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.
3683 Commits
32 Branches
293MB
Tree: 155ec6edf8
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '155ec6edf8'
${ noResults }
FFmpeg/libavcodec/dv.c

972 lines
29KB
Raw Blame History 

  1. /*

  2.  * DV decoder

  3.  * Copyright (c) 2002 Fabrice Bellard.

  4.  * Copyright (c) 2004 Roman Shaposhnik.

  5.  *

  6.  * DV encoder 

  7.  * Copyright (c) 2003 Roman Shaposhnik.

  8.  *

  9.  * Many thanks to Dan Dennedy <dan@dennedy.org> for providing wealth

  10.  * of DV technical info.

  11.  *

  12.  * This library is free software; you can redistribute it and/or

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

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

  15.  * version 2 of the License, or (at your option) any later version.

  16.  *

  17.  * This library is distributed in the hope that it will be useful,

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

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

  20.  * Lesser General Public License for more details.

  21.  *

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

  23.  * License along with this library; if not, write to the Free Software

  24.  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

  25.  */

  26. 

  27. /**

  28.  * @file dv.c

  29.  * DV codec.

  30.  */

  31. #include "avcodec.h"

  32. #include "dsputil.h"

  33. #include "mpegvideo.h"

  34. #include "simple_idct.h"

  35. #include "dvdata.h"

  36. 

  37. typedef struct DVVideoContext {

  38.     const DVprofile* sys;

  39.     AVFrame picture;

  40.     uint8_t *buf;

  41.     

  42.     uint8_t dv_zigzag[2][64];

  43.     uint8_t dv_idct_shift[2][2][22][64];

  44.   

  45.     void (*get_pixels)(DCTELEM *block, const uint8_t *pixels, int line_size);

  46.     void (*fdct[2])(DCTELEM *block);

  47.     void (*idct_put[2])(uint8_t *dest, int line_size, DCTELEM *block);

  48. } DVVideoContext;

  49. 

  50. #define TEX_VLC_BITS 9

  51. 

  52. #ifdef DV_CODEC_TINY_TARGET

  53. #define DV_VLC_MAP_RUN_SIZE 15

  54. #define DV_VLC_MAP_LEV_SIZE 23

  55. #else

  56. #define DV_VLC_MAP_RUN_SIZE  64 

  57. #define DV_VLC_MAP_LEV_SIZE 512

  58. #endif

  59. 

  60. /* MultiThreading */

  61. static uint8_t** dv_anchor;

  62. 

  63. /* XXX: also include quantization */

  64. static RL_VLC_ELEM *dv_rl_vlc;

  65. /* VLC encoding lookup table */

  66. static struct dv_vlc_pair {

  67.    uint32_t vlc;

  68.    uint8_t  size;

  69. } (*dv_vlc_map)[DV_VLC_MAP_LEV_SIZE] = NULL;

  70. 

  71. static void dv_build_unquantize_tables(DVVideoContext *s, uint8_t* perm)

  72. {

  73.     int i, q, j;

  74. 

  75.     /* NOTE: max left shift is 6 */

  76.     for(q = 0; q < 22; q++) {

  77.         /* 88DCT */

  78.         for(i = 1; i < 64; i++) {

  79.             /* 88 table */

  80.             j = perm[i];

  81.             s->dv_idct_shift[0][0][q][j] =

  82.                 dv_quant_shifts[q][dv_88_areas[i]] + 1;

  83. 	    s->dv_idct_shift[1][0][q][j] = s->dv_idct_shift[0][0][q][j] + 1;

  84.         }

  85.         

  86.         /* 248DCT */

  87.         for(i = 1; i < 64; i++) {

  88.             /* 248 table */

  89.             s->dv_idct_shift[0][1][q][i] =  

  90.                 dv_quant_shifts[q][dv_248_areas[i]] + 1;

  91. 	    s->dv_idct_shift[1][1][q][i] = s->dv_idct_shift[0][1][q][i] + 1;

  92.         }

  93.     }

  94. }

  95. 

  96. static int dvvideo_init(AVCodecContext *avctx)

  97. {

  98.     DVVideoContext *s = avctx->priv_data;

  99.     DSPContext dsp;

  100.     static int done=0;

  101.     int i, j;

  102. 

  103.     if (!done) {

  104.         VLC dv_vlc;

  105.         uint16_t new_dv_vlc_bits[NB_DV_VLC*2];

  106.         uint8_t new_dv_vlc_len[NB_DV_VLC*2];

  107.         uint8_t new_dv_vlc_run[NB_DV_VLC*2];

  108.         int16_t new_dv_vlc_level[NB_DV_VLC*2];

  109. 

  110.         done = 1;

  111. 

  112.         dv_vlc_map = av_mallocz(DV_VLC_MAP_LEV_SIZE*DV_VLC_MAP_RUN_SIZE*sizeof(struct dv_vlc_pair));

  113. 	if (!dv_vlc_map)

  114. 	    return -ENOMEM;

  115. 

  116. 	/* dv_anchor lets each thread know its Id */

  117. 	dv_anchor = av_malloc(12*27*sizeof(void*));

  118. 	if (!dv_anchor) {

  119. 	    av_free(dv_vlc_map);

  120. 	    return -ENOMEM;

  121. 	}

  122. 	for (i=0; i<12*27; i++)

  123. 	    dv_anchor[i] = (void*)(size_t)i;

  124. 

