xref: /third_party/ffmpeg/libavcodec/h264_mb.c (revision cabdff1a)
1/*
2 * H.26L/H.264/AVC/JVT/14496-10/... decoder
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 macroblock decoding
25 */
26
27#include <stdint.h>
28
29#include "config.h"
30
31#include "libavutil/common.h"
32#include "libavutil/intreadwrite.h"
33#include "avcodec.h"
34#include "h264dec.h"
35#include "h264_ps.h"
36#include "qpeldsp.h"
37#include "threadframe.h"
38
39static inline int get_lowest_part_list_y(H264SliceContext *sl,
40                                         int n, int height, int y_offset, int list)
41{
42    int raw_my             = sl->mv_cache[list][scan8[n]][1];
43    int filter_height_down = (raw_my & 3) ? 3 : 0;
44    int full_my            = (raw_my >> 2) + y_offset;
45    int bottom             = full_my + filter_height_down + height;
46
47    av_assert2(height >= 0);
48
49    return FFMAX(0, bottom);
50}
51
52static inline void get_lowest_part_y(const H264Context *h, H264SliceContext *sl,
53                                     int16_t refs[2][48], int n,
54                                     int height, int y_offset, int list0,
55                                     int list1, int *nrefs)
56{
57    int my;
58
59    y_offset += 16 * (sl->mb_y >> MB_FIELD(sl));
60
61    if (list0) {
62        int ref_n = sl->ref_cache[0][scan8[n]];
63        H264Ref *ref = &sl->ref_list[0][ref_n];
64
65        // Error resilience puts the current picture in the ref list.
66        // Don't try to wait on these as it will cause a deadlock.
67        // Fields can wait on each other, though.
68        if (ref->parent->tf.progress->data != h->cur_pic.tf.progress->data ||
69            (ref->reference & 3) != h->picture_structure) {
70            my = get_lowest_part_list_y(sl, n, height, y_offset, 0);
71            if (refs[0][ref_n] < 0)
72                nrefs[0] += 1;
73            refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
74        }
75    }
76
77    if (list1) {
78        int ref_n    = sl->ref_cache[1][scan8[n]];
79        H264Ref *ref = &sl->ref_list[1][ref_n];
80
81        if (ref->parent->tf.progress->data != h->cur_pic.tf.progress->data ||
82            (ref->reference & 3) != h->picture_structure) {
83            my = get_lowest_part_list_y(sl, n, height, y_offset, 1);
84            if (refs[1][ref_n] < 0)
85                nrefs[1] += 1;
86            refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
87        }
88    }
89}
90
91/**
92 * Wait until all reference frames are available for MC operations.
93 *
94 * @param h the H.264 context
95 */
96static void await_references(const H264Context *h, H264SliceContext *sl)
97{
98    const int mb_xy   = sl->mb_xy;
99    const int mb_type = h->cur_pic.mb_type[mb_xy];
100    int16_t refs[2][48];
101    int nrefs[2] = { 0 };
102    int ref, list;
103
104    memset(refs, -1, sizeof(refs));
105
106    if (IS_16X16(mb_type)) {
107        get_lowest_part_y(h, sl, refs, 0, 16, 0,
108                          IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
109    } else if (IS_16X8(mb_type)) {
110        get_lowest_part_y(h, sl, refs, 0, 8, 0,
111                          IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
112        get_lowest_part_y(h, sl, refs, 8, 8, 8,
113                          IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
114    } else if (IS_8X16(mb_type)) {
115        get_lowest_part_y(h, sl, refs, 0, 16, 0,
116                          IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
117        get_lowest_part_y(h, sl, refs, 4, 16, 0,
118                          IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
119    } else {
120        int i;
121
122        av_assert2(IS_8X8(mb_type));
123
124        for (i = 0; i < 4; i++) {
125            const int sub_mb_type = sl->sub_mb_type[i];
126            const int n           = 4 * i;
127            int y_offset          = (i & 2) << 2;
128
129            if (IS_SUB_8X8(sub_mb_type)) {
130                get_lowest_part_y(h, sl, refs, n, 8, y_offset,
131                                  IS_DIR(sub_mb_type, 0, 0),
132                                  IS_DIR(sub_mb_type, 0, 1),
133                                  nrefs);
134            } else if (IS_SUB_8X4(sub_mb_type)) {
135                get_lowest_part_y(h, sl, refs, n, 4, y_offset,
136                                  IS_DIR(sub_mb_type, 0, 0),
137                                  IS_DIR(sub_mb_type, 0, 1),
138                                  nrefs);
139                get_lowest_part_y(h, sl, refs, n + 2, 4, y_offset + 4,
140                                  IS_DIR(sub_mb_type, 0, 0),
141                                  IS_DIR(sub_mb_type, 0, 1),
