| 1 | /* |
| 2 | * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding |
| 3 | * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at> |
| 4 | * |
| 5 | * This file is part of FFmpeg. |
| 6 | * |
| 7 | * FFmpeg is free software; you can redistribute it and/or |
| 8 | * modify it under the terms of the GNU Lesser General Public |
| 9 | * License as published by the Free Software Foundation; either |
| 10 | * version 2.1 of the License, or (at your option) any later version. |
| 11 | * |
| 12 | * FFmpeg is distributed in the hope that it will be useful, |
| 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 15 | * Lesser General Public License for more details. |
| 16 | * |
| 17 | * You should have received a copy of the GNU Lesser General Public |
| 18 | * License along with FFmpeg; if not, write to the Free Software |
| 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 20 | */ |
| 21 | |
| 22 | /** |
| 23 | * @file |
| 24 | * H.264 / AVC / MPEG4 part10 direct mb/block decoding. |
| 25 | * @author Michael Niedermayer <michaelni@gmx.at> |
| 26 | */ |
| 27 | |
| 28 | #include "internal.h" |
| 29 | #include "avcodec.h" |
| 30 | #include "h264.h" |
| 31 | #include "mpegutils.h" |
| 32 | #include "rectangle.h" |
| 33 | #include "thread.h" |
| 34 | |
| 35 | #include <assert.h> |
| 36 | |
| 37 | static int get_scale_factor(H264Context *const h, int poc, int poc1, int i) |
| 38 | { |
| 39 | int poc0 = h->ref_list[0][i].poc; |
| 40 | int td = av_clip(poc1 - poc0, -128, 127); |
| 41 | if (td == 0 || h->ref_list[0][i].long_ref) { |
| 42 | return 256; |
| 43 | } else { |
| 44 | int tb = av_clip(poc - poc0, -128, 127); |
| 45 | int tx = (16384 + (FFABS(td) >> 1)) / td; |
| 46 | return av_clip((tb * tx + 32) >> 6, -1024, 1023); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | void ff_h264_direct_dist_scale_factor(H264Context *const h) |
| 51 | { |
| 52 | const int poc = FIELD_PICTURE(h) ? h->cur_pic_ptr->field_poc[h->picture_structure == PICT_BOTTOM_FIELD] |
| 53 | : h->cur_pic_ptr->poc; |
| 54 | const int poc1 = h->ref_list[1][0].poc; |
| 55 | int i, field; |
| 56 | |
| 57 | if (FRAME_MBAFF(h)) |
| 58 | for (field = 0; field < 2; field++) { |
| 59 | const int poc = h->cur_pic_ptr->field_poc[field]; |
| 60 | const int poc1 = h->ref_list[1][0].field_poc[field]; |
| 61 | for (i = 0; i < 2 * h->ref_count[0]; i++) |
| 62 | h->dist_scale_factor_field[field][i ^ field] = |
| 63 | get_scale_factor(h, poc, poc1, i + 16); |
| 64 | } |
| 65 | |
| 66 | for (i = 0; i < h->ref_count[0]; i++) |
| 67 | h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i); |
| 68 | } |
| 69 | |
| 70 | static void fill_colmap(H264Context *h, int map[2][16 + 32], int list, |
| 71 | int field, int colfield, int mbafi) |
| 72 | { |
| 73 | H264Picture *const ref1 = &h->ref_list[1][0]; |
| 74 | int j, old_ref, rfield; |
| 75 | int start = mbafi ? 16 : 0; |
| 76 | int end = mbafi ? 16 + 2 * h->ref_count[0] : h->ref_count[0]; |
| 77 | int interl = mbafi || h->picture_structure != PICT_FRAME; |
| 78 | |
| 79 | /* bogus; fills in for missing frames */ |
| 80 | memset(map[list], 0, sizeof(map[list])); |
| 81 | |
| 82 | for (rfield = 0; rfield < 2; rfield++) { |
| 83 | for (old_ref = 0; old_ref < ref1->ref_count[colfield][list]; old_ref++) { |
| 84 | int poc = ref1->ref_poc[colfield][list][old_ref]; |
| 85 | |
| 86 | if (!interl) |
| 87 | poc |= 3; |
