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.
30552 Commits
32 Branches
283MB
Tree: b2e56e08c9
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'b2e56e08c9'
${ noResults }
FFmpeg/libavcodec/vp8.c

1834 lines
66KB
Raw Blame History 

  1. /**

  2.  * VP8 compatible video decoder

  3.  *

  4.  * Copyright (C) 2010 David Conrad

  5.  * Copyright (C) 2010 Ronald S. Bultje

  6.  * Copyright (C) 2010 Jason Garrett-Glaser

  7.  *

  8.  * This file is part of Libav.

  9.  *

  10.  * Libav is free software; you can redistribute it and/or

  11.  * modify it under the terms of the GNU Lesser General Public

  12.  * License as published by the Free Software Foundation; either

  13.  * version 2.1 of the License, or (at your option) any later version.

  14.  *

  15.  * Libav is distributed in the hope that it will be useful,

  16.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  17.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  18.  * Lesser General Public License for more details.

  19.  *

  20.  * You should have received a copy of the GNU Lesser General Public

  21.  * License along with Libav; if not, write to the Free Software

  22.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  23.  */

  24. 

  25. #include "libavutil/imgutils.h"

  26. #include "avcodec.h"

  27. #include "vp8.h"

  28. #include "vp8data.h"

  29. #include "rectangle.h"

  30. #include "thread.h"

  31. 

  32. #if ARCH_ARM

  33. #   include "arm/vp8.h"

  34. #endif

  35. 

  36. static void free_buffers(VP8Context *s)

  37. {

  38.     av_freep(&s->macroblocks_base);

  39.     av_freep(&s->filter_strength);

  40.     av_freep(&s->intra4x4_pred_mode_top);

  41.     av_freep(&s->top_nnz);

  42.     av_freep(&s->edge_emu_buffer);

  43.     av_freep(&s->top_border);

  44. 

  45.     s->macroblocks = NULL;

  46. }

  47. 

  48. static int vp8_alloc_frame(VP8Context *s, AVFrame *f)

  49. {

  50.     int ret;

  51.     if ((ret = ff_thread_get_buffer(s->avctx, f)) < 0)

  52.         return ret;

  53.     if (!s->maps_are_invalid && s->num_maps_to_be_freed) {

  54.         f->ref_index[0] = s->segmentation_maps[--s->num_maps_to_be_freed];

  55.     } else if (!(f->ref_index[0] = av_mallocz(s->mb_width * s->mb_height))) {

  56.         ff_thread_release_buffer(s->avctx, f);

  57.         return AVERROR(ENOMEM);

  58.     }

  59.     return 0;

  60. }

  61. 

  62. static void vp8_release_frame(VP8Context *s, AVFrame *f, int is_close)

  63. {

  64.     if (!is_close) {

  65.         if (f->ref_index[0]) {

  66.             assert(s->num_maps_to_be_freed < FF_ARRAY_ELEMS(s->segmentation_maps));

  67.             s->segmentation_maps[s->num_maps_to_be_freed++] = f->ref_index[0];

  68.             f->ref_index[0] = NULL;

  69.         }

  70.     } else {

  71.         av_freep(&f->ref_index[0]);

  72.     }

  73.     ff_thread_release_buffer(s->avctx, f);

  74. }

  75. 

  76. static void vp8_decode_flush_impl(AVCodecContext *avctx, int force, int is_close)

  77. {

  78.     VP8Context *s = avctx->priv_data;

  79.     int i;

  80. 

  81.     if (!avctx->is_copy || force) {

  82.         for (i = 0; i < 5; i++)

  83.             if (s->frames[i].data[0])

  84.                 vp8_release_frame(s, &s->frames[i], is_close);

  85.     }

  86.     memset(s->framep, 0, sizeof(s->framep));

  87. 

  88.     free_buffers(s);

  89.     s->maps_are_invalid = 1;

  90. }

  91. 

  92. static void vp8_decode_flush(AVCodecContext *avctx)

  93. {

  94.     vp8_decode_flush_impl(avctx, 0, 0);

  95. }

  96. 

  97. static int update_dimensions(VP8Context *s, int width, int height)

  98. {

  99.     if (width  != s->avctx->width ||

  100.         height != s->avctx->height) {

  101.         if (av_image_check_size(width, height, 0, s->avctx))

  102.             return AVERROR_INVALIDDATA;

  103. 

  104.         vp8_decode_flush_impl(s->avctx, 1, 0);

  105. 

  106.         avcodec_set_dimensions(s->avctx, width, height);

  107.     }

  108. 

  109.     s->mb_width  = (s->avctx->coded_width +15) / 16;

  110.     s->mb_height = (s->avctx->coded_height+15) / 16;

  111. 

  112.     s->macroblocks_base        = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks));

  113.     s->filter_strength         = av_mallocz(s->mb_width*sizeof(*s->filter_strength));

  114.     s->intra4x4_pred_mode_top  = av_mallocz(s->mb_width*4);

  115.     s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));

  116.     s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));

  117. 

  118.     if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||

  119.         !s->top_nnz || !s->top_border)

  120.         return AVERROR(ENOMEM);

  121. 

  122.     s->macroblocks        = s->macroblocks_base + 1;

  123. 

  124.     return 0;

  125. }

  126. 

  127. static void parse_segment_info(VP8Context *s)

  128. {

  129.     VP56RangeCoder *c = &s->c;

  130.     int i;

  131. 

  132.     s->segmentation.update_map = vp8_rac_get(c);

  133. 

  134.     if (vp8_rac_get(c)) { // update segment feature data

  135.         s->segmentation.absolute_vals = vp8_rac_get(c);

  136. 

  137.         for (i = 0; i < 4; i++)

  138.             s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);

  139. 

  140.         for (i = 0; i < 4; i++)

  141.             s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);

  142.     }

  143.     if (s->segmentation.update_map)

  144.         for (i = 0; i < 3; i++)

  145.             s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;

  146. }

  147. 

  148. static void update_lf_deltas(VP8Context *s)

  149. {

  150.     VP56RangeCoder *c = &s->c;

  151.     int i;

  152. 

  153.     for (i = 0; i < 4; i++)

  154.         s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);

  155. 

  156.     for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++)

  157.         s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);

  158. }

  159. 

  160. static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)

  161. {

  162.     const uint8_t *sizes = buf;

  163.     int i;

  164. 

  165.     s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);

  166. 

  167.     buf      += 3*(s->num_coeff_partitions-1);

  168.     buf_size -= 3*(s->num_coeff_partitions-1);

  169.     if (buf_size < 0)

  170.         return -1;

  171. 

  172.     for (i = 0; i < s->num_coeff_partitions-1; i++) {

  173.         int size = AV_RL24(sizes + 3*i);

  174.         if (buf_size - size < 0)

  175.             return -1;

  176. 

  177.         ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size);

  178.         buf      += size;

  179.         buf_size -= size;

  180.     }

  181.     ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);

  182. 

  183.     return 0;

  184. }

  185. 

  186. static void get_quants(VP8Context *s)

  187. {

  188.     VP56RangeCoder *c = &s->c;

  189.     int i, base_qi;

  190. 

  191.     int yac_qi     = vp8_rac_get_uint(c, 7);

  192.     int ydc_delta  = vp8_rac_get_sint(c, 4);

  193.     int y2dc_delta = vp8_rac_get_sint(c, 4);

  194.     int y2ac_delta = vp8_rac_get_sint(c, 4);

  195.     int uvdc_delta = vp8_rac_get_sint(c, 4);

  196.     int uvac_delta = vp8_rac_get_sint(c, 4);

  197. 

  198.     for (i = 0; i < 4; i++) {

  199.         if (s->segmentation.enabled) {

  200.             base_qi = s->segmentation.base_quant[i];

  201.             if (!s->segmentation.absolute_vals)

  202.                 base_qi += yac_qi;

  203.         } else

  204.             base_qi = yac_qi;

  205. 

  206.         s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip_uintp2(base_qi + ydc_delta , 7)];

  207.         s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip_uintp2(base_qi             , 7)];

  208.         s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip_uintp2(base_qi + y2dc_delta, 7)];

  209.         s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip_uintp2(base_qi + y2ac_delta, 7)] / 100;

  210.         s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip_uintp2(base_qi + uvdc_delta, 7)];

  211.         s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip_uintp2(base_qi + uvac_delta, 7)];

  212. 

  213.         s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);

  214.         s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);

  215.     }

  216. }

  217. 

  218. /**

  219.  * Determine which buffers golden and altref should be updated with after this frame.

  220.  * The spec isn't clear here, so I'm going by my understanding of what libvpx does

  221.  *

  222.  * Intra frames update all 3 references

  223.  * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set

  224.  * If the update (golden|altref) flag is set, it's updated with the current frame

  225.  *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.

