vp9_loopfilter.c 13.4 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

John Koleszar's avatar
John Koleszar committed
19
static void lf_init_lut(loop_filter_info_n *lfi) {
20 21 22 23 24 25 26 27 28 29 30 31 32 33
  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;
  lfi->mode_lf_lut[D27_PRED] = 0;
  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
34 35
}

36
static void update_sharpness(loop_filter_info_n *const lfi, int sharpness_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
37
  int lvl;
John Koleszar's avatar
John Koleszar committed
38

Dmitry Kovalev's avatar
Dmitry Kovalev committed
39 40 41 42
  // 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
43

John Koleszar's avatar
John Koleszar committed
44 45 46
    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
John Koleszar's avatar
John Koleszar committed
47
    }
John Koleszar's avatar
John Koleszar committed
48 49 50 51

    if (block_inside_limit < 1)
      block_inside_limit = 1;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
52 53 54
    vpx_memset(lfi->lim[lvl], block_inside_limit, SIMD_WIDTH);
    vpx_memset(lfi->blim[lvl], (2 * lvl + block_inside_limit), SIMD_WIDTH);
    vpx_memset(lfi->mblim[lvl], (2 * (lvl + 2) + block_inside_limit),
John Koleszar's avatar
John Koleszar committed
55 56
               SIMD_WIDTH);
  }
Attila Nagy's avatar
Attila Nagy committed
57
}
John Koleszar's avatar
John Koleszar committed
58

59
void vp9_loop_filter_init(VP9_COMMON *cm) {
John Koleszar's avatar
John Koleszar committed
60 61
  loop_filter_info_n *lfi = &cm->lf_info;
  int i;
Attila Nagy's avatar
Attila Nagy committed
62

63
  // init limits for given sharpness
64
  update_sharpness(lfi, cm->sharpness_level);
John Koleszar's avatar
John Koleszar committed
65
  cm->last_sharpness_level = cm->sharpness_level;
Attila Nagy's avatar
Attila Nagy committed
66

67
  // init LUT for lvl  and hev thr picking
John Koleszar's avatar
John Koleszar committed
68
  lf_init_lut(lfi);
Attila Nagy's avatar
Attila Nagy committed
69

70 71
  // init hev threshold const vectors
  for (i = 0; i < 4; i++)
John Koleszar's avatar
John Koleszar committed
72
    vpx_memset(lfi->hev_thr[i], i, SIMD_WIDTH);
John Koleszar's avatar
John Koleszar committed
73 74
}

75 76
static void loop_filter_frame_init(VP9_COMMON *const cm, MACROBLOCKD *const xd,
                                   int default_filt_lvl) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
77
  int seg;
78 79 80
  // 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
81 82
  const int n_shift = default_filt_lvl >> 5;
  loop_filter_info_n *const lfi = &cm->lf_info;
John Koleszar's avatar
John Koleszar committed
83

Dmitry Kovalev's avatar
Dmitry Kovalev committed
84
  // update limits if sharpness has changed
John Koleszar's avatar
John Koleszar committed
85
  if (cm->last_sharpness_level != cm->sharpness_level) {
86
    update_sharpness(lfi, cm->sharpness_level);
John Koleszar's avatar
John Koleszar committed
87 88 89 90
    cm->last_sharpness_level = cm->sharpness_level;
  }

  for (seg = 0; seg < MAX_MB_SEGMENTS; seg++) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
91
    int lvl_seg = default_filt_lvl, ref, mode, intra_lvl;
John Koleszar's avatar
John Koleszar committed
92 93

    // Set the baseline filter values for each segment
94 95 96
    if (vp9_segfeature_active(&xd->seg, seg, SEG_LVL_ALT_LF)) {
      const int data = vp9_get_segdata(&xd->seg, seg, SEG_LVL_ALT_LF);
      lvl_seg = xd->seg.abs_delta == SEGMENT_ABSDATA
Dmitry Kovalev's avatar
Dmitry Kovalev committed
97 98
                  ? data
                  : clamp(default_filt_lvl + data, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
99
    }
John Koleszar's avatar
John Koleszar committed
100