  125. 	/* it's faster to include sign bit in a generic VLC parsing scheme */

  126. 	for (i=0, j=0; i<NB_DV_VLC; i++, j++) {

  127. 	    new_dv_vlc_bits[j] = dv_vlc_bits[i];

  128. 	    new_dv_vlc_len[j] = dv_vlc_len[i];

  129. 	    new_dv_vlc_run[j] = dv_vlc_run[i];

  130. 	    new_dv_vlc_level[j] = dv_vlc_level[i];

  131. 	    

  132. 	    if (dv_vlc_level[i]) {

  133. 	        new_dv_vlc_bits[j] <<= 1;

  134. 		new_dv_vlc_len[j]++;

  135. 

  136. 		j++;

  137. 		new_dv_vlc_bits[j] = (dv_vlc_bits[i] << 1) | 1;

  138. 		new_dv_vlc_len[j] = dv_vlc_len[i] + 1;

  139. 		new_dv_vlc_run[j] = dv_vlc_run[i];

  140. 		new_dv_vlc_level[j] = -dv_vlc_level[i];

  141. 	    }

  142. 	}

  143.              

  144.         /* NOTE: as a trick, we use the fact the no codes are unused

  145.            to accelerate the parsing of partial codes */

  146.         init_vlc(&dv_vlc, TEX_VLC_BITS, j, 

  147.                  new_dv_vlc_len, 1, 1, new_dv_vlc_bits, 2, 2);

  148. 

  149.         dv_rl_vlc = av_malloc(dv_vlc.table_size * sizeof(RL_VLC_ELEM));

  150. 	if (!dv_rl_vlc) {

  151. 	    av_free(dv_anchor);

  152. 	    av_free(dv_vlc_map);

  153. 	    return -ENOMEM;

  154. 	}

  155.         for(i = 0; i < dv_vlc.table_size; i++){

  156.             int code= dv_vlc.table[i][0];

  157.             int len = dv_vlc.table[i][1];

  158.             int level, run;

  159.         

  160.             if(len<0){ //more bits needed

  161.                 run= 0;

  162.                 level= code;

  163.             } else {

  164.                 run=   new_dv_vlc_run[code] + 1;

  165.                 level= new_dv_vlc_level[code];

  166.             }

  167.             dv_rl_vlc[i].len = len;

  168.             dv_rl_vlc[i].level = level;

  169.             dv_rl_vlc[i].run = run;

  170.         }

  171. 	free_vlc(&dv_vlc);

  172. 

  173. 	for (i = 0; i < NB_DV_VLC - 1; i++) {

  174.            if (dv_vlc_run[i] >= DV_VLC_MAP_RUN_SIZE)

  175. 	       continue;

  176. #ifdef DV_CODEC_TINY_TARGET

  177.            if (dv_vlc_level[i] >= DV_VLC_MAP_LEV_SIZE)

  178. 	       continue;

  179. #endif

  180. 	   

  181. 	   if (dv_vlc_map[dv_vlc_run[i]][dv_vlc_level[i]].size != 0)

  182. 	       continue;

  183. 	       

  184. 	   dv_vlc_map[dv_vlc_run[i]][dv_vlc_level[i]].vlc = dv_vlc_bits[i] << 

  185. 	                                                    (!!dv_vlc_level[i]);

  186. 	   dv_vlc_map[dv_vlc_run[i]][dv_vlc_level[i]].size = dv_vlc_len[i] + 

  187. 	                                                     (!!dv_vlc_level[i]);

  188. 	}

  189. 	for (i = 0; i < DV_VLC_MAP_RUN_SIZE; i++) {

  190. #ifdef DV_CODEC_TINY_TARGET

  191. 	   for (j = 1; j < DV_VLC_MAP_LEV_SIZE; j++) {

  192. 	      if (dv_vlc_map[i][j].size == 0) {

  193. 	          dv_vlc_map[i][j].vlc = dv_vlc_map[0][j].vlc |

  194. 		            (dv_vlc_map[i-1][0].vlc << (dv_vlc_map[0][j].size));

  195. 	          dv_vlc_map[i][j].size = dv_vlc_map[i-1][0].size + 

  196. 		                          dv_vlc_map[0][j].size;

  197. 	      }

  198. 	   }

  199. #else

  200. 	   for (j = 1; j < DV_VLC_MAP_LEV_SIZE/2; j++) {

  201. 	      if (dv_vlc_map[i][j].size == 0) {

  202. 	          dv_vlc_map[i][j].vlc = dv_vlc_map[0][j].vlc |

  203. 		            (dv_vlc_map[i-1][0].vlc << (dv_vlc_map[0][j].size));

  204. 	          dv_vlc_map[i][j].size = dv_vlc_map[i-1][0].size + 

  205. 		                          dv_vlc_map[0][j].size;

  206. 	      }

  207. 	      dv_vlc_map[i][((uint16_t)(-j))&0x1ff].vlc = 

  208. 	                                    dv_vlc_map[i][j].vlc | 1;

  209. 	      dv_vlc_map[i][((uint16_t)(-j))&0x1ff].size = 

  210. 	                                    dv_vlc_map[i][j].size;

  211. 	   }

  212. #endif

  213. 	}

  214.     }

  215. 

  216.     /* Generic DSP setup */

  217.     dsputil_init(&dsp, avctx);

  218.     s->get_pixels = dsp.get_pixels;

  219. 

  220.     /* 88DCT setup */

  221.     s->fdct[0] = dsp.fdct;

  222.     s->idct_put[0] = dsp.idct_put;

  223.     for (i=0; i<64; i++)

  224.        s->dv_zigzag[0][i] = dsp.idct_permutation[ff_zigzag_direct[i]];

  225. 

