blob: a7a107c8c216361c811b8d3aff28477c22d15074
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 / MPEG-4 part10 direct mb/block decoding. |
25 | * @author Michael Niedermayer <michaelni@gmx.at> |
26 | */ |
27 | |
28 | #include "internal.h" |
29 | #include "avcodec.h" |
30 | #include "h264dec.h" |
31 | #include "h264_ps.h" |
32 | #include "mpegutils.h" |
33 | #include "rectangle.h" |
34 | #include "thread.h" |
35 | |
36 | #include <assert.h> |
37 | |
38 | static int get_scale_factor(H264SliceContext *sl, |
39 | int poc, int poc1, int i) |
40 | { |
41 | int poc0 = sl->ref_list[0][i].poc; |
42 | int64_t pocdiff = poc1 - (int64_t)poc0; |
43 | int td = av_clip_int8(pocdiff); |
44 | |
45 | if (pocdiff != (int)pocdiff) |
46 | avpriv_request_sample(sl->h264->avctx, "pocdiff overflow\n"); |
47 | |
48 | if (td == 0 || sl->ref_list[0][i].parent->long_ref) { |
49 | return 256; |
50 | } else { |
51 | int64_t pocdiff0 = poc - (int64_t)poc0; |
52 | int tb = av_clip_int8(pocdiff0); |
53 | int tx = (16384 + (FFABS(td) >> 1)) / td; |
54 | |
55 | if (pocdiff0 != (int)pocdiff0) |
56 | av_log(sl->h264->avctx, AV_LOG_DEBUG, "pocdiff0 overflow\n"); |
57 | |
58 | return av_clip_intp2((tb * tx + 32) >> 6, 10); |
59 | } |
60 | } |
61 | |
62 | void ff_h264_direct_dist_scale_factor(const H264Context *const h, |
63 | H264SliceContext *sl) |
64 | { |
65 | const int poc = FIELD_PICTURE(h) ? h->cur_pic_ptr->field_poc[h->picture_structure == PICT_BOTTOM_FIELD] |
66 | : h->cur_pic_ptr->poc; |
67 | const int poc1 = sl->ref_list[1][0].poc; |
68 | int i, field; |
69 | |
70 | if (FRAME_MBAFF(h)) |
71 | for (field = 0; field < 2; field++) { |
72 | const int poc = h->cur_pic_ptr->field_poc[field]; |
73 | const int poc1 = sl->ref_list[1][0].parent->field_poc[field]; |
74 | for (i = 0; i < 2 * sl->ref_count[0]; i++) |
75 | sl->dist_scale_factor_field[field][i ^ field] = |
76 | get_scale_factor(sl, poc, poc1, i + 16); |
77 | } |
78 | |
79 | for (i = 0; i < sl->ref_count[0]; i++) |
80 | sl->dist_scale_factor[i] = get_scale_factor(sl, poc, poc1, i); |
81 | } |
82 | |
83 | static void fill_colmap(const H264Context *h, H264SliceContext *sl, |
84 | int map[2][16 + 32], int list, |
85 | int field, int colfield, int mbafi) |
86 | { |
87 | H264Picture *const ref1 = sl->ref_list[1][0].parent; |
88 | int j, old_ref, rfield; |
89 | int start = mbafi ? 16 : 0; |
90 | int end = mbafi ? 16 + 2 * sl->ref_count[0] : sl->ref_count[0]; |
91 | int interl = mbafi || h->picture_structure != PICT_FRAME; |
92 | |
93 | /* bogus; fills in for missing frames */ |
94 | memset(map[list], 0, sizeof(map[list])); |
95 | |
96 | for (rfield = 0; rfield < 2; rfield++) { |
97 | for (old_ref = 0; old_ref < ref1->ref_count[colfield][list]; old_ref++) { |
98 | int poc = ref1->ref_poc[colfield][list][old_ref]; |
99 | |
100 | if (!interl) |
101 | poc |= 3; |
102 | // FIXME: store all MBAFF references so this is not needed |
103 | else if (interl && (poc & 3) == 3) |
104 | poc = (poc & ~3) + rfield + 1; |
105 | |
106 | for (j = start; j < end; j++) { |
107 | if (4 * sl->ref_list[0][j].parent->frame_num + |
108 | (sl->ref_list[0][j].reference & 3) == poc) { |
109 | int cur_ref = mbafi ? (j - 16) ^ field : j; |
110 | if (ref1->mbaff) |
111 | map[list][2 * old_ref + (rfield ^ field) + 16] = cur_ref; |