  226.  * If the flag is not set, the number read means:

  227.  *      0: no update

  228.  *      1: VP56_FRAME_PREVIOUS

  229.  *      2: update golden with altref, or update altref with golden

  230.  */

  231. static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)

  232. {

  233.     VP56RangeCoder *c = &s->c;

  234. 

  235.     if (update)

  236.         return VP56_FRAME_CURRENT;

  237. 

  238.     switch (vp8_rac_get_uint(c, 2)) {

  239.     case 1:

  240.         return VP56_FRAME_PREVIOUS;

  241.     case 2:

  242.         return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;

  243.     }

  244.     return VP56_FRAME_NONE;

  245. }

  246. 

  247. static void update_refs(VP8Context *s)

  248. {

  249.     VP56RangeCoder *c = &s->c;

  250. 

  251.     int update_golden = vp8_rac_get(c);

  252.     int update_altref = vp8_rac_get(c);

  253. 

  254.     s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);

  255.     s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);

  256. }

  257. 

  258. static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)

  259. {

  260.     VP56RangeCoder *c = &s->c;

  261.     int header_size, hscale, vscale, i, j, k, l, m, ret;

  262.     int width  = s->avctx->width;

  263.     int height = s->avctx->height;

  264. 

  265.     s->keyframe  = !(buf[0] & 1);

  266.     s->profile   =  (buf[0]>>1) & 7;

  267.     s->invisible = !(buf[0] & 0x10);

  268.     header_size  = AV_RL24(buf) >> 5;

  269.     buf      += 3;

  270.     buf_size -= 3;

  271. 

  272.     if (s->profile > 3)

  273.         av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);

  274. 

  275.     if (!s->profile)

  276.         memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));

  277.     else    // profile 1-3 use bilinear, 4+ aren't defined so whatever

  278.         memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));

  279. 

  280.     if (header_size > buf_size - 7*s->keyframe) {

  281.         av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");

  282.         return AVERROR_INVALIDDATA;

  283.     }

  284. 

  285.     if (s->keyframe) {

  286.         if (AV_RL24(buf) != 0x2a019d) {

  287.             av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));

  288.             return AVERROR_INVALIDDATA;

  289.         }

  290.         width  = AV_RL16(buf+3) & 0x3fff;

  291.         height = AV_RL16(buf+5) & 0x3fff;

  292.         hscale = buf[4] >> 6;

  293.         vscale = buf[6] >> 6;

  294.         buf      += 7;

  295.         buf_size -= 7;

  296. 

  297.         if (hscale || vscale)

  298.             av_log_missing_feature(s->avctx, "Upscaling", 1);

  299. 

  300.         s->update_golden = s->update_altref = VP56_FRAME_CURRENT;

  301.         for (i = 0; i < 4; i++)

  302.             for (j = 0; j < 16; j++)

  303.                 memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],

  304.                        sizeof(s->prob->token[i][j]));

  305.         memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));

  306.         memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));

  307.         memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));

  308.         memset(&s->segmentation, 0, sizeof(s->segmentation));

  309.     }

  310. 

  311.     if (!s->macroblocks_base || /* first frame */

  312.         width != s->avctx->width || height != s->avctx->height) {

  313.         if ((ret = update_dimensions(s, width, height)) < 0)

  314.             return ret;

  315.     }

  316. 

  317.     ff_vp56_init_range_decoder(c, buf, header_size);

  318.     buf      += header_size;

  319.     buf_size -= header_size;

  320. 

  321.     if (s->keyframe) {

  322.         if (vp8_rac_get(c))

  323.             av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");

  324.         vp8_rac_get(c); // whether we can skip clamping in dsp functions

  325.     }

  326. 

  327.     if ((s->segmentation.enabled = vp8_rac_get(c)))

  328.         parse_segment_info(s);

  329.     else

  330.         s->segmentation.update_map = 0; // FIXME: move this to some init function?

  331. 

  332.     s->filter.simple    = vp8_rac_get(c);

  333.     s->filter.level     = vp8_rac_get_uint(c, 6);

  334.     s->filter.sharpness = vp8_rac_get_uint(c, 3);

  335. 

  336.     if ((s->lf_delta.enabled = vp8_rac_get(c)))

  337.         if (vp8_rac_get(c))

  338.             update_lf_deltas(s);

  339. 

  340.     if (setup_partitions(s, buf, buf_size)) {

  341.         av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");

  342.         return AVERROR_INVALIDDATA;

  343.     }

  344. 

  345.     get_quants(s);

  346. 

  347.     if (!s->keyframe) {

  348.         update_refs(s);

  349.         s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);

  350.         s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);

  351.     }

  352. 

  353.     // if we aren't saving this frame's probabilities for future frames,

  354.     // make a copy of the current probabilities

  355.     if (!(s->update_probabilities = vp8_rac_get(c)))

  356.         s->prob[1] = s->prob[0];

  357. 

  358.     s->update_last = s->keyframe || vp8_rac_get(c);

  359. 

  360.     for (i = 0; i < 4; i++)

  361.         for (j = 0; j < 8; j++)

  362.             for (k = 0; k < 3; k++)

  363.                 for (l = 0; l < NUM_DCT_TOKENS-1; l++)

  364.                     if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {

  365.                         int prob = vp8_rac_get_uint(c, 8);

  366.                         for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)

  367.                             s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;

  368.                     }

  369. 

  370.     if ((s->mbskip_enabled = vp8_rac_get(c)))

  371.         s->prob->mbskip = vp8_rac_get_uint(c, 8);

  372. 

  373.     if (!s->keyframe) {

  374.         s->prob->intra  = vp8_rac_get_uint(c, 8);

  375.         s->prob->last   = vp8_rac_get_uint(c, 8);

  376.         s->prob->golden = vp8_rac_get_uint(c, 8);

  377. 

  378.         if (vp8_rac_get(c))

  379.             for (i = 0; i < 4; i++)

  380.                 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);

  381.         if (vp8_rac_get(c))

  382.             for (i = 0; i < 3; i++)

  383.                 s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);

  384. 

  385.         // 17.2 MV probability update

  386.         for (i = 0; i < 2; i++)

  387.             for (j = 0; j < 19; j++)

  388.                 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))

  389.                     s->prob->mvc[i][j] = vp8_rac_get_nn(c);

  390.     }

  391. 

  392.     return 0;

  393. }

  394. 

  395. static av_always_inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src)

  396. {

  397.     dst->x = av_clip(src->x, s->mv_min.x, s->mv_max.x);

  398.     dst->y = av_clip(src->y, s->mv_min.y, s->mv_max.y);

  399. }

  400. 

  401. /**

  402.  * Motion vector coding, 17.1.

  403.  */

  404. static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)

  405. {

  406.     int bit, x = 0;

  407. 

  408.     if (vp56_rac_get_prob_branchy(c, p[0])) {

  409.         int i;

  410. 

  411.         for (i = 0; i < 3; i++)

  412.             x += vp56_rac_get_prob(c, p[9 + i]) << i;

  413.         for (i = 9; i > 3; i--)

  414.             x += vp56_rac_get_prob(c, p[9 + i]) << i;

  415.         if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))

  416.             x += 8;

  417.     } else {

  418.         // small_mvtree

  419.         const uint8_t *ps = p+2;

  420.         bit = vp56_rac_get_prob(c, *ps);

  421.         ps += 1 + 3*bit;

  422.         x  += 4*bit;

  423.         bit = vp56_rac_get_prob(c, *ps);

  424.         ps += 1 + bit;

  425.         x  += 2*bit;

  426.         x  += vp56_rac_get_prob(c, *ps);

  427.     }

  428. 

  429.     return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;

  430. }

  431. 

  432. static av_always_inline

  433. const uint8_t *get_submv_prob(uint32_t left, uint32_t top)

  434. {

  435.     if (left == top)

  436.         return vp8_submv_prob[4-!!left];

  437.     if (!top)

  438.         return vp8_submv_prob[2];

  439.     return vp8_submv_prob[1-!!left];

  440. }

  441. 

  442. /**

  443.  * Split motion vector prediction, 16.4.

  444.  * @returns the number of motion vectors parsed (2, 4 or 16)

  445.  */

  446. static av_always_inline

  447. int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)

  448. {

  449.     int part_idx;

  450.     int n, num;

  451.     VP8Macroblock *top_mb  = &mb[2];

  452.     VP8Macroblock *left_mb = &mb[-1];

  453.     const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],

  454.                   *mbsplits_top = vp8_mbsplits[top_mb->partitioning],

  455.                   *mbsplits_cur, *firstidx;

  456.     VP56mv *top_mv  = top_mb->bmv;

  457.     VP56mv *left_mv = left_mb->bmv;

  458.     VP56mv *cur_mv  = mb->bmv;

  459. 

