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.
40948 Commits
32 Branches
935MB
Tree: bc75b64cff
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bc75b64cff'
${ noResults }
FFmpeg/libavcodec/h264_direct.c

699 lines
28KB
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(H264Context *const h, int poc, int poc1, int i)

  38. {

  39.     int poc0 = h->ref_list[0][i].poc;

  40.     int td = av_clip(poc1 - poc0, -128, 127);

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

  42.         return 256;

  43.     } else {

  44.         int tb = av_clip(poc - poc0, -128, 127);

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

  46.         return av_clip((tb * tx + 32) >> 6, -1024, 1023);

  47.     }

  48. }

  49. 

  50. void ff_h264_direct_dist_scale_factor(H264Context *const h)

  51. {

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

  53.                                       : h->cur_pic_ptr->poc;

  54.     const int poc1 = h->ref_list[1][0].poc;

  55.     int i, field;

  56. 

  57.     if (FRAME_MBAFF(h))

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

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

  60.             const int poc1 = h->ref_list[1][0].field_poc[field];

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

  62.                 h->dist_scale_factor_field[field][i ^ field] =

  63.                     get_scale_factor(h, poc, poc1, i + 16);

  64.         }

  65. 

  66.     for (i = 0; i < h->ref_count[0]; i++)

  67.         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);

  68. }

  69. 

  70. static void fill_colmap(H264Context *h, int map[2][16 + 32], int list,

  71.                         int field, int colfield, int mbafi)

  72. {

  73.     H264Picture *const ref1 = &h->ref_list[1][0];

  74.     int j, old_ref, rfield;

  75.     int start  = mbafi ? 16                       : 0;

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

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

  78. 

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

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

  81. 

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

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

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

  85. 

  86.             if (!interl)

  87.                 poc |= 3;

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

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

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

  91. 

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

  93.                 if (4 * h->ref_list[0][j].frame_num +

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

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

  96.                     if (ref1->mbaff)

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

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

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

  100.                     break;

  101.                 }

  102.             }

  103.         }

  104.     }

  105. }

  106. 

  107. void ff_h264_direct_ref_list_init(H264Context *const h)

  108. {

  109.     H264Picture *const ref1 = &h->ref_list[1][0];

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

  111.     int list, j, field;

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

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

  114. 

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

  116.         cur->ref_count[sidx][list] = h->ref_count[list];

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

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

  119.                                           (h->ref_list[list][j].reference & 3);

  120.     }

  121. 

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

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

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

  125.     }

  126. 

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

  128. 

  129.     h->col_fieldoff = 0;

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

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

  132.         int *col_poc = h->ref_list[1]->field_poc;

  133.         h->col_parity = (FFABS(col_poc[0] - cur_poc) >=

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

  135.         ref1sidx =

  136.         sidx     = h->col_parity;

  137.     // FL -> FL & differ parity

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

  139.                !h->ref_list[1][0].mbaff) {

  140.         h->col_fieldoff = 2 * h->ref_list[1][0].reference - 3;

  141.     }

  142. 

  143.     if (h->slice_type_nos != AV_PICTURE_TYPE_B || h->direct_spatial_mv_pred)

  144.         return;

  145. 

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

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

  148.         if (FRAME_MBAFF(h))

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

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

  151.                             field, 1);

  152.     }

  153. }

  154. 

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

  156.                                    int mb_y)

  157. {

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

  159.     int ref_field_picture = ref->field_picture;

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

  161. 

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

  163.         return;

  164. 

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

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

  167. 

  168.     ff_thread_await_progress(&ref->tf,

  169.                              FFMIN(16 * mb_y >> ref_field_picture,

  170.                                    ref_height - 1),

  171.                              ref_field_picture && ref_field);

  172. }

  173. 

  174. static void pred_spatial_direct_motion(H264Context *const h, int *mb_type)

  175. {

  176.     int b8_stride = 2;

  177.     int b4_stride = h->b_stride;

  178.     int mb_xy = h->mb_xy, mb_y = h->mb_y;

  179.     int mb_type_col[2];

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

  181.     const int8_t *l1ref0, *l1ref1;

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

  183.     unsigned int sub_mb_type = MB_TYPE_L0L1;

  184.     int i8, i4;

  185.     int ref[2];

  186.     int mv[2];

  187.     int list;

  188. 

  189.     assert(h->ref_list[1][0].reference & 3);

  190. 

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

  192.                            h->mb_y + !!IS_INTERLACED(*mb_type));

  193. 

  194. #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \

  195.                                 MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM)

  196. 

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

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

  199.         int left_ref     = h->ref_cache[list][scan8[0] - 1];

  200.         int top_ref      = h->ref_cache[list][scan8[0] - 8];

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

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

  203.         if (refc == PART_NOT_AVAILABLE) {

  204.             refc = h->ref_cache[list][scan8[0] - 8 - 1];

  205.             C    = h->mv_cache[list][scan8[0]  - 8 - 1];

  206.         }

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

  208.                            (unsigned)top_ref,

  209.                            (unsigned)refc);

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

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

  212.              * cannot happen for direct blocks. */

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

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

  215. 

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

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

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

  219. 