112 | if (rfield == field || !interl) |
113 | map[list][old_ref] = cur_ref; |
114 | break; |
115 | } |
116 | } |
117 | } |
118 | } |
119 | } |
120 | |
121 | void ff_h264_direct_ref_list_init(const H264Context *const h, H264SliceContext *sl) |
122 | { |
123 | H264Ref *const ref1 = &sl->ref_list[1][0]; |
124 | H264Picture *const cur = h->cur_pic_ptr; |
125 | int list, j, field; |
126 | int sidx = (h->picture_structure & 1) ^ 1; |
127 | int ref1sidx = (ref1->reference & 1) ^ 1; |
128 | |
129 | for (list = 0; list < sl->list_count; list++) { |
130 | cur->ref_count[sidx][list] = sl->ref_count[list]; |
131 | for (j = 0; j < sl->ref_count[list]; j++) |
132 | cur->ref_poc[sidx][list][j] = 4 * sl->ref_list[list][j].parent->frame_num + |
133 | (sl->ref_list[list][j].reference & 3); |
134 | } |
135 | |
136 | if (h->picture_structure == PICT_FRAME) { |
137 | memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0])); |
138 | memcpy(cur->ref_poc[1], cur->ref_poc[0], sizeof(cur->ref_poc[0])); |
139 | } |
140 | |
141 | if (h->current_slice == 0) { |
142 | cur->mbaff = FRAME_MBAFF(h); |
143 | } else { |
144 | av_assert0(cur->mbaff == FRAME_MBAFF(h)); |
145 | } |
146 | |
147 | sl->col_fieldoff = 0; |
148 | |
149 | if (sl->list_count != 2 || !sl->ref_count[1]) |
150 | return; |
151 | |
152 | if (h->picture_structure == PICT_FRAME) { |
153 | int cur_poc = h->cur_pic_ptr->poc; |
154 | int *col_poc = sl->ref_list[1][0].parent->field_poc; |
155 | if (col_poc[0] == INT_MAX && col_poc[1] == INT_MAX) { |
156 | av_log(h->avctx, AV_LOG_ERROR, "co located POCs unavailable\n"); |
157 | sl->col_parity = 1; |
158 | } else |
159 | sl->col_parity = (FFABS(col_poc[0] - cur_poc) >= |
160 | FFABS(col_poc[1] - cur_poc)); |
161 | ref1sidx = |
162 | sidx = sl->col_parity; |
163 | // FL -> FL & differ parity |
164 | } else if (!(h->picture_structure & sl->ref_list[1][0].reference) && |
165 | !sl->ref_list[1][0].parent->mbaff) { |
166 | sl->col_fieldoff = 2 * sl->ref_list[1][0].reference - 3; |
167 | } |
168 | |
169 | if (sl->slice_type_nos != AV_PICTURE_TYPE_B || sl->direct_spatial_mv_pred) |
170 | return; |
171 | |
172 | for (list = 0; list < 2; list++) { |
173 | fill_colmap(h, sl, sl->map_col_to_list0, list, sidx, ref1sidx, 0); |
174 | if (FRAME_MBAFF(h)) |
175 | for (field = 0; field < 2; field++) |
176 | fill_colmap(h, sl, sl->map_col_to_list0_field[field], list, field, |
177 | field, 1); |
178 | } |
179 | } |
180 | |
181 | static void await_reference_mb_row(const H264Context *const h, H264Ref *ref, |
182 | int mb_y) |
183 | { |
184 | int ref_field = ref->reference - 1; |
185 | int ref_field_picture = ref->parent->field_picture; |
186 | int ref_height = 16 * h->mb_height >> ref_field_picture; |
187 | |
188 | if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_FRAME)) |
189 | return; |
190 | |
191 | /* FIXME: It can be safe to access mb stuff |
192 | * even if pixels aren't deblocked yet. */ |
193 | |
194 | ff_thread_await_progress(&ref->parent->tf, |
195 | FFMIN(16 * mb_y >> ref_field_picture, |
196 | ref_height - 1), |
197 | ref_field_picture && ref_field); |
198 | } |
199 | |
200 | static void pred_spatial_direct_motion(const H264Context *const h, H264SliceContext *sl, |
201 | int *mb_type) |
202 | { |
203 | int b8_stride = 2; |
204 | int b4_stride = h->b_stride; |
