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.
44689 Commits
32 Branches
293MB
Tree: c5714b51aa
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c5714b51aa'
${ noResults }
FFmpeg/libavcodec/h264_direct.c

710 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 / MPEG-4 part10 direct mb/block decoding.

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

  26.  */

  27. 

  28. #include "internal.h"

  29. #include "avcodec.h"

  30. #include "h264dec.h"

  31. #include "h264_ps.h"

  32. #include "mpegutils.h"

  33. #include "rectangle.h"

  34. #include "thread.h"

  35. 

  36. #include <assert.h>

  37. 

  38. static int get_scale_factor(H264SliceContext *sl,

  39.                             int poc, int poc1, int i)

  40. {

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

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

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

  44.         return 256;

  45.     } else {

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

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

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

  49.     }

  50. }

  51. 

  52. void ff_h264_direct_dist_scale_factor(const H264Context *const h,

  53.                                       H264SliceContext *sl)

  54. {

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

  56.                                       : h->cur_pic_ptr->poc;

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

  58.     int i, field;

  59. 

  60.     if (FRAME_MBAFF(h))

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

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

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

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

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

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

  67.         }

  68. 

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

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

  71. }

  72. 

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

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

  75.                         int field, int colfield, int mbafi)

  76. {

  77.     H264Picture *const ref1 = sl->ref_list[1][0].parent;

  78.     int j, old_ref, rfield;

  79.     int start  = mbafi ? 16                       : 0;

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

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

  82. 

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

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

  85. 

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

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

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

  89. 

  90.             if (!interl)

  91.                 poc |= 3;

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

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

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

  95. 

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

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

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

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

  100.                     if (ref1->mbaff)

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

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

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

  104.                     break;

  105.                 }

  106.             }

  107.         }

  108.     }

  109. }

  110. 

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

  112. {

  113.     H264Ref *const ref1 = &sl->ref_list[1][0];

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

  115.     int list, j, field;

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

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

  118. 

  119.     for (list = 0; list < sl->list_count; list++) {

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

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

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

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

  124.     }

  125. 

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

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

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

  129.     }

  130. 

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

  132. 

  133.     sl->col_fieldoff = 0;

  134. 

  135.     if (sl->list_count != 2 || !sl->ref_count[1])

  136.         return;

  137. 

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

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

  140.         int *col_poc = sl->ref_list[1][0].parent->field_poc;

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

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

  143.         ref1sidx =

  144.         sidx     = sl->col_parity;

  145.     // FL -> FL & differ parity

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

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

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

  149.     }

  150. 

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

  152.         return;

  153. 

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

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

  156.         if (FRAME_MBAFF(h))

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

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

  159.                             field, 1);

  160.     }

  161. }

  162. 

  163. static void await_reference_mb_row(const H264Context *const h, H264Ref *ref,

  164.                                    int mb_y)

  165. {

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

  167.     int ref_field_picture = ref->parent->field_picture;

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

  169. 

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

  171.         return;

  172. 

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

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

  175. 

  176.     ff_thread_await_progress(&ref->parent->tf,

  177.                              FFMIN(16 * mb_y >> ref_field_picture,

  178.                                    ref_height - 1),

  179.                              ref_field_picture && ref_field);

  180. }

  181. 

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

  183.                                        int *mb_type)

  184. {

  185.     int b8_stride = 2;

  186.     int b4_stride = h->b_stride;

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

  188.     int mb_type_col[2];

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

  190.     const int8_t *l1ref0, *l1ref1;

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

  192.     unsigned int sub_mb_type = MB_TYPE_L0L1;

  193.     int i8, i4;

  194.     int ref[2];

  195.     int mv[2];

  196.     int list;

  197. 

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

  199. 

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

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

  202. 

  203. #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \

  204.                                 MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM)

  205. 

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

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

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

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

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

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

  212.         if (refc == PART_NOT_AVAILABLE) {

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

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

  215.         }

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

  217.                            (unsigned)top_ref,

  218.                            (unsigned)refc);

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

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

  221.              * cannot happen for direct blocks. */

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

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

  224. 

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

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

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

  228. 

