falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
41570 Commits
32 Branches
698MB
Tree: d27b8248a6
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'd27b8248a6'
${ noResults }
FFmpeg/libavcodec/h264_direct.c

705 lines
29KB
Raw Blame History 

  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 Libav.

  6.  *

  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.

  11.  *

  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.

  16.  *

  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

  20.  */

  21. 

  22. /**

  23.  * @file

  24.  * H.264 / AVC / MPEG4 part10 direct mb/block decoding.

  25.  * @author Michael Niedermayer <michaelni@gmx.at>

  26.  */

  27. 

  28. #include "internal.h"

  29. #include "avcodec.h"

  30. #include "h264.h"

  31. #include "mpegutils.h"

  32. #include "rectangle.h"

  33. #include "thread.h"

  34. 

  35. #include <assert.h>

  36. 

  37. static int get_scale_factor(H264SliceContext *sl,

  38.                             int poc, int poc1, int i)

  39. {

  40.     int poc0 = sl->ref_list[0][i].poc;

  41.     int td = av_clip_int8(poc1 - poc0);

  42.     if (td == 0 || sl->ref_list[0][i].long_ref) {

  43.         return 256;

  44.     } else {

  45.         int tb = av_clip_int8(poc - poc0);

  46.         int tx = (16384 + (FFABS(td) >> 1)) / td;

  47.         return av_clip_intp2((tb * tx + 32) >> 6, 10);

  48.     }

  49. }

  50. 

  51. void ff_h264_direct_dist_scale_factor(const H264Context *const h,

  52.                                       H264SliceContext *sl)

  53. {

  54.     const int poc  = FIELD_PICTURE(h) ? h->cur_pic_ptr->field_poc[h->picture_structure == PICT_BOTTOM_FIELD]

  55.                                       : h->cur_pic_ptr->poc;

  56.     const int poc1 = sl->ref_list[1][0].poc;

  57.     int i, field;

  58. 

  59.     if (FRAME_MBAFF(h))

  60.         for (field = 0; field < 2; field++) {

  61.             const int poc  = h->cur_pic_ptr->field_poc[field];

  62.             const int poc1 = sl->ref_list[1][0].field_poc[field];

  63.             for (i = 0; i < 2 * sl->ref_count[0]; i++)

  64.                 sl->dist_scale_factor_field[field][i ^ field] =

  65.                     get_scale_factor(sl, poc, poc1, i + 16);

  66.         }

  67. 

  68.     for (i = 0; i < sl->ref_count[0]; i++)

  69.         sl->dist_scale_factor[i] = get_scale_factor(sl, poc, poc1, i);

  70. }

  71. 

  72. static void fill_colmap(const H264Context *h, H264SliceContext *sl,

  73.                         int map[2][16 + 32], int list,

  74.                         int field, int colfield, int mbafi)

  75. {

  76.     H264Picture *const ref1 = &sl->ref_list[1][0];

  77.     int j, old_ref, rfield;

  78.     int start  = mbafi ? 16                       : 0;

  79.     int end    = mbafi ? 16 + 2 * sl->ref_count[0] : sl->ref_count[0];

  80.     int interl = mbafi || h->picture_structure != PICT_FRAME;

  81. 

  82.     /* bogus; fills in for missing frames */

  83.     memset(map[list], 0, sizeof(map[list]));

  84. 

  85.     for (rfield = 0; rfield < 2; rfield++) {

  86.         for (old_ref = 0; old_ref < ref1->ref_count[colfield][list]; old_ref++) {

  87.             int poc = ref1->ref_poc[colfield][list][old_ref];

  88. 

  89.             if (!interl)

  90.                 poc |= 3;

  91.             // FIXME: store all MBAFF references so this is not needed

  92.             else if (interl && (poc & 3) == 3)

  93.                 poc = (poc & ~3) + rfield + 1;

  94. 

