2 * H.26L/H.264/AVC/JVT/14496-10/... decoder
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
5 * This file is part of Libav.
7 * Libav 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.
12 * Libav 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.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * H.264 / AVC / MPEG4 part10 macroblock decoding
31 #include "libavutil/common.h"
32 #include "libavutil/intreadwrite.h"
39 static inline int get_lowest_part_list_y(H264Context *h, H264SliceContext *sl,
40 H264Picture *pic, int n,
41 int height, int y_offset, int list)
43 int raw_my = sl->mv_cache[list][scan8[n]][1];
44 int filter_height_up = (raw_my & 3) ? 2 : 0;
45 int filter_height_down = (raw_my & 3) ? 3 : 0;
46 int full_my = (raw_my >> 2) + y_offset;
47 int top = full_my - filter_height_up;
48 int bottom = full_my + filter_height_down + height;
50 return FFMAX(abs(top), bottom);
53 static inline void get_lowest_part_y(H264Context *h, H264SliceContext *sl,
54 int refs[2][48], int n,
55 int height, int y_offset, int list0,
56 int list1, int *nrefs)
60 y_offset += 16 * (h->mb_y >> MB_FIELD(h));
63 int ref_n = sl->ref_cache[0][scan8[n]];
64 H264Picture *ref = &sl->ref_list[0][ref_n];
66 // Error resilience puts the current picture in the ref list.
67 // Don't try to wait on these as it will cause a deadlock.
68 // Fields can wait on each other, though.
69 if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
70 (ref->reference & 3) != h->picture_structure) {
71 my = get_lowest_part_list_y(h, sl, ref, n, height, y_offset, 0);
72 if (refs[0][ref_n] < 0)
74 refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
79 int ref_n = sl->ref_cache[1][scan8[n]];
80 H264Picture *ref = &sl->ref_list[1][ref_n];
82 if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
83 (ref->reference & 3) != h->picture_structure) {
84 my = get_lowest_part_list_y(h, sl, ref, n, height, y_offset, 1);
85 if (refs[1][ref_n] < 0)
87 refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
93 * Wait until all reference frames are available for MC operations.
95 * @param h the H264 context
97 static void await_references(H264Context *h, H264SliceContext *sl)
99 const int mb_xy = h->mb_xy;
100 const int mb_type = h->cur_pic.mb_type[mb_xy];
102 int nrefs[2] = { 0 };
105 memset(refs, -1, sizeof(refs));
107 if (IS_16X16(mb_type)) {
108 get_lowest_part_y(h, sl, refs, 0, 16, 0,
109 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
110 } else if (IS_16X8(mb_type)) {
111 get_lowest_part_y(h, sl, refs, 0, 8, 0,
112 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
113 get_lowest_part_y(h, sl, refs, 8, 8, 8,
114 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
115 } else if (IS_8X16(mb_type)) {
116 get_lowest_part_y(h, sl, refs, 0, 16, 0,
117 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
118 get_lowest_part_y(h, sl, refs, 4, 16, 0,
119 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
123 assert(IS_8X8(mb_type));
125 for (i = 0; i < 4; i++) {
126 const int sub_mb_type = sl->sub_mb_type[i];
128 int y_offset = (i & 2) << 2;
130 if (IS_SUB_8X8(sub_mb_type)) {
131 get_lowest_part_y(h, sl, refs, n, 8, y_offset,
132 IS_DIR(sub_mb_type, 0, 0),
