1/* 2 * VP9 compatible video decoder 3 * 4 * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com> 5 * Copyright (C) 2013 Clément Bœsch <u pkh me> 6 * 7 * This file is part of FFmpeg. 8 * 9 * FFmpeg is free software; you can redistribute it and/or 10 * modify it under the terms of the GNU Lesser General Public 11 * License as published by the Free Software Foundation; either 12 * version 2.1 of the License, or (at your option) any later version. 13 * 14 * FFmpeg is distributed in the hope that it will be useful, 15 * but WITHOUT ANY WARRANTY; without even the implied warranty of 16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 17 * Lesser General Public License for more details. 18 * 19 * You should have received a copy of the GNU Lesser General Public 20 * License along with FFmpeg; if not, write to the Free Software 21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 22 */ 23 24#include "threadframe.h" 25#include "vp56.h" 26#include "vp9.h" 27#include "vp9data.h" 28#include "vp9dec.h" 29 30static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src, 31 VP9TileData *td) 32{ 33 dst->x = av_clip(src->x, td->min_mv.x, td->max_mv.x); 34 dst->y = av_clip(src->y, td->min_mv.y, td->max_mv.y); 35} 36 37static void find_ref_mvs(VP9TileData *td, 38 VP56mv *pmv, int ref, int z, int idx, int sb) 39{ 40 static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = { 41 [BS_64x64] = { { 3, -1 }, { -1, 3 }, { 4, -1 }, { -1, 4 }, 42 { -1, -1 }, { 0, -1 }, { -1, 0 }, { 6, -1 } }, 43 [BS_64x32] = { { 0, -1 }, { -1, 0 }, { 4, -1 }, { -1, 2 }, 44 { -1, -1 }, { 0, -3 }, { -3, 0 }, { 2, -1 } }, 45 [BS_32x64] = { { -1, 0 }, { 0, -1 }, { -1, 4 }, { 2, -1 }, 46 { -1, -1 }, { -3, 0 }, { 0, -3 }, { -1, 2 } }, 47 [BS_32x32] = { { 1, -1 }, { -1, 1 }, { 2, -1 }, { -1, 2 }, 48 { -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } }, 49 [BS_32x16] = { { 0, -1 }, { -1, 0 }, { 2, -1 }, { -1, -1 }, 50 { -1, 1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } }, 51 [BS_16x32] = { { -1, 0 }, { 0, -1 }, { -1, 2 }, { -1, -1 }, 52 { 1, -1 }, { -3, 0 }, { 0, -3 }, { -3, -3 } }, 53 [BS_16x16] = { { 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, 1 }, 54 { -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 } }, 55 [BS_16x8] = { { 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, -1 }, 56 { 0, -2 }, { -2, 0 }, { -2, -1 }, { -1, -2 } }, 57 [BS_8x16] = { { -1, 0 }, { 0, -1 }, { -1, 1 }, { -1, -1 }, 58 { -2, 0 }, { 0, -2 }, { -1, -2 }, { -2, -1 } }, 59 [BS_8x8] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 }, 60 { -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } }, 61 [BS_8x4] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 }, 62 { -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } }, 63 [BS_4x8] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 }, 64 { -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } }, 65 [BS_4x4] = { { 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 }, 66 { -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } }, 67 }; 68 VP9Context *s = td->s; 69 VP9Block *b = td->b; 70 int row = td->row, col = td->col, row7 = td->row7; 71 const int8_t (*p)[2] = mv_ref_blk_off[b->bs]; 72#define INVALID_MV 0x80008000U 73 uint32_t mem = INVALID_MV, mem_sub8x8 = INVALID_MV; 74 int i; 75 76#define RETURN_DIRECT_MV(mv) \ 77 do { \ 78 uint32_t m = AV_RN32A(&mv); \ 79 if (!idx) { \ 80 AV_WN32A(pmv, m); \ 81 return; \ 82 } else if (mem == INVALID_MV) { \ 83 mem = m; \ 84 } else if (m != mem) { \ 85 AV_WN32A(pmv, m); \ 86 return; \ 87 } \ 88 } while (0) 89 90 if (sb >= 0) { 91 if (sb == 2 || sb == 1) { 92 RETURN_DIRECT_MV(b->mv[0][z]); 93 } else if (sb == 3) { 94 RETURN_DIRECT_MV(b->mv[2][z]); 95 RETURN_DIRECT_MV(b->mv[1][z]); 96 RETURN_DIRECT_MV(b->mv[0][z]); 97 } 98 99#define RETURN_MV(mv) \ 