  95.             for (j = start; j < end; j++) {

  96.                 if (4 * sl->ref_list[0][j].frame_num +

  97.                     (sl->ref_list[0][j].reference & 3) == poc) {

  98.                     int cur_ref = mbafi ? (j - 16) ^ field : j;

  99.                     if (ref1->mbaff)

  100.                         map[list][2 * old_ref + (rfield ^ field) + 16] = cur_ref;

  101.                     if (rfield == field || !interl)

  102.                         map[list][old_ref] = cur_ref;

  103.                     break;

  104.                 }

  105.             }

  106.         }

  107.     }

  108. }

  109. 

  110. void ff_h264_direct_ref_list_init(const H264Context *const h, H264SliceContext *sl)

  111. {

  112.     H264Picture *const ref1 = &sl->ref_list[1][0];

  113.     H264Picture *const cur = h->cur_pic_ptr;

  114.     int list, j, field;

  115.     int sidx     = (h->picture_structure & 1) ^ 1;

  116.     int ref1sidx = (ref1->reference      & 1) ^ 1;

  117. 

  118.     for (list = 0; list < 2; list++) {

  119.         cur->ref_count[sidx][list] = sl->ref_count[list];

  120.         for (j = 0; j < sl->ref_count[list]; j++)

  121.             cur->ref_poc[sidx][list][j] = 4 * sl->ref_list[list][j].frame_num +

  122.                                           (sl->ref_list[list][j].reference & 3);

  123.     }

  124. 

  125.     if (h->picture_structure == PICT_FRAME) {

  126.         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));

  127.         memcpy(cur->ref_poc[1],   cur->ref_poc[0],   sizeof(cur->ref_poc[0]));

  128.     }

  129. 

  130.     cur->mbaff = FRAME_MBAFF(h);

  131. 

  132.     sl->col_fieldoff = 0;

  133.     if (h->picture_structure == PICT_FRAME) {

  134.         int cur_poc  = h->cur_pic_ptr->poc;

  135.         int *col_poc = sl->ref_list[1]->field_poc;

  136.         sl->col_parity = (FFABS(col_poc[0] - cur_poc) >=

  137.                           FFABS(col_poc[1] - cur_poc));

  138.         ref1sidx =

  139.         sidx     = sl->col_parity;

  140.     // FL -> FL & differ parity

  141.     } else if (!(h->picture_structure & sl->ref_list[1][0].reference) &&

  142.                !sl->ref_list[1][0].mbaff) {

  143.         sl->col_fieldoff = 2 * sl->ref_list[1][0].reference - 3;

  144.     }

  145. 

  146.     if (sl->slice_type_nos != AV_PICTURE_TYPE_B || sl->direct_spatial_mv_pred)

  147.         return;

  148. 

  149.     for (list = 0; list < 2; list++) {

  150.         fill_colmap(h, sl, sl->map_col_to_list0, list, sidx, ref1sidx, 0);

  151.         if (FRAME_MBAFF(h))

  152.             for (field = 0; field < 2; field++)

  153.                 fill_colmap(h, sl, sl->map_col_to_list0_field[field], list, field,

  154.                             field, 1);

  155.     }

  156. }

  157. 

  158. static void await_reference_mb_row(const H264Context *const h, H264Picture *ref,

  159.                                    int mb_y)

  160. {

  161.     int ref_field         = ref->reference - 1;

  162.     int ref_field_picture = ref->field_picture;

  163.     int ref_height        = 16 * h->mb_height >> ref_field_picture;

  164. 

  165.     if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_FRAME))

  166.         return;

  167. 

  168.     /* FIXME: It can be safe to access mb stuff

  169.      * even if pixels aren't deblocked yet. */

  170. 

  171.     ff_thread_await_progress(&ref->tf,

  172.                              FFMIN(16 * mb_y >> ref_field_picture,

  173.                                    ref_height - 1),

  174.                              ref_field_picture && ref_field);

  175. }

  176. 