133 IS_DIR(sub_mb_type, 0, 1),
135 } else if (IS_SUB_8X4(sub_mb_type)) {
136 get_lowest_part_y(h, sl, refs, n, 4, y_offset,
137 IS_DIR(sub_mb_type, 0, 0),
138 IS_DIR(sub_mb_type, 0, 1),
140 get_lowest_part_y(h, sl, refs, n + 2, 4, y_offset + 4,
141 IS_DIR(sub_mb_type, 0, 0),
142 IS_DIR(sub_mb_type, 0, 1),
144 } else if (IS_SUB_4X8(sub_mb_type)) {
145 get_lowest_part_y(h, sl, refs, n, 8, y_offset,
146 IS_DIR(sub_mb_type, 0, 0),
147 IS_DIR(sub_mb_type, 0, 1),
149 get_lowest_part_y(h, sl, refs, n + 1, 8, y_offset,
150 IS_DIR(sub_mb_type, 0, 0),
151 IS_DIR(sub_mb_type, 0, 1),
155 assert(IS_SUB_4X4(sub_mb_type));
156 for (j = 0; j < 4; j++) {
157 int sub_y_offset = y_offset + 2 * (j & 2);
158 get_lowest_part_y(h, sl, refs, n + j, 4, sub_y_offset,
159 IS_DIR(sub_mb_type, 0, 0),
160 IS_DIR(sub_mb_type, 0, 1),
167 for (list = sl->list_count - 1; list >= 0; list--)
168 for (ref = 0; ref < 48 && nrefs[list]; ref++) {
169 int row = refs[list][ref];
171 H264Picture *ref_pic = &sl->ref_list[list][ref];
172 int ref_field = ref_pic->reference - 1;
173 int ref_field_picture = ref_pic->field_picture;
174 int pic_height = 16 * h->mb_height >> ref_field_picture;
179 if (!FIELD_PICTURE(h) && ref_field_picture) { // frame referencing two fields
180 ff_thread_await_progress(&ref_pic->tf,
181 FFMIN((row >> 1) - !(row & 1),
184 ff_thread_await_progress(&ref_pic->tf,
185 FFMIN((row >> 1), pic_height - 1),
187 } else if (FIELD_PICTURE(h) && !ref_field_picture) { // field referencing one field of a frame
188 ff_thread_await_progress(&ref_pic->tf,
189 FFMIN(row * 2 + ref_field,
192 } else if (FIELD_PICTURE(h)) {
193 ff_thread_await_progress(&ref_pic->tf,
194 FFMIN(row, pic_height - 1),
197 ff_thread_await_progress(&ref_pic->tf,
198 FFMIN(row, pic_height - 1),
205 static av_always_inline void mc_dir_part(H264Context *h, H264SliceContext *sl,
207 int n, int square, int height,
209 uint8_t *dest_y, uint8_t *dest_cb,
211 int src_x_offset, int src_y_offset,
212 qpel_mc_func *qpix_op,
213 h264_chroma_mc_func chroma_op,
214 int pixel_shift, int chroma_idc)
216 const int mx = sl->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
217 int my = sl->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
218 const int luma_xy = (mx & 3) + ((my & 3) << 2);
219 ptrdiff_t offset = ((mx >> 2) << pixel_shift) + (my >> 2) * sl->mb_linesize;
220 uint8_t *src_y = pic->f.data[0] + offset;
221 uint8_t *src_cb, *src_cr;
223 int extra_height = 0;
225 const int full_mx = mx >> 2;
226 const int full_my = my >> 2;
227 const int pic_width = 16 * h->mb_width;
228 const int pic_height = 16 * h->mb_height >> MB_FIELD(h);
236 if (full_mx < 0 - extra_width ||
237 full_my < 0 - extra_height ||
238 full_mx + 16 /*FIXME*/ > pic_width + extra_width ||
239 full_my + 16 /*FIXME*/ > pic_height + extra_height) {
240 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
241 src_y - (2 << pixel_shift) - 2 * sl->mb_linesize,
242 sl->mb_linesize, sl->mb_linesize,
243 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
244 full_my - 2, pic_width, pic_height);
245 src_y = h->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
249 qpix_op[luma_xy](dest_y, src_y, sl->mb_linesize); // FIXME try variable height perhaps?