100 do { \ 101 if (sb > 0) { \ 102 VP56mv tmp; \ 103 uint32_t m; \ 104 av_assert2(idx == 1); \ 105 av_assert2(mem != INVALID_MV); \ 106 if (mem_sub8x8 == INVALID_MV) { \ 107 clamp_mv(&tmp, &mv, td); \ 108 m = AV_RN32A(&tmp); \ 109 if (m != mem) { \ 110 AV_WN32A(pmv, m); \ 111 return; \ 112 } \ 113 mem_sub8x8 = AV_RN32A(&mv); \ 114 } else if (mem_sub8x8 != AV_RN32A(&mv)) { \ 115 clamp_mv(&tmp, &mv, td); \ 116 m = AV_RN32A(&tmp); \ 117 if (m != mem) { \ 118 AV_WN32A(pmv, m); \ 119 } else { \ 120 /* BUG I'm pretty sure this isn't the intention */ \ 121 AV_WN32A(pmv, 0); \ 122 } \ 123 return; \ 124 } \ 125 } else { \ 126 uint32_t m = AV_RN32A(&mv); \ 127 if (!idx) { \ 128 clamp_mv(pmv, &mv, td); \ 129 return; \ 130 } else if (mem == INVALID_MV) { \ 131 mem = m; \ 132 } else if (m != mem) { \ 133 clamp_mv(pmv, &mv, td); \ 134 return; \ 135 } \ 136 } \ 137 } while (0) 138 139 if (row > 0) { 140 VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col]; 141 if (mv->ref[0] == ref) 142 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]); 143 else if (mv->ref[1] == ref) 144 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]); 145 } 146 if (col > td->tile_col_start) { 147 VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1]; 148 if (mv->ref[0] == ref) 149 RETURN_MV(td->left_mv_ctx[2 * row7 + (sb >> 1)][0]); 150 else if (mv->ref[1] == ref) 151 RETURN_MV(td->left_mv_ctx[2 * row7 + (sb >> 1)][1]); 152 } 153 i = 2; 154 } else { 155 i = 0; 156 } 157 158 // previously coded MVs in this neighborhood, using same reference frame 159 for (; i < 8; i++) { 160 int c = p[i][0] + col, r = p[i][1] + row; 161 162 if (c >= td->tile_col_start && c < s->cols && 163 r >= 0 && r < s->rows) { 164 VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c]; 165 166 if (mv->ref[0] == ref) 167 RETURN_MV(mv->mv[0]); 168 else if (mv->ref[1] == ref) 169 RETURN_MV(mv->mv[1]); 170 } 171 } 172 173 // MV at this position in previous frame, using same reference frame 174 if (s->s.h.use_last_frame_mvs) { 175 VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col]; 176 177 if (!s->s.frames[REF_FRAME_MVPAIR].uses_2pass) 178 ff_thread_await_progress(&s->s.frames[REF_FRAME_MVPAIR].tf, row >> 3, 0); 179 if (mv->ref[0] == ref) 180 RETURN_MV(mv->mv[0]); 181 else if (mv->ref[1] == ref) 182 RETURN_MV(mv->mv[1]); 183 } 184 185#define RETURN_SCALE_MV(mv, scale) \ 186 do { \ 187 if (scale) { \ 188 VP56mv mv_temp = { -mv.x, -mv.y }; \ 189 RETURN_MV(mv_temp); \ 190 } else { \ 191 RETURN_MV(mv); \ 192 } \ 193 } while (0) 194 195 // previously coded MVs in this neighborhood, using different reference frame 196 for (i = 0; i < 8; i++) { 197 int c = p[i][0] + col, r = p[i][1] + row; 198 199 if (c >= td->tile_col_start && c < s->cols && r >= 0 && r < s->rows) { 200 VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c]; 201 202 if (mv->ref[0] != ref && mv->ref[0] >= 0) 203 RETURN_SCALE_MV(mv->mv[0], 204 s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]); 205 if (mv->ref[1] != ref && mv->ref[1] >= 0 && 206 // BUG - libvpx has this condition regardless of whether 207 // we used the first ref MV and pre-scaling 208 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) { 209 RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]); 210 } 211 } 212 } 213 214 // MV at this position in previous frame, using different reference frame 215 if (s->s.h.use_last_frame_mvs) { 216 VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col]; 217 218 // no need to await_progress, because we already did that above 219 if (mv->ref[0] != ref && mv->ref[0] >= 0) 220 RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]); 221 if (mv->ref[1] != ref && mv->ref[1] >= 0 && 222 // BUG - libvpx has this condition regardless of whether 223 // we used the first ref MV and pre-scaling 224 