  177. static void pred_spatial_direct_motion(const H264Context *const h, H264SliceContext *sl,

  178.                                        int *mb_type)

  179. {

  180.     int b8_stride = 2;

  181.     int b4_stride = h->b_stride;

  182.     int mb_xy = sl->mb_xy, mb_y = sl->mb_y;

  183.     int mb_type_col[2];

  184.     const int16_t (*l1mv0)[2], (*l1mv1)[2];

  185.     const int8_t *l1ref0, *l1ref1;

  186.     const int is_b8x8 = IS_8X8(*mb_type);

  187.     unsigned int sub_mb_type = MB_TYPE_L0L1;

  188.     int i8, i4;

  189.     int ref[2];

  190.     int mv[2];

  191.     int list;

  192. 

  193.     assert(sl->ref_list[1][0].reference & 3);

  194. 

  195.     await_reference_mb_row(h, &sl->ref_list[1][0],

  196.                            sl->mb_y + !!IS_INTERLACED(*mb_type));

  197. 

  198. #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \

  199.                                 MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM)

  200. 

  201.     /* ref = min(neighbors) */

  202.     for (list = 0; list < 2; list++) {

  203.         int left_ref     = sl->ref_cache[list][scan8[0] - 1];

  204.         int top_ref      = sl->ref_cache[list][scan8[0] - 8];

  205.         int refc         = sl->ref_cache[list][scan8[0] - 8 + 4];

  206.         const int16_t *C = sl->mv_cache[list][scan8[0]  - 8 + 4];

  207.         if (refc == PART_NOT_AVAILABLE) {

  208.             refc = sl->ref_cache[list][scan8[0] - 8 - 1];

  209.             C    = sl->mv_cache[list][scan8[0]  - 8 - 1];

  210.         }

  211.         ref[list] = FFMIN3((unsigned)left_ref,

  212.                            (unsigned)top_ref,

  213.                            (unsigned)refc);

  214.         if (ref[list] >= 0) {

  215.             /* This is just pred_motion() but with the cases removed that

  216.              * cannot happen for direct blocks. */

  217.             const int16_t *const A = sl->mv_cache[list][scan8[0] - 1];

  218.             const int16_t *const B = sl->mv_cache[list][scan8[0] - 8];

  219. 

  220.             int match_count = (left_ref == ref[list]) +

  221.                               (top_ref  == ref[list]) +

  222.                               (refc     == ref[list]);

  223. 

  224.             if (match_count > 1) { // most common

  225.                 mv[list] = pack16to32(mid_pred(A[0], B[0], C[0]),

  226.                                       mid_pred(A[1], B[1], C[1]));

  227.             } else {

  228.                 assert(match_count == 1);

  229.                 if (left_ref == ref[list])

  230.                     mv[list] = AV_RN32A(A);

  231.                 else if (top_ref == ref[list])

  232.                     mv[list] = AV_RN32A(B);

  233.                 else

  234.                     mv[list] = AV_RN32A(C);

  235.             }

  236.         } else {

  237.             int mask = ~(MB_TYPE_L0 << (2 * list));

  238.             mv[list]  = 0;

  239.             ref[list] = -1;

  240.             if (!is_b8x8)

  241.                 *mb_type &= mask;

  242.             sub_mb_type &= mask;

  243.         }

  244.     }

  245.     if (ref[0] < 0 && ref[1] < 0) {

  246.         ref[0] = ref[1] = 0;

  247.         if (!is_b8x8)

  248.             *mb_type |= MB_TYPE_L0L1;

  249.         sub_mb_type |= MB_TYPE_L0L1;

  250.     }

  251. 

  252.     if (!(is_b8x8 | mv[0] | mv[1])) {

  253.         fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);

  254.         fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);

  255.         fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);

  256.         fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);

  257.         *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |

  258.                                  MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  259.                    MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  260.         return;

  261.     }

  262. 