| 88 | // FIXME: store all MBAFF references so this is not needed |
| 89 | else if (interl && (poc & 3) == 3) |
| 90 | poc = (poc & ~3) + rfield + 1; |
| 91 | |
| 92 | for (j = start; j < end; j++) { |
| 93 | if (4 * h->ref_list[0][j].frame_num + |
| 94 | (h->ref_list[0][j].reference & 3) == poc) { |
| 95 | int cur_ref = mbafi ? (j - 16) ^ field : j; |
| 96 | if (ref1->mbaff) |
| 97 | map[list][2 * old_ref + (rfield ^ field) + 16] = cur_ref; |
| 98 | if (rfield == field || !interl) |
| 99 | map[list][old_ref] = cur_ref; |
| 100 | break; |
| 101 | } |
| 102 | } |
| 103 | } |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | void ff_h264_direct_ref_list_init(H264Context *const h) |
| 108 | { |
| 109 | H264Picture *const ref1 = &h->ref_list[1][0]; |
| 110 | H264Picture *const cur = h->cur_pic_ptr; |
| 111 | int list, j, field; |
| 112 | int sidx = (h->picture_structure & 1) ^ 1; |
| 113 | int ref1sidx = (ref1->reference & 1) ^ 1; |
| 114 | |
| 115 | for (list = 0; list < 2; list++) { |
| 116 | cur->ref_count[sidx][list] = h->ref_count[list]; |
| 117 | for (j = 0; j < h->ref_count[list]; j++) |
| 118 | cur->ref_poc[sidx][list][j] = 4 * h->ref_list[list][j].frame_num + |
| 119 | (h->ref_list[list][j].reference & 3); |
| 120 | } |
| 121 | |
| 122 | if (h->picture_structure == PICT_FRAME) { |
| 123 | memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0])); |
| 124 | memcpy(cur->ref_poc[1], cur->ref_poc[0], sizeof(cur->ref_poc[0])); |
| 125 | } |
| 126 | |
| 127 | cur->mbaff = FRAME_MBAFF(h); |
| 128 | |
| 129 | h->col_fieldoff = 0; |
| 130 | if (h->picture_structure == PICT_FRAME) { |
| 131 | int cur_poc = h->cur_pic_ptr->poc; |
| 132 | int *col_poc = h->ref_list[1]->field_poc; |
| 133 | h->col_parity = (FFABS(col_poc[0] - cur_poc) >= |
| 134 | FFABS(col_poc[1] - cur_poc)); |
| 135 | ref1sidx = |
| 136 | sidx = h->col_parity; |
| 137 | // FL -> FL & differ parity |
| 138 | } else if (!(h->picture_structure & h->ref_list[1][0].reference) && |
| 139 | !h->ref_list[1][0].mbaff) { |
| 140 | h->col_fieldoff = 2 * h->ref_list[1][0].reference - 3; |
| 141 | } |
| 142 | |
| 143 | if (h->slice_type_nos != AV_PICTURE_TYPE_B || h->direct_spatial_mv_pred) |
| 144 | return; |
| 145 | |
| 146 | for (list = 0; list < 2; list++) { |
| 147 | fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0); |
| 148 | if (FRAME_MBAFF(h)) |
| 149 | for (field = 0; field < 2; field++) |
| 150 | fill_colmap(h, h->map_col_to_list0_field[field], list, field, |
| 151 | field, 1); |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | static void await_reference_mb_row(H264Context *const h, H264Picture *ref, |
| 156 | int mb_y) |
| 157 | { |
| 158 | int ref_field = ref->reference - 1; |
| 159 | int ref_field_picture = ref->field_picture; |
| 160 | int ref_height = 16 * h->mb_height >> ref_field_picture; |
| 161 | |
| 162 | if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_FRAME)) |
| 163 | return; |
| 164 | |
| 165 | /* FIXME: It can be safe to access mb stuff |
| 166 | * even if pixels aren't deblocked yet. */ |
| 167 | |
| 168 | ff_thread_await_progress(&ref->tf, |
| 169 | FFMIN(16 * mb_y >> ref_field_picture, |
| 170 | ref_height - 1), |
| 171 | ref_field_picture && ref_field); |
| 172 | } |
| 173 | |
| 174 | static void pred_spatial_direct_motion(H264Context *const h, int *mb_type) |
| 175 | { |
| 176 | int b8_stride = 2; |
| 177 | int b4_stride = h->b_stride; |
| 178 | int mb_xy = h->mb_xy, mb_y = h->mb_y; |