251 qpix_op[luma_xy](dest_y + delta, src_y + delta, sl->mb_linesize);
253 if (CONFIG_GRAY && h->flags & CODEC_FLAG_GRAY)
256 if (chroma_idc == 3 /* yuv444 */) {
257 src_cb = pic->f.data[1] + offset;
259 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
260 src_cb - (2 << pixel_shift) - 2 * sl->mb_linesize,
261 sl->mb_linesize, sl->mb_linesize,
262 16 + 5, 16 + 5 /*FIXME*/,
263 full_mx - 2, full_my - 2,
264 pic_width, pic_height);
265 src_cb = h->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
267 qpix_op[luma_xy](dest_cb, src_cb, sl->mb_linesize); // FIXME try variable height perhaps?
269 qpix_op[luma_xy](dest_cb + delta, src_cb + delta, sl->mb_linesize);
271 src_cr = pic->f.data[2] + offset;
273 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
274 src_cr - (2 << pixel_shift) - 2 * sl->mb_linesize,
275 sl->mb_linesize, sl->mb_linesize,
276 16 + 5, 16 + 5 /*FIXME*/,
277 full_mx - 2, full_my - 2,
278 pic_width, pic_height);
279 src_cr = h->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
281 qpix_op[luma_xy](dest_cr, src_cr, sl->mb_linesize); // FIXME try variable height perhaps?
283 qpix_op[luma_xy](dest_cr + delta, src_cr + delta, sl->mb_linesize);
287 ysh = 3 - (chroma_idc == 2 /* yuv422 */);
288 if (chroma_idc == 1 /* yuv420 */ && MB_FIELD(h)) {
289 // chroma offset when predicting from a field of opposite parity
290 my += 2 * ((h->mb_y & 1) - (pic->reference - 1));
291 emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
294 src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
295 (my >> ysh) * sl->mb_uvlinesize;
296 src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
297 (my >> ysh) * sl->mb_uvlinesize;
300 h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cb,
301 sl->mb_uvlinesize, sl->mb_uvlinesize,
302 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
303 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
304 src_cb = h->edge_emu_buffer;
306 chroma_op(dest_cb, src_cb, sl->mb_uvlinesize,
307 height >> (chroma_idc == 1 /* yuv420 */),
308 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
311 h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cr,
312 sl->mb_uvlinesize, sl->mb_uvlinesize,
313 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
314 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
315 src_cr = h->edge_emu_buffer;
317 chroma_op(dest_cr, src_cr, sl->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
318 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
321 static av_always_inline void mc_part_std(H264Context *h, H264SliceContext *sl,
323 int height, int delta,
324 uint8_t *dest_y, uint8_t *dest_cb,
326 int x_offset, int y_offset,
327 qpel_mc_func *qpix_put,
328 h264_chroma_mc_func chroma_put,
329 qpel_mc_func *qpix_avg,
330 h264_chroma_mc_func chroma_avg,
331 int list0, int list1,
332 int pixel_shift, int chroma_idc)
334 qpel_mc_func *qpix_op = qpix_put;
335 h264_chroma_mc_func chroma_op = chroma_put;
337 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
338 if (chroma_idc == 3 /* yuv444 */) {
339 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
340 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
341 } else if (chroma_idc == 2 /* yuv422 */) {
342 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
343 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
344 } else { /* yuv420 */
345 dest_cb += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
346 dest_cr += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
348 x_offset += 8 * h->mb_x;