  263.     if (IS_INTERLACED(sl->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL

  264.         if (!IS_INTERLACED(*mb_type)) {                    //     AFR/FR    -> AFL/FL

  265.             mb_y  = (sl->mb_y & ~1) + sl->col_parity;

  266.             mb_xy = sl->mb_x +

  267.                     ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride;

  268.             b8_stride = 0;

  269.         } else {

  270.             mb_y  += sl->col_fieldoff;

  271.             mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity

  272.         }

  273.         goto single_col;

  274.     } else {                                             // AFL/AFR/FR/FL -> AFR/FR

  275.         if (IS_INTERLACED(*mb_type)) {                   // AFL       /FL -> AFR/FR

  276.             mb_y           =  sl->mb_y & ~1;

  277.             mb_xy          = (sl->mb_y & ~1) * h->mb_stride + sl->mb_x;

  278.             mb_type_col[0] = sl->ref_list[1][0].mb_type[mb_xy];

  279.             mb_type_col[1] = sl->ref_list[1][0].mb_type[mb_xy + h->mb_stride];

  280.             b8_stride      = 2 + 4 * h->mb_stride;

  281.             b4_stride     *= 6;

  282.             if (IS_INTERLACED(mb_type_col[0]) !=

  283.                 IS_INTERLACED(mb_type_col[1])) {

  284.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  285.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  286.             }

  287. 

  288.             sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */

  289.             if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) &&

  290.                 (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) &&

  291.                 !is_b8x8) {

  292.                 *mb_type |= MB_TYPE_16x8 | MB_TYPE_DIRECT2;  /* B_16x8 */

  293.             } else {

  294.                 *mb_type |= MB_TYPE_8x8;

  295.             }

  296.         } else {                                         //     AFR/FR    -> AFR/FR

  297. single_col:

  298.             mb_type_col[0] =

  299.             mb_type_col[1] = sl->ref_list[1][0].mb_type[mb_xy];

  300. 

  301.             sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */

  302.             if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) {

  303.                 *mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_16x16 */

  304.             } else if (!is_b8x8 &&

  305.                        (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) {

  306.                 *mb_type |= MB_TYPE_DIRECT2 |

  307.                             (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16));

  308.             } else {

  309.                 if (!h->sps.direct_8x8_inference_flag) {

  310.                     /* FIXME: Save sub mb types from previous frames (or derive

  311.                      * from MVs) so we know exactly what block size to use. */

  312.                     sub_mb_type += (MB_TYPE_8x8 - MB_TYPE_16x16); /* B_SUB_4x4 */

  313.                 }

  314.                 *mb_type |= MB_TYPE_8x8;

  315.             }

  316.         }

  317.     }

  318. 

  319.     await_reference_mb_row(h, &sl->ref_list[1][0], mb_y);

  320. 

  321.     l1mv0  = &sl->ref_list[1][0].motion_val[0][h->mb2b_xy[mb_xy]];

  322.     l1mv1  = &sl->ref_list[1][0].motion_val[1][h->mb2b_xy[mb_xy]];

  323.     l1ref0 = &sl->ref_list[1][0].ref_index[0][4 * mb_xy];

  324.     l1ref1 = &sl->ref_list[1][0].ref_index[1][4 * mb_xy];

  325.     if (!b8_stride) {

  326.         if (sl->mb_y & 1) {

  327.             l1ref0 += 2;

  328.             l1ref1 += 2;

  329.             l1mv0  += 2 * b4_stride;

  330.             l1mv1  += 2 * b4_stride;

  331.         }

  332.     }

  333. 

  334.     if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) {

  335.         int n = 0;

  336.         for (i8 = 0; i8 < 4; i8++) {

  337.             int x8  = i8 & 1;

  338.             int y8  = i8 >> 1;

  339.             int xy8 = x8     + y8 * b8_stride;

  340.             int xy4 = x8 * 3 + y8 * b4_stride;

  341.             int a, b;

  342. 