| 179 | int mb_type_col[2]; |
| 180 | const int16_t (*l1mv0)[2], (*l1mv1)[2]; |
| 181 | const int8_t *l1ref0, *l1ref1; |
| 182 | const int is_b8x8 = IS_8X8(*mb_type); |
| 183 | unsigned int sub_mb_type = MB_TYPE_L0L1; |
| 184 | int i8, i4; |
| 185 | int ref[2]; |
| 186 | int mv[2]; |
| 187 | int list; |
| 188 | |
| 189 | assert(h->ref_list[1][0].reference & 3); |
| 190 | |
| 191 | await_reference_mb_row(h, &h->ref_list[1][0], |
| 192 | h->mb_y + !!IS_INTERLACED(*mb_type)); |
| 193 | |
| 194 | #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \ |
| 195 | MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM) |
| 196 | |
| 197 | /* ref = min(neighbors) */ |
| 198 | for (list = 0; list < 2; list++) { |
| 199 | int left_ref = h->ref_cache[list][scan8[0] - 1]; |
| 200 | int top_ref = h->ref_cache[list][scan8[0] - 8]; |
| 201 | int refc = h->ref_cache[list][scan8[0] - 8 + 4]; |
| 202 | const int16_t *C = h->mv_cache[list][scan8[0] - 8 + 4]; |
| 203 | if (refc == PART_NOT_AVAILABLE) { |
| 204 | refc = h->ref_cache[list][scan8[0] - 8 - 1]; |
| 205 | C = h->mv_cache[list][scan8[0] - 8 - 1]; |
| 206 | } |
| 207 | ref[list] = FFMIN3((unsigned)left_ref, |
| 208 | (unsigned)top_ref, |
| 209 | (unsigned)refc); |
| 210 | if (ref[list] >= 0) { |
| 211 | /* This is just pred_motion() but with the cases removed that |
| 212 | * cannot happen for direct blocks. */ |
| 213 | const int16_t *const A = h->mv_cache[list][scan8[0] - 1]; |
| 214 | const int16_t *const B = h->mv_cache[list][scan8[0] - 8]; |
| 215 | |
| 216 | int match_count = (left_ref == ref[list]) + |
| 217 | (top_ref == ref[list]) + |
| 218 | (refc == ref[list]); |
| 219 | |
| 220 | if (match_count > 1) { // most common |
| 221 | mv[list] = pack16to32(mid_pred(A[0], B[0], C[0]), |
| 222 | mid_pred(A[1], B[1], C[1])); |
| 223 | } else { |
| 224 | assert(match_count == 1); |
| 225 | if (left_ref == ref[list]) |
| 226 | mv[list] = AV_RN32A(A); |
| 227 | else if (top_ref == ref[list]) |
| 228 | mv[list] = AV_RN32A(B); |
| 229 | else |
| 230 | mv[list] = AV_RN32A(C); |
| 231 | } |
| 232 | av_assert2(ref[list] < (h->ref_count[list] << !!FRAME_MBAFF(h))); |
| 233 | } else { |
| 234 | int mask = ~(MB_TYPE_L0 << (2 * list)); |
| 235 | mv[list] = 0; |
| 236 | ref[list] = -1; |
| 237 | if (!is_b8x8) |
| 238 | *mb_type &= mask; |
| 239 | sub_mb_type &= mask; |
| 240 | } |
| 241 | } |
| 242 | if (ref[0] < 0 && ref[1] < 0) { |
| 243 | ref[0] = ref[1] = 0; |
| 244 | if (!is_b8x8) |
| 245 | *mb_type |= MB_TYPE_L0L1; |
| 246 | sub_mb_type |= MB_TYPE_L0L1; |
| 247 | } |
| 248 | |
| 249 | if (!(is_b8x8 | mv[0] | mv[1])) { |
| 250 | fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); |
| 251 | fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); |
| 252 | fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4); |
| 253 | fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4); |
| 254 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
| 255 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
| 256 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
| 257 | return; |
| 258 | } |
| 259 | |
| 260 | if (IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL |
| 261 | if (!IS_INTERLACED(*mb_type)) { // AFR/FR -> AFL/FL |
| 262 | mb_y = (h->mb_y & ~1) + h->col_parity; |
| 263 | mb_xy = h->mb_x + |
| 264 | ((h->mb_y & ~1) + h->col_parity) * h->mb_stride; |
| 265 | b8_stride = 0; |