John Koleszar's avatar
John Koleszar committed
101
    if (!xd->mode_ref_lf_delta_enabled) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
102 103
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
John Koleszar's avatar
John Koleszar committed
104 105 106
      vpx_memset(lfi->lvl[seg][0], lvl_seg, 4 * 4);
      continue;
    }
John Koleszar's avatar
John Koleszar committed
107

Dmitry Kovalev's avatar
Dmitry Kovalev committed
108 109
    intra_lvl = lvl_seg + (xd->ref_lf_deltas[INTRA_FRAME] << n_shift);
    lfi->lvl[seg][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
110

Dmitry Kovalev's avatar
Dmitry Kovalev committed
111 112 113 114 115
    for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref)
      for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
        const int inter_lvl = lvl_seg + (xd->ref_lf_deltas[ref] << n_shift)
                                      + (xd->mode_lf_deltas[mode] << n_shift);
        lfi->lvl[seg][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
John Koleszar's avatar
John Koleszar committed
116 117
      }
  }
John Koleszar's avatar
John Koleszar committed
118 119
}

120 121
static int build_lfi(const VP9_COMMON *cm, const MB_MODE_INFO *mbmi,
                      struct loop_filter_info *lfi) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
122 123 124 125 126 127 128
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
  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) {
129 130 131
    lfi->mblim = lfi_n->mblim[filter_level];
    lfi->blim = lfi_n->blim[filter_level];
    lfi->lim = lfi_n->lim[filter_level];
Dmitry Kovalev's avatar
Dmitry Kovalev committed
132
    lfi->hev_thr = lfi_n->hev_thr[filter_level >> 4];
133
    return 1;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
134 135
  } else {
    return 0;
136 137 138 139 140 141 142
  }
}

static void filter_selectively_vert(uint8_t *s, int pitch,
                                    unsigned int mask_16x16,
                                    unsigned int mask_8x8,
                                    unsigned int mask_4x4,
143
                                    unsigned int mask_4x4_int,
144 145 146
                                    const struct loop_filter_info *lfi) {
  unsigned int mask;

John Koleszar's avatar
John Koleszar committed
147 148
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
       mask; mask >>= 1) {
149 150 151
    if (mask & 1) {
      if (mask_16x16 & 1) {
        vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim,
152
                                   lfi->hev_thr);
153 154
        assert(!(mask_8x8 & 1));
        assert(!(mask_4x4 & 1));
155
        assert(!(mask_4x4_int & 1));
156 157 158 159 160 161 162 163 164 165 166 167
      } 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));
      }
    }
168 169 170
    if (mask_4x4_int & 1)
      vp9_loop_filter_vertical_edge(s + 4, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, 1);
171 172 173 174 175
    s += 8;
    lfi++;
    mask_16x16 >>= 1;
    mask_8x8 >>= 1;
    mask_4x4 >>= 1;
176
    mask_4x4_int >>= 1;
177 178 179 180 181 182 183
  }
}

static void filter_selectively_horiz(uint8_t *s, int pitch,
                                     unsigned int mask_16x16,
                                     unsigned int mask_8x8,
                                     unsigned int mask_4x4,
184
                                     unsigned int mask_4x4_int,
185 186 187
                                     int only_4x4_1,
                                     const struct loop_filter_info *lfi) {
  unsigned int mask;
188
  int count;
189

John Koleszar's avatar
John Koleszar committed
190
  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
191 192
       mask; mask >>= count) {
    count =1;
193 194 195
    if (mask & 1) {
      if (!only_4x4_1) {
        if (mask_16x16 & 1) {
196 197 198 199 200 201 202 203
          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);
          }
204 205
          assert(!(mask_8x8 & 1));
          assert(!(mask_4x4 & 1));
206
          assert(!(mask_4x4_int & 1));
207 208 209 210 211 212 213 214 215 216 217 218 219
        } 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));
        }
      }

220
      if (mask_4x4_int & 1)
221 222 223
        vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim,
                                        lfi->lim, lfi->hev_thr, 1);
    }
224 225 226 227 228 229
    s += 8 * count;
    lfi += count;
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
230 231 232 233 234 235
  }
}

static void filter_block_plane(VP9_COMMON *cm, MACROBLOCKD *xd,
                               int plane, int mi_row, int mi_col) {
  const int ss_x = xd->plane[plane].subsampling_x;
236
  const int ss_y = xd->plane[plane].subsampling_y;
237 238 239 240
  const int row_step = 1 << xd->plane[plane].subsampling_y;
  const int col_step = 1 << xd->plane[plane].subsampling_x;
  struct buf_2d * const dst = &xd->plane[plane].dst;
  uint8_t* const dst0 = dst->buf;
241 242 243 244 245
  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];
246
  int r, c;
247 248
  MODE_INFO *mi = xd->mode_info_context;
  int row_step_stride = cm->mode_info_stride * row_step;
249