205 | int mb_xy = sl->mb_xy, mb_y = sl->mb_y; |
206 | int mb_type_col[2]; |
207 | const int16_t (*l1mv0)[2], (*l1mv1)[2]; |
208 | const int8_t *l1ref0, *l1ref1; |
209 | const int is_b8x8 = IS_8X8(*mb_type); |
210 | unsigned int sub_mb_type = MB_TYPE_L0L1; |
211 | int i8, i4; |
212 | int ref[2]; |
213 | int mv[2]; |
214 | int list; |
215 | |
216 | assert(sl->ref_list[1][0].reference & 3); |
217 | |
218 | await_reference_mb_row(h, &sl->ref_list[1][0], |
219 | sl->mb_y + !!IS_INTERLACED(*mb_type)); |
220 | |
221 | #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \ |
222 | MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM) |
223 | |
224 | /* ref = min(neighbors) */ |
225 | for (list = 0; list < 2; list++) { |
226 | int left_ref = sl->ref_cache[list][scan8[0] - 1]; |
227 | int top_ref = sl->ref_cache[list][scan8[0] - 8]; |
228 | int refc = sl->ref_cache[list][scan8[0] - 8 + 4]; |
229 | const int16_t *C = sl->mv_cache[list][scan8[0] - 8 + 4]; |
230 | if (refc == PART_NOT_AVAILABLE) { |
231 | refc = sl->ref_cache[list][scan8[0] - 8 - 1]; |
232 | C = sl->mv_cache[list][scan8[0] - 8 - 1]; |
233 | } |
234 | ref[list] = FFMIN3((unsigned)left_ref, |
235 | (unsigned)top_ref, |
236 | (unsigned)refc); |
237 | if (ref[list] >= 0) { |
238 | /* This is just pred_motion() but with the cases removed that |
239 | * cannot happen for direct blocks. */ |
240 | const int16_t *const A = sl->mv_cache[list][scan8[0] - 1]; |
241 | const int16_t *const B = sl->mv_cache[list][scan8[0] - 8]; |
242 | |
243 | int match_count = (left_ref == ref[list]) + |
244 | (top_ref == ref[list]) + |
245 | (refc == ref[list]); |
246 | |
247 | if (match_count > 1) { // most common |
248 | mv[list] = pack16to32(mid_pred(A[0], B[0], C[0]), |
249 | mid_pred(A[1], B[1], C[1])); |
250 | } else { |
251 | assert(match_count == 1); |
252 | if (left_ref == ref[list]) |
253 | mv[list] = AV_RN32A(A); |
254 | else if (top_ref == ref[list]) |
255 | mv[list] = AV_RN32A(B); |
256 | else |
257 | mv[list] = AV_RN32A(C); |
258 | } |
259 | av_assert2(ref[list] < (sl->ref_count[list] << !!FRAME_MBAFF(h))); |
260 | } else { |
261 | int mask = ~(MB_TYPE_L0 << (2 * list)); |
262 | mv[list] = 0; |
263 | ref[list] = -1; |
264 | if (!is_b8x8) |
265 | *mb_type &= mask; |
266 | sub_mb_type &= mask; |
267 | } |
268 | } |
269 | if (ref[0] < 0 && ref[1] < 0) { |
270 | ref[0] = ref[1] = 0; |
271 | if (!is_b8x8) |
272 | *mb_type |= MB_TYPE_L0L1; |
273 | sub_mb_type |= MB_TYPE_L0L1; |
274 | } |
275 | |
276 | if (!(is_b8x8 | mv[0] | mv[1])) { |
277 | fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); |
278 | fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); |
279 | fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4); |
280 | fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4); |
281 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
282 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
283 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
284 | return; |
285 | } |
286 | |
287 | if (IS_INTERLACED(sl->ref_list[1][0].parent->mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL |
288 | if (!IS_INTERLACED(*mb_type)) { // AFR/FR -> AFL/FL |
289 | mb_y = (sl->mb_y & ~1) + sl->col_parity; |
290 | mb_xy = sl->mb_x + |
291 | ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride; |