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

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

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

  223.             } else {

  224.                 assert(match_count == 1);

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

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

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

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

  229.                 else

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

  231.             }

  232.         } else {

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

  234.             mv[list]  = 0;

  235.             ref[list] = -1;

  236.             if (!is_b8x8)

  237.                 *mb_type &= mask;

  238.             sub_mb_type &= mask;

  239.         }

  240.     }

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

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

  243.         if (!is_b8x8)

  244.             *mb_type |= MB_TYPE_L0L1;

  245.         sub_mb_type |= MB_TYPE_L0L1;

  246.     }

  247. 

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

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

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

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

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

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

  254.                                  MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  255.                    MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  256.         return;

  257.     }

  258. 

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

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

  261.             mb_y  = (h->mb_y & ~1) + h->col_parity;

  262.             mb_xy = h->mb_x +

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

  264.             b8_stride = 0;

  265.         } else {

  266.             mb_y  += h->col_fieldoff;

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

  268.         }

  269.         goto single_col;

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

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

  272.             mb_y           =  h->mb_y & ~1;

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

  274.             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];

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

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

  277.             b4_stride     *= 6;

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

  279.                 IS_INTERLACED(mb_type_col[1])) {

  280.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  281.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  282.             }

  283. 

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

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

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

  287.                 !is_b8x8) {

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

  289.             } else {

  290.                 *mb_type |= MB_TYPE_8x8;

  291.             }

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

  293. single_col:

  294.             mb_type_col[0] =

  295.             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];

  296. 

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

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

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

  300.             } else if (!is_b8x8 &&

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

  302.                 *mb_type |= MB_TYPE_DIRECT2 |

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

  304.             } else {

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

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

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

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

  309.                 }

  310.                 *mb_type |= MB_TYPE_8x8;

  311.             }

  312.         }

  313.     }

  314. 

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

  316. 

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

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

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

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

  321.     if (!b8_stride) {

  322.         if (h->mb_y & 1) {

  323.             l1ref0 += 2;

  324.             l1ref1 += 2;

  325.             l1mv0  += 2 * b4_stride;

  326.             l1mv1  += 2 * b4_stride;

  327.         }

  328.     }

  329. 

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

  331.         int n = 0;

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

  333.             int x8  = i8 & 1;

  334.             int y8  = i8 >> 1;

  335.             int xy8 = x8     + y8 * b8_stride;

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

  337.             int a, b;

  338. 

  339.             if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))

  340.                 continue;

  341.             h->sub_mb_type[i8] = sub_mb_type;

  342. 

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

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

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

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

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

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

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

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

  351.                  (l1ref0[xy8] < 0 &&

  352.                   l1ref1[xy8] == 0 &&

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

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

  355.                 a =

  356.                 b = 0;

  357.                 if (ref[0] > 0)

  358.                     a = mv[0];

  359.                 if (ref[1] > 0)

  360.                     b = mv[1];

  361.                 n++;

  362.             } else {

  363.                 a = mv[0];

  364.                 b = mv[1];

  365.             }

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

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

  368.         }

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

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

  371.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  372.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

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

  374.         int a, b;

  375. 

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

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

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

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

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

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

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

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

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

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

  386.             a = b = 0;

  387.             if (ref[0] > 0)

  388.                 a = mv[0];

  389.             if (ref[1] > 0)

  390.                 b = mv[1];

  391.         } else {

  392.             a = mv[0];

  393.             b = mv[1];

  394.         }

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

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

  397.     } else {

  398.         int n = 0;

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

  400.             const int x8 = i8 & 1;

  401.             const int y8 = i8 >> 1;

  402. 

  403.             if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))

  404.                 continue;

  405.             h->sub_mb_type[i8] = sub_mb_type;

  406. 

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

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

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

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

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

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

  413. 

  414.             assert(b8_stride == 2);

  415.             /* col_zero_flag */

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

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

  418.                  (l1ref0[i8] < 0 &&

  419.                   l1ref1[i8] == 0 &&

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

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

  422.                 if (IS_SUB_8X8(sub_mb_type)) {

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

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

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

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

  427.                                            8, 0, 4);

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

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

  430.                                            8, 0, 4);

  431.                         n += 4;

  432.                     }

  433.                 } else {

  434.                     int m = 0;

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

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

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

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

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

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

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

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

  443.                             m++;

  444.                         }

  445.                     }

  446.                     if (!(m & 3))

  447.                         h->sub_mb_type[i8] += MB_TYPE_16x16 - MB_TYPE_8x8;

  448.                     n += m;

  449.                 }

  450.             }

  451.         }

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

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

  454.                                      MB_TYPE_P1L0 | MB_TYPE_P1L1)) |

  455.                        MB_TYPE_16x16 | MB_TYPE_DIRECT2;

  456.     }

  457. }

  458. 

  459. static void pred_temp_direct_motion(H264Context *const h, int *mb_type)

  460. {

  461.     int b8_stride = 2;

  462.     int b4_stride = h->b_stride;

  463.     int mb_xy = h->mb_xy, mb_y = h->mb_y;

  464.     int mb_type_col[2];

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

  466.     const int8_t *l1ref0, *l1ref1;

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

  468.     unsigned int sub_mb_type;

  469.     int i8, i4;

  470. 