349 y_offset += 8 * (h->mb_y >> MB_FIELD(h));
352 H264Picture *ref = &sl->ref_list[0][sl->ref_cache[0][scan8[n]]];
353 mc_dir_part(h, sl, ref, n, square, height, delta, 0,
354 dest_y, dest_cb, dest_cr, x_offset, y_offset,
355 qpix_op, chroma_op, pixel_shift, chroma_idc);
358 chroma_op = chroma_avg;
362 H264Picture *ref = &sl->ref_list[1][sl->ref_cache[1][scan8[n]]];
363 mc_dir_part(h, sl, ref, n, square, height, delta, 1,
364 dest_y, dest_cb, dest_cr, x_offset, y_offset,
365 qpix_op, chroma_op, pixel_shift, chroma_idc);
369 static av_always_inline void mc_part_weighted(H264Context *h, H264SliceContext *sl,
371 int height, int delta,
372 uint8_t *dest_y, uint8_t *dest_cb,
374 int x_offset, int y_offset,
375 qpel_mc_func *qpix_put,
376 h264_chroma_mc_func chroma_put,
377 h264_weight_func luma_weight_op,
378 h264_weight_func chroma_weight_op,
379 h264_biweight_func luma_weight_avg,
380 h264_biweight_func chroma_weight_avg,
381 int list0, int list1,
382 int pixel_shift, int chroma_idc)
386 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
387 if (chroma_idc == 3 /* yuv444 */) {
388 chroma_height = height;
389 chroma_weight_avg = luma_weight_avg;
390 chroma_weight_op = luma_weight_op;
391 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
392 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
393 } else if (chroma_idc == 2 /* yuv422 */) {
394 chroma_height = height;
395 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
396 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
397 } else { /* yuv420 */
398 chroma_height = height >> 1;
399 dest_cb += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
400 dest_cr += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
402 x_offset += 8 * h->mb_x;
403 y_offset += 8 * (h->mb_y >> MB_FIELD(h));
405 if (list0 && list1) {
406 /* don't optimize for luma-only case, since B-frames usually
407 * use implicit weights => chroma too. */
408 uint8_t *tmp_cb = h->bipred_scratchpad;
409 uint8_t *tmp_cr = h->bipred_scratchpad + (16 << pixel_shift);
410 uint8_t *tmp_y = h->bipred_scratchpad + 16 * sl->mb_uvlinesize;
411 int refn0 = sl->ref_cache[0][scan8[n]];
412 int refn1 = sl->ref_cache[1][scan8[n]];
414 mc_dir_part(h, sl, &sl->ref_list[0][refn0], n, square, height, delta, 0,
415 dest_y, dest_cb, dest_cr,
416 x_offset, y_offset, qpix_put, chroma_put,
417 pixel_shift, chroma_idc);
418 mc_dir_part(h, sl, &sl->ref_list[1][refn1], n, square, height, delta, 1,
419 tmp_y, tmp_cb, tmp_cr,
420 x_offset, y_offset, qpix_put, chroma_put,
421 pixel_shift, chroma_idc);
423 if (sl->use_weight == 2) {
424 int weight0 = sl->implicit_weight[refn0][refn1][h->mb_y & 1];
425 int weight1 = 64 - weight0;
426 luma_weight_avg(dest_y, tmp_y, sl->mb_linesize,
427 height, 5, weight0, weight1, 0);
428 chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize,
429 chroma_height, 5, weight0, weight1, 0);
430 chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize,
431 chroma_height, 5, weight0, weight1, 0);
433 luma_weight_avg(dest_y, tmp_y, sl->mb_linesize, height,
434 sl->luma_log2_weight_denom,
435 sl->luma_weight[refn0][0][0],
436 sl->luma_weight[refn1][1][0],
437 sl->luma_weight[refn0][0][1] +
438 sl->luma_weight[refn1][1][1]);
439 chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize, chroma_height,
440 sl->chroma_log2_weight_denom,
441 sl->chroma_weight[refn0][0][0][0],
442 sl->chroma_weight[refn1][1][0][0],
443 sl->chroma_weight[refn0][0][0][1] +
444 sl->chroma_weight[refn1][1][0][1]);