142                                  nrefs);
143            } else if (IS_SUB_4X8(sub_mb_type)) {
144                get_lowest_part_y(h, sl, refs, n, 8, y_offset,
145                                  IS_DIR(sub_mb_type, 0, 0),
146                                  IS_DIR(sub_mb_type, 0, 1),
147                                  nrefs);
148                get_lowest_part_y(h, sl, refs, n + 1, 8, y_offset,
149                                  IS_DIR(sub_mb_type, 0, 0),
150                                  IS_DIR(sub_mb_type, 0, 1),
151                                  nrefs);
152            } else {
153                int j;
154                av_assert2(IS_SUB_4X4(sub_mb_type));
155                for (j = 0; j < 4; j++) {
156                    int sub_y_offset = y_offset + 2 * (j & 2);
157                    get_lowest_part_y(h, sl, refs, n + j, 4, sub_y_offset,
158                                      IS_DIR(sub_mb_type, 0, 0),
159                                      IS_DIR(sub_mb_type, 0, 1),
160                                      nrefs);
161                }
162            }
163        }
164    }
165
166    for (list = sl->list_count - 1; list >= 0; list--)
167        for (ref = 0; ref < 48 && nrefs[list]; ref++) {
168            int row = refs[list][ref];
169            if (row >= 0) {
170                H264Ref *ref_pic  = &sl->ref_list[list][ref];
171                int ref_field         = ref_pic->reference - 1;
172                int ref_field_picture = ref_pic->parent->field_picture;
173                int pic_height        = 16 * h->mb_height >> ref_field_picture;
174
175                row <<= MB_MBAFF(sl);
176                nrefs[list]--;
177
178                if (!FIELD_PICTURE(h) && ref_field_picture) { // frame referencing two fields
179                    av_assert2((ref_pic->parent->reference & 3) == 3);
180                    ff_thread_await_progress(&ref_pic->parent->tf,
181                                             FFMIN((row >> 1) - !(row & 1),
182                                                   pic_height - 1),
183                                             1);
184                    ff_thread_await_progress(&ref_pic->parent->tf,
185                                             FFMIN((row >> 1), pic_height - 1),
186                                             0);
187                } else if (FIELD_PICTURE(h) && !ref_field_picture) { // field referencing one field of a frame
188                    ff_thread_await_progress(&ref_pic->parent->tf,
189                                             FFMIN(row * 2 + ref_field,
190                                                   pic_height - 1),
191                                             0);
192                } else if (FIELD_PICTURE(h)) {
193                    ff_thread_await_progress(&ref_pic->parent->tf,
194                                             FFMIN(row, pic_height - 1),
195                                             ref_field);
196                } else {
197                    ff_thread_await_progress(&ref_pic->parent->tf,
198                                             FFMIN(row, pic_height - 1),
199                                             0);
200                }
201            }
202        }
203}
204
205static av_always_inline void mc_dir_part(const H264Context *h, H264SliceContext *sl,
206                                         H264Ref *pic,
207                                         int n, int square, int height,
208                                         int delta, int list,
209                                         uint8_t *dest_y, uint8_t *dest_cb,
210                                         uint8_t *dest_cr,
211                                         int src_x_offset, int src_y_offset,
212                                         const qpel_mc_func *qpix_op,
213                                         h264_chroma_mc_func chroma_op,
214                                         int pixel_shift, int chroma_idc)
215{
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) * (1 << pixel_shift) + (my >> 2) * sl->mb_linesize;
220    uint8_t *src_y    = pic->data[0] + offset;
221    uint8_t *src_cb, *src_cr;
222    int extra_width  = 0;
223    int extra_height = 0;
224    int emu = 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(sl);
229    int ysh;
230
231    if (mx & 7)
232        extra_width -= 3;
233    if (my & 7)
234        extra_height -= 3;
235
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(sl->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 = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
246        emu   = 1;
247    }
248
249    qpix_op[luma_xy](dest_y, src_y, sl->mb_linesize); // FIXME try variable height perhaps?