| 266 | } else { |
| 267 | mb_y += h->col_fieldoff; |
| 268 | mb_xy += h->mb_stride * h->col_fieldoff; // non-zero for FL -> FL & differ parity |
| 269 | } |
| 270 | goto single_col; |
| 271 | } else { // AFL/AFR/FR/FL -> AFR/FR |
| 272 | if (IS_INTERLACED(*mb_type)) { // AFL /FL -> AFR/FR |
| 273 | mb_y = h->mb_y & ~1; |
| 274 | mb_xy = (h->mb_y & ~1) * h->mb_stride + h->mb_x; |
| 275 | mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy]; |
| 276 | mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + h->mb_stride]; |
| 277 | b8_stride = 2 + 4 * h->mb_stride; |
| 278 | b4_stride *= 6; |
| 279 | if (IS_INTERLACED(mb_type_col[0]) != |
| 280 | IS_INTERLACED(mb_type_col[1])) { |
| 281 | mb_type_col[0] &= ~MB_TYPE_INTERLACED; |
| 282 | mb_type_col[1] &= ~MB_TYPE_INTERLACED; |
| 283 | } |
| 284 | |
| 285 | sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
| 286 | if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) && |
| 287 | (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) && |
| 288 | !is_b8x8) { |
| 289 | *mb_type |= MB_TYPE_16x8 | MB_TYPE_DIRECT2; /* B_16x8 */ |
| 290 | } else { |
| 291 | *mb_type |= MB_TYPE_8x8; |
| 292 | } |
| 293 | } else { // AFR/FR -> AFR/FR |
| 294 | single_col: |
| 295 | mb_type_col[0] = |
| 296 | mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy]; |
| 297 | |
| 298 | sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
| 299 | if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) { |
| 300 | *mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_16x16 */ |
| 301 | } else if (!is_b8x8 && |
| 302 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) { |
| 303 | *mb_type |= MB_TYPE_DIRECT2 | |
| 304 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16)); |
| 305 | } else { |
| 306 | if (!h->sps.direct_8x8_inference_flag) { |
| 307 | /* FIXME: Save sub mb types from previous frames (or derive |
| 308 | * from MVs) so we know exactly what block size to use. */ |
| 309 | sub_mb_type += (MB_TYPE_8x8 - MB_TYPE_16x16); /* B_SUB_4x4 */ |
| 310 | } |
| 311 | *mb_type |= MB_TYPE_8x8; |
| 312 | } |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | await_reference_mb_row(h, &h->ref_list[1][0], mb_y); |
| 317 | |
| 318 | l1mv0 = (void*)&h->ref_list[1][0].motion_val[0][h->mb2b_xy[mb_xy]]; |
| 319 | l1mv1 = (void*)&h->ref_list[1][0].motion_val[1][h->mb2b_xy[mb_xy]]; |
| 320 | l1ref0 = &h->ref_list[1][0].ref_index[0][4 * mb_xy]; |
| 321 | l1ref1 = &h->ref_list[1][0].ref_index[1][4 * mb_xy]; |
| 322 | if (!b8_stride) { |
| 323 | if (h->mb_y & 1) { |
| 324 | l1ref0 += 2; |
| 325 | l1ref1 += 2; |
| 326 | l1mv0 += 2 * b4_stride; |
| 327 | l1mv1 += 2 * b4_stride; |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) { |
| 332 | int n = 0; |
| 333 | for (i8 = 0; i8 < 4; i8++) { |
| 334 | int x8 = i8 & 1; |
| 335 | int y8 = i8 >> 1; |
| 336 | int xy8 = x8 + y8 * b8_stride; |
| 337 | int xy4 = x8 * 3 + y8 * b4_stride; |
| 338 | int a, b; |
| 339 | |
| 340 | if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) |
| 341 | continue; |
| 342 | h->sub_mb_type[i8] = sub_mb_type; |
| 343 | |
| 344 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 345 | (uint8_t)ref[0], 1); |
| 346 | fill_rectangle(&h->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, |
| 347 | (uint8_t)ref[1], 1); |
| 348 | if (!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref && |
| 349 | ((l1ref0[xy8] == 0 && |