  460.     if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {

  461.         if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {

  462.             part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);

  463.         } else {

  464.             part_idx = VP8_SPLITMVMODE_8x8;

  465.         }

  466.     } else {

  467.         part_idx = VP8_SPLITMVMODE_4x4;

  468.     }

  469. 

  470.     num = vp8_mbsplit_count[part_idx];

  471.     mbsplits_cur = vp8_mbsplits[part_idx],

  472.     firstidx = vp8_mbfirstidx[part_idx];

  473.     mb->partitioning = part_idx;

  474. 

  475.     for (n = 0; n < num; n++) {

  476.         int k = firstidx[n];

  477.         uint32_t left, above;

  478.         const uint8_t *submv_prob;

  479. 

  480.         if (!(k & 3))

  481.             left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);

  482.         else

  483.             left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);

  484.         if (k <= 3)

  485.             above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);

  486.         else

  487.             above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);

  488. 

  489.         submv_prob = get_submv_prob(left, above);

  490. 

  491.         if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {

  492.             if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {

  493.                 if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {

  494.                     mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);

  495.                     mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);

  496.                 } else {

  497.                     AV_ZERO32(&mb->bmv[n]);

  498.                 }

  499.             } else {

  500.                 AV_WN32A(&mb->bmv[n], above);

  501.             }

  502.         } else {

  503.             AV_WN32A(&mb->bmv[n], left);

  504.         }

  505.     }

  506. 

  507.     return num;

  508. }

  509. 

  510. static av_always_inline

  511. void decode_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y)

  512. {

  513.     VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,

  514.                                   mb - 1 /* left */,

  515.                                   mb + 1 /* top-left */ };

  516.     enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };

  517.     enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT };

  518.     int idx = CNT_ZERO;

  519.     int cur_sign_bias = s->sign_bias[mb->ref_frame];

  520.     int8_t *sign_bias = s->sign_bias;

  521.     VP56mv near_mv[4];

  522.     uint8_t cnt[4] = { 0 };

  523.     VP56RangeCoder *c = &s->c;

  524. 

  525.     AV_ZERO32(&near_mv[0]);

  526.     AV_ZERO32(&near_mv[1]);

  527.     AV_ZERO32(&near_mv[2]);

  528. 

  529.     /* Process MB on top, left and top-left */

  530.     #define MV_EDGE_CHECK(n)\

  531.     {\

  532.         VP8Macroblock *edge = mb_edge[n];\

  533.         int edge_ref = edge->ref_frame;\

  534.         if (edge_ref != VP56_FRAME_CURRENT) {\

  535.             uint32_t mv = AV_RN32A(&edge->mv);\

  536.             if (mv) {\

  537.                 if (cur_sign_bias != sign_bias[edge_ref]) {\

  538.                     /* SWAR negate of the values in mv. */\

  539.                     mv = ~mv;\

  540.                     mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\

  541.                 }\

  542.                 if (!n || mv != AV_RN32A(&near_mv[idx]))\

  543.                     AV_WN32A(&near_mv[++idx], mv);\

  544.                 cnt[idx]      += 1 + (n != 2);\

  545.             } else\

  546.                 cnt[CNT_ZERO] += 1 + (n != 2);\

  547.         }\

  548.     }

  549. 

  550.     MV_EDGE_CHECK(0)

  551.     MV_EDGE_CHECK(1)

  552.     MV_EDGE_CHECK(2)

  553. 

  554.     mb->partitioning = VP8_SPLITMVMODE_NONE;

  555.     if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_ZERO]][0])) {

  556.         mb->mode = VP8_MVMODE_MV;

  557. 

  558.         /* If we have three distinct MVs, merge first and last if they're the same */

  559.         if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1 + VP8_EDGE_TOP]) == AV_RN32A(&near_mv[1 + VP8_EDGE_TOPLEFT]))

  560.             cnt[CNT_NEAREST] += 1;

  561. 

  562.         /* Swap near and nearest if necessary */

  563.         if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {

  564.             FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);

  565.             FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);

  566.         }

  567. 

  568.         if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAREST]][1])) {

  569.             if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAR]][2])) {

  570. 

  571.                 /* Choose the best mv out of 0,0 and the nearest mv */

  572.                 clamp_mv(s, &mb->mv, &near_mv[CNT_ZERO + (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])]);

  573.                 cnt[CNT_SPLITMV] = ((mb_edge[VP8_EDGE_LEFT]->mode    == VP8_MVMODE_SPLIT) +

  574.                                     (mb_edge[VP8_EDGE_TOP]->mode     == VP8_MVMODE_SPLIT)) * 2 +

  575.                                     (mb_edge[VP8_EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);

  576. 

  577.                 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_SPLITMV]][3])) {

  578.                     mb->mode = VP8_MVMODE_SPLIT;

  579.                     mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];

  580.                 } else {

  581.                     mb->mv.y += read_mv_component(c, s->prob->mvc[0]);

  582.                     mb->mv.x += read_mv_component(c, s->prob->mvc[1]);

  583.                     mb->bmv[0] = mb->mv;

  584.                 }

  585.             } else {

  586.                 clamp_mv(s, &mb->mv, &near_mv[CNT_NEAR]);

  587.                 mb->bmv[0] = mb->mv;

  588.             }

  589.         } else {

  590.             clamp_mv(s, &mb->mv, &near_mv[CNT_NEAREST]);

  591.             mb->bmv[0] = mb->mv;

  592.         }

  593.     } else {

  594.         mb->mode = VP8_MVMODE_ZERO;

  595.         AV_ZERO32(&mb->mv);

  596.         mb->bmv[0] = mb->mv;

  597.     }

  598. }

  599. 

  600. static av_always_inline

  601. void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,

  602.                            int mb_x, int keyframe)

  603. {

  604.     uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  605.     if (keyframe) {

  606.         int x, y;

  607.         uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;

  608.         uint8_t* const left = s->intra4x4_pred_mode_left;

  609.         for (y = 0; y < 4; y++) {

  610.             for (x = 0; x < 4; x++) {

  611.                 const uint8_t *ctx;

  612.                 ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];

  613.                 *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);

  614.                 left[y] = top[x] = *intra4x4;

  615.                 intra4x4++;

  616.             }

  617.         }

  618.     } else {

  619.         int i;

  620.         for (i = 0; i < 16; i++)

  621.             intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);

  622.     }

  623. }

  624. 

  625. static av_always_inline

  626. void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment, uint8_t *ref)

  627. {

  628.     VP56RangeCoder *c = &s->c;

  629. 

  630.     if (s->segmentation.update_map)

  631.         *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

  632.     else

  633.         *segment = ref ? *ref : *segment;

  634.     s->segment = *segment;

  635. 

  636.     mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;

  637. 

  638.     if (s->keyframe) {

  639.         mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);

  640. 

  641.         if (mb->mode == MODE_I4x4) {

  642.             decode_intra4x4_modes(s, c, mb_x, 1);

  643.         } else {

  644.             const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;

  645.             AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);

  646.             AV_WN32A(s->intra4x4_pred_mode_left, modes);

  647.         }

  648. 

  649.         s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);

  650.         mb->ref_frame = VP56_FRAME_CURRENT;

  651.     } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {

  652.         // inter MB, 16.2

  653.         if (vp56_rac_get_prob_branchy(c, s->prob->last))

  654.             mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?

  655.                 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;

  656.         else

  657.             mb->ref_frame = VP56_FRAME_PREVIOUS;

  658.         s->ref_count[mb->ref_frame-1]++;

  659. 

  660.         // motion vectors, 16.3

  661.         decode_mvs(s, mb, mb_x, mb_y);

  662.     } else {

  663.         // intra MB, 16.1

  664.         mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);

  665. 

  666.         if (mb->mode == MODE_I4x4)

  667.             decode_intra4x4_modes(s, c, mb_x, 0);

  668. 

  669.         s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);

  670.         mb->ref_frame = VP56_FRAME_CURRENT;

  671.         mb->partitioning = VP8_SPLITMVMODE_NONE;

  672.         AV_ZERO32(&mb->bmv[0]);

  673.     }

  674. }

  675. 

  676. #ifndef decode_block_coeffs_internal

  677. /**

  678.  * @param c arithmetic bitstream reader context

  679.  * @param block destination for block coefficients

  680.  * @param probs probabilities to use when reading trees from the bitstream

  681.  * @param i initial coeff index, 0 unless a separate DC block is coded

  682.  * @param qmul array holding the dc/ac dequant factor at position 0/1

  683.  * @return 0 if no coeffs were decoded

  684.  *         otherwise, the index of the last coeff decoded plus one

  685.  */

  686. static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16],

  687.                                         uint8_t probs[16][3][NUM_DCT_TOKENS-1],

  688.                                         int i, uint8_t *token_prob, int16_t qmul[2])

  689. {

  690.     goto skip_eob;

  691.     do {

  692.         int coeff;

  693.         if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB

  694.             return i;

  695. 