250    if (!square)
251        qpix_op[luma_xy](dest_y + delta, src_y + delta, sl->mb_linesize);
252
253    if (CONFIG_GRAY && h->flags & AV_CODEC_FLAG_GRAY)
254        return;
255
256    if (chroma_idc == 3 /* yuv444 */) {
257        src_cb = pic->data[1] + offset;
258        if (emu) {
259            h->vdsp.emulated_edge_mc(sl->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 = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
266        }
267        qpix_op[luma_xy](dest_cb, src_cb, sl->mb_linesize); // FIXME try variable height perhaps?
268        if (!square)
269            qpix_op[luma_xy](dest_cb + delta, src_cb + delta, sl->mb_linesize);
270
271        src_cr = pic->data[2] + offset;
272        if (emu) {
273            h->vdsp.emulated_edge_mc(sl->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 = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
280        }
281        qpix_op[luma_xy](dest_cr, src_cr, sl->mb_linesize); // FIXME try variable height perhaps?
282        if (!square)
283            qpix_op[luma_xy](dest_cr + delta, src_cr + delta, sl->mb_linesize);
284        return;
285    }
286
287    ysh = 3 - (chroma_idc == 2 /* yuv422 */);
288    if (chroma_idc == 1 /* yuv420 */ && MB_FIELD(sl)) {
289        // chroma offset when predicting from a field of opposite parity
290        my  += 2 * ((sl->mb_y & 1) - (pic->reference - 1));
291        emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
292    }
293
294    src_cb = pic->data[1] + ((mx >> 3) * (1 << pixel_shift)) +
295             (my >> ysh) * sl->mb_uvlinesize;
296    src_cr = pic->data[2] + ((mx >> 3) * (1 << pixel_shift)) +
297             (my >> ysh) * sl->mb_uvlinesize;
298
299    if (emu) {
300        h->vdsp.emulated_edge_mc(sl->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 = sl->edge_emu_buffer;
305    }
306    chroma_op(dest_cb, src_cb, sl->mb_uvlinesize,
307              height >> (chroma_idc == 1 /* yuv420 */),
308              mx & 7, ((unsigned)my << (chroma_idc == 2 /* yuv422 */)) & 7);
309
310    if (emu) {
311        h->vdsp.emulated_edge_mc(sl->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 = sl->edge_emu_buffer;
316    }
317    chroma_op(dest_cr, src_cr, sl->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
318              mx & 7, ((unsigned)my << (chroma_idc == 2 /* yuv422 */)) & 7);
319}
320
321static av_always_inline void mc_part_std(const H264Context *h, H264SliceContext *sl,
322                                         int n, int square,
323                                         int height, int delta,
324                                         uint8_t *dest_y, uint8_t *dest_cb,
325                                         uint8_t *dest_cr,
326                                         int x_offset, int y_offset,
327                                         const qpel_mc_func *qpix_put,
328                                         h264_chroma_mc_func chroma_put,
329                                         const qpel_mc_func *qpix_avg,
330                                         h264_chroma_mc_func chroma_avg,
331                                         int list0, int list1,
332                                         int pixel_shift, int chroma_idc)
333{
334    const qpel_mc_func *qpix_op   = qpix_put;
335    h264_chroma_mc_func chroma_op = chroma_put;
336
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;
347    }
348    x_offset += 8 * sl->mb_x;
349    y_offset += 8 * (sl->mb_y >> MB_FIELD(sl));
350
351    if (list0) {
352        H264Ref *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);
356
357        qpix_op   = qpix_avg;
358        chroma_op = chroma_avg;
359    }
360
361    if (list1) {
362        H264Ref *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);
366    }
367}
368
369static av_always_inline void mc_part_weighted(const H264Context *h, H264SliceContext *sl,
370                                              int n, int square,
371                                              int height, int delta,
372                                              uint8_t *dest_y, uint8_t *dest_cb,
373                                              uint8_t *dest_cr,
374                                              int x_offset, int y_offset,
375                                              const 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)
383{
384    int chroma_height;
385
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;
401    }
402    x_offset += 8 * sl->mb_x;
403    y_offset += 8 * (sl->mb_y >> MB_FIELD(sl));
404
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 = sl->bipred_scratchpad;
409        uint8_t *tmp_cr = sl->bipred_scratchpad + (16 << pixel_shift);
410        uint8_t *tmp_y  = sl->bipred_scratchpad + 16 * sl->mb_uvlinesize;