  226.     /* 248DCT setup */

  227.     s->fdct[1] = dsp.fdct248;

  228.     s->idct_put[1] = simple_idct248_put;  // FIXME: need to add it to DSP

  229.     memcpy(s->dv_zigzag[1], ff_zigzag248_direct, 64);

  230. 

  231.     /* XXX: do it only for constant case */

  232.     dv_build_unquantize_tables(s, dsp.idct_permutation);

  233. 

  234.     /* FIXME: I really don't think this should be here */

  235.     if (dv_codec_profile(avctx))

  236. 	avctx->pix_fmt = dv_codec_profile(avctx)->pix_fmt; 

  237.     avctx->coded_frame = &s->picture;

  238.     

  239.     return 0;

  240. }

  241. 

  242. // #define VLC_DEBUG

  243. // #define printf(...) av_log(NULL, AV_LOG_ERROR, __VA_ARGS__)

  244. 

  245. typedef struct BlockInfo {

  246.     const uint8_t *shift_table;

  247.     const uint8_t *scan_table;

  248.     uint8_t pos; /* position in block */

  249.     uint8_t dct_mode;

  250.     uint8_t partial_bit_count;

  251.     uint16_t partial_bit_buffer;

  252.     int shift_offset;

  253. } BlockInfo;

  254. 

  255. /* block size in bits */

  256. static const uint16_t block_sizes[6] = {

  257.     112, 112, 112, 112, 80, 80

  258. };

  259. /* bit budget for AC only in 5 MBs */

  260. static const int vs_total_ac_bits = (100 * 4 + 68*2) * 5;

  261. /* see dv_88_areas and dv_248_areas for details */

  262. static const int mb_area_start[5] = { 1, 6, 21, 43, 64 }; 

  263. 

  264. #ifndef ALT_BITSTREAM_READER

  265. #warning only works with ALT_BITSTREAM_READER

  266. #endif

  267. 

  268. static inline int get_bits_left(GetBitContext *s)

  269. {

  270.     return s->size_in_bits - get_bits_count(s);

  271. }

  272. 

  273. static inline int get_bits_size(GetBitContext *s)

  274. {

  275.     return s->size_in_bits;

  276. }

  277. 

  278. static inline int put_bits_left(PutBitContext* s)

  279. {

  280.     return (s->buf_end - s->buf) * 8 - put_bits_count(s);

  281. }

  282. 

  283. /* decode ac coefs */

  284. static void dv_decode_ac(GetBitContext *gb, BlockInfo *mb, DCTELEM *block)

  285. {

  286.     int last_index = get_bits_size(gb);

  287.     const uint8_t *scan_table = mb->scan_table;

  288.     const uint8_t *shift_table = mb->shift_table;

  289.     int pos = mb->pos;

  290.     int partial_bit_count = mb->partial_bit_count;

  291.     int level, pos1, run, vlc_len, index;

  292.     

  293.     OPEN_READER(re, gb);

  294.     UPDATE_CACHE(re, gb);

  295.     

  296.     /* if we must parse a partial vlc, we do it here */

  297.     if (partial_bit_count > 0) {

  298.         re_cache = ((unsigned)re_cache >> partial_bit_count) |

  299. 	           (mb->partial_bit_buffer << (sizeof(re_cache)*8 - partial_bit_count));

  300. 	re_index -= partial_bit_count;

  301. 	mb->partial_bit_count = 0;

  302.     }

  303. 

  304.     /* get the AC coefficients until last_index is reached */

  305.     for(;;) {

  306. #ifdef VLC_DEBUG

  307.         printf("%2d: bits=%04x index=%d\n", pos, SHOW_UBITS(re, gb, 16), re_index);

  308. #endif

  309.         /* our own optimized GET_RL_VLC */

  310.         index = NEG_USR32(re_cache, TEX_VLC_BITS);

  311. 	vlc_len = dv_rl_vlc[index].len;

  312.         if (vlc_len < 0) {

  313.             index = NEG_USR32((unsigned)re_cache << TEX_VLC_BITS, -vlc_len) + dv_rl_vlc[index].level;

  314.             vlc_len = TEX_VLC_BITS - vlc_len;

  315.         }

  316.         level = dv_rl_vlc[index].level;

  317. 	run = dv_rl_vlc[index].run;

  318. 	

  319. 	/* gotta check if we're still within gb boundaries */

  320. 	if (re_index + vlc_len > last_index) {

  321. 	    /* should be < 16 bits otherwise a codeword could have been parsed */

  322. 	    mb->partial_bit_count = last_index - re_index;

  323. 	    mb->partial_bit_buffer = NEG_USR32(re_cache, mb->partial_bit_count);

  324. 	    re_index = last_index;

  325. 	    break;

  326. 	}

  327. 	re_index += vlc_len;

  328. 

  329. #ifdef VLC_DEBUG

  330. 	printf("run=%d level=%d\n", run, level);

  331. #endif

  332. 	pos += run; 	

  333. 	if (pos >= 64)

  334. 	    break;

  335.         

  336. 	if (level) {

  337.             pos1 = scan_table[pos];

  338.             block[pos1] = level << shift_table[pos1];

  339.         } 

  340. 

  341.         UPDATE_CACHE(re, gb);

  342.     }

  343.     CLOSE_READER(re, gb);

  344.     mb->pos = pos;

  345. }

  346. 

  347. static inline void bit_copy(PutBitContext *pb, GetBitContext *gb)

  348. {

  349.     int bits_left = get_bits_left(gb);

  350.     while (bits_left >= 16) {

  351.         put_bits(pb, 16, get_bits(gb, 16));

  352.         bits_left -= 16;

  353.     }

  354.     if (bits_left > 0) {

  355.         put_bits(pb, bits_left, get_bits(gb, bits_left));

  356.     }

  357. }

  358. 