  343.             if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))

  344.                 continue;

  345.             sl->sub_mb_type[i8] = sub_mb_type;

  346. 

  347.             fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,

  348.                            (uint8_t)ref[0], 1);

  349.             fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8,

  350.                            (uint8_t)ref[1], 1);

  351.             if (!IS_INTRA(mb_type_col[y8]) && !sl->ref_list[1][0].long_ref &&

  352.                 ((l1ref0[xy8] == 0 &&

  353.                   FFABS(l1mv0[xy4][0]) <= 1 &&

  354.                   FFABS(l1mv0[xy4][1]) <= 1) ||

  355.                  (l1ref0[xy8] < 0 &&

  356.                   l1ref1[xy8] == 0 &&

  357.                   FFABS(l1mv1[xy4][0]) <= 1 &&

  358.                   FFABS(l1mv1[xy4][1]) <= 1))) {

  359.                 a =

  360.                 b = 0;

  361.                 if (ref[0] > 0)

  362.                     a = mv[0];

  363.                 if (ref[1] > 0)

  364.                     b = mv[1];

  365.                 n++;

  366.             } else {

  367.                 a = mv[0];

  368.                 b = mv[1];

  369.             }

  370.             fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, a, 4);

  371.             fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, b, 4);

  372.         }

  373.         if (!is_b8x8 && !(n & 3))

  374.             *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |

  375.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  376.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  377.     } else if (IS_16X16(*mb_type)) {

  378.         int a, b;

  379. 

  380.         fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);

  381.         fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);

  382.         if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].long_ref &&

  383.             ((l1ref0[0] == 0 &&

  384.               FFABS(l1mv0[0][0]) <= 1 &&

  385.               FFABS(l1mv0[0][1]) <= 1) ||

  386.              (l1ref0[0] < 0 && !l1ref1[0] &&

  387.               FFABS(l1mv1[0][0]) <= 1 &&

  388.               FFABS(l1mv1[0][1]) <= 1 &&

  389.               h->x264_build > 33U))) {

  390.             a = b = 0;

  391.             if (ref[0] > 0)

  392.                 a = mv[0];

  393.             if (ref[1] > 0)

  394.                 b = mv[1];

  395.         } else {

  396.             a = mv[0];

  397.             b = mv[1];

  398.         }

  399.         fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);

  400.         fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);

  401.     } else {

  402.         int n = 0;

  403.         for (i8 = 0; i8 < 4; i8++) {

  404.             const int x8 = i8 & 1;

  405.             const int y8 = i8 >> 1;

  406. 

  407.             if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))

  408.                 continue;

  409.             sl->sub_mb_type[i8] = sub_mb_type;

  410. 

  411.             fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, mv[0], 4);

  412.             fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, mv[1], 4);

  413.             fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,

  414.                            (uint8_t)ref[0], 1);

  415.             fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8,

  416.                            (uint8_t)ref[1], 1);

  417. 

  418.             assert(b8_stride == 2);

  419.             /* col_zero_flag */

  420.             if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].long_ref &&

  421.                 (l1ref0[i8] == 0 ||

  422.                  (l1ref0[i8] < 0 &&

  423.                   l1ref1[i8] == 0 &&

  424.                   h->x264_build > 33U))) {

  425.                 const int16_t (*l1mv)[2] = l1ref0[i8] == 0 ? l1mv0 : l1mv1;

  426.                 if (IS_SUB_8X8(sub_mb_type)) {

  427.                     const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride];

  428.                     if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) {

  429.                         if (ref[0] == 0)

  430.                             fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2,

  431.                                            8, 0, 4);

  432.                         if (ref[1] == 0)

  433.                             fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2,

  434.                                            8, 0, 4);

  435.                         n += 4;

  436.                     }

  437.                 } else {

  438.                     int m = 0;