  471.     assert(h->ref_list[1][0].reference & 3);

  472. 

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

  474.                            h->mb_y + !!IS_INTERLACED(*mb_type));

  475. 

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

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

  478.             mb_y  = (h->mb_y & ~1) + h->col_parity;

  479.             mb_xy = h->mb_x +

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

  481.             b8_stride = 0;

  482.         } else {

  483.             mb_y  += h->col_fieldoff;

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

  485.         }

  486.         goto single_col;

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

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

  489.             mb_y           = h->mb_y & ~1;

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

  491.             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];

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

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

  494.             b4_stride     *= 6;

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

  496.                 IS_INTERLACED(mb_type_col[1])) {

  497.                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;

  498.                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;

  499.             }

  500. 

  501.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  502.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

  503. 

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

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

  506.                 !is_b8x8) {

  507.                 *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 |

  508.                             MB_TYPE_DIRECT2;                /* B_16x8 */

  509.             } else {

  510.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  511.             }

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

  513. single_col:

  514.             mb_type_col[0]     =

  515.                 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];

  516. 

  517.             sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  518.                           MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */

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

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

  521.                             MB_TYPE_DIRECT2;                /* B_16x16 */

  522.             } else if (!is_b8x8 &&

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

  524.                 *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 |

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

  526.             } else {

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

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

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

  530.                     sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |

  531.                                   MB_TYPE_DIRECT2;          /* B_SUB_4x4 */

  532.                 }

  533.                 *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;

  534.             }

  535.         }

  536.     }

  537. 

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

  539. 

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

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

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

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

  544.     if (!b8_stride) {

  545.         if (h->mb_y & 1) {

  546.             l1ref0 += 2;

  547.             l1ref1 += 2;

  548.             l1mv0  += 2 * b4_stride;

  549.             l1mv1  += 2 * b4_stride;

  550.         }

  551.     }

  552. 

  553.     {

  554.         const int *map_col_to_list0[2] = { h->map_col_to_list0[0],

  555.                                            h->map_col_to_list0[1] };

  556.         const int *dist_scale_factor = h->dist_scale_factor;

  557.         int ref_offset;

  558. 

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

  560.             map_col_to_list0[0] = h->map_col_to_list0_field[h->mb_y & 1][0];

  561.             map_col_to_list0[1] = h->map_col_to_list0_field[h->mb_y & 1][1];

  562.             dist_scale_factor   = h->dist_scale_factor_field[h->mb_y & 1];

  563.         }

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

  565. 

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

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

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

  569. 

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

  571.                 const int x8 = i8 & 1;

  572.                 const int y8 = i8 >> 1;

  573.                 int ref0, scale;

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

  575. 

  576.                 if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))

  577.                     continue;

  578.                 h->sub_mb_type[i8] = sub_mb_type;

  579. 

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

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

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

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

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

  585.                     continue;

  586.                 }

  587. 

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

  589.                 if (ref0 >= 0)

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

  591.                 else {

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

  593.                                                ref_offset];

  594.                     l1mv = l1mv1;

  595.                 }

  596.                 scale = dist_scale_factor[ref0];

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

  598.                                ref0, 1);

  599. 

  600.                 {

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

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

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

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

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

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

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

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

  609.                 }

  610.             }

  611.             return;

  612.         }

  613. 

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

  615. 

  616.         if (IS_16X16(*mb_type)) {

  617.             int ref, mv0, mv1;

  618. 

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

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

  621.                 ref = mv0 = mv1 = 0;

  622.             } else {

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

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

  625.                 const int scale = dist_scale_factor[ref0];

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

  627.                 int mv_l0[2];

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

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

  630.                 ref      = ref0;

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

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

  633.             }

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

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

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

  637.         } else {

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

  639.                 const int x8 = i8 & 1;

  640.                 const int y8 = i8 >> 1;

  641.                 int ref0, scale;

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

  643. 

  644.                 if (is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))

  645.                     continue;

  646.                 h->sub_mb_type[i8] = sub_mb_type;

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

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

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

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

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

  652.                     continue;

  653.                 }

  654. 

  655.                 assert(b8_stride == 2);

  656.                 ref0 = l1ref0[i8];

  657.                 if (ref0 >= 0)

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

  659.                 else {

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

  661.                     l1mv = l1mv1;

  662.                 }

  663.                 scale = dist_scale_factor[ref0];

  664. 

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

  666.                                ref0, 1);

  667.                 if (IS_SUB_8X8(sub_mb_type)) {

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

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

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

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

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

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

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

  675.                 } else {

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

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

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

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

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

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

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

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

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

  685.                     }

  686.                 }

  687.             }

  688.         }

  689.     }

  690. }

  691. 

  692. void ff_h264_pred_direct_motion(H264Context *const h, int *mb_type)

  693. {

  694.     if (h->direct_spatial_mv_pred)

  695.         pred_spatial_direct_motion(h, mb_type);

  696.     else

  697.         pred_temp_direct_motion(h, mb_type);

  698. }