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

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

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

  232.             } else {

  233.                 assert(match_count == 1);

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

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

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

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

  238.                 else

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

  240.             }

  241.         } else {

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

  243.             mv[list]  = 0;

  244.             ref[list] = -1;

  245.             if (!is_b8x8)

  246.                 *mb_type &= mask;

  247.             sub_mb_type &= mask;

  248.         }

  249.     }

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

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

  252.         if (!is_b8x8)

  253.             *mb_type |= MB_TYPE_L0L1;

  254.         sub_mb_type |= MB_TYPE_L0L1;

  255.     }

  256. 

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

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

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

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

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

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

  263.                                  MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  264.                    MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  265.         return;

  266.     }

  267. 

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

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

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

  271.             mb_xy = sl->mb_x +

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

  273.             b8_stride = 0;

  274.         } else {

  275.             mb_y  += sl->col_fieldoff;

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

  277.         }

  278.         goto single_col;

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

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

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

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

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

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

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

  286.             b4_stride     *= 6;

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

  288.                 IS_INTERLACED(mb_type_col[1])) {

  289.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  290.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  291.             }

  292. 

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

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

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

  296.                 !is_b8x8) {

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

  298.             } else {

  299.                 *mb_type |= MB_TYPE_8x8;

  300.             }

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

  302. single_col:

  303.             mb_type_col[0] =

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

  305. 

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

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

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

  309.             } else if (!is_b8x8 &&

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

  311.                 *mb_type |= MB_TYPE_DIRECT2 |

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

  313.             } else {

  314.                 if (!h->ps.sps->direct_8x8_inference_flag) {

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

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

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

  318.                 }

  319.                 *mb_type |= MB_TYPE_8x8;

  320.             }

  321.         }

  322.     }

  323. 

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

  325. 

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

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

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

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

  330.     if (!b8_stride) {

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

  332.             l1ref0 += 2;

  333.             l1ref1 += 2;

  334.             l1mv0  += 2 * b4_stride;

  335.             l1mv1  += 2 * b4_stride;

  336.         }

  337.     }

  338. 

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

  340.         int n = 0;

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

  342.             int x8  = i8 & 1;

  343.             int y8  = i8 >> 1;

  344.             int xy8 = x8     + y8 * b8_stride;

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

  346.             int a, b;

  347. 

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

  349.                 continue;

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

  351. 

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

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

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

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

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

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

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

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

  360.                  (l1ref0[xy8] < 0 &&

  361.                   l1ref1[xy8] == 0 &&

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

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

  364.                 a =

  365.                 b = 0;

  366.                 if (ref[0] > 0)

  367.                     a = mv[0];

  368.                 if (ref[1] > 0)

  369.                     b = mv[1];

  370.                 n++;

  371.             } else {

  372.                 a = mv[0];

  373.                 b = mv[1];

  374.             }

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

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

  377.         }

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

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

  380.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  381.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

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

  383.         int a, b;

  384. 

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

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

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

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

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

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

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

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

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

  394.               h->sei.unregistered.x264_build > 33U))) {

  395.             a = b = 0;

  396.             if (ref[0] > 0)

  397.                 a = mv[0];

  398.             if (ref[1] > 0)

  399.                 b = mv[1];

  400.         } else {

  401.             a = mv[0];

  402.             b = mv[1];

  403.         }

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

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

  406.     } else {

  407.         int n = 0;

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

  409.             const int x8 = i8 & 1;

  410.             const int y8 = i8 >> 1;

  411. 

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

  413.                 continue;

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

  415. 

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

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

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

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

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

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

  422. 

  423.             assert(b8_stride == 2);

  424.             /* col_zero_flag */

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

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

  427.                  (l1ref0[i8] < 0 &&

  428.                   l1ref1[i8] == 0 &&

  429.                   h->sei.unregistered.x264_build > 33U))) {

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

  431.                 if (IS_SUB_8X8(sub_mb_type)) {

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

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

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

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

  436.                                            8, 0, 4);

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

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

  439.                                            8, 0, 4);

  440.                         n += 4;

  441.                     }

  442.                 } else {

  443.                     int m = 0;

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

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

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

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

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

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

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

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

  452.                             m++;

  453.                         }

  454.                     }

  455.                     if (!(m & 3))

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

  457.                     n += m;

  458.                 }

  459.             }

  460.         }

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

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

  463.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  464.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  465.     }

  466. }

  467. 