411        int refn0       = sl->ref_cache[0][scan8[n]];
412        int refn1       = sl->ref_cache[1][scan8[n]];
413
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);
422
423        if (sl->pwt.use_weight == 2) {
424            int weight0 = sl->pwt.implicit_weight[refn0][refn1][sl->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            if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
429                chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize,
430                                  chroma_height, 5, weight0, weight1, 0);
431                chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize,
432                                  chroma_height, 5, weight0, weight1, 0);
433            }
434        } else {
435            luma_weight_avg(dest_y, tmp_y, sl->mb_linesize, height,
436                            sl->pwt.luma_log2_weight_denom,
437                            sl->pwt.luma_weight[refn0][0][0],
438                            sl->pwt.luma_weight[refn1][1][0],
439                            sl->pwt.luma_weight[refn0][0][1] +
440                            sl->pwt.luma_weight[refn1][1][1]);
441            if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
442                chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize, chroma_height,
443                                  sl->pwt.chroma_log2_weight_denom,
444                                  sl->pwt.chroma_weight[refn0][0][0][0],
445                                  sl->pwt.chroma_weight[refn1][1][0][0],
446                                  sl->pwt.chroma_weight[refn0][0][0][1] +
447                                  sl->pwt.chroma_weight[refn1][1][0][1]);
448                chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize, chroma_height,
449                                  sl->pwt.chroma_log2_weight_denom,
450                                  sl->pwt.chroma_weight[refn0][0][1][0],
451                                  sl->pwt.chroma_weight[refn1][1][1][0],
452                                  sl->pwt.chroma_weight[refn0][0][1][1] +
453                                  sl->pwt.chroma_weight[refn1][1][1][1]);
454            }
455        }
456    } else {
457        int list     = list1 ? 1 : 0;
458        int refn     = sl->ref_cache[list][scan8[n]];
459        H264Ref *ref = &sl->ref_list[list][refn];
460        mc_dir_part(h, sl, ref, n, square, height, delta, list,
461                    dest_y, dest_cb, dest_cr, x_offset, y_offset,
462                    qpix_put, chroma_put, pixel_shift, chroma_idc);
463
464        luma_weight_op(dest_y, sl->mb_linesize, height,
465                       sl->pwt.luma_log2_weight_denom,
466                       sl->pwt.luma_weight[refn][list][0],
467                       sl->pwt.luma_weight[refn][list][1]);
468        if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
469            if (sl->pwt.use_weight_chroma) {
470                chroma_weight_op(dest_cb, sl->mb_uvlinesize, chroma_height,
471                                 sl->pwt.chroma_log2_weight_denom,
472                                 sl->pwt.chroma_weight[refn][list][0][0],
473                                 sl->pwt.chroma_weight[refn][list][0][1]);
474                chroma_weight_op(dest_cr, sl->mb_uvlinesize, chroma_height,
475                                 sl->pwt.chroma_log2_weight_denom,
476                                 sl->pwt.chroma_weight[refn][list][1][0],
477                                 sl->pwt.chroma_weight[refn][list][1][1]);
478            }
479        }
480    }
481}
482
483static av_always_inline void prefetch_motion(const H264Context *h, H264SliceContext *sl,
484                                             int list, int pixel_shift,
485                                             int chroma_idc)
486{
487    /* fetch pixels for estimated mv 4 macroblocks ahead
488     * optimized for 64byte cache lines */
489    const int refn = sl->ref_cache[list][scan8[0]];
490    if (refn >= 0) {
491        const int mx  = (sl->mv_cache[list][scan8[0]][0] >> 2) + 16 * sl->mb_x + 8;
492        const int my  = (sl->mv_cache[list][scan8[0]][1] >> 2) + 16 * sl->mb_y;
493        uint8_t **src = sl->ref_list[list][refn].data;
494        int off       =  mx * (1<< pixel_shift) +
495                        (my + (sl->mb_x & 3) * 4) * sl->mb_linesize +
496                        (64 << pixel_shift);
497        h->vdsp.prefetch(src[0] + off, sl->linesize, 4);
498        if (chroma_idc == 3 /* yuv444 */) {
499            h->vdsp.prefetch(src[1] + off, sl->linesize, 4);
500            h->vdsp.prefetch(src[2] + off, sl->linesize, 4);
501        } else {
502            off= ((mx>>1)+64) * (1<<pixel_shift) + ((my>>1) + (sl->mb_x&7))*sl->uvlinesize;