  439.                     for (i4 = 0; i4 < 4; i4++) {

  440.                         const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) +

  441.                                                      (y8 * 2 + (i4 >> 1)) * b4_stride];

  442.                         if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) {

  443.                             if (ref[0] == 0)

  444.                                 AV_ZERO32(sl->mv_cache[0][scan8[i8 * 4 + i4]]);

  445.                             if (ref[1] == 0)

  446.                                 AV_ZERO32(sl->mv_cache[1][scan8[i8 * 4 + i4]]);

  447.                             m++;

  448.                         }

  449.                     }

  450.                     if (!(m & 3))

  451.                         sl->sub_mb_type[i8] += MB_TYPE_16x16 - MB_TYPE_8x8;

  452.                     n += m;

  453.                 }

  454.             }

  455.         }

  456.         if (!is_b8x8 && !(n & 15))

  457.             *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |

  458.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  459.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  460.     }

  461. }

  462. 

  463. static void pred_temp_direct_motion(const H264Context *const h, H264SliceContext *sl,

  464.                                     int *mb_type)

  465. {

  466.     int b8_stride = 2;

  467.     int b4_stride = h->b_stride;

  468.     int mb_xy = sl->mb_xy, mb_y = sl->mb_y;

  469.     int mb_type_col[2];

  470.     const int16_t (*l1mv0)[2], (*l1mv1)[2];

  471.     const int8_t *l1ref0, *l1ref1;

  472.     const int is_b8x8 = IS_8X8(*mb_type);

  473.     unsigned int sub_mb_type;

  474.     int i8, i4;

  475. 

  476.     assert(sl->ref_list[1][0].reference & 3);

  477. 

  478.     await_reference_mb_row(h, &sl->ref_list[1][0],

  479.                            sl->mb_y + !!IS_INTERLACED(*mb_type));

  480. 

  481.     if (IS_INTERLACED(sl->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL

  482.         if (!IS_INTERLACED(*mb_type)) {                    //     AFR/FR    -> AFL/FL

  483.             mb_y  = (sl->mb_y & ~1) + sl->col_parity;

  484.             mb_xy = sl->mb_x +

  485.                     ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride;

  486.             b8_stride = 0;

  487.         } else {

  488.             mb_y  += sl->col_fieldoff;

  489.             mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity

  490.         }

  491.         goto single_col;

  492.     } else {                                        // AFL/AFR/FR/FL -> AFR/FR

  493.         if (IS_INTERLACED(*mb_type)) {              // AFL       /FL -> AFR/FR

  494.             mb_y           = sl->mb_y & ~1;

  495.             mb_xy          = sl->mb_x + (sl->mb_y & ~1) * h->mb_stride;

  496.             mb_type_col[0] = sl->ref_list[1][0].mb_type[mb_xy];

  497.             mb_type_col[1] = sl->ref_list[1][0].mb_type[mb_xy + h->mb_stride];

  498.             b8_stride      = 2 + 4 * h->mb_stride;

  499.             b4_stride     *= 6;

  500.             if (IS_INTERLACED(mb_type_col[0]) !=

  501.                 IS_INTERLACED(mb_type_col[1])) {

  502.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  503.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  504.             }

  505. 

  506.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  507.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

  508. 

  509.             if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) &&

  510.                 (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) &&

  511.                 !is_b8x8) {

  512.                 *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 |

  513.                             MB_TYPE_DIRECT2;                /* B_16x8 */

  514.             } else {

  515.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  516.             }

  517.         } else {                                    //     AFR/FR    -> AFR/FR

  518. single_col:

  519.             mb_type_col[0]     =

  520.                 mb_type_col[1] = sl->ref_list[1][0].mb_type[mb_xy];

  521. 