| 350 | FFABS(l1mv0[xy4][0]) <= 1 && |
| 351 | FFABS(l1mv0[xy4][1]) <= 1) || |
| 352 | (l1ref0[xy8] < 0 && |
| 353 | l1ref1[xy8] == 0 && |
| 354 | FFABS(l1mv1[xy4][0]) <= 1 && |
| 355 | FFABS(l1mv1[xy4][1]) <= 1))) { |
| 356 | a = |
| 357 | b = 0; |
| 358 | if (ref[0] > 0) |
| 359 | a = mv[0]; |
| 360 | if (ref[1] > 0) |
| 361 | b = mv[1]; |
| 362 | n++; |
| 363 | } else { |
| 364 | a = mv[0]; |
| 365 | b = mv[1]; |
| 366 | } |
| 367 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, a, 4); |
| 368 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, b, 4); |
| 369 | } |
| 370 | if (!is_b8x8 && !(n & 3)) |
| 371 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
| 372 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
| 373 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
| 374 | } else if (IS_16X16(*mb_type)) { |
| 375 | int a, b; |
| 376 | |
| 377 | fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); |
| 378 | fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); |
| 379 | if (!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && |
| 380 | ((l1ref0[0] == 0 && |
| 381 | FFABS(l1mv0[0][0]) <= 1 && |
| 382 | FFABS(l1mv0[0][1]) <= 1) || |
| 383 | (l1ref0[0] < 0 && !l1ref1[0] && |
| 384 | FFABS(l1mv1[0][0]) <= 1 && |
| 385 | FFABS(l1mv1[0][1]) <= 1 && |
| 386 | h->x264_build > 33U))) { |
| 387 | a = b = 0; |
| 388 | if (ref[0] > 0) |
| 389 | a = mv[0]; |
| 390 | if (ref[1] > 0) |
| 391 | b = mv[1]; |
| 392 | } else { |
| 393 | a = mv[0]; |
| 394 | b = mv[1]; |
| 395 | } |
| 396 | fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4); |
| 397 | fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4); |
| 398 | } else { |
| 399 | int n = 0; |
| 400 | for (i8 = 0; i8 < 4; i8++) { |
| 401 | const int x8 = i8 & 1; |
| 402 | const int y8 = i8 >> 1; |
| 403 | |
| 404 | if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) |
| 405 | continue; |
| 406 | h->sub_mb_type[i8] = sub_mb_type; |
| 407 | |
| 408 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, mv[0], 4); |
| 409 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, mv[1], 4); |
| 410 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 411 | (uint8_t)ref[0], 1); |
| 412 | fill_rectangle(&h->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, |
| 413 | (uint8_t)ref[1], 1); |
| 414 | |
| 415 | assert(b8_stride == 2); |
| 416 | /* col_zero_flag */ |
| 417 | if (!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && |
| 418 | (l1ref0[i8] == 0 || |
| 419 | (l1ref0[i8] < 0 && |
| 420 | l1ref1[i8] == 0 && |
| 421 | h->x264_build > 33U))) { |
| 422 | const int16_t (*l1mv)[2] = l1ref0[i8] == 0 ? l1mv0 : l1mv1; |
| 423 | if (IS_SUB_8X8(sub_mb_type)) { |
| 424 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride]; |
| 425 | if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) { |
| 426 | if (ref[0] == 0) |
| 427 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, |
| 428 | 8, 0, 4); |
| 429 | if (ref[1] == 0) |
| 430 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, |
| 431 | 8, 0, 4); |
| 432 | n += 4; |
| 433 | } |
| 434 | } else { |
| 435 | int m = 0; |
| 436 | for (i4 = 0; i4 < 4; i4++) { |
| 437 | const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) + |
| 438 | (y8 * 2 + (i4 >> 1)) * b4_stride]; |
| 439 | if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) { |
| 440 | if (ref[0] == 0) |
| 441 | AV_ZERO32(h->mv_cache[0][scan8[i8 * 4 + i4]]); |