503            h->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
504        }
505    }
506}
507
508static av_always_inline void xchg_mb_border(const H264Context *h, H264SliceContext *sl,
509                                            uint8_t *src_y,
510                                            uint8_t *src_cb, uint8_t *src_cr,
511                                            int linesize, int uvlinesize,
512                                            int xchg, int chroma444,
513                                            int simple, int pixel_shift)
514{
515    int deblock_topleft;
516    int deblock_top;
517    int top_idx = 1;
518    uint8_t *top_border_m1;
519    uint8_t *top_border;
520
521    if (!simple && FRAME_MBAFF(h)) {
522        if (sl->mb_y & 1) {
523            if (!MB_MBAFF(sl))
524                return;
525        } else {
526            top_idx = MB_MBAFF(sl) ? 0 : 1;
527        }
528    }
529
530    if (sl->deblocking_filter == 2) {
531        deblock_topleft = h->slice_table[sl->mb_xy - 1 - h->mb_stride] == sl->slice_num;
532        deblock_top     = sl->top_type;
533    } else {
534        deblock_topleft = (sl->mb_x > 0);
535        deblock_top     = (sl->mb_y > !!MB_FIELD(sl));
536    }
537
538    src_y  -= linesize   + 1 + pixel_shift;
539    src_cb -= uvlinesize + 1 + pixel_shift;
540    src_cr -= uvlinesize + 1 + pixel_shift;
541
542    top_border_m1 = sl->top_borders[top_idx][sl->mb_x - 1];
543    top_border    = sl->top_borders[top_idx][sl->mb_x];
544
545#define XCHG(a, b, xchg)                        \
546    if (pixel_shift) {                          \
547        if (xchg) {                             \
548            AV_SWAP64(b + 0, a + 0);            \
549            AV_SWAP64(b + 8, a + 8);            \
550        } else {                                \
551            AV_COPY128(b, a);                   \
552        }                                       \
553    } else if (xchg)                            \
554        AV_SWAP64(b, a);                        \
555    else                                        \
556        AV_COPY64(b, a);
557
558    if (deblock_top) {
559        if (deblock_topleft) {
560            XCHG(top_border_m1 + (8 << pixel_shift),
561                 src_y - (7 << pixel_shift), 1);
562        }
563        XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
564        XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
565        if (sl->mb_x + 1 < h->mb_width) {
566            XCHG(sl->top_borders[top_idx][sl->mb_x + 1],
567                 src_y + (17 << pixel_shift), 1);
568        }
569        if (simple || !CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
570            if (chroma444) {
571                if (deblock_topleft) {
572                    XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
573                    XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
574                }
575                XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
576                XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
577                XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
578                XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
579                if (sl->mb_x + 1 < h->mb_width) {
580                    XCHG(sl->top_borders[top_idx][sl->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
581                    XCHG(sl->top_borders[top_idx][sl->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
582                }
583            } else {
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);
587                }
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);
590            }
591        }
592    }
593}
594
595static av_always_inline int dctcoef_get(int16_t *mb, int high_bit_depth,
596                                        int index)
597{
598    if (high_bit_depth) {
599        return AV_RN32A(((int32_t *)mb) + index);
600    } else
601        return AV_RN16A(mb + index);
602}
603
604static av_always_inline void dctcoef_set(int16_t *mb, int high_bit_depth,
605                                         int index, int value)
606{
607    if (high_bit_depth) {
608        AV_WN32A(((int32_t *)mb) + index, value);
609    } else
610        AV_WN16A(mb + index, value);
611}
612
613static av_always_inline void hl_decode_mb_predict_luma(const H264Context *h,
614                                                       H264SliceContext *sl,
615                                                       int mb_type, int simple,
616                                                       int transform_bypass,
617                                                       int pixel_shift,
618                                                       const int *block_offset,