445 chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize, chroma_height,
446 sl->chroma_log2_weight_denom,
447 sl->chroma_weight[refn0][0][1][0],
448 sl->chroma_weight[refn1][1][1][0],
449 sl->chroma_weight[refn0][0][1][1] +
450 sl->chroma_weight[refn1][1][1][1]);
453 int list = list1 ? 1 : 0;
454 int refn = sl->ref_cache[list][scan8[n]];
455 H264Picture *ref = &sl->ref_list[list][refn];
456 mc_dir_part(h, sl, ref, n, square, height, delta, list,
457 dest_y, dest_cb, dest_cr, x_offset, y_offset,
458 qpix_put, chroma_put, pixel_shift, chroma_idc);
460 luma_weight_op(dest_y, sl->mb_linesize, height,
461 sl->luma_log2_weight_denom,
462 sl->luma_weight[refn][list][0],
463 sl->luma_weight[refn][list][1]);
464 if (sl->use_weight_chroma) {
465 chroma_weight_op(dest_cb, sl->mb_uvlinesize, chroma_height,
466 sl->chroma_log2_weight_denom,
467 sl->chroma_weight[refn][list][0][0],
468 sl->chroma_weight[refn][list][0][1]);
469 chroma_weight_op(dest_cr, sl->mb_uvlinesize, chroma_height,
470 sl->chroma_log2_weight_denom,
471 sl->chroma_weight[refn][list][1][0],
472 sl->chroma_weight[refn][list][1][1]);
477 static av_always_inline void prefetch_motion(H264Context *h, H264SliceContext *sl,
478 int list, int pixel_shift,
481 /* fetch pixels for estimated mv 4 macroblocks ahead
482 * optimized for 64byte cache lines */
483 const int refn = sl->ref_cache[list][scan8[0]];
485 const int mx = (sl->mv_cache[list][scan8[0]][0] >> 2) + 16 * h->mb_x + 8;
486 const int my = (sl->mv_cache[list][scan8[0]][1] >> 2) + 16 * h->mb_y;
487 uint8_t **src = sl->ref_list[list][refn].f.data;
488 int off = (mx << pixel_shift) +
489 (my + (h->mb_x & 3) * 4) * sl->mb_linesize +
491 h->vdsp.prefetch(src[0] + off, h->linesize, 4);
492 if (chroma_idc == 3 /* yuv444 */) {
493 h->vdsp.prefetch(src[1] + off, h->linesize, 4);
494 h->vdsp.prefetch(src[2] + off, h->linesize, 4);
496 off = ((mx >> 1) << pixel_shift) +
497 ((my >> 1) + (h->mb_x & 7)) * h->uvlinesize +
499 h->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
504 static av_always_inline void xchg_mb_border(H264Context *h, H264SliceContext *sl,
506 uint8_t *src_cb, uint8_t *src_cr,
507 int linesize, int uvlinesize,
508 int xchg, int chroma444,
509 int simple, int pixel_shift)
514 uint8_t *top_border_m1;
517 if (!simple && FRAME_MBAFF(h)) {
522 top_idx = MB_MBAFF(h) ? 0 : 1;
526 if (sl->deblocking_filter == 2) {
527 deblock_topleft = h->slice_table[h->mb_xy - 1 - h->mb_stride] == sl->slice_num;
528 deblock_top = sl->top_type;
530 deblock_topleft = (h->mb_x > 0);
531 deblock_top = (h->mb_y > !!MB_FIELD(h));
534 src_y -= linesize + 1 + pixel_shift;
535 src_cb -= uvlinesize + 1 + pixel_shift;
536 src_cr -= uvlinesize + 1 + pixel_shift;
538 top_border_m1 = h->top_borders[top_idx][h->mb_x - 1];
539 top_border = h->top_borders[top_idx][h->mb_x];
541 #define XCHG(a, b, xchg) \
544 AV_SWAP64(b + 0, a + 0); \
545 AV_SWAP64(b + 8, a + 8); \
555 if (deblock_topleft) {
556 XCHG(top_border_m1 + (8 << pixel_shift),
557 src_y - (7 << pixel_shift), 1);
559 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
560 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
561 if (h->mb_x + 1 < h->mb_width) {
562 XCHG(h->top_borders[top_idx][h->mb_x + 1],
563 src_y + (17 << pixel_shift), 1);
566 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
569 if (deblock_topleft) {
570 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
571 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
573 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