250
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
251 252 253 254 255 256
    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
257
    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
258
      const int skip_this = mi[c].mbmi.mb_skip_coeff
Yaowu Xu's avatar
Yaowu Xu committed
259
                            && mi[c].mbmi.ref_frame[0] != INTRA_FRAME;
260
      // left edge of current unit is block/partition edge -> no skip
261 262
      const int block_edge_left = b_width_log2(mi[c].mbmi.sb_type) ?
          !(c & ((1 << (b_width_log2(mi[c].mbmi.sb_type)-1)) - 1)) : 1;
263
      const int skip_this_c = skip_this && !block_edge_left;
264
      // top edge of current unit is block/partition edge -> no skip
265 266
      const int block_edge_above = b_height_log2(mi[c].mbmi.sb_type) ?
          !(r & ((1 << (b_height_log2(mi[c].mbmi.sb_type)-1)) - 1)) : 1;
267
      const int skip_this_r = skip_this && !block_edge_above;
268 269
      const TX_SIZE tx_size = plane ? get_uv_tx_size(&mi[c].mbmi)
                                    : mi[c].mbmi.txfm_size;
270
      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
271
      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
272 273 274 275 276 277 278 279

      // Filter level can vary per MI
      if (!build_lfi(cm, &mi[c].mbmi,
                     lfi[r] + (c >> xd->plane[plane].subsampling_x)))
        continue;

      // Build masks based on the transform size of each block
      if (tx_size == TX_32X32) {
280 281 282 283 284 285 286 287 288 289 290 291
        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);
        }
292
      } else if (tx_size == TX_16X16) {
293 294 295 296 297 298 299 300 301 302 303 304
        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);
        }
305 306 307
      } else {
        // force 8x8 filtering on 32x32 boundaries
        if (!skip_this_c) {
308
          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
309 310 311 312 313 314
            mask_8x8_c |= 1 << (c >> ss_x);
          else
            mask_4x4_c |= 1 << (c >> ss_x);
        }

        if (!skip_this_r) {
315
          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
316 317 318 319 320
            mask_8x8[r] |= 1 << (c >> ss_x);
          else
            mask_4x4[r] |= 1 << (c >> ss_x);
        }

321 322
        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
          mask_4x4_int[r] |= 1 << (c >> ss_x);
323 324 325 326 327 328 329 330 331
      }
    }

    // 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,
332
                            mask_4x4_int[r], lfi[r]);
333
    dst->buf += 8 * dst->stride;
334
    mi += row_step_stride;
335 336 337 338
  }

  // Now do horizontal pass
  dst->buf = dst0;
339
  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
340 341 342
    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];

343 344 345 346
    filter_selectively_horiz(dst->buf, dst->stride,
                             mask_16x16[r],
                             mask_8x8[r],
                             mask_4x4[r],
347
                             mask_4x4_int_r, mi_row + r == 0, lfi[r]);
348 349 350 351
    dst->buf += 8 * dst->stride;
  }
}

352 353
void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
                           int frame_filter_level, int y_only) {
354 355 356
  int mi_row, mi_col;

  // Initialize the loop filter for this frame.
357
  loop_filter_frame_init(cm, xd, frame_filter_level);
358

359
  for (mi_row = 0; mi_row < cm->mi_rows; mi_row += MI_BLOCK_SIZE) {
360 361
    MODE_INFO* const mi = cm->mi + mi_row * cm->mode_info_stride;

362
    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
363 364 365 366 367 368 369 370 371 372
      int plane;

      setup_dst_planes(xd, cm->frame_to_show, mi_row, mi_col);
      for (plane = 0; plane < (y_only ? 1 : MAX_MB_PLANE); plane++) {
        xd->mode_info_context = mi + mi_col;
        filter_block_plane(cm, xd, plane, mi_row, mi_col);
      }
    }
  }
}