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

  469.                                     int *mb_type)

  470. {

  471.     int b8_stride = 2;

  472.     int b4_stride = h->b_stride;

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

  474.     int mb_type_col[2];

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

  476.     const int8_t *l1ref0, *l1ref1;

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

  478.     unsigned int sub_mb_type;

  479.     int i8, i4;

  480. 

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

  482. 

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

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

  485. 

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

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

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

  489.             mb_xy = sl->mb_x +

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

  491.             b8_stride = 0;

  492.         } else {

  493.             mb_y  += sl->col_fieldoff;

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

  495.         }

  496.         goto single_col;

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

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

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

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

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

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

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

  504.             b4_stride     *= 6;

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

  506.                 IS_INTERLACED(mb_type_col[1])) {

  507.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  508.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  509.             }

  510. 

  511.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  512.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

  513. 

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

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

  516.                 !is_b8x8) {

  517.                 *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 |

  518.                             MB_TYPE_DIRECT2;                /* B_16x8 */

  519.             } else {

  520.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  521.             }

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

  523. single_col:

  524.             mb_type_col[0]     =

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

  526. 

  527.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  528.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

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

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

  531.                             MB_TYPE_DIRECT2;                /* B_16x16 */

  532.             } else if (!is_b8x8 &&

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

  534.                 *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 |

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

  536.             } else {

  537.                 if (!h->ps.sps->direct_8x8_inference_flag) {

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

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

  540.                     sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  541.                                   MB_TYPE_DIRECT2;          /* B_SUB_4x4 */

  542.                 }

  543.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  544.             }

  545.         }

  546.     }

  547. 

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

  549. 

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

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

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

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

  554.     if (!b8_stride) {

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

  556.             l1ref0 += 2;

  557.             l1ref1 += 2;

  558.             l1mv0  += 2 * b4_stride;

  559.             l1mv1  += 2 * b4_stride;

  560.         }

  561.     }

  562. 

  563.     {

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

  565.                                            sl->map_col_to_list0[1] };

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

  567.         int ref_offset;

  568. 

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

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

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

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

  573.         }

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

  575. 

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

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

  578.             assert(h->ps.sps->direct_8x8_inference_flag);

  579. 

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

  581.                 const int x8 = i8 & 1;

  582.                 const int y8 = i8 >> 1;

  583.                 int ref0, scale;

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

  585. 

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

  587.                     continue;

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

  589. 

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

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

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

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

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

  595.                     continue;

  596.                 }

  597. 

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

  599.                 if (ref0 >= 0)

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

  601.                 else {

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

  603.                                                ref_offset];

  604.                     l1mv = l1mv1;

  605.                 }

  606.                 scale = dist_scale_factor[ref0];

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

  608.                                ref0, 1);

  609. 

  610.                 {

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

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

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

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

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

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

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

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

  619.                 }

  620.             }

  621.             return;

  622.         }

  623. 

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

  625. 

  626.         if (IS_16X16(*mb_type)) {

  627.             int ref, mv0, mv1;

  628. 

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

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

  631.                 ref = mv0 = mv1 = 0;

  632.             } else {

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

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

  635.                 const int scale = dist_scale_factor[ref0];

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

  637.                 int mv_l0[2];

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

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

  640.                 ref      = ref0;

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

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

  643.             }

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

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

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

  647.         } else {

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

  649.                 const int x8 = i8 & 1;

  650.                 const int y8 = i8 >> 1;

  651.                 int ref0, scale;

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

  653. 

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

  655.                     continue;

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

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

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

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

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

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

  662.                     continue;

  663.                 }

  664. 

  665.                 assert(b8_stride == 2);

  666.                 ref0 = l1ref0[i8];

  667.                 if (ref0 >= 0)

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

  669.                 else {

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

  671.                     l1mv = l1mv1;

  672.                 }

  673.                 scale = dist_scale_factor[ref0];

  674. 

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

  676.                                ref0, 1);

  677.                 if (IS_SUB_8X8(sub_mb_type)) {

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

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

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

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

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

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

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

  685.                 } else {

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

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

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

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

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

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

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

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

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

  695.                     }

  696.                 }

  697.             }

  698.         }

  699.     }

  700. }

  701. 

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

  703.                                 int *mb_type)

  704. {

  705.     if (sl->direct_spatial_mv_pred)

  706.         pred_spatial_direct_motion(h, sl, mb_type);

  707.     else

  708.         pred_temp_direct_motion(h, sl, mb_type);

  709. }