  359. /* mb_x and mb_y are in units of 8 pixels */

  360. static inline void dv_decode_video_segment(DVVideoContext *s, 

  361.                                            uint8_t *buf_ptr1, 

  362.                                            const uint16_t *mb_pos_ptr)

  363. {

  364.     int quant, dc, dct_mode, class1, j;

  365.     int mb_index, mb_x, mb_y, v, last_index;

  366.     DCTELEM *block, *block1;

  367.     int c_offset;

  368.     uint8_t *y_ptr;

  369.     void (*idct_put)(uint8_t *dest, int line_size, DCTELEM *block);

  370.     uint8_t *buf_ptr;

  371.     PutBitContext pb, vs_pb;

  372.     GetBitContext gb;

  373.     BlockInfo mb_data[5 * 6], *mb, *mb1;

  374.     DCTELEM sblock[5*6][64] __align8;

  375.     uint8_t mb_bit_buffer[80 + 4]; /* allow some slack */

  376.     uint8_t vs_bit_buffer[5 * 80 + 4]; /* allow some slack */

  377. 	    

  378.     memset(sblock, 0, sizeof(sblock));

  379. 

  380.     /* pass 1 : read DC and AC coefficients in blocks */

  381.     buf_ptr = buf_ptr1;

  382.     block1 = &sblock[0][0];

  383.     mb1 = mb_data;

  384.     init_put_bits(&vs_pb, vs_bit_buffer, 5 * 80);

  385.     for(mb_index = 0; mb_index < 5; mb_index++, mb1 += 6, block1 += 6 * 64) {

  386.         /* skip header */

  387.         quant = buf_ptr[3] & 0x0f;

  388.         buf_ptr += 4;

  389.         init_put_bits(&pb, mb_bit_buffer, 80);

  390.         mb = mb1;

  391.         block = block1;

  392.         for(j = 0;j < 6; j++) {

  393.             last_index = block_sizes[j];

  394. 	    init_get_bits(&gb, buf_ptr, last_index);

  395.             

  396.             /* get the dc */

  397.             dc = get_bits(&gb, 9);

  398.             dc = (dc << (32 - 9)) >> (32 - 9);

  399.             dct_mode = get_bits1(&gb);

  400.             mb->dct_mode = dct_mode;

  401.             mb->scan_table = s->dv_zigzag[dct_mode];

  402.             class1 = get_bits(&gb, 2);

  403.             mb->shift_table = s->dv_idct_shift[class1 == 3][dct_mode]

  404.                 [quant + dv_quant_offset[class1]];

  405.             dc = dc << 2;

  406.             /* convert to unsigned because 128 is not added in the

  407.                standard IDCT */

  408.             dc += 1024;

  409.             block[0] = dc;

  410.             buf_ptr += last_index >> 3;

  411.             mb->pos = 0;

  412.             mb->partial_bit_count = 0;

  413. 

  414. #ifdef VLC_DEBUG

  415.             printf("MB block: %d, %d ", mb_index, j);

  416. #endif

  417.             dv_decode_ac(&gb, mb, block);

  418. 

  419.             /* write the remaining bits  in a new buffer only if the

  420.                block is finished */

  421.             if (mb->pos >= 64)

  422.                 bit_copy(&pb, &gb);

  423.             

  424.             block += 64;

  425.             mb++;

  426.         }

  427.         

  428.         /* pass 2 : we can do it just after */

  429. #ifdef VLC_DEBUG

  430.         printf("***pass 2 size=%d MB#=%d\n", put_bits_count(&pb), mb_index);

  431. #endif

  432.         block = block1;

  433.         mb = mb1;

  434.         init_get_bits(&gb, mb_bit_buffer, put_bits_count(&pb));

  435. 	flush_put_bits(&pb);

  436.         for(j = 0;j < 6; j++, block += 64, mb++) {

  437.             if (mb->pos < 64 && get_bits_left(&gb) > 0) {

  438.                 dv_decode_ac(&gb, mb, block);

  439.                 /* if still not finished, no need to parse other blocks */

  440.                 if (mb->pos < 64)

  441.                     break;

  442.             }

  443.         }

  444.         /* all blocks are finished, so the extra bytes can be used at

  445.            the video segment level */

  446.         if (j >= 6)

  447. 	    bit_copy(&vs_pb, &gb);

  448.     }

  449. 

  450.     /* we need a pass other the whole video segment */

  451. #ifdef VLC_DEBUG

  452.     printf("***pass 3 size=%d\n", put_bits_count(&vs_pb));

  453. #endif

  454.     block = &sblock[0][0];

  455.     mb = mb_data;

  456.     init_get_bits(&gb, vs_bit_buffer, put_bits_count(&vs_pb));

  457.     flush_put_bits(&vs_pb);

  458.     for(mb_index = 0; mb_index < 5; mb_index++) {

  459.         for(j = 0;j < 6; j++) {

  460.             if (mb->pos < 64) {

  461. #ifdef VLC_DEBUG

  462.                 printf("start %d:%d\n", mb_index, j);

  463. #endif

  464.                 dv_decode_ac(&gb, mb, block);

  465.             }

  466. 	    if (mb->pos >= 64 && mb->pos < 127)

  467. 		av_log(NULL, AV_LOG_ERROR, "AC EOB marker is absent pos=%d\n", mb->pos);

  468.             block += 64;

  469.             mb++;

  470.         }

  471.     }

  472.     