  696. skip_eob:

  697.         if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0

  698.             if (++i == 16)

  699.                 return i; // invalid input; blocks should end with EOB

  700.             token_prob = probs[i][0];

  701.             goto skip_eob;

  702.         }

  703. 

  704.         if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1

  705.             coeff = 1;

  706.             token_prob = probs[i+1][1];

  707.         } else {

  708.             if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4

  709.                 coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);

  710.                 if (coeff)

  711.                     coeff += vp56_rac_get_prob(c, token_prob[5]);

  712.                 coeff += 2;

  713.             } else {

  714.                 // DCT_CAT*

  715.                 if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {

  716.                     if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1

  717.                         coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);

  718.                     } else {                                    // DCT_CAT2

  719.                         coeff  = 7;

  720.                         coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;

  721.                         coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);

  722.                     }

  723.                 } else {    // DCT_CAT3 and up

  724.                     int a = vp56_rac_get_prob(c, token_prob[8]);

  725.                     int b = vp56_rac_get_prob(c, token_prob[9+a]);

  726.                     int cat = (a<<1) + b;

  727.                     coeff  = 3 + (8<<cat);

  728.                     coeff += vp8_rac_get_coeff(c, ff_vp8_dct_cat_prob[cat]);

  729.                 }

  730.             }

  731.             token_prob = probs[i+1][2];

  732.         }

  733.         block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];

  734.     } while (++i < 16);

  735. 

  736.     return i;

  737. }

  738. #endif

  739. 

  740. /**

  741.  * @param c arithmetic bitstream reader context

  742.  * @param block destination for block coefficients

  743.  * @param probs probabilities to use when reading trees from the bitstream

  744.  * @param i initial coeff index, 0 unless a separate DC block is coded

  745.  * @param zero_nhood the initial prediction context for number of surrounding

  746.  *                   all-zero blocks (only left/top, so 0-2)

  747.  * @param qmul array holding the dc/ac dequant factor at position 0/1

  748.  * @return 0 if no coeffs were decoded

  749.  *         otherwise, the index of the last coeff decoded plus one

  750.  */

  751. static av_always_inline

  752. int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],

  753.                         uint8_t probs[16][3][NUM_DCT_TOKENS-1],

  754.                         int i, int zero_nhood, int16_t qmul[2])

  755. {

  756.     uint8_t *token_prob = probs[i][zero_nhood];

  757.     if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB

  758.         return 0;

  759.     return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);

  760. }

  761. 

  762. static av_always_inline

  763. void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,

  764.                       uint8_t t_nnz[9], uint8_t l_nnz[9])

  765. {

  766.     int i, x, y, luma_start = 0, luma_ctx = 3;

  767.     int nnz_pred, nnz, nnz_total = 0;

  768.     int segment = s->segment;

  769.     int block_dc = 0;

  770. 

  771.     if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  772.         nnz_pred = t_nnz[8] + l_nnz[8];

  773. 

  774.         // decode DC values and do hadamard

  775.         nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,

  776.                                   s->qmat[segment].luma_dc_qmul);

  777.         l_nnz[8] = t_nnz[8] = !!nnz;

  778.         if (nnz) {

  779.             nnz_total += nnz;

  780.             block_dc = 1;

  781.             if (nnz == 1)

  782.                 s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);

  783.             else

  784.                 s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);

  785.         }

  786.         luma_start = 1;

  787.         luma_ctx = 0;

  788.     }

  789. 

  790.     // luma blocks

  791.     for (y = 0; y < 4; y++)

  792.         for (x = 0; x < 4; x++) {

  793.             nnz_pred = l_nnz[y] + t_nnz[x];

  794.             nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,

  795.                                       nnz_pred, s->qmat[segment].luma_qmul);

  796.             // nnz+block_dc may be one more than the actual last index, but we don't care

  797.             s->non_zero_count_cache[y][x] = nnz + block_dc;

  798.             t_nnz[x] = l_nnz[y] = !!nnz;

  799.             nnz_total += nnz;

  800.         }

  801. 

  802.     // chroma blocks

  803.     // TODO: what to do about dimensions? 2nd dim for luma is x,

  804.     // but for chroma it's (y<<1)|x

  805.     for (i = 4; i < 6; i++)

  806.         for (y = 0; y < 2; y++)

  807.             for (x = 0; x < 2; x++) {

  808.                 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];

  809.                 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,

  810.                                           nnz_pred, s->qmat[segment].chroma_qmul);

  811.                 s->non_zero_count_cache[i][(y<<1)+x] = nnz;

  812.                 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;

  813.                 nnz_total += nnz;

  814.             }

  815. 

  816.     // if there were no coded coeffs despite the macroblock not being marked skip,

  817.     // we MUST not do the inner loop filter and should not do IDCT

  818.     // Since skip isn't used for bitstream prediction, just manually set it.

  819.     if (!nnz_total)

  820.         mb->skip = 1;

  821. }

  822. 

  823. static av_always_inline

  824. void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,

  825.                       int linesize, int uvlinesize, int simple)

  826. {

  827.     AV_COPY128(top_border, src_y + 15*linesize);

  828.     if (!simple) {

  829.         AV_COPY64(top_border+16, src_cb + 7*uvlinesize);

  830.         AV_COPY64(top_border+24, src_cr + 7*uvlinesize);

  831.     }

  832. }

  833. 

  834. static av_always_inline

  835. void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,

  836.                     int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,

  837.                     int simple, int xchg)

  838. {

  839.     uint8_t *top_border_m1 = top_border-32;     // for TL prediction

  840.     src_y  -=   linesize;

  841.     src_cb -= uvlinesize;

  842.     src_cr -= uvlinesize;

  843. 

  844. #define XCHG(a,b,xchg) do {                     \

  845.         if (xchg) AV_SWAP64(b,a);               \

  846.         else      AV_COPY64(b,a);               \

  847.     } while (0)

  848. 

  849.     XCHG(top_border_m1+8, src_y-8, xchg);

  850.     XCHG(top_border,      src_y,   xchg);

  851.     XCHG(top_border+8,    src_y+8, 1);

  852.     if (mb_x < mb_width-1)

  853.         XCHG(top_border+32, src_y+16, 1);

  854. 

  855.     // only copy chroma for normal loop filter

  856.     // or to initialize the top row to 127

  857.     if (!simple || !mb_y) {

  858.         XCHG(top_border_m1+16, src_cb-8, xchg);

  859.         XCHG(top_border_m1+24, src_cr-8, xchg);

  860.         XCHG(top_border+16,    src_cb, 1);

  861.         XCHG(top_border+24,    src_cr, 1);

  862.     }

  863. }

  864. 

  865. static av_always_inline

  866. int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)

  867. {

  868.     if (!mb_x) {

  869.         return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;

  870.     } else {

  871.         return mb_y ? mode : LEFT_DC_PRED8x8;

  872.     }

  873. }

  874. 

  875. static av_always_inline

  876. int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)

  877. {

  878.     if (!mb_x) {

  879.         return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;

  880.     } else {

  881.         return mb_y ? mode : HOR_PRED8x8;

  882.     }

  883. }

  884. 

  885. static av_always_inline

  886. int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)

  887. {

  888.     if (mode == DC_PRED8x8) {

  889.         return check_dc_pred8x8_mode(mode, mb_x, mb_y);

  890.     } else {

  891.         return mode;

  892.     }

  893. }

  894. 

  895. static av_always_inline

  896. int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)

  897. {

  898.     switch (mode) {

  899.     case DC_PRED8x8:

  900.         return check_dc_pred8x8_mode(mode, mb_x, mb_y);

  901.     case VERT_PRED8x8:

  902.         return !mb_y ? DC_127_PRED8x8 : mode;

  903.     case HOR_PRED8x8:

  904.         return !mb_x ? DC_129_PRED8x8 : mode;

  905.     case PLANE_PRED8x8 /*TM*/:

  906.         return check_tm_pred8x8_mode(mode, mb_x, mb_y);

  907.     }

  908.     return mode;

  909. }

  910. 

  911. static av_always_inline

  912. int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)

  913. {

  914.     if (!mb_x) {

  915.         return mb_y ? VERT_VP8_PRED : DC_129_PRED;

  916.     } else {

  917.         return mb_y ? mode : HOR_VP8_PRED;

  918.     }

  919. }

  920. 