292 | b8_stride = 0; |
293 | } else { |
294 | mb_y += sl->col_fieldoff; |
295 | mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity |
296 | } |
297 | goto single_col; |
298 | } else { // AFL/AFR/FR/FL -> AFR/FR |
299 | if (IS_INTERLACED(*mb_type)) { // AFL /FL -> AFR/FR |
300 | mb_y = sl->mb_y & ~1; |
301 | mb_xy = (sl->mb_y & ~1) * h->mb_stride + sl->mb_x; |
302 | mb_type_col[0] = sl->ref_list[1][0].parent->mb_type[mb_xy]; |
303 | mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy + h->mb_stride]; |
304 | b8_stride = 2 + 4 * h->mb_stride; |
305 | b4_stride *= 6; |
306 | if (IS_INTERLACED(mb_type_col[0]) != |
307 | IS_INTERLACED(mb_type_col[1])) { |
308 | mb_type_col[0] &= ~MB_TYPE_INTERLACED; |
309 | mb_type_col[1] &= ~MB_TYPE_INTERLACED; |
310 | } |
311 | |
312 | sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
313 | if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) && |
314 | (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) && |
315 | !is_b8x8) { |
316 | *mb_type |= MB_TYPE_16x8 | MB_TYPE_DIRECT2; /* B_16x8 */ |
317 | } else { |
318 | *mb_type |= MB_TYPE_8x8; |
319 | } |
320 | } else { // AFR/FR -> AFR/FR |
321 | single_col: |
322 | mb_type_col[0] = |
323 | mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy]; |
324 | |
325 | sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
326 | if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) { |
327 | *mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_16x16 */ |
328 | } else if (!is_b8x8 && |
329 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) { |
330 | *mb_type |= MB_TYPE_DIRECT2 | |
331 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16)); |
332 | } else { |
333 | if (!h->ps.sps->direct_8x8_inference_flag) { |
334 | /* FIXME: Save sub mb types from previous frames (or derive |
335 | * from MVs) so we know exactly what block size to use. */ |
336 | sub_mb_type += (MB_TYPE_8x8 - MB_TYPE_16x16); /* B_SUB_4x4 */ |
337 | } |
338 | *mb_type |= MB_TYPE_8x8; |
339 | } |
340 | } |
341 | } |
342 | |
343 | await_reference_mb_row(h, &sl->ref_list[1][0], mb_y); |
344 | |
345 | l1mv0 = (void*)&sl->ref_list[1][0].parent->motion_val[0][h->mb2b_xy[mb_xy]]; |
346 | l1mv1 = (void*)&sl->ref_list[1][0].parent->motion_val[1][h->mb2b_xy[mb_xy]]; |
347 | l1ref0 = &sl->ref_list[1][0].parent->ref_index[0][4 * mb_xy]; |
348 | l1ref1 = &sl->ref_list[1][0].parent->ref_index[1][4 * mb_xy]; |
349 | if (!b8_stride) { |
350 | if (sl->mb_y & 1) { |
351 | l1ref0 += 2; |
352 | l1ref1 += 2; |
353 | l1mv0 += 2 * b4_stride; |
354 | l1mv1 += 2 * b4_stride; |
355 | } |
356 | } |
357 | |
358 | if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) { |
359 | int n = 0; |
360 | for (i8 = 0; i8 < 4; i8++) { |
361 | int x8 = i8 & 1; |
362 | int y8 = i8 >> 1; |
363 | int xy8 = x8 + y8 * b8_stride; |
364 | int xy4 = x8 * 3 + y8 * b4_stride; |
365 | int a, b; |
366 | |
367 | if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8])) |
368 | continue; |
369 | sl->sub_mb_type[i8] = sub_mb_type; |
370 | |
371 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
372 | (uint8_t)ref[0], 1); |
373 | fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, |
374 | (uint8_t)ref[1], 1); |
375 | if (!IS_INTRA(mb_type_col[y8]) && !sl->ref_list[1][0].parent->long_ref && |