  473.     /* compute idct and place blocks */

  474.     block = &sblock[0][0];

  475.     mb = mb_data;

  476.     for(mb_index = 0; mb_index < 5; mb_index++) {

  477.         v = *mb_pos_ptr++;

  478.         mb_x = v & 0xff;

  479.         mb_y = v >> 8;

  480.         y_ptr = s->picture.data[0] + (mb_y * s->picture.linesize[0] * 8) + (mb_x * 8);

  481.         if (s->sys->pix_fmt == PIX_FMT_YUV411P)

  482.             c_offset = (mb_y * s->picture.linesize[1] * 8) + ((mb_x >> 2) * 8);

  483.         else

  484.             c_offset = ((mb_y >> 1) * s->picture.linesize[1] * 8) + ((mb_x >> 1) * 8);

  485.         for(j = 0;j < 6; j++) {

  486.             idct_put = s->idct_put[mb->dct_mode];

  487.             if (j < 4) {

  488.                 if (s->sys->pix_fmt == PIX_FMT_YUV411P && mb_x < (704 / 8)) {

  489.                     /* NOTE: at end of line, the macroblock is handled as 420 */

  490.                     idct_put(y_ptr + (j * 8), s->picture.linesize[0], block);

  491.                 } else {

  492.                     idct_put(y_ptr + ((j & 1) * 8) + ((j >> 1) * 8 * s->picture.linesize[0]),

  493.                              s->picture.linesize[0], block);

  494.                 }

  495.             } else {

  496.                 if (s->sys->pix_fmt == PIX_FMT_YUV411P && mb_x >= (704 / 8)) {

  497.                     uint64_t aligned_pixels[64/8];

  498.                     uint8_t *pixels= (uint8_t*)aligned_pixels;

  499. 		    uint8_t *c_ptr, *c_ptr1, *ptr;

  500.                     int y, linesize;

  501.                     /* NOTE: at end of line, the macroblock is handled as 420 */

  502.                     idct_put(pixels, 8, block);

  503.                     linesize = s->picture.linesize[6 - j];

  504.                     c_ptr = s->picture.data[6 - j] + c_offset;

  505.                     ptr = pixels;

  506.                     for(y = 0;y < 8; y++) {

  507.                         /* convert to 411P */

  508.                         c_ptr1 = c_ptr + 8*linesize;

  509.                         c_ptr[0]= ptr[0]; c_ptr1[0]= ptr[4];

  510.                         c_ptr[1]= ptr[1]; c_ptr1[1]= ptr[5];

  511.                         c_ptr[2]= ptr[2]; c_ptr1[2]= ptr[6];

  512.                         c_ptr[3]= ptr[3]; c_ptr1[3]= ptr[7];

  513.                         c_ptr += linesize;

  514.                         ptr += 8;

  515.                     }

  516.                 } else {

  517.                     /* don't ask me why they inverted Cb and Cr ! */

  518.                     idct_put(s->picture.data[6 - j] + c_offset, 

  519.                              s->picture.linesize[6 - j], block);

  520.                 }

  521.             }

  522.             block += 64;

  523.             mb++;

  524.         }

  525.     }

  526. }

  527. 

  528. #ifdef DV_CODEC_TINY_TARGET

  529. /* Converts run and level (where level != 0) pair into vlc, returning bit size */

  530. static always_inline int dv_rl2vlc(int run, int l, uint32_t* vlc)

  531. {

  532.     int sign = l >> 8;

  533.     int level = (l ^ sign) - sign;

  534.     int size;

  535.     

  536.     sign = (sign & 1);

  537. 

  538.     if (run < DV_VLC_MAP_RUN_SIZE && level < DV_VLC_MAP_LEV_SIZE) {

  539.         *vlc = dv_vlc_map[run][level].vlc | sign;

  540. 	size = dv_vlc_map[run][level].size;

  541.     }

  542.     else { 

  543.         if (level < DV_VLC_MAP_LEV_SIZE) {

  544. 	    *vlc = dv_vlc_map[0][level].vlc | sign;

  545. 	    size = dv_vlc_map[0][level].size;

  546. 	} else {

  547.             *vlc = 0xfe00 | (level << 1) | sign;

  548. 	    size = 16;

  549. 	}

  550. 	if (run) {

  551. 	    *vlc |= ((run < 16) ? dv_vlc_map[run-1][0].vlc : 

  552. 	                          (0x1f80 | (run - 1))) << size;

  553. 	    size += (run < 16) ? dv_vlc_map[run-1][0].size : 13;

  554. 	}

  555.     }

  556.     

  557.     return size;

  558. }

  559. 

  560. static always_inline int dv_rl2vlc_size(int run, int l)

  561. {

  562.     int level = (l ^ (l >> 8)) - (l >> 8);

  563.     int size;

  564.     

  565.     if (run < DV_VLC_MAP_RUN_SIZE && level < DV_VLC_MAP_LEV_SIZE) {

  566. 	size = dv_vlc_map[run][level].size; 

  567.     }

  568.     else { 

  569. 	size = (level < DV_VLC_MAP_LEV_SIZE) ? dv_vlc_map[0][level].size : 16;

  570. 	if (run) {

  571. 	    size += (run < 16) ? dv_vlc_map[run-1][0].size : 13;

  572. 	}

  573.     }

  574.     return size;

  575. }

  576. #else

  577. static always_inline int dv_rl2vlc(int run, int l, uint32_t* vlc)

  578. {

  579.     *vlc = dv_vlc_map[run][((uint16_t)l)&0x1ff].vlc;

  580.     return dv_vlc_map[run][((uint16_t)l)&0x1ff].size;

  581. }

  582. 