  921. static av_always_inline

  922. int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)

  923. {

  924.     switch (mode) {

  925.     case VERT_PRED:

  926.         if (!mb_x && mb_y) {

  927.             *copy_buf = 1;

  928.             return mode;

  929.         }

  930.         /* fall-through */

  931.     case DIAG_DOWN_LEFT_PRED:

  932.     case VERT_LEFT_PRED:

  933.         return !mb_y ? DC_127_PRED : mode;

  934.     case HOR_PRED:

  935.         if (!mb_y) {

  936.             *copy_buf = 1;

  937.             return mode;

  938.         }

  939.         /* fall-through */

  940.     case HOR_UP_PRED:

  941.         return !mb_x ? DC_129_PRED : mode;

  942.     case TM_VP8_PRED:

  943.         return check_tm_pred4x4_mode(mode, mb_x, mb_y);

  944.     case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC

  945.     case DIAG_DOWN_RIGHT_PRED:

  946.     case VERT_RIGHT_PRED:

  947.     case HOR_DOWN_PRED:

  948.         if (!mb_y || !mb_x)

  949.             *copy_buf = 1;

  950.         return mode;

  951.     }

  952.     return mode;

  953. }

  954. 

  955. static av_always_inline

  956. void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  957.                    int mb_x, int mb_y)

  958. {

  959.     AVCodecContext *avctx = s->avctx;

  960.     int x, y, mode, nnz;

  961.     uint32_t tr;

  962. 

  963.     // for the first row, we need to run xchg_mb_border to init the top edge to 127

  964.     // otherwise, skip it if we aren't going to deblock

  965.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  966.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  967.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  968.                        s->filter.simple, 1);

  969. 

  970.     if (mb->mode < MODE_I4x4) {

  971.         if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested

  972.             mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);

  973.         } else {

  974.             mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);

  975.         }

  976.         s->hpc.pred16x16[mode](dst[0], s->linesize);

  977.     } else {

  978.         uint8_t *ptr = dst[0];

  979.         uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  980.         uint8_t tr_top[4] = { 127, 127, 127, 127 };

  981. 

  982.         // all blocks on the right edge of the macroblock use bottom edge

  983.         // the top macroblock for their topright edge

  984.         uint8_t *tr_right = ptr - s->linesize + 16;

  985. 

  986.         // if we're on the right edge of the frame, said edge is extended

  987.         // from the top macroblock

  988.         if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&

  989.             mb_x == s->mb_width-1) {

  990.             tr = tr_right[-1]*0x01010101u;

  991.             tr_right = (uint8_t *)&tr;

  992.         }

  993. 

  994.         if (mb->skip)

  995.             AV_ZERO128(s->non_zero_count_cache);

  996. 

  997.         for (y = 0; y < 4; y++) {

  998.             uint8_t *topright = ptr + 4 - s->linesize;

  999.             for (x = 0; x < 4; x++) {

  1000.                 int copy = 0, linesize = s->linesize;

  1001.                 uint8_t *dst = ptr+4*x;

  1002.                 DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];

  1003. 

  1004.                 if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1005.                     topright = tr_top;

  1006.                 } else if (x == 3)

  1007.                     topright = tr_right;

  1008. 

  1009.                 if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works

  1010.                     mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);

  1011.                     if (copy) {

  1012.                         dst = copy_dst + 12;

  1013.                         linesize = 8;

  1014.                         if (!(mb_y + y)) {

  1015.                             copy_dst[3] = 127U;

  1016.                             AV_WN32A(copy_dst+4, 127U * 0x01010101U);

  1017.                         } else {

  1018.                             AV_COPY32(copy_dst+4, ptr+4*x-s->linesize);

  1019.                             if (!(mb_x + x)) {

  1020.                                 copy_dst[3] = 129U;

  1021.                             } else {

  1022.                                 copy_dst[3] = ptr[4*x-s->linesize-1];

  1023.                             }

  1024.                         }

  1025.                         if (!(mb_x + x)) {

  1026.                             copy_dst[11] =

  1027.                             copy_dst[19] =

  1028.                             copy_dst[27] =

  1029.                             copy_dst[35] = 129U;

  1030.                         } else {

  1031.                             copy_dst[11] = ptr[4*x              -1];

  1032.                             copy_dst[19] = ptr[4*x+s->linesize  -1];

  1033.                             copy_dst[27] = ptr[4*x+s->linesize*2-1];

  1034.                             copy_dst[35] = ptr[4*x+s->linesize*3-1];

  1035.                         }

  1036.                     }

  1037.                 } else {

  1038.                     mode = intra4x4[x];

  1039.                 }

  1040.                 s->hpc.pred4x4[mode](dst, topright, linesize);

  1041.                 if (copy) {

  1042.                     AV_COPY32(ptr+4*x              , copy_dst+12);

  1043.                     AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);

  1044.                     AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);

  1045.                     AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);

  1046.                 }

  1047. 

  1048.                 nnz = s->non_zero_count_cache[y][x];

  1049.                 if (nnz) {

  1050.                     if (nnz == 1)

  1051.                         s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);

  1052.                     else

  1053.                         s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);

  1054.                 }

  1055.                 topright += 4;

  1056.             }

  1057. 

  1058.             ptr   += 4*s->linesize;

  1059.             intra4x4 += 4;

  1060.         }

  1061.     }

  1062. 

  1063.     if (avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1064.         mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);

  1065.     } else {

  1066.         mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);

  1067.     }

  1068.     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);

  1069.     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);

  1070. 

  1071.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  1072.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  1073.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  1074.                        s->filter.simple, 0);

  1075. }

  1076. 

  1077. static const uint8_t subpel_idx[3][8] = {

  1078.     { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,

  1079.                                 // also function pointer index

  1080.     { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required

  1081.     { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels

  1082. };

  1083. 

  1084. /**

  1085.  * luma MC function

  1086.  *

  1087.  * @param s VP8 decoding context

  1088.  * @param dst target buffer for block data at block position

  1089.  * @param ref reference picture buffer at origin (0, 0)

  1090.  * @param mv motion vector (relative to block position) to get pixel data from

  1091.  * @param x_off horizontal position of block from origin (0, 0)

  1092.  * @param y_off vertical position of block from origin (0, 0)

  1093.  * @param block_w width of block (16, 8 or 4)

  1094.  * @param block_h height of block (always same as block_w)

  1095.  * @param width width of src/dst plane data

  1096.  * @param height height of src/dst plane data

  1097.  * @param linesize size of a single line of plane data, including padding

  1098.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1099.  */

  1100. static av_always_inline

  1101. void vp8_mc_luma(VP8Context *s, uint8_t *dst, AVFrame *ref, const VP56mv *mv,

  1102.                  int x_off, int y_off, int block_w, int block_h,

  1103.                  int width, int height, int linesize,

  1104.                  vp8_mc_func mc_func[3][3])

  1105. {

  1106.     uint8_t *src = ref->data[0];

  1107. 

  1108.     if (AV_RN32A(mv)) {

  1109. 

  1110.         int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];

  1111.         int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];

  1112. 

  1113.         x_off += mv->x >> 2;

  1114.         y_off += mv->y >> 2;

  1115. 

  1116.         // edge emulation

  1117.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 4, 0);

  1118.         src += y_off * linesize + x_off;

  1119.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1120.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1121.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,

  1122.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1123.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1124.             src = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1125.         }

  1126.         mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);

  1127.     } else {

  1128.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0);

  1129.         mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1130.     }

  1131. }

  1132. 

  1133. /**

  1134.  * chroma MC function

  1135.  *

  1136.  * @param s VP8 decoding context

  1137.  * @param dst1 target buffer for block data at block position (U plane)

  1138.  * @param dst2 target buffer for block data at block position (V plane)

  1139.  * @param ref reference picture buffer at origin (0, 0)

  1140.  * @param mv motion vector (relative to block position) to get pixel data from

  1141.  * @param x_off horizontal position of block from origin (0, 0)

  1142.  * @param y_off vertical position of block from origin (0, 0)

  1143.  * @param block_w width of block (16, 8 or 4)

  1144.  * @param block_h height of block (always same as block_w)

  1145.  * @param width width of src/dst plane data

  1146.  * @param height height of src/dst plane data

  1147.  * @param linesize size of a single line of plane data, including padding

  1148.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1149.  */

  1150. static av_always_inline

  1151. void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, AVFrame *ref,

  1152.                    const VP56mv *mv, int x_off, int y_off,

  1153.                    int block_w, int block_h, int width, int height, int linesize,

  1154.                    vp8_mc_func mc_func[3][3])

  1155. {

  1156.     uint8_t *src1 = ref->data[1], *src2 = ref->data[2];

  1157. 

  1158.     if (AV_RN32A(mv)) {

  1159.         int mx = mv->x&7, mx_idx = subpel_idx[0][mx];

  1160.         int my = mv->y&7, my_idx = subpel_idx[0][my];

  1161. 