| 442 | if (ref[1] == 0) |
| 443 | AV_ZERO32(h->mv_cache[1][scan8[i8 * 4 + i4]]); |
| 444 | m++; |
| 445 | } |
| 446 | } |
| 447 | if (!(m & 3)) |
| 448 | h->sub_mb_type[i8] += MB_TYPE_16x16 - MB_TYPE_8x8; |
| 449 | n += m; |
| 450 | } |
| 451 | } |
| 452 | } |
| 453 | if (!is_b8x8 && !(n & 15)) |
| 454 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
| 455 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
| 456 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | static void pred_temp_direct_motion(H264Context *const h, int *mb_type) |
| 461 | { |
| 462 | int b8_stride = 2; |
| 463 | int b4_stride = h->b_stride; |
| 464 | int mb_xy = h->mb_xy, mb_y = h->mb_y; |
| 465 | int mb_type_col[2]; |
| 466 | const int16_t (*l1mv0)[2], (*l1mv1)[2]; |
| 467 | const int8_t *l1ref0, *l1ref1; |
| 468 | const int is_b8x8 = IS_8X8(*mb_type); |
| 469 | unsigned int sub_mb_type; |
| 470 | int i8, i4; |
| 471 | |
| 472 | assert(h->ref_list[1][0].reference & 3); |
| 473 | |
| 474 | await_reference_mb_row(h, &h->ref_list[1][0], |
| 475 | h->mb_y + !!IS_INTERLACED(*mb_type)); |
| 476 | |
| 477 | if (IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL |
| 478 | if (!IS_INTERLACED(*mb_type)) { // AFR/FR -> AFL/FL |
| 479 | mb_y = (h->mb_y & ~1) + h->col_parity; |
| 480 | mb_xy = h->mb_x + |
| 481 | ((h->mb_y & ~1) + h->col_parity) * h->mb_stride; |
| 482 | b8_stride = 0; |
| 483 | } else { |
| 484 | mb_y += h->col_fieldoff; |
| 485 | mb_xy += h->mb_stride * h->col_fieldoff; // non-zero for FL -> FL & differ parity |
| 486 | } |
| 487 | goto single_col; |
| 488 | } else { // AFL/AFR/FR/FL -> AFR/FR |
| 489 | if (IS_INTERLACED(*mb_type)) { // AFL /FL -> AFR/FR |
| 490 | mb_y = h->mb_y & ~1; |
| 491 | mb_xy = h->mb_x + (h->mb_y & ~1) * h->mb_stride; |
| 492 | mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy]; |
| 493 | mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + h->mb_stride]; |
| 494 | b8_stride = 2 + 4 * h->mb_stride; |
| 495 | b4_stride *= 6; |
| 496 | if (IS_INTERLACED(mb_type_col[0]) != |
| 497 | IS_INTERLACED(mb_type_col[1])) { |
| 498 | mb_type_col[0] &= ~MB_TYPE_INTERLACED; |
| 499 | mb_type_col[1] &= ~MB_TYPE_INTERLACED; |
| 500 | } |
| 501 | |
| 502 | sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
| 503 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
| 504 | |
| 505 | if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) && |
| 506 | (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) && |
| 507 | !is_b8x8) { |
| 508 | *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 | |
| 509 | MB_TYPE_DIRECT2; /* B_16x8 */ |
| 510 | } else { |
| 511 | *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1; |
| 512 | } |
| 513 | } else { // AFR/FR -> AFR/FR |
| 514 | single_col: |
| 515 | mb_type_col[0] = |
| 516 | mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy]; |
| 517 | |
| 518 | sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
| 519 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
| 520 | if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) { |
| 521 | *mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
| 522 | MB_TYPE_DIRECT2; /* B_16x16 */ |
| 523 | } else if (!is_b8x8 && |
| 524 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) { |
| 525 | *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 | |
| 526 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16)); |