  583. static always_inline int dv_rl2vlc_size(int run, int l)

  584. {

  585.     return dv_vlc_map[run][((uint16_t)l)&0x1ff].size;

  586. }

  587. #endif

  588. 

  589. typedef struct EncBlockInfo {

  590.     int area_q[4];

  591.     int bit_size[4];

  592.     int prev_run[4];

  593.     int cur_ac;

  594.     int cno;

  595.     int dct_mode;

  596.     DCTELEM *mb;

  597.     uint8_t partial_bit_count;

  598.     uint32_t partial_bit_buffer; /* we can't use uint16_t here */

  599. } EncBlockInfo;

  600. 

  601. static always_inline void dv_encode_ac(EncBlockInfo* bi, PutBitContext* pb_pool, 

  602.                                        int pb_size)

  603. {

  604.     int run;

  605.     int bits_left;

  606.     PutBitContext* pb = pb_pool;

  607.     int size = bi->partial_bit_count;

  608.     uint32_t vlc = bi->partial_bit_buffer;

  609.     

  610.     bi->partial_bit_count = bi->partial_bit_buffer = 0;

  611. vlc_loop:

  612.        /* Find suitable storage space */

  613.        for (; size > (bits_left = put_bits_left(pb)); pb++) {

  614.           if (bits_left) {

  615.               size -= bits_left;

  616. 	      put_bits(pb, bits_left, vlc >> size);

  617. 	      vlc = vlc & ((1<<size)-1);

  618. 	  }

  619. 	  if (pb_size == 1) {

  620. 	      bi->partial_bit_count = size;

  621. 	      bi->partial_bit_buffer = vlc;

  622. 	      return;

  623. 	  }

  624. 	  --pb_size;

  625.        }

  626.        

  627.        /* Store VLC */

  628.        put_bits(pb, size, vlc);

  629.        

  630.        /* Construct the next VLC */

  631.        run = 0;

  632.        for (; bi->cur_ac < 64; bi->cur_ac++, run++) {

  633.            if (bi->mb[bi->cur_ac]) {

  634. 	       size = dv_rl2vlc(run, bi->mb[bi->cur_ac], &vlc);

  635. 	       bi->cur_ac++;

  636. 	       goto vlc_loop;

  637. 	   }

  638.        }

  639.    

  640.        if (bi->cur_ac == 64) {

  641.            size = 4; vlc = 6; /* End Of Block stamp */

  642. 	   bi->cur_ac++;

  643. 	   goto vlc_loop;

  644.        }

  645. }

  646. 

  647. static always_inline void dv_set_class_number(DCTELEM* blk, EncBlockInfo* bi, 

  648.                                               const uint8_t* zigzag_scan, int bias)

  649. {

  650.     int i, area;

  651.     int run;

  652.     int classes[] = {12, 24, 36, 0xffff};

  653. 

  654.     run = 0;

  655.     bi->mb[0] = blk[0]; 

  656.     bi->cno = 0;

  657.     for (area = 0; area < 4; area++) {

  658.        bi->prev_run[area] = run;

  659.        bi->bit_size[area] = 0;

  660.        for (i=mb_area_start[area]; i<mb_area_start[area+1]; i++) {

  661.           bi->mb[i] = (blk[zigzag_scan[i]] / 16);

  662.           while ((bi->mb[i] ^ (bi->mb[i] >> 8)) > classes[bi->cno])

  663.               bi->cno++;

  664.        

  665.           if (bi->mb[i]) {

  666.               bi->bit_size[area] += dv_rl2vlc_size(run, bi->mb[i]);

  667. 	      run = 0;

  668.           } else

  669.               ++run;

  670.        }

  671.     }

  672.     bi->bit_size[3] += 4; /* EOB marker */

  673.     bi->cno += bias;

  674.     

  675.     if (bi->cno >= 3) { /* FIXME: we have to recreate bit_size[], prev_run[] */

  676.         bi->cno = 3;

  677. 	for (i=1; i<64; i++)

  678. 	   bi->mb[i] /= 2;

  679.     }

  680. }

  681. 

  682. #define SC(x, y) ((s[x] - s[y]) ^ ((s[x] - s[y]) >> 7))

  683. static always_inline int dv_guess_dct_mode(DCTELEM *blk) {

  684.     DCTELEM *s;

  685.     int score88 = 0;

  686.     int score248 = 0;

  687.     int i;

  688.     

  689.     /* Compute 8-8 score (small values give a better chance for 8-8 DCT) */

  690.     s = blk;

  691.     for(i=0; i<7; i++) {

  692.         score88 += SC(0,  8) + SC(1, 9) + SC(2, 10) + SC(3, 11) + 

  693. 	           SC(4, 12) + SC(5,13) + SC(6, 14) + SC(7, 15);

  694.         s += 8;

  695.     }

  696.     /* Compute 2-4-8 score (small values give a better chance for 2-4-8 DCT) */

  697.     s = blk;

  698.     for(i=0; i<6; i++) {

  699.         score248 += SC(0, 16) + SC(1,17) + SC(2, 18) + SC(3, 19) +

  700. 	            SC(4, 20) + SC(5,21) + SC(6, 22) + SC(7, 23);

  701.         s += 8;

  702.     }

  703. 

  704.     return (score88 - score248 > -10);

  705. }

  706. 

  707. static inline void dv_guess_qnos(EncBlockInfo* blks, int* qnos)

  708. {

  709.     int size[5];

  710.     int i, j, k, a, run;

  711.     EncBlockInfo* b;

  712.     