  1162.         x_off += mv->x >> 3;

  1163.         y_off += mv->y >> 3;

  1164. 

  1165.         // edge emulation

  1166.         src1 += y_off * linesize + x_off;

  1167.         src2 += y_off * linesize + x_off;

  1168.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0);

  1169.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1170.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1171.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,

  1172.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1173.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1174.             src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1175.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1176. 

  1177.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,

  1178.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1179.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1180.             src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1181.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1182.         } else {

  1183.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1184.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1185.         }

  1186.     } else {

  1187.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 3, 0);

  1188.         mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1189.         mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1190.     }

  1191. }

  1192. 

  1193. static av_always_inline

  1194. void vp8_mc_part(VP8Context *s, uint8_t *dst[3],

  1195.                  AVFrame *ref_frame, int x_off, int y_off,

  1196.                  int bx_off, int by_off,

  1197.                  int block_w, int block_h,

  1198.                  int width, int height, VP56mv *mv)

  1199. {

  1200.     VP56mv uvmv = *mv;

  1201. 

  1202.     /* Y */

  1203.     vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,

  1204.                 ref_frame, mv, x_off + bx_off, y_off + by_off,

  1205.                 block_w, block_h, width, height, s->linesize,

  1206.                 s->put_pixels_tab[block_w == 8]);

  1207. 

  1208.     /* U/V */

  1209.     if (s->profile == 3) {

  1210.         uvmv.x &= ~7;

  1211.         uvmv.y &= ~7;

  1212.     }

  1213.     x_off   >>= 1; y_off   >>= 1;

  1214.     bx_off  >>= 1; by_off  >>= 1;

  1215.     width   >>= 1; height  >>= 1;

  1216.     block_w >>= 1; block_h >>= 1;

  1217.     vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,

  1218.                   dst[2] + by_off * s->uvlinesize + bx_off, ref_frame,

  1219.                   &uvmv, x_off + bx_off, y_off + by_off,

  1220.                   block_w, block_h, width, height, s->uvlinesize,

  1221.                   s->put_pixels_tab[1 + (block_w == 4)]);

  1222. }

  1223. 

  1224. /* Fetch pixels for estimated mv 4 macroblocks ahead.

  1225.  * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */

  1226. static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)

  1227. {

  1228.     /* Don't prefetch refs that haven't been used very often this frame. */

  1229.     if (s->ref_count[ref-1] > (mb_xy >> 5)) {

  1230.         int x_off = mb_x << 4, y_off = mb_y << 4;

  1231.         int mx = (mb->mv.x>>2) + x_off + 8;

  1232.         int my = (mb->mv.y>>2) + y_off;

  1233.         uint8_t **src= s->framep[ref]->data;

  1234.         int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

  1235.         /* For threading, a ff_thread_await_progress here might be useful, but

  1236.          * it actually slows down the decoder. Since a bad prefetch doesn't

  1237.          * generate bad decoder output, we don't run it here. */

  1238.         s->dsp.prefetch(src[0]+off, s->linesize, 4);

  1239.         off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;

  1240.         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);

  1241.     }

  1242. }

  1243. 

  1244. /**

  1245.  * Apply motion vectors to prediction buffer, chapter 18.

  1246.  */

  1247. static av_always_inline

  1248. void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  1249.                    int mb_x, int mb_y)

  1250. {

  1251.     int x_off = mb_x << 4, y_off = mb_y << 4;

  1252.     int width = 16*s->mb_width, height = 16*s->mb_height;

  1253.     AVFrame *ref = s->framep[mb->ref_frame];

  1254.     VP56mv *bmv = mb->bmv;

  1255. 

  1256.     switch (mb->partitioning) {

  1257.     case VP8_SPLITMVMODE_NONE:

  1258.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1259.                     0, 0, 16, 16, width, height, &mb->mv);

  1260.         break;

  1261.     case VP8_SPLITMVMODE_4x4: {

  1262.         int x, y;

  1263.         VP56mv uvmv;

  1264. 

  1265.         /* Y */

  1266.         for (y = 0; y < 4; y++) {

  1267.             for (x = 0; x < 4; x++) {

  1268.                 vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,

  1269.                             ref, &bmv[4*y + x],

  1270.                             4*x + x_off, 4*y + y_off, 4, 4,

  1271.                             width, height, s->linesize,

  1272.                             s->put_pixels_tab[2]);

  1273.             }

  1274.         }

  1275. 

  1276.         /* U/V */

  1277.         x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;

  1278.         for (y = 0; y < 2; y++) {

  1279.             for (x = 0; x < 2; x++) {

  1280.                 uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +

  1281.                          mb->bmv[ 2*y    * 4 + 2*x+1].x +

  1282.                          mb->bmv[(2*y+1) * 4 + 2*x  ].x +

  1283.                          mb->bmv[(2*y+1) * 4 + 2*x+1].x;

  1284.                 uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +

  1285.                          mb->bmv[ 2*y    * 4 + 2*x+1].y +

  1286.                          mb->bmv[(2*y+1) * 4 + 2*x  ].y +

  1287.                          mb->bmv[(2*y+1) * 4 + 2*x+1].y;

  1288.                 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;

  1289.                 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;

  1290.                 if (s->profile == 3) {

  1291.                     uvmv.x &= ~7;

  1292.                     uvmv.y &= ~7;

  1293.                 }

  1294.                 vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,

  1295.                               dst[2] + 4*y*s->uvlinesize + x*4, ref, &uvmv,

  1296.                               4*x + x_off, 4*y + y_off, 4, 4,

  1297.                               width, height, s->uvlinesize,

  1298.                               s->put_pixels_tab[2]);

  1299.             }

  1300.         }

  1301.         break;

  1302.     }

  1303.     case VP8_SPLITMVMODE_16x8:

  1304.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1305.                     0, 0, 16, 8, width, height, &bmv[0]);

  1306.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1307.                     0, 8, 16, 8, width, height, &bmv[1]);

  1308.         break;

  1309.     case VP8_SPLITMVMODE_8x16:

  1310.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1311.                     0, 0, 8, 16, width, height, &bmv[0]);

  1312.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1313.                     8, 0, 8, 16, width, height, &bmv[1]);

  1314.         break;

  1315.     case VP8_SPLITMVMODE_8x8:

  1316.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1317.                     0, 0, 8, 8, width, height, &bmv[0]);

  1318.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1319.                     8, 0, 8, 8, width, height, &bmv[1]);

  1320.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1321.                     0, 8, 8, 8, width, height, &bmv[2]);

  1322.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1323.                     8, 8, 8, 8, width, height, &bmv[3]);

  1324.         break;

  1325.     }

  1326. }

  1327. 

  1328. static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)

  1329. {

  1330.     int x, y, ch;

  1331. 

  1332.     if (mb->mode != MODE_I4x4) {

  1333.         uint8_t *y_dst = dst[0];

  1334.         for (y = 0; y < 4; y++) {

  1335.             uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[y]);

  1336.             if (nnz4) {

  1337.                 if (nnz4&~0x01010101) {

  1338.                     for (x = 0; x < 4; x++) {

  1339.                         if ((uint8_t)nnz4 == 1)

  1340.                             s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);

  1341.                         else if((uint8_t)nnz4 > 1)

  1342.                             s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);

  1343.                         nnz4 >>= 8;

  1344.                         if (!nnz4)

  1345.                             break;

  1346.                     }

  1347.                 } else {

  1348.                     s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);

  1349.                 }

  1350.             }

  1351.             y_dst += 4*s->linesize;

  1352.         }

  1353.     }

  1354. 

  1355.     for (ch = 0; ch < 2; ch++) {

  1356.         uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[4+ch]);

  1357.         if (nnz4) {

  1358.             uint8_t *ch_dst = dst[1+ch];

  1359.             if (nnz4&~0x01010101) {

  1360.                 for (y = 0; y < 2; y++) {

  1361.                     for (x = 0; x < 2; x++) {

  1362.                         if ((uint8_t)nnz4 == 1)

  1363.                             s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1364.                         else if((uint8_t)nnz4 > 1)

  1365.                             s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1366.                         nnz4 >>= 8;

  1367.                         if (!nnz4)

  1368.                             goto chroma_idct_end;

  1369.                     }

  1370.                     ch_dst += 4*s->uvlinesize;

  1371.                 }

  1372.             } else {

  1373.                 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);

  1374.             }

  1375.         }

  1376. chroma_idct_end: ;

  1377.     }

  1378. }

  1379. 

  1380. static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )

  1381. {

  1382.     int interior_limit, filter_level;

  1383. 