574 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
575 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
576 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
577 if (h->mb_x + 1 < h->mb_width) {
578 XCHG(h->top_borders[top_idx][h->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
579 XCHG(h->top_borders[top_idx][h->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
584 if (deblock_topleft) {
585 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
586 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
588 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
589 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
595 static av_always_inline int dctcoef_get(int16_t *mb, int high_bit_depth,
598 if (high_bit_depth) {
599 return AV_RN32A(((int32_t *)mb) + index);
601 return AV_RN16A(mb + index);
604 static av_always_inline void dctcoef_set(int16_t *mb, int high_bit_depth,
605 int index, int value)
607 if (high_bit_depth) {
608 AV_WN32A(((int32_t *)mb) + index, value);
610 AV_WN16A(mb + index, value);
613 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
614 H264SliceContext *sl,
615 int mb_type, int is_h264,
617 int transform_bypass,
621 uint8_t *dest_y, int p)
623 void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
624 void (*idct_dc_add)(uint8_t *dst, int16_t *block, int stride);
626 int qscale = p == 0 ? sl->qscale : sl->chroma_qp[p - 1];
627 block_offset += 16 * p;
628 if (IS_INTRA4x4(mb_type)) {
629 if (IS_8x8DCT(mb_type)) {
630 if (transform_bypass) {
632 idct_add = h->h264dsp.h264_add_pixels8_clear;
634 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
635 idct_add = h->h264dsp.h264_idct8_add;
637 for (i = 0; i < 16; i += 4) {
638 uint8_t *const ptr = dest_y + block_offset[i];
639 const int dir = sl->intra4x4_pred_mode_cache[scan8[i]];
640 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
641 h->hpc.pred8x8l_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
643 const int nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
644 h->hpc.pred8x8l[dir](ptr, (sl->topleft_samples_available << i) & 0x8000,
645 (sl->topright_samples_available << i) & 0x4000, linesize);
647 if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
648 idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
650 idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
655 if (transform_bypass) {
657 idct_add = h->h264dsp.h264_add_pixels4_clear;
659 idct_dc_add = h->h264dsp.h264_idct_dc_add;
660 idct_add = h->h264dsp.h264_idct_add;
662 for (i = 0; i < 16; i++) {
663 uint8_t *const ptr = dest_y + block_offset[i];
664 const int dir = sl->intra4x4_pred_mode_cache[scan8[i]];
666 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
667 h->hpc.pred4x4_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
672 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
673 const int topright_avail = (sl->topright_samples_available << i) & 0x8000;
674 assert(h->mb_y || linesize <= block_offset[i]);
675 if (!topright_avail) {
677 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
678 topright = (uint8_t *)&tr_high;
680 tr = ptr[3 - linesize] * 0x01010101u;
681 topright = (uint8_t *)&tr;
684 topright = ptr + (4 << pixel_shift) - linesize;
688 h->hpc.pred4x4[dir](ptr, topright, linesize);
689 nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
692 if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
693 idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
695 idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
696 } else if (CONFIG_SVQ3_DECODER)