376 | ((l1ref0[xy8] == 0 && |
377 | FFABS(l1mv0[xy4][0]) <= 1 && |
378 | FFABS(l1mv0[xy4][1]) <= 1) || |
379 | (l1ref0[xy8] < 0 && |
380 | l1ref1[xy8] == 0 && |
381 | FFABS(l1mv1[xy4][0]) <= 1 && |
382 | FFABS(l1mv1[xy4][1]) <= 1))) { |
383 | a = |
384 | b = 0; |
385 | if (ref[0] > 0) |
386 | a = mv[0]; |
387 | if (ref[1] > 0) |
388 | b = mv[1]; |
389 | n++; |
390 | } else { |
391 | a = mv[0]; |
392 | b = mv[1]; |
393 | } |
394 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, a, 4); |
395 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, b, 4); |
396 | } |
397 | if (!is_b8x8 && !(n & 3)) |
398 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
399 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
400 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
401 | } else if (IS_16X16(*mb_type)) { |
402 | int a, b; |
403 | |
404 | fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); |
405 | fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); |
406 | if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].parent->long_ref && |
407 | ((l1ref0[0] == 0 && |
408 | FFABS(l1mv0[0][0]) <= 1 && |
409 | FFABS(l1mv0[0][1]) <= 1) || |
410 | (l1ref0[0] < 0 && !l1ref1[0] && |
411 | FFABS(l1mv1[0][0]) <= 1 && |
412 | FFABS(l1mv1[0][1]) <= 1 && |
413 | h->sei.unregistered.x264_build > 33U))) { |
414 | a = b = 0; |
415 | if (ref[0] > 0) |
416 | a = mv[0]; |
417 | if (ref[1] > 0) |
418 | b = mv[1]; |
419 | } else { |
420 | a = mv[0]; |
421 | b = mv[1]; |
422 | } |
423 | fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, a, 4); |
424 | fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, b, 4); |
425 | } else { |
426 | int n = 0; |
427 | for (i8 = 0; i8 < 4; i8++) { |
428 | const int x8 = i8 & 1; |
429 | const int y8 = i8 >> 1; |
430 | |
431 | if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8])) |
432 | continue; |
433 | sl->sub_mb_type[i8] = sub_mb_type; |
434 | |
435 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, mv[0], 4); |
436 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, mv[1], 4); |
437 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
438 | (uint8_t)ref[0], 1); |
439 | fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, |
440 | (uint8_t)ref[1], 1); |
441 | |
442 | assert(b8_stride == 2); |
443 | /* col_zero_flag */ |
444 | if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].parent->long_ref && |
445 | (l1ref0[i8] == 0 || |
446 | (l1ref0[i8] < 0 && |
447 | l1ref1[i8] == 0 && |
448 | h->sei.unregistered.x264_build > 33U))) { |
449 | const int16_t (*l1mv)[2] = l1ref0[i8] == 0 ? l1mv0 : l1mv1; |
450 | if (IS_SUB_8X8(sub_mb_type)) { |
451 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride]; |
452 | if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) { |
453 | if (ref[0] == 0) |
454 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, |
455 | 8, 0, 4); |
456 | if (ref[1] == 0) |
457 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, |
458 | 8, 0, 4); |
459 | n += 4; |
460 | } |
461 | } else { |
462 | int m = 0; |
463 | for (i4 = 0; i4 < 4; i4++) { |
464 | const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) + |
465 | (y8 * 2 + (i4 >> 1)) * b4_stride]; |
466 | if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) { |
467 | if (ref[0] == 0) |