619                                                       int linesize,
620                                                       uint8_t *dest_y, int p)
621{
622    void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
623    void (*idct_dc_add)(uint8_t *dst, int16_t *block, int stride);
624    int i;
625    int qscale = p == 0 ? sl->qscale : sl->chroma_qp[p - 1];
626    block_offset += 16 * p;
627    if (IS_INTRA4x4(mb_type)) {
628        if (IS_8x8DCT(mb_type)) {
629            if (transform_bypass) {
630                idct_dc_add =
631                idct_add    = h->h264dsp.h264_add_pixels8_clear;
632            } else {
633                idct_dc_add = h->h264dsp.h264_idct8_dc_add;
634                idct_add    = h->h264dsp.h264_idct8_add;
635            }
636            for (i = 0; i < 16; i += 4) {
637                uint8_t *const ptr = dest_y + block_offset[i];
638                const int dir      = sl->intra4x4_pred_mode_cache[scan8[i]];
639                if (transform_bypass && h->ps.sps->profile_idc == 244 && dir <= 1) {
640                    if (h->x264_build < 151U) {
641                        h->hpc.pred8x8l_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
642                    } else
643                        h->hpc.pred8x8l_filter_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift),
644                                                        (sl-> topleft_samples_available << i) & 0x8000,
645                                                        (sl->topright_samples_available << i) & 0x4000, linesize);
646                } else {
647                    const int nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
648                    h->hpc.pred8x8l[dir](ptr, (sl->topleft_samples_available << i) & 0x8000,
649                                         (sl->topright_samples_available << i) & 0x4000, linesize);
650                    if (nnz) {
651                        if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
652                            idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
653                        else
654                            idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
655                    }
656                }
657            }
658        } else {
659            if (transform_bypass) {
660                idct_dc_add  =
661                idct_add     = h->h264dsp.h264_add_pixels4_clear;
662            } else {
663                idct_dc_add = h->h264dsp.h264_idct_dc_add;
664                idct_add    = h->h264dsp.h264_idct_add;
665            }
666            for (i = 0; i < 16; i++) {
667                uint8_t *const ptr = dest_y + block_offset[i];
668                const int dir      = sl->intra4x4_pred_mode_cache[scan8[i]];
669
670                if (transform_bypass && h->ps.sps->profile_idc == 244 && dir <= 1) {
671                    h->hpc.pred4x4_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
672                } else {
673                    uint8_t *topright;
674                    int nnz, tr;
675                    uint64_t tr_high;
676                    if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
677                        const int topright_avail = (sl->topright_samples_available << i) & 0x8000;
678                        av_assert2(sl->mb_y || linesize <= block_offset[i]);
679                        if (!topright_avail) {
680                            if (pixel_shift) {
681                                tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
682                                topright = (uint8_t *)&tr_high;
683                            } else {
684                                tr       = ptr[3 - linesize] * 0x01010101u;
685                                topright = (uint8_t *)&tr;
686                            }
687                        } else
688                            topright = ptr + (4 << pixel_shift) - linesize;
689                    } else
690                        topright = NULL;
691
692                    h->hpc.pred4x4[dir](ptr, topright, linesize);
693                    nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
694                    if (nnz) {
695                        if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
696                            idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
697                        else
698                            idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
699                    }
700                }
701            }
702        }
703    } else {
704        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),
708                                                     sl->mb_luma_dc[p],