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) { 225 RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]); 226 } 227 } 228 229 AV_ZERO32(pmv); 230 clamp_mv(pmv, pmv, td); 231#undef INVALID_MV 232#undef RETURN_MV 233#undef RETURN_SCALE_MV 234} 235 236static av_always_inline int read_mv_component(VP9TileData *td, int idx, int hp) 237{ 238 VP9Context *s = td->s; 239 int bit, sign = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].sign); 240 int n, c = vp8_rac_get_tree(td->c, ff_vp9_mv_class_tree, 241 s->prob.p.mv_comp[idx].classes); 242 243 td->counts.mv_comp[idx].sign[sign]++; 244 td->counts.mv_comp[idx].classes[c]++; 245 if (c) { 246 int m; 247 248 for (n = 0, m = 0; m < c; m++) { 249 bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].bits[m]); 250 n |= bit << m; 251 td->counts.mv_comp[idx].bits[m][bit]++; 252 } 253 n <<= 3; 254 bit = vp8_rac_get_tree(td->c, ff_vp9_mv_fp_tree, 255 s->prob.p.mv_comp[idx].fp); 256 n |= bit << 1; 257 td->counts.mv_comp[idx].fp[bit]++; 258 if (hp) { 259 bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].hp); 260 td->counts.mv_comp[idx].hp[bit]++; 261 n |= bit; 262 } else { 263 n |= 1; 264 // bug in libvpx - we count for bw entropy purposes even if the 265 // bit wasn't coded 266 td->counts.mv_comp[idx].hp[1]++; 267 } 268 n += 8 << c; 269 } else { 270 n = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].class0); 271 td->counts.mv_comp[idx].class0[n]++; 272 bit = vp8_rac_get_tree(td->c, ff_vp9_mv_fp_tree, 273 s->prob.p.mv_comp[idx].class0_fp[n]); 274 td->counts.mv_comp[idx].class0_fp[n][bit]++; 275 n = (n << 3) | (bit << 1); 276 if (hp) { 277 bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].class0_hp); 278 td->counts.mv_comp[idx].class0_hp[bit]++; 279 n |= bit; 280 } else { 281 n |= 1; 282 // bug in libvpx - we count for bw entropy purposes even if the 283 // bit wasn't coded 284 td->counts.mv_comp[idx].class0_hp[1]++; 285 } 286 } 287 288 return sign ? -(n + 1) : (n + 1); 289} 290 291void ff_vp9_fill_mv(VP9TileData *td, VP56mv *mv, int mode, int sb) 292{ 293 VP9Context *s = td->s; 294 VP9Block *b = td->b; 295 296 if (mode == ZEROMV) { 297 AV_ZERO64(mv); 298 } else { 299 int hp; 300 301 // FIXME cache this value and reuse for other subblocks 302 find_ref_mvs(td, &mv[0], b->ref[0], 0, mode == NEARMV, 303 mode == NEWMV ? -1 : sb); 304 // FIXME maybe move this code into find_ref_mvs() 305 if ((mode == NEWMV || sb == -1) && 306 !(hp = s->s.h.highprecisionmvs && 307 abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) { 308 if (mv[0].y & 1) { 309 if (mv[0].y < 0) 310 mv[0].y++; 311 else 312 mv[0].y--; 313 } 314 if (mv[0].x & 1) { 315 if (mv[0].x < 0) 316 mv[0].x++; 317 else 318 mv[0].x--; 319 } 320 } 321 if (mode == NEWMV) { 322 enum MVJoint j = vp8_rac_get_tree(td->c, ff_vp9_mv_joint_tree, 323 s->prob.p.mv_joint); 324 325 td->counts.mv_joint[j]++; 326 if (j >= MV_JOINT_V) 327 mv[0].y += read_mv_component(td, 0, hp); 328 if (j & 1) 329 mv[0].x += read_mv_component(td, 1, hp); 330 } 331 332 if (b->comp) { 333 // FIXME cache this value and reuse for other subblocks 334 find_ref_mvs(td, &mv[1], b->ref[1], 1, mode == NEARMV, 335 mode == NEWMV ? -1 : sb); 336 if ((mode == NEWMV || sb == -1) && 337 !(hp = s->s.h.highprecisionmvs && 338 abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) { 339 if (mv[1].y & 1) { 340 if (mv[1].y < 0) 341 mv[1].y++; 342 else 343 mv[1].y--; 344 } 345 if (mv[1].x & 1) { 346 if (mv[1].x < 0) 347 mv[1].x++; 348 else 349 mv[1].x--; 350 } 351 } 352 if (mode == NEWMV) { 353 enum MVJoint j = vp8_rac_get_tree(td->c, ff_vp9_mv_joint_tree, 354 s->prob.p.mv_joint); 355 356 td->counts.mv_joint[j]++; 357 if (j >= MV_JOINT_V) 358 mv[1].y += read_mv_component(td, 0, hp); 359 if (j & 1) 360 mv[1].x += read_mv_component(td, 1, hp); 361 } 362 } 363 } 364} 365