468 | AV_ZERO32(sl->mv_cache[0][scan8[i8 * 4 + i4]]); |
469 | if (ref[1] == 0) |
470 | AV_ZERO32(sl->mv_cache[1][scan8[i8 * 4 + i4]]); |
471 | m++; |
472 | } |
473 | } |
474 | if (!(m & 3)) |
475 | sl->sub_mb_type[i8] += MB_TYPE_16x16 - MB_TYPE_8x8; |
476 | n += m; |
477 | } |
478 | } |
479 | } |
480 | if (!is_b8x8 && !(n & 15)) |
481 | *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 | |
482 | MB_TYPE_P1L0 | MB_TYPE_P1L1)) | |
483 | MB_TYPE_16x16 | MB_TYPE_DIRECT2; |
484 | } |
485 | } |
486 | |
487 | static void pred_temp_direct_motion(const H264Context *const h, H264SliceContext *sl, |
488 | int *mb_type) |
489 | { |
490 | int b8_stride = 2; |
491 | int b4_stride = h->b_stride; |
492 | int mb_xy = sl->mb_xy, mb_y = sl->mb_y; |
493 | int mb_type_col[2]; |
494 | const int16_t (*l1mv0)[2], (*l1mv1)[2]; |
495 | const int8_t *l1ref0, *l1ref1; |
496 | const int is_b8x8 = IS_8X8(*mb_type); |
497 | unsigned int sub_mb_type; |
498 | int i8, i4; |
499 | |
500 | assert(sl->ref_list[1][0].reference & 3); |
501 | |
502 | await_reference_mb_row(h, &sl->ref_list[1][0], |
503 | sl->mb_y + !!IS_INTERLACED(*mb_type)); |
504 | |
505 | if (IS_INTERLACED(sl->ref_list[1][0].parent->mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL |
506 | if (!IS_INTERLACED(*mb_type)) { // AFR/FR -> AFL/FL |
507 | mb_y = (sl->mb_y & ~1) + sl->col_parity; |
508 | mb_xy = sl->mb_x + |
509 | ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride; |
510 | b8_stride = 0; |
511 | } else { |
512 | mb_y += sl->col_fieldoff; |
513 | mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity |
514 | } |
515 | goto single_col; |
516 | } else { // AFL/AFR/FR/FL -> AFR/FR |
517 | if (IS_INTERLACED(*mb_type)) { // AFL /FL -> AFR/FR |
518 | mb_y = sl->mb_y & ~1; |
519 | mb_xy = sl->mb_x + (sl->mb_y & ~1) * h->mb_stride; |
520 | mb_type_col[0] = sl->ref_list[1][0].parent->mb_type[mb_xy]; |
521 | mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy + h->mb_stride]; |
522 | b8_stride = 2 + 4 * h->mb_stride; |
523 | b4_stride *= 6; |
524 | if (IS_INTERLACED(mb_type_col[0]) != |
525 | IS_INTERLACED(mb_type_col[1])) { |
526 | mb_type_col[0] &= ~MB_TYPE_INTERLACED; |
527 | mb_type_col[1] &= ~MB_TYPE_INTERLACED; |
528 | } |
529 | |
530 | sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
531 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
532 | |
533 | if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) && |
534 | (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) && |
535 | !is_b8x8) { |
536 | *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 | |
537 | MB_TYPE_DIRECT2; /* B_16x8 */ |
538 | } else { |
539 | *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1; |
540 | } |
541 | } else { // AFR/FR -> AFR/FR |
542 | single_col: |
543 | mb_type_col[0] = |
544 | mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy]; |
545 | |
546 | sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
547 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */ |
548 | if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) { |
549 | *mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
550 | MB_TYPE_DIRECT2; /* B_16x16 */ |
551 | } else if (!is_b8x8 && |
552 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) { |
553 | *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 | |