  1384.     if (s->segmentation.enabled) {

  1385.         filter_level = s->segmentation.filter_level[s->segment];

  1386.         if (!s->segmentation.absolute_vals)

  1387.             filter_level += s->filter.level;

  1388.     } else

  1389.         filter_level = s->filter.level;

  1390. 

  1391.     if (s->lf_delta.enabled) {

  1392.         filter_level += s->lf_delta.ref[mb->ref_frame];

  1393.         filter_level += s->lf_delta.mode[mb->mode];

  1394.     }

  1395. 

  1396.     filter_level = av_clip_uintp2(filter_level, 6);

  1397. 

  1398.     interior_limit = filter_level;

  1399.     if (s->filter.sharpness) {

  1400.         interior_limit >>= (s->filter.sharpness + 3) >> 2;

  1401.         interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);

  1402.     }

  1403.     interior_limit = FFMAX(interior_limit, 1);

  1404. 

  1405.     f->filter_level = filter_level;

  1406.     f->inner_limit = interior_limit;

  1407.     f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;

  1408. }

  1409. 

  1410. static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)

  1411. {

  1412.     int mbedge_lim, bedge_lim, hev_thresh;

  1413.     int filter_level = f->filter_level;

  1414.     int inner_limit = f->inner_limit;

  1415.     int inner_filter = f->inner_filter;

  1416.     int linesize = s->linesize;

  1417.     int uvlinesize = s->uvlinesize;

  1418.     static const uint8_t hev_thresh_lut[2][64] = {

  1419.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1420.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1421.           3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,

  1422.           3, 3, 3, 3 },

  1423.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1424.           1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,

  1425.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1426.           2, 2, 2, 2 }

  1427.     };

  1428. 

  1429.     if (!filter_level)

  1430.         return;

  1431. 

  1432.      bedge_lim = 2*filter_level + inner_limit;

  1433.     mbedge_lim = bedge_lim + 4;

  1434. 

  1435.     hev_thresh = hev_thresh_lut[s->keyframe][filter_level];

  1436. 

  1437.     if (mb_x) {

  1438.         s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,

  1439.                                        mbedge_lim, inner_limit, hev_thresh);

  1440.         s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1441.                                        mbedge_lim, inner_limit, hev_thresh);

  1442.     }

  1443. 

  1444.     if (inner_filter) {

  1445.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,

  1446.                                              inner_limit, hev_thresh);

  1447.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,

  1448.                                              inner_limit, hev_thresh);

  1449.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,

  1450.                                              inner_limit, hev_thresh);

  1451.         s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,

  1452.                                              uvlinesize,  bedge_lim,

  1453.                                              inner_limit, hev_thresh);

  1454.     }

  1455. 

  1456.     if (mb_y) {

  1457.         s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,

  1458.                                        mbedge_lim, inner_limit, hev_thresh);

  1459.         s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1460.                                        mbedge_lim, inner_limit, hev_thresh);

  1461.     }

  1462. 

  1463.     if (inner_filter) {

  1464.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,

  1465.                                              linesize,    bedge_lim,

  1466.                                              inner_limit, hev_thresh);

  1467.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,

  1468.                                              linesize,    bedge_lim,

  1469.                                              inner_limit, hev_thresh);

  1470.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,

  1471.                                              linesize,    bedge_lim,

  1472.                                              inner_limit, hev_thresh);

  1473.         s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,

  1474.                                              dst[2] + 4 * uvlinesize,

  1475.                                              uvlinesize,  bedge_lim,

  1476.                                              inner_limit, hev_thresh);

  1477.     }

  1478. }

  1479. 

  1480. static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)

  1481. {

  1482.     int mbedge_lim, bedge_lim;

  1483.     int filter_level = f->filter_level;

  1484.     int inner_limit = f->inner_limit;

  1485.     int inner_filter = f->inner_filter;

  1486.     int linesize = s->linesize;

  1487. 

  1488.     if (!filter_level)

  1489.         return;

  1490. 

  1491.      bedge_lim = 2*filter_level + inner_limit;

  1492.     mbedge_lim = bedge_lim + 4;

  1493. 

  1494.     if (mb_x)

  1495.         s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);

  1496.     if (inner_filter) {

  1497.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);

  1498.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);

  1499.         s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);

  1500.     }

  1501. 

  1502.     if (mb_y)

  1503.         s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);

  1504.     if (inner_filter) {

  1505.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);

  1506.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);

  1507.         s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);

  1508.     }

  1509. }

  1510. 

  1511. static void filter_mb_row(VP8Context *s, AVFrame *curframe, int mb_y)

  1512. {

  1513.     VP8FilterStrength *f = s->filter_strength;

  1514.     uint8_t *dst[3] = {

  1515.         curframe->data[0] + 16*mb_y*s->linesize,

  1516.         curframe->data[1] +  8*mb_y*s->uvlinesize,

  1517.         curframe->data[2] +  8*mb_y*s->uvlinesize

  1518.     };

  1519.     int mb_x;

  1520. 

  1521.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1522.         backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);

  1523.         filter_mb(s, dst, f++, mb_x, mb_y);

  1524.         dst[0] += 16;

  1525.         dst[1] += 8;

  1526.         dst[2] += 8;

  1527.     }

  1528. }

  1529. 

  1530. static void filter_mb_row_simple(VP8Context *s, AVFrame *curframe, int mb_y)

  1531. {

  1532.     VP8FilterStrength *f = s->filter_strength;

  1533.     uint8_t *dst = curframe->data[0] + 16*mb_y*s->linesize;

  1534.     int mb_x;

  1535. 

  1536.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1537.         backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);

  1538.         filter_mb_simple(s, dst, f++, mb_x, mb_y);

  1539.         dst += 16;

  1540.     }

  1541. }

  1542. 

  1543. static void release_queued_segmaps(VP8Context *s, int is_close)

  1544. {

  1545.     int leave_behind = is_close ? 0 : !s->maps_are_invalid;

  1546.     while (s->num_maps_to_be_freed > leave_behind)

  1547.         av_freep(&s->segmentation_maps[--s->num_maps_to_be_freed]);

  1548.     s->maps_are_invalid = 0;

  1549. }

  1550. 

  1551. static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,

  1552.                             AVPacket *avpkt)

  1553. {

  1554.     VP8Context *s = avctx->priv_data;

  1555.     int ret, mb_x, mb_y, i, y, referenced;

  1556.     enum AVDiscard skip_thresh;

  1557.     AVFrame *av_uninit(curframe), *prev_frame = s->framep[VP56_FRAME_CURRENT];

  1558. 

  1559.     release_queued_segmaps(s, 0);

  1560. 

  1561.     if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)

  1562.         return ret;

  1563. 

  1564.     referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT

  1565.                                 || s->update_altref == VP56_FRAME_CURRENT;

  1566. 

  1567.     skip_thresh = !referenced ? AVDISCARD_NONREF :

  1568.                     !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;

  1569. 

  1570.     if (avctx->skip_frame >= skip_thresh) {

  1571.         s->invisible = 1;

  1572.         goto skip_decode;

  1573.     }

  1574.     s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;

  1575. 

  1576.     // release no longer referenced frames

  1577.     for (i = 0; i < 5; i++)

  1578.         if (s->frames[i].data[0] &&

  1579.             &s->frames[i] != prev_frame &&

  1580.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1581.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1582.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])

  1583.             vp8_release_frame(s, &s->frames[i], 0);

  1584. 

  1585.     // find a free buffer

  1586.     for (i = 0; i < 5; i++)

  1587.         if (&s->frames[i] != prev_frame &&

  1588.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1589.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1590.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {

  1591.             curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];

  1592.             break;

  1593.         }

  1594.     if (i == 5) {

  1595.         av_log(avctx, AV_LOG_FATAL, "Ran out of free frames!\n");

  1596.         abort();

  1597.     }

  1598.     if (curframe->data[0])

  1599.         ff_thread_release_buffer(avctx, curframe);

  1600. 

  1601.     curframe->key_frame = s->keyframe;

  1602.     curframe->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  1603.     curframe->reference = referenced ? 3 : 0;

  1604.     if ((ret = vp8_alloc_frame(s, curframe))) {

  1605.         av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");

  1606.         return ret;

  1607.     }

  1608. 

  1609.     // check if golden and altref are swapped

  1610.     if (s->update_altref != VP56_FRAME_NONE) {

  1611.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[s->update_altref];

  1612.     } else {

  1613.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[VP56_FRAME_GOLDEN2];

  1614.     }

  1615.     if (s->update_golden != VP56_FRAME_NONE) {

  1616.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[s->update_golden];

  1617.     } else {

  1618.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[VP56_FRAME_GOLDEN];

  1619.     }

  1620.     if (s->update_last) {

  1621.         s->next_framep[VP56_FRAME_PREVIOUS] = curframe;

  1622.     } else {

  1623.         s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS];

  1624.     }

  1625.     s->next_framep[VP56_FRAME_CURRENT]      = curframe;

  1626. 