  522.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  523.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

  524.             if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) {

  525.                 *mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  526.                             MB_TYPE_DIRECT2;                /* B_16x16 */

  527.             } else if (!is_b8x8 &&

  528.                        (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) {

  529.                 *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 |

  530.                             (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16));

  531.             } else {

  532.                 if (!h->sps.direct_8x8_inference_flag) {

  533.                     /* FIXME: save sub mb types from previous frames (or derive

  534.                      * from MVs) so we know exactly what block size to use */

  535.                     sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  536.                                   MB_TYPE_DIRECT2;          /* B_SUB_4x4 */

  537.                 }

  538.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  539.             }

  540.         }

  541.     }

  542. 

  543.     await_reference_mb_row(h, &sl->ref_list[1][0], mb_y);

  544. 

  545.     l1mv0  = &sl->ref_list[1][0].motion_val[0][h->mb2b_xy[mb_xy]];

  546.     l1mv1  = &sl->ref_list[1][0].motion_val[1][h->mb2b_xy[mb_xy]];

  547.     l1ref0 = &sl->ref_list[1][0].ref_index[0][4 * mb_xy];

  548.     l1ref1 = &sl->ref_list[1][0].ref_index[1][4 * mb_xy];

  549.     if (!b8_stride) {

  550.         if (sl->mb_y & 1) {

  551.             l1ref0 += 2;

  552.             l1ref1 += 2;

  553.             l1mv0  += 2 * b4_stride;

  554.             l1mv1  += 2 * b4_stride;

  555.         }

  556.     }

  557. 

  558.     {

  559.         const int *map_col_to_list0[2] = { sl->map_col_to_list0[0],

  560.                                            sl->map_col_to_list0[1] };

  561.         const int *dist_scale_factor = sl->dist_scale_factor;

  562.         int ref_offset;

  563. 

  564.         if (FRAME_MBAFF(h) && IS_INTERLACED(*mb_type)) {

  565.             map_col_to_list0[0] = sl->map_col_to_list0_field[sl->mb_y & 1][0];

  566.             map_col_to_list0[1] = sl->map_col_to_list0_field[sl->mb_y & 1][1];

  567.             dist_scale_factor   = sl->dist_scale_factor_field[sl->mb_y & 1];

  568.         }

  569.         ref_offset = (sl->ref_list[1][0].mbaff << 4) & (mb_type_col[0] >> 3);

  570. 

  571.         if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) {

  572.             int y_shift = 2 * !IS_INTERLACED(*mb_type);

  573.             assert(h->sps.direct_8x8_inference_flag);

  574. 

  575.             for (i8 = 0; i8 < 4; i8++) {

  576.                 const int x8 = i8 & 1;

  577.                 const int y8 = i8 >> 1;

  578.                 int ref0, scale;

  579.                 const int16_t (*l1mv)[2] = l1mv0;

  580. 

  581.                 if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))

  582.                     continue;

  583.                 sl->sub_mb_type[i8] = sub_mb_type;

  584. 

  585.                 fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1);

  586.                 if (IS_INTRA(mb_type_col[y8])) {

  587.                     fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1);

  588.                     fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4);

  589.                     fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4);

  590.                     continue;

  591.                 }

  592. 

  593.                 ref0 = l1ref0[x8 + y8 * b8_stride];

  594.                 if (ref0 >= 0)

  595.                     ref0 = map_col_to_list0[0][ref0 + ref_offset];

  596.                 else {

  597.                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8 * b8_stride] +

  598.                                                ref_offset];

  599.                     l1mv = l1mv1;

  600.                 }

  601.                 scale = dist_scale_factor[ref0];

  602.                 fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,

  603.                                ref0, 1);

  604. 

  605.                 {

  606.                     const int16_t *mv_col = l1mv[x8 * 3 + y8 * b4_stride];

  607.                     int my_col            = (mv_col[1] << y_shift) / 2;

  608.                     int mx                = (scale * mv_col[0] + 128) >> 8;

  609.                     int my                = (scale * my_col    + 128) >> 8;

  610.                     fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8,

  611.                                    pack16to32(mx, my), 4);

  612.                     fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8,

  613.                                    pack16to32(mx - mv_col[0], my - my_col), 4);

  614.                 }

  615.             }

  616.             return;

  617.         }

  618. 