554 | (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16)); |
555 | } else { |
556 | if (!h->ps.sps->direct_8x8_inference_flag) { |
557 | /* FIXME: save sub mb types from previous frames (or derive |
558 | * from MVs) so we know exactly what block size to use */ |
559 | sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 | |
560 | MB_TYPE_DIRECT2; /* B_SUB_4x4 */ |
561 | } |
562 | *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1; |
563 | } |
564 | } |
565 | } |
566 | |
567 | await_reference_mb_row(h, &sl->ref_list[1][0], mb_y); |
568 | |
569 | l1mv0 = (void*)&sl->ref_list[1][0].parent->motion_val[0][h->mb2b_xy[mb_xy]]; |
570 | l1mv1 = (void*)&sl->ref_list[1][0].parent->motion_val[1][h->mb2b_xy[mb_xy]]; |
571 | l1ref0 = &sl->ref_list[1][0].parent->ref_index[0][4 * mb_xy]; |
572 | l1ref1 = &sl->ref_list[1][0].parent->ref_index[1][4 * mb_xy]; |
573 | if (!b8_stride) { |
574 | if (sl->mb_y & 1) { |
575 | l1ref0 += 2; |
576 | l1ref1 += 2; |
577 | l1mv0 += 2 * b4_stride; |
578 | l1mv1 += 2 * b4_stride; |
579 | } |
580 | } |
581 | |
582 | { |
583 | const int *map_col_to_list0[2] = { sl->map_col_to_list0[0], |
584 | sl->map_col_to_list0[1] }; |
585 | const int *dist_scale_factor = sl->dist_scale_factor; |
586 | int ref_offset; |
587 | |
588 | if (FRAME_MBAFF(h) && IS_INTERLACED(*mb_type)) { |
589 | map_col_to_list0[0] = sl->map_col_to_list0_field[sl->mb_y & 1][0]; |
590 | map_col_to_list0[1] = sl->map_col_to_list0_field[sl->mb_y & 1][1]; |
591 | dist_scale_factor = sl->dist_scale_factor_field[sl->mb_y & 1]; |
592 | } |
593 | ref_offset = (sl->ref_list[1][0].parent->mbaff << 4) & (mb_type_col[0] >> 3); |
594 | |
595 | if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) { |
596 | int y_shift = 2 * !IS_INTERLACED(*mb_type); |
597 | assert(h->ps.sps->direct_8x8_inference_flag); |
598 | |
599 | for (i8 = 0; i8 < 4; i8++) { |
600 | const int x8 = i8 & 1; |
601 | const int y8 = i8 >> 1; |
602 | int ref0, scale; |
603 | const int16_t (*l1mv)[2] = l1mv0; |
604 | |
605 | if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8])) |
606 | continue; |
607 | sl->sub_mb_type[i8] = sub_mb_type; |
608 | |
609 | fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
610 | if (IS_INTRA(mb_type_col[y8])) { |
611 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
612 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
613 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
614 | continue; |
615 | } |
616 | |
617 | ref0 = l1ref0[x8 + y8 * b8_stride]; |
618 | if (ref0 >= 0) |
619 | ref0 = map_col_to_list0[0][ref0 + ref_offset]; |
620 | else { |
621 | ref0 = map_col_to_list0[1][l1ref1[x8 + y8 * b8_stride] + |
622 | ref_offset]; |
623 | l1mv = l1mv1; |
624 | } |
625 | scale = dist_scale_factor[ref0]; |
626 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
627 | ref0, 1); |
628 | |
629 | { |
630 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * b4_stride]; |
631 | int my_col = (mv_col[1] * (1 << y_shift)) / 2; |
632 | int mx = (scale * mv_col[0] + 128) >> 8; |
633 | int my = (scale * my_col + 128) >> 8; |
634 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, |
635 | pack16to32(mx, my), 4); |
636 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, |
637 | pack16to32(mx - mv_col[0], my - my_col), 4); |
638 | } |