  1627.     ff_thread_finish_setup(avctx);

  1628. 

  1629.     // Given that arithmetic probabilities are updated every frame, it's quite likely

  1630.     // that the values we have on a random interframe are complete junk if we didn't

  1631.     // start decode on a keyframe. So just don't display anything rather than junk.

  1632.     if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||

  1633.                          !s->framep[VP56_FRAME_GOLDEN] ||

  1634.                          !s->framep[VP56_FRAME_GOLDEN2])) {

  1635.         av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");

  1636.         return AVERROR_INVALIDDATA;

  1637.     }

  1638. 

  1639.     s->linesize   = curframe->linesize[0];

  1640.     s->uvlinesize = curframe->linesize[1];

  1641. 

  1642.     if (!s->edge_emu_buffer)

  1643.         s->edge_emu_buffer = av_malloc(21*s->linesize);

  1644. 

  1645.     memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));

  1646. 

  1647.     /* Zero macroblock structures for top/top-left prediction from outside the frame. */

  1648.     memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));

  1649. 

  1650.     // top edge of 127 for intra prediction

  1651.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1652.         s->top_border[0][15] = s->top_border[0][23] = 127;

  1653.         memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);

  1654.     }

  1655.     memset(s->ref_count, 0, sizeof(s->ref_count));

  1656.     if (s->keyframe)

  1657.         memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);

  1658. 

  1659. #define MARGIN (16 << 2)

  1660.     s->mv_min.y = -MARGIN;

  1661.     s->mv_max.y = ((s->mb_height - 1) << 6) + MARGIN;

  1662. 

  1663.     for (mb_y = 0; mb_y < s->mb_height; mb_y++) {

  1664.         VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];

  1665.         VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;

  1666.         int mb_xy = mb_y*s->mb_width;

  1667.         uint8_t *dst[3] = {

  1668.             curframe->data[0] + 16*mb_y*s->linesize,

  1669.             curframe->data[1] +  8*mb_y*s->uvlinesize,

  1670.             curframe->data[2] +  8*mb_y*s->uvlinesize

  1671.         };

  1672. 

  1673.         memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock

  1674.         memset(s->left_nnz, 0, sizeof(s->left_nnz));

  1675.         AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);

  1676. 

  1677.         // left edge of 129 for intra prediction

  1678.         if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1679.             for (i = 0; i < 3; i++)

  1680.                 for (y = 0; y < 16>>!!i; y++)

  1681.                     dst[i][y*curframe->linesize[i]-1] = 129;

  1682.             if (mb_y == 1) // top left edge is also 129

  1683.                 s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;

  1684.         }

  1685. 

  1686.         s->mv_min.x = -MARGIN;

  1687.         s->mv_max.x = ((s->mb_width  - 1) << 6) + MARGIN;

  1688.         if (prev_frame && s->segmentation.enabled && !s->segmentation.update_map)

  1689.             ff_thread_await_progress(prev_frame, mb_y, 0);

  1690. 

  1691.         for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {

  1692.             /* Prefetch the current frame, 4 MBs ahead */

  1693.             s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);

  1694.             s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

  1695. 

  1696.             decode_mb_mode(s, mb, mb_x, mb_y, curframe->ref_index[0] + mb_xy,

  1697.                            prev_frame && prev_frame->ref_index[0] ? prev_frame->ref_index[0] + mb_xy : NULL);

  1698. 

  1699.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);

  1700. 

  1701.             if (!mb->skip)

  1702.                 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

  1703. 

  1704.             if (mb->mode <= MODE_I4x4)

  1705.                 intra_predict(s, dst, mb, mb_x, mb_y);

  1706.             else

  1707.                 inter_predict(s, dst, mb, mb_x, mb_y);

  1708. 

  1709.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);

  1710. 

  1711.             if (!mb->skip) {

  1712.                 idct_mb(s, dst, mb);

  1713.             } else {

  1714.                 AV_ZERO64(s->left_nnz);

  1715.                 AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned

  1716. 

  1717.                 // Reset DC block predictors if they would exist if the mb had coefficients

  1718.                 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  1719.                     s->left_nnz[8]      = 0;

  1720.                     s->top_nnz[mb_x][8] = 0;

  1721.                 }

  1722.             }

  1723. 

  1724.             if (s->deblock_filter)

  1725.                 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

  1726. 

  1727.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);

  1728. 

  1729.             dst[0] += 16;

  1730.             dst[1] += 8;

  1731.             dst[2] += 8;

  1732.             s->mv_min.x -= 64;

  1733.             s->mv_max.x -= 64;

  1734.         }

  1735.         if (s->deblock_filter) {

  1736.             if (s->filter.simple)

  1737.                 filter_mb_row_simple(s, curframe, mb_y);

  1738.             else

  1739.                 filter_mb_row(s, curframe, mb_y);

  1740.         }

  1741.         s->mv_min.y -= 64;

  1742.         s->mv_max.y -= 64;

  1743. 

  1744.         ff_thread_report_progress(curframe, mb_y, 0);

  1745.     }

  1746. 

  1747.     ff_thread_report_progress(curframe, INT_MAX, 0);

  1748. skip_decode:

  1749.     // if future frames don't use the updated probabilities,

  1750.     // reset them to the values we saved

  1751.     if (!s->update_probabilities)

  1752.         s->prob[0] = s->prob[1];

  1753. 

  1754.     memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4);

  1755. 

  1756.     if (!s->invisible) {

  1757.         *(AVFrame*)data = *curframe;

  1758.         *data_size = sizeof(AVFrame);

  1759.     }

  1760. 

  1761.     return avpkt->size;

  1762. }

  1763. 

  1764. static av_cold int vp8_decode_init(AVCodecContext *avctx)

  1765. {

  1766.     VP8Context *s = avctx->priv_data;

  1767. 

  1768.     s->avctx = avctx;

  1769.     avctx->pix_fmt = PIX_FMT_YUV420P;

  1770. 

  1771.     dsputil_init(&s->dsp, avctx);

  1772.     ff_h264_pred_init(&s->hpc, CODEC_ID_VP8, 8, 1);

  1773.     ff_vp8dsp_init(&s->vp8dsp);

  1774. 

  1775.     return 0;

  1776. }

  1777. 

  1778. static av_cold int vp8_decode_free(AVCodecContext *avctx)

  1779. {

  1780.     vp8_decode_flush_impl(avctx, 0, 1);

  1781.     release_queued_segmaps(avctx->priv_data, 1);

  1782.     return 0;

  1783. }

  1784. 

  1785. static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx)

  1786. {

  1787.     VP8Context *s = avctx->priv_data;

  1788. 

  1789.     s->avctx = avctx;

  1790. 

  1791.     return 0;

  1792. }

  1793. 

  1794. #define REBASE(pic) \

  1795.     pic ? pic - &s_src->frames[0] + &s->frames[0] : NULL

  1796. 

  1797. static int vp8_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  1798. {

  1799.     VP8Context *s = dst->priv_data, *s_src = src->priv_data;

  1800. 

  1801.     if (s->macroblocks_base &&

  1802.         (s_src->mb_width != s->mb_width || s_src->mb_height != s->mb_height)) {

  1803.         free_buffers(s);

  1804.     }

  1805. 

  1806.     s->prob[0] = s_src->prob[!s_src->update_probabilities];

  1807.     s->segmentation = s_src->segmentation;

  1808.     s->lf_delta = s_src->lf_delta;

  1809.     memcpy(s->sign_bias, s_src->sign_bias, sizeof(s->sign_bias));

  1810. 

  1811.     memcpy(&s->frames, &s_src->frames, sizeof(s->frames));

  1812.     s->framep[0] = REBASE(s_src->next_framep[0]);

  1813.     s->framep[1] = REBASE(s_src->next_framep[1]);

  1814.     s->framep[2] = REBASE(s_src->next_framep[2]);

  1815.     s->framep[3] = REBASE(s_src->next_framep[3]);

  1816. 

  1817.     return 0;

  1818. }

  1819. 

  1820. AVCodec ff_vp8_decoder = {

  1821.     .name           = "vp8",

  1822.     .type           = AVMEDIA_TYPE_VIDEO,

  1823.     .id             = CODEC_ID_VP8,

  1824.     .priv_data_size = sizeof(VP8Context),

  1825.     .init           = vp8_decode_init,

  1826.     .close          = vp8_decode_free,

  1827.     .decode         = vp8_decode_frame,

  1828.     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  1829.     .flush = vp8_decode_flush,

  1830.     .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),

  1831.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy),

  1832.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context),

  1833. };