697 ff_svq3_add_idct_c(ptr, sl->mb + i * 16 + p * 256, linesize, qscale, 0);
703 h->hpc.pred16x16[sl->intra16x16_pred_mode](dest_y, linesize);
705 if (sl->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
706 if (!transform_bypass)
707 h->h264dsp.h264_luma_dc_dequant_idct(sl->mb + (p * 256 << pixel_shift),
709 h->dequant4_coeff[p][qscale][0]);
711 static const uint8_t dc_mapping[16] = {
712 0 * 16, 1 * 16, 4 * 16, 5 * 16,
713 2 * 16, 3 * 16, 6 * 16, 7 * 16,
714 8 * 16, 9 * 16, 12 * 16, 13 * 16,
715 10 * 16, 11 * 16, 14 * 16, 15 * 16
717 for (i = 0; i < 16; i++)
718 dctcoef_set(sl->mb + (p * 256 << pixel_shift),
719 pixel_shift, dc_mapping[i],
720 dctcoef_get(sl->mb_luma_dc[p],
724 } else if (CONFIG_SVQ3_DECODER)
725 ff_svq3_luma_dc_dequant_idct_c(sl->mb + p * 256,
726 sl->mb_luma_dc[p], qscale);
730 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, H264SliceContext *sl,
732 int is_h264, int simple,
733 int transform_bypass,
737 uint8_t *dest_y, int p)
739 void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
741 block_offset += 16 * p;
742 if (!IS_INTRA4x4(mb_type)) {
744 if (IS_INTRA16x16(mb_type)) {
745 if (transform_bypass) {
746 if (h->sps.profile_idc == 244 &&
747 (sl->intra16x16_pred_mode == VERT_PRED8x8 ||
748 sl->intra16x16_pred_mode == HOR_PRED8x8)) {
749 h->hpc.pred16x16_add[sl->intra16x16_pred_mode](dest_y, block_offset,
750 sl->mb + (p * 256 << pixel_shift),
753 for (i = 0; i < 16; i++)
754 if (sl->non_zero_count_cache[scan8[i + p * 16]] ||
755 dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
756 h->h264dsp.h264_add_pixels4_clear(dest_y + block_offset[i],
757 sl->mb + (i * 16 + p * 256 << pixel_shift),
761 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
762 sl->mb + (p * 256 << pixel_shift),
764 sl->non_zero_count_cache + p * 5 * 8);
766 } else if (sl->cbp & 15) {
767 if (transform_bypass) {
768 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
769 idct_add = IS_8x8DCT(mb_type) ? h->h264dsp.h264_add_pixels8_clear
770 : h->h264dsp.h264_add_pixels4_clear;
771 for (i = 0; i < 16; i += di)
772 if (sl->non_zero_count_cache[scan8[i + p * 16]])
773 idct_add(dest_y + block_offset[i],
774 sl->mb + (i * 16 + p * 256 << pixel_shift),
777 if (IS_8x8DCT(mb_type))
778 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
779 sl->mb + (p * 256 << pixel_shift),
781 sl->non_zero_count_cache + p * 5 * 8);
783 h->h264dsp.h264_idct_add16(dest_y, block_offset,
784 sl->mb + (p * 256 << pixel_shift),
786 sl->non_zero_count_cache + p * 5 * 8);
789 } else if (CONFIG_SVQ3_DECODER) {
790 for (i = 0; i < 16; i++)
791 if (sl->non_zero_count_cache[scan8[i + p * 16]] || sl->mb[i * 16 + p * 256]) {
792 // FIXME benchmark weird rule, & below
793 uint8_t *const ptr = dest_y + block_offset[i];
794 ff_svq3_add_idct_c(ptr, sl->mb + i * 16 + p * 256, linesize,
795 sl->qscale, IS_INTRA(mb_type) ? 1 : 0);
803 #include "h264_mb_template.c"
807 #include "h264_mb_template.c"
811 #include "h264_mb_template.c"
813 void ff_h264_hl_decode_mb(H264Context *h, H264SliceContext *sl)
815 const int mb_xy = h->mb_xy;
816 const int mb_type = h->cur_pic.mb_type[mb_xy];
817 int is_complex = CONFIG_SMALL || h->is_complex ||
818 IS_INTRA_PCM(mb_type) || sl->qscale == 0;
821 if (is_complex || h->pixel_shift)
822 hl_decode_mb_444_complex(h, sl);
824 hl_decode_mb_444_simple_8(h, sl);
825 } else if (is_complex) {
826 hl_decode_mb_complex(h, sl);
827 } else if (h->pixel_shift) {
828 hl_decode_mb_simple_16(h, sl);
830 hl_decode_mb_simple_8(h, sl);