639 | } |
640 | return; |
641 | } |
642 | |
643 | /* one-to-one mv scaling */ |
644 | |
645 | if (IS_16X16(*mb_type)) { |
646 | int ref, mv0, mv1; |
647 | |
648 | fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1); |
649 | if (IS_INTRA(mb_type_col[0])) { |
650 | ref = mv0 = mv1 = 0; |
651 | } else { |
652 | const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset] |
653 | : map_col_to_list0[1][l1ref1[0] + ref_offset]; |
654 | const int scale = dist_scale_factor[ref0]; |
655 | const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0]; |
656 | int mv_l0[2]; |
657 | mv_l0[0] = (scale * mv_col[0] + 128) >> 8; |
658 | mv_l0[1] = (scale * mv_col[1] + 128) >> 8; |
659 | ref = ref0; |
660 | mv0 = pack16to32(mv_l0[0], mv_l0[1]); |
661 | mv1 = pack16to32(mv_l0[0] - mv_col[0], mv_l0[1] - mv_col[1]); |
662 | } |
663 | fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1); |
664 | fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4); |
665 | fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4); |
666 | } else { |
667 | for (i8 = 0; i8 < 4; i8++) { |
668 | const int x8 = i8 & 1; |
669 | const int y8 = i8 >> 1; |
670 | int ref0, scale; |
671 | const int16_t (*l1mv)[2] = l1mv0; |
672 | |
673 | if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8])) |
674 | continue; |
675 | sl->sub_mb_type[i8] = sub_mb_type; |
676 | fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
677 | if (IS_INTRA(mb_type_col[0])) { |
678 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1); |
679 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
680 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4); |
681 | continue; |
682 | } |
683 | |
684 | assert(b8_stride == 2); |
685 | ref0 = l1ref0[i8]; |
686 | if (ref0 >= 0) |
687 | ref0 = map_col_to_list0[0][ref0 + ref_offset]; |
688 | else { |
689 | ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset]; |
690 | l1mv = l1mv1; |
691 | } |
692 | scale = dist_scale_factor[ref0]; |
693 | |
694 | fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, |
695 | ref0, 1); |
696 | if (IS_SUB_8X8(sub_mb_type)) { |
697 | const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride]; |
698 | int mx = (scale * mv_col[0] + 128) >> 8; |
699 | int my = (scale * mv_col[1] + 128) >> 8; |
700 | fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, |
701 | pack16to32(mx, my), 4); |
702 | fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, |
703 | pack16to32(mx - mv_col[0], my - mv_col[1]), 4); |
704 | } else { |
705 | for (i4 = 0; i4 < 4; i4++) { |
706 | const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) + |
707 | (y8 * 2 + (i4 >> 1)) * b4_stride]; |
708 | int16_t *mv_l0 = sl->mv_cache[0][scan8[i8 * 4 + i4]]; |
709 | mv_l0[0] = (scale * mv_col[0] + 128) >> 8; |
710 | mv_l0[1] = (scale * mv_col[1] + 128) >> 8; |
711 | AV_WN32A(sl->mv_cache[1][scan8[i8 * 4 + i4]], |
712 | pack16to32(mv_l0[0] - mv_col[0], |
713 | mv_l0[1] - mv_col[1])); |
714 | } |
715 | } |
716 | } |
717 | } |
718 | } |
719 | } |
720 | |
721 | void ff_h264_pred_direct_motion(const H264Context *const h, H264SliceContext *sl, |
722 | int *mb_type) |
723 | { |
724 | if (sl->direct_spatial_mv_pred) |
725 | pred_spatial_direct_motion(h, sl, mb_type); |
726 | else |
727 | pred_temp_direct_motion(h, sl, mb_type); |
728 | } |
729 |