709                                                     h->ps.pps->dequant4_coeff[p][qscale][0]);
710            else {
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
716                };
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],
721                                            pixel_shift, i));
722            }
723        }
724    }
725}
726
727static av_always_inline void hl_decode_mb_idct_luma(const H264Context *h, H264SliceContext *sl,
728                                                    int mb_type, int simple,
729                                                    int transform_bypass,
730                                                    int pixel_shift,
731                                                    const int *block_offset,
732                                                    int linesize,
733                                                    uint8_t *dest_y, int p)
734{
735    void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
736    int i;
737    block_offset += 16 * p;
738    if (!IS_INTRA4x4(mb_type)) {
739        if (IS_INTRA16x16(mb_type)) {
740            if (transform_bypass) {
741                if (h->ps.sps->profile_idc == 244 &&
742                    (sl->intra16x16_pred_mode == VERT_PRED8x8 ||
743                     sl->intra16x16_pred_mode == HOR_PRED8x8)) {
744                    h->hpc.pred16x16_add[sl->intra16x16_pred_mode](dest_y, block_offset,
745                                                                   sl->mb + (p * 256 << pixel_shift),
746                                                                   linesize);
747                } else {
748                    for (i = 0; i < 16; i++)
749                        if (sl->non_zero_count_cache[scan8[i + p * 16]] ||
750                            dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
751                            h->h264dsp.h264_add_pixels4_clear(dest_y + block_offset[i],
752                                                              sl->mb + (i * 16 + p * 256 << pixel_shift),
753                                                              linesize);
754                }
755            } else {
756                h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
757                                                sl->mb + (p * 256 << pixel_shift),
758                                                linesize,
759                                                sl->non_zero_count_cache + p * 5 * 8);
760            }
761        } else if (sl->cbp & 15) {
762            if (transform_bypass) {
763                const int di = IS_8x8DCT(mb_type) ? 4 : 1;
764                idct_add = IS_8x8DCT(mb_type) ? h->h264dsp.h264_add_pixels8_clear
765                    : h->h264dsp.h264_add_pixels4_clear;
766                for (i = 0; i < 16; i += di)
767                    if (sl->non_zero_count_cache[scan8[i + p * 16]])
768                        idct_add(dest_y + block_offset[i],
769                                 sl->mb + (i * 16 + p * 256 << pixel_shift),
770                                 linesize);
771            } else {
772                if (IS_8x8DCT(mb_type))
773                    h->h264dsp.h264_idct8_add4(dest_y, block_offset,
774                                               sl->mb + (p * 256 << pixel_shift),
775                                               linesize,
776                                               sl->non_zero_count_cache + p * 5 * 8);
777                else
778                    h->h264dsp.h264_idct_add16(dest_y, block_offset,
779                                               sl->mb + (p * 256 << pixel_shift),
780                                               linesize,
781                                               sl->non_zero_count_cache + p * 5 * 8);
782            }
783        }
784    }
785}
786
787#define BITS   8
788#define SIMPLE 1
789#include "h264_mb_template.c"
790
791#undef  BITS
792#define BITS   16
793#include "h264_mb_template.c"
794
795#undef  SIMPLE
796#define SIMPLE 0
797#include "h264_mb_template.c"
798
799void ff_h264_hl_decode_mb(const H264Context *h, H264SliceContext *sl)
800{
801    const int mb_xy   = sl->mb_xy;
802    const int mb_type = h->cur_pic.mb_type[mb_xy];
803    int is_complex    = CONFIG_SMALL || sl->is_complex ||
804                        IS_INTRA_PCM(mb_type) || sl->qscale == 0;
805
806    if (CHROMA444(h)) {
807        if (is_complex || h->pixel_shift)
808            hl_decode_mb_444_complex(h, sl);
809        else
810            hl_decode_mb_444_simple_8(h, sl);
811    } else if (is_complex) {
812        hl_decode_mb_complex(h, sl);
813    } else if (h->pixel_shift) {
814        hl_decode_mb_simple_16(h, sl);
815    } else
816        hl_decode_mb_simple_8(h, sl);
817}
818