| 527 | } else { |
| 528 | if (!h->sps.direct_8x8_inference_flag) { |
| 529 | /* FIXME: save sub mb types from previous frames (or derive |
| 530 | * from MVs) so we know exactly what block size to use */ |
| 531 | sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
| 532 | MB_TYPE_DIRECT2; /* B_SUB_4x4 */ |
| 533 | } |
| 534 | *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1; |
| 535 | } |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | await_reference_mb_row(h, &h->ref_list[1][0], mb_y); |
| 540 | |
| 541 | l1mv0 = (void*)&h->ref_list[1][0].motion_val[0][h->mb2b_xy[mb_xy]]; |
| 542 | l1mv1 = (void*)&h->ref_list[1][0].motion_val[1][h->mb2b_xy[mb_xy]]; |
| 543 | l1ref0 = &h->ref_list[1][0].ref_index[0][4 * mb_xy]; |
| 544 | l1ref1 = &h->ref_list[1][0].ref_index[1][4 * mb_xy]; |
| 545 | if (!b8_stride) { |
| 546 | if (h->mb_y & 1) { |
| 547 | l1ref0 += 2; |
| 548 | l1ref1 += 2; |
| 549 | l1mv0 += 2 * b4_stride; |
| 550 | l1mv1 += 2 * b4_stride; |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | { |
| 555 | const int *map_col_to_list0[2] = { h->map_col_to_list0[0], |
| 556 | h->map_col_to_list0[1] }; |
| 557 | const int *dist_scale_factor = h->dist_scale_factor; |
| 558 | int ref_offset; |
| 559 | |
| 560 | if (FRAME_MBAFF(h) && IS_INTERLACED(*mb_type)) { |
| 561 | map_col_to_list0[0] = h->map_col_to_list0_field[h->mb_y & 1][0]; |
| 562 | map_col_to_list0[1] = h->map_col_to_list0_field[h->mb_y & 1][1]; |
| 563 | dist_scale_factor = h->dist_scale_factor_field[h->mb_y & 1]; |
| 564 | } |
| 565 | ref_offset = (h->ref_list[1][0].mbaff << 4) & (mb_type_col[0] >> 3); |
| 566 | |
| 567 | if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) { |
| 568 | int y_shift = 2 * !IS_INTERLACED(*mb_type); |
| 569 | assert(h->sps.direct_8x8_inference_flag); |
| 570 | |
| 571 | for (i8 = 0; i8 < 4; i8++) { |
| 572 | const int x8 = i8 & 1; |
| 573 | const int y8 = i8 >> 1; |
| 574 | int ref0, scale; |
| 575 | const int16_t (*l1mv)[2] = l1mv0; |
| 576 | |
| 577 | if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) |
| 578 | continue; |
| 579 | h->sub_mb_type[i8] = sub_mb_type; |
| 580 | |
| 581 | fill_rectangle(&h->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
| 582 | if (IS_INTRA(mb_type_col[y8])) { |
| 583 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
| 584 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
| 585 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
| 586 | continue; |
| 587 | } |
| 588 | |
| 589 | ref0 = l1ref0[x8 + y8 * b8_stride]; |
| 590 | if (ref0 >= 0) |
| 591 | ref0 = map_col_to_list0[0][ref0 + ref_offset]; |
| 592 | else { |
| 593 | ref0 = map_col_to_list0[1][l1ref1[x8 + y8 * b8_stride] + |
| 594 | ref_offset]; |
| 595 | l1mv = l1mv1; |
| 596 | } |
| 597 | scale = dist_scale_factor[ref0]; |
| 598 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 599 | ref0, 1); |
| 600 | |
| 601 | { |
| 602 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * b4_stride]; |
| 603 | int my_col = (mv_col[1] << y_shift) / 2; |
| 604 | int mx = (scale * mv_col[0] + 128) >> 8; |
| 605 | int my = (scale * my_col + 128) >> 8; |
| 606 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 607 | pack16to32(mx, my), 4); |
| 608 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, |
| 609 | pack16to32(mx - mv_col[0], my - my_col), 4); |
| 610 | } |
| 611 | } |
| 612 | return; |
| 613 | } |
| 614 | |