  713.     do {

  714.        b = blks;

  715.        for (i=0; i<5; i++) {

  716.           if (!qnos[i])

  717. 	      continue;

  718. 	  

  719. 	  qnos[i]--;

  720. 	  size[i] = 0;

  721.           for (j=0; j<6; j++, b++) {

  722. 	     for (a=0; a<4; a++) {

  723. 	        if (b->area_q[a] != dv_quant_shifts[qnos[i] + dv_quant_offset[b->cno]][a]) {

  724. 		    b->bit_size[a] = (a==3)?4:0;

  725. 		    b->area_q[a]++;

  726. 		    run = b->prev_run[a];

  727. 		    for (k=mb_area_start[a]; k<mb_area_start[a+1]; k++) {

  728. 		       b->mb[k] /= 2;

  729. 		       if (b->mb[k]) {

  730.                            b->bit_size[a] += dv_rl2vlc_size(run, b->mb[k]);

  731. 	                   run = 0;

  732.                        } else

  733.                            ++run;

  734. 		    }

  735. 		}

  736. 		size[i] += b->bit_size[a];

  737. 	     }

  738. 	  }

  739.        }

  740.     } while ((vs_total_ac_bits < size[0] + size[1] + size[2] + size[3] + size[4]) && 

  741.              (qnos[0]|qnos[1]|qnos[2]|qnos[3]|qnos[4]));

  742. }

  743. 

  744. /*

  745.  * This is a very rough initial implementaion. The performance is

  746.  * horrible and the weighting is missing. But it's missing from the 

  747.  * decoding step also -- so at least we're on the same page with decoder ;-)

  748.  */

  749. static inline void dv_encode_video_segment(DVVideoContext *s, 

  750.                                            uint8_t *dif, 

  751.                                            const uint16_t *mb_pos_ptr)

  752. {

  753.     int mb_index, i, j, v;

  754.     int mb_x, mb_y, c_offset, linesize; 

  755.     uint8_t*  y_ptr;

  756.     uint8_t*  data;

  757.     uint8_t*  ptr;

  758.     int       do_edge_wrap;

  759.     DCTELEM   block[64] __align8;

  760.     DCTELEM   sblock[5*6][64] __align8;

  761.     EncBlockInfo  enc_blks[5*6];

  762.     PutBitContext pbs[5*6];

  763.     PutBitContext* pb; 

  764.     EncBlockInfo* enc_blk;

  765.     int       vs_bit_size = 0;

  766.     int       qnos[5];

  767.    

  768.     enc_blk = &enc_blks[0];

  769.     pb = &pbs[0];

  770.     for(mb_index = 0; mb_index < 5; mb_index++) {

  771.         v = *mb_pos_ptr++;

  772.         mb_x = v & 0xff;

  773.         mb_y = v >> 8;

  774.         y_ptr = s->picture.data[0] + (mb_y * s->picture.linesize[0] * 8) + (mb_x * 8);

  775. 	c_offset = (s->sys->pix_fmt == PIX_FMT_YUV411P) ?

  776. 	           ((mb_y * s->picture.linesize[1] * 8) + ((mb_x >> 2) * 8)) :

  777. 		   (((mb_y >> 1) * s->picture.linesize[1] * 8) + ((mb_x >> 1) * 8));

  778. 	do_edge_wrap = 0;

  779. 	qnos[mb_index] = 15; /* No quantization */

  780.         ptr = dif + mb_index*80 + 4;

  781.         for(j = 0;j < 6; j++) {

  782.             if (j < 4) {  /* Four Y blocks */

  783. 		/* NOTE: at end of line, the macroblock is handled as 420 */

  784. 		if (s->sys->pix_fmt == PIX_FMT_YUV411P && mb_x < (704 / 8)) {

  785.                     data = y_ptr + (j * 8);

  786.                 } else {

  787.                     data = y_ptr + ((j & 1) * 8) + ((j >> 1) * 8 * s->picture.linesize[0]);

  788.                 }

  789. 		linesize = s->picture.linesize[0];

  790.             } else {      /* Cr and Cb blocks */

  791. 	        /* don't ask Fabrice why they inverted Cb and Cr ! */

  792. 	        data = s->picture.data[6 - j] + c_offset;

  793. 		linesize = s->picture.linesize[6 - j];

  794. 		if (s->sys->pix_fmt == PIX_FMT_YUV411P && mb_x >= (704 / 8))

  795. 		    do_edge_wrap = 1;

  796. 	    }	

  797.             

  798. 	    /* Everything is set up -- now just copy data -> DCT block */

  799. 	    if (do_edge_wrap) {  /* Edge wrap copy: 4x16 -> 8x8 */

  800. 		uint8_t* d;

  801. 		DCTELEM *b = block;

  802. 	        for (i=0;i<8;i++) {

  803. 		   d = data + 8 * linesize;

  804. 		   b[0] = data[0]; b[1] = data[1]; b[2] = data[2]; b[3] = data[3];

  805.                    b[4] =    d[0]; b[5] =    d[1]; b[6] =    d[2]; b[7] =    d[3];

  806. 		   data += linesize;

  807. 		   b += 8;

  808. 		}

  809. 	    } else {             /* Simple copy: 8x8 -> 8x8 */

  810. 	        s->get_pixels(block, data, linesize);

  811. 	    }

  812. 	  