  619.         /* one-to-one mv scaling */

  620. 

  621.         if (IS_16X16(*mb_type)) {

  622.             int ref, mv0, mv1;

  623. 

  624.             fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);

  625.             if (IS_INTRA(mb_type_col[0])) {

  626.                 ref = mv0 = mv1 = 0;

  627.             } else {

  628.                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]

  629.                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];

  630.                 const int scale = dist_scale_factor[ref0];

  631.                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];

  632.                 int mv_l0[2];

  633.                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;

  634.                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;

  635.                 ref      = ref0;

  636.                 mv0      = pack16to32(mv_l0[0], mv_l0[1]);

  637.                 mv1      = pack16to32(mv_l0[0] - mv_col[0], mv_l0[1] - mv_col[1]);

  638.             }

  639.             fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);

  640.             fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);

  641.             fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);

  642.         } else {

  643.             for (i8 = 0; i8 < 4; i8++) {

  644.                 const int x8 = i8 & 1;

  645.                 const int y8 = i8 >> 1;

  646.                 int ref0, scale;

  647.                 const int16_t (*l1mv)[2] = l1mv0;

  648. 

  649.                 if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))

  650.                     continue;

  651.                 sl->sub_mb_type[i8] = sub_mb_type;

  652.                 fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1);

  653.                 if (IS_INTRA(mb_type_col[0])) {

  654.                     fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1);

  655.                     fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4);

  656.                     fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4);

  657.                     continue;

  658.                 }

  659. 

  660.                 assert(b8_stride == 2);

  661.                 ref0 = l1ref0[i8];

  662.                 if (ref0 >= 0)

  663.                     ref0 = map_col_to_list0[0][ref0 + ref_offset];

  664.                 else {

  665.                     ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];

  666.                     l1mv = l1mv1;

  667.                 }

  668.                 scale = dist_scale_factor[ref0];

  669. 

  670.                 fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,

  671.                                ref0, 1);

  672.                 if (IS_SUB_8X8(sub_mb_type)) {

  673.                     const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride];

  674.                     int mx                = (scale * mv_col[0] + 128) >> 8;

  675.                     int my                = (scale * mv_col[1] + 128) >> 8;

  676.                     fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8,

  677.                                    pack16to32(mx, my), 4);

  678.                     fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8,

  679.                                    pack16to32(mx - mv_col[0], my - mv_col[1]), 4);

  680.                 } else {

  681.                     for (i4 = 0; i4 < 4; i4++) {

  682.                         const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) +

  683.                                                      (y8 * 2 + (i4 >> 1)) * b4_stride];

  684.                         int16_t *mv_l0 = sl->mv_cache[0][scan8[i8 * 4 + i4]];

  685.                         mv_l0[0] = (scale * mv_col[0] + 128) >> 8;

  686.                         mv_l0[1] = (scale * mv_col[1] + 128) >> 8;

  687.                         AV_WN32A(sl->mv_cache[1][scan8[i8 * 4 + i4]],

  688.                                  pack16to32(mv_l0[0] - mv_col[0],

  689.                                             mv_l0[1] - mv_col[1]));

  690.                     }

  691.                 }

  692.             }

  693.         }

  694.     }

  695. }

  696. 

  697. void ff_h264_pred_direct_motion(const H264Context *const h, H264SliceContext *sl,

  698.                                 int *mb_type)

  699. {

  700.     if (sl->direct_spatial_mv_pred)

  701.         pred_spatial_direct_motion(h, sl, mb_type);

  702.     else

  703.         pred_temp_direct_motion(h, sl, mb_type);

  704. }