| 615 | /* one-to-one mv scaling */ |
| 616 | |
| 617 | if (IS_16X16(*mb_type)) { |
| 618 | int ref, mv0, mv1; |
| 619 | |
| 620 | fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1); |
| 621 | if (IS_INTRA(mb_type_col[0])) { |
| 622 | ref = mv0 = mv1 = 0; |
| 623 | } else { |
| 624 | const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset] |
| 625 | : map_col_to_list0[1][l1ref1[0] + ref_offset]; |
| 626 | const int scale = dist_scale_factor[ref0]; |
| 627 | const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0]; |
| 628 | int mv_l0[2]; |
| 629 | mv_l0[0] = (scale * mv_col[0] + 128) >> 8; |
| 630 | mv_l0[1] = (scale * mv_col[1] + 128) >> 8; |
| 631 | ref = ref0; |
| 632 | mv0 = pack16to32(mv_l0[0], mv_l0[1]); |
| 633 | mv1 = pack16to32(mv_l0[0] - mv_col[0], mv_l0[1] - mv_col[1]); |
| 634 | } |
| 635 | fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1); |
| 636 | fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4); |
| 637 | fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4); |
| 638 | } else { |
| 639 | for (i8 = 0; i8 < 4; i8++) { |
| 640 | const int x8 = i8 & 1; |
| 641 | const int y8 = i8 >> 1; |
| 642 | int ref0, scale; |
| 643 | const int16_t (*l1mv)[2] = l1mv0; |
| 644 | |
| 645 | if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) |
| 646 | continue; |
| 647 | h->sub_mb_type[i8] = sub_mb_type; |
| 648 | fill_rectangle(&h->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
| 649 | if (IS_INTRA(mb_type_col[0])) { |
| 650 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
| 651 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
| 652 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
| 653 | continue; |
| 654 | } |
| 655 | |
| 656 | assert(b8_stride == 2); |
| 657 | ref0 = l1ref0[i8]; |
| 658 | if (ref0 >= 0) |
| 659 | ref0 = map_col_to_list0[0][ref0 + ref_offset]; |
| 660 | else { |
| 661 | ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset]; |
| 662 | l1mv = l1mv1; |
| 663 | } |
| 664 | scale = dist_scale_factor[ref0]; |
| 665 | |
| 666 | fill_rectangle(&h->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 667 | ref0, 1); |
| 668 | if (IS_SUB_8X8(sub_mb_type)) { |
| 669 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride]; |
| 670 | int mx = (scale * mv_col[0] + 128) >> 8; |
| 671 | int my = (scale * mv_col[1] + 128) >> 8; |
| 672 | fill_rectangle(&h->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, |
| 673 | pack16to32(mx, my), 4); |
| 674 | fill_rectangle(&h->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, |
| 675 | pack16to32(mx - mv_col[0], my - mv_col[1]), 4); |
| 676 | } else { |
| 677 | for (i4 = 0; i4 < 4; i4++) { |
| 678 | const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) + |
| 679 | (y8 * 2 + (i4 >> 1)) * b4_stride]; |
| 680 | int16_t *mv_l0 = h->mv_cache[0][scan8[i8 * 4 + i4]]; |
| 681 | mv_l0[0] = (scale * mv_col[0] + 128) >> 8; |
| 682 | mv_l0[1] = (scale * mv_col[1] + 128) >> 8; |
| 683 | AV_WN32A(h->mv_cache[1][scan8[i8 * 4 + i4]], |
| 684 | pack16to32(mv_l0[0] - mv_col[0], |
| 685 | mv_l0[1] - mv_col[1])); |
| 686 | } |
| 687 | } |
| 688 | } |
| 689 | } |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | void ff_h264_pred_direct_motion(H264Context *const h, int *mb_type) |
| 694 | { |
| 695 | if (h->direct_spatial_mv_pred) |
| 696 | pred_spatial_direct_motion(h, mb_type); |
| 697 | else |
| 698 | pred_temp_direct_motion(h, mb_type); |
| 699 | } |