int mi_row, int mi_col,
                               LOOP_FILTER_MASK *lfm) {
  const int ss_x = plane->subsampling_x;
  const int ss_y = plane->subsampling_y;
  const int row_step = 1 << ss_x;
  const int col_step = 1 << ss_y;
  const int row_step_stride = cm->mode_info_stride * row_step;
  struct buf_2d *const dst = &plane->dst;
  uint8_t* const dst0 = dst->buf;
  unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
  struct loop_filter_info lfi[MI_BLOCK_SIZE][MI_BLOCK_SIZE];
  int r, c;
  int row_shift = 3 - ss_x;
  int row_mask = 0xff >> (ss_x << 2);

#define MASK_ROW(value) ((value >> (r_sampled << row_shift)) & row_mask)

  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
    int r_sampled = r >> ss_x;

    // Determine the vertical edges that need filtering
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
940 941
      const MODE_INFO *mi = mi_8x8[c];
      if (!build_lfi(&cm->lf_info, &mi[0].mbmi, lfi[r] + (c >> ss_x)))
Jim Bankoski's avatar
Jim Bankoski committed
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963
        continue;
    }
    if (!plane->plane_type) {
      mask_4x4_int[r] = MASK_ROW(lfm->int_4x4_y);
      // Disable filtering on the leftmost column
      filter_selectively_vert(dst->buf, dst->stride,
                              MASK_ROW(lfm->left_y[TX_16X16]),
                              MASK_ROW(lfm->left_y[TX_8X8]),
                              MASK_ROW(lfm->left_y[TX_4X4]),
                              MASK_ROW(lfm->int_4x4_y),
                              lfi[r]);
    } else {
      mask_4x4_int[r] = MASK_ROW(lfm->int_4x4_uv);
      // Disable filtering on the leftmost column
      filter_selectively_vert(dst->buf, dst->stride,
                              MASK_ROW(lfm->left_uv[TX_16X16]),
                              MASK_ROW(lfm->left_uv[TX_8X8]),
                              MASK_ROW(lfm->left_uv[TX_4X4]),
                              MASK_ROW(lfm->int_4x4_uv),
                              lfi[r]);
    }
    dst->buf += 8 * dst->stride;
964
    mi_8x8 += row_step_stride;
Jim Bankoski's avatar
Jim Bankoski committed
965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993
  }

  // Now do horizontal pass
  dst->buf = dst0;
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
    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];
    int r_sampled = r >> ss_x;

    if (!plane->plane_type) {
      filter_selectively_horiz(dst->buf, dst->stride,
                               MASK_ROW(lfm->above_y[TX_16X16]),
                               MASK_ROW(lfm->above_y[TX_8X8]),
                               MASK_ROW(lfm->above_y[TX_4X4]),
                               MASK_ROW(lfm->int_4x4_y),
                               mi_row + r == 0, lfi[r]);
    } else {
      filter_selectively_horiz(dst->buf, dst->stride,
                               MASK_ROW(lfm->above_uv[TX_16X16]),
                               MASK_ROW(lfm->above_uv[TX_8X8]),
                               MASK_ROW(lfm->above_uv[TX_4X4]),
                               mask_4x4_int_r,
                               mi_row + r == 0, lfi[r]);
    }
    dst->buf += 8 * dst->stride;
  }
#undef MASK_ROW
}

994 995 996
void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
                          VP9_COMMON *cm, MACROBLOCKD *xd,
                          int start, int stop, int y_only) {
997
  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
998
  int mi_row, mi_col;
Jim Bankoski's avatar
Jim Bankoski committed
999
  LOOP_FILTER_MASK lfm;
1000
#if CONFIG_NON420
Jim Bankoski's avatar
Jim Bankoski committed
1001 1002
  int use_420 = y_only || (xd->plane[1].subsampling_y == 1 &&
      xd->plane[1].subsampling_x == 1);
1003
#endif
1004

1005
  for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
1006
    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mode_info_stride;
1007

1008
    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
1009 1010
      int plane;

1011
      setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
Jim Bankoski's avatar
Jim Bankoski committed
1012 1013

      // TODO(JBB): Make setup_mask work for non 420.
1014
#if CONFIG_NON420
Jim Bankoski's avatar
Jim Bankoski committed
1015
      if (use_420)
1016
#endif
1017 1018
        setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mode_info_stride,
                   &lfm);
Jim Bankoski's avatar
Jim Bankoski committed
1019

1020
      for (plane = 0; plane < num_planes; ++plane) {
1021
#if CONFIG_NON420
Jim Bankoski's avatar
Jim Bankoski committed
1022
        if (use_420)
1023
#endif
1024 1025
          filter_block_plane(cm, &xd->plane[plane], mi_8x8 + mi_col, mi_row,
                             mi_col, &lfm);
1026
#if CONFIG_NON420
Jim Bankoski's avatar
Jim Bankoski committed
1027
        else
1028
          filter_block_plane_non420(cm, &xd->plane[plane], mi_8x8 + mi_col,
Jim Bankoski's avatar
Jim Bankoski committed
1029
                                    mi_row, mi_col);
1030
#endif
1031 1032 1033 1034
      }
    }
  }
}
1035 1036

void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
1037 1038 1039
                           int frame_filter_level,
                           int y_only, int partial) {
  int start_mi_row, end_mi_row, mi_rows_to_filter;
1040
  if (!frame_filter_level) return;
1041 1042 1043 1044 1045 1046 1047 1048
  start_mi_row = 0;
  mi_rows_to_filter = cm->mi_rows;
  if (partial && cm->mi_rows > 8) {
    start_mi_row = cm->mi_rows >> 1;
    start_mi_row &= 0xfffffff8;
    mi_rows_to_filter = MAX(cm->mi_rows / 8, 8);
  }
  end_mi_row = start_mi_row + mi_rows_to_filter;
1049
  vp9_loop_filter_frame_init(cm, frame_filter_level);
1050
  vp9_loop_filter_rows(cm->frame_to_show, cm, xd,
1051 1052
                       start_mi_row, end_mi_row,
                       y_only);
1053
}
1054 1055 1056 1057 1058 1059 1060 1061

int vp9_loop_filter_worker(void *arg1, void *arg2) {
  LFWorkerData *const lf_data = (LFWorkerData*)arg1;
  (void)arg2;
  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
                       lf_data->start, lf_data->stop, lf_data->y_only);
  return 1;
}