  813.             enc_blk->dct_mode = dv_guess_dct_mode(block);

  814. 	    enc_blk->mb = &sblock[mb_index*6+j][0];

  815. 	    enc_blk->area_q[0] = enc_blk->area_q[1] = enc_blk->area_q[2] = enc_blk->area_q[3] = 0;

  816. 	    enc_blk->partial_bit_count = 0;

  817. 	    enc_blk->partial_bit_buffer = 0;

  818. 	    enc_blk->cur_ac = 1;

  819. 	    

  820. 	    s->fdct[enc_blk->dct_mode](block);

  821. 	    

  822. 	    dv_set_class_number(block, enc_blk, 

  823. 	                        enc_blk->dct_mode ? ff_zigzag248_direct : ff_zigzag_direct,

  824. 				j/4*(j%2));

  825.            

  826.             init_put_bits(pb, ptr, block_sizes[j]/8);

  827. 	    put_bits(pb, 9, (uint16_t)(((enc_blk->mb[0] >> 3) - 1024) >> 2));

  828. 	    put_bits(pb, 1, enc_blk->dct_mode);

  829. 	    put_bits(pb, 2, enc_blk->cno);

  830. 	    

  831. 	    vs_bit_size += enc_blk->bit_size[0] + enc_blk->bit_size[1] +

  832. 	                   enc_blk->bit_size[2] + enc_blk->bit_size[3];

  833. 	    ++enc_blk;

  834. 	    ++pb;

  835. 	    ptr += block_sizes[j]/8;

  836.         }

  837.     }

  838. 

  839.     if (vs_total_ac_bits < vs_bit_size)

  840.         dv_guess_qnos(&enc_blks[0], &qnos[0]);

  841. 

  842.     for (i=0; i<5; i++) {

  843.        dif[i*80 + 3] = qnos[i];

  844.     }

  845. 

  846.     /* First pass over individual cells only */

  847.     for (j=0; j<5*6; j++)

  848.        dv_encode_ac(&enc_blks[j], &pbs[j], 1);

  849. 

  850.     /* Second pass over each MB space */

  851.     for (j=0; j<5*6; j++) {

  852.        if (enc_blks[j].cur_ac < 65 || enc_blks[j].partial_bit_count)

  853.            dv_encode_ac(&enc_blks[j], &pbs[(j/6)*6], 6);

  854.     }

  855. 

  856.     /* Third and final pass over the whole vides segment space */

  857.     for (j=0; j<5*6; j++) {

  858.        if (enc_blks[j].cur_ac < 65 || enc_blks[j].partial_bit_count)

  859.            dv_encode_ac(&enc_blks[j], &pbs[0], 6*5);

  860.     }

  861. 

  862.     for (j=0; j<5*6; j++)

  863.        flush_put_bits(&pbs[j]);

  864. }

  865. 

  866. static int dv_decode_mt(AVCodecContext *avctx, void* sl)

  867. {

  868.     DVVideoContext *s = avctx->priv_data;

  869.     int slice = (size_t)sl;

  870.     dv_decode_video_segment(s, &s->buf[((slice/27)*6+(slice/3)+slice*5+7)*80],

  871. 	                    &s->sys->video_place[slice*5]);

  872.     return 0;

  873. }

  874. 

  875. static int dv_encode_mt(AVCodecContext *avctx, void* sl)

  876. {

  877.     DVVideoContext *s = avctx->priv_data;

  878.     int slice = (size_t)sl;

  879.     dv_encode_video_segment(s, &s->buf[((slice/27)*6+(slice/3)+slice*5+7)*80],

  880. 	                    &s->sys->video_place[slice*5]);

  881.     return 0;

  882. }

  883. 

  884. /* NOTE: exactly one frame must be given (120000 bytes for NTSC,

  885.    144000 bytes for PAL) */

  886. static int dvvideo_decode_frame(AVCodecContext *avctx, 

  887.                                  void *data, int *data_size,

  888.                                  uint8_t *buf, int buf_size)

  889. {

  890.     DVVideoContext *s = avctx->priv_data;

  891.   

  892.     /* special case for last picture */

  893.     if(buf_size==0)

  894.         return 0;

  895.     

  896.     s->sys = dv_frame_profile(buf);

  897.     if (!s->sys || buf_size < s->sys->frame_size)

  898.         return -1; /* NOTE: we only accept several full frames */

  899. 

  900.     if(s->picture.data[0])

  901.         avctx->release_buffer(avctx, &s->picture);

  902.     

  903.     s->picture.reference = 0;

  904.     avctx->pix_fmt = s->sys->pix_fmt;

  905.     avctx->width = s->sys->width;

  906.     avctx->height = s->sys->height;

  907.     if(avctx->get_buffer(avctx, &s->picture) < 0) {

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

  909.         return -1;

  910.     }

  911.     s->picture.interlaced_frame = 1;

  912.     s->picture.top_field_first = 0;

  913. 

  914.     s->buf = buf;

  915.     avctx->execute(avctx, dv_decode_mt, (void**)&dv_anchor[0], NULL, 

  916. 	           s->sys->difseg_size * 27);

  917.     

  918.     emms_c();

  919. 

  920.     /* return image */

  921.     *data_size = sizeof(AVFrame);

  922.     *(AVFrame*)data= s->picture;

  923.     

  924.     return s->sys->frame_size;

  925. }

  926. 

  927. static int dvvideo_encode_frame(AVCodecContext *c, uint8_t *buf, int buf_size, 

  928.                                 void *data)

  929. {

  930.     DVVideoContext *s = c->priv_data;

  931. 

  932.     s->sys = dv_codec_profile(c);

  933.     if (!s->sys)

  934. 	return -1;

  935.     

  936.     c->pix_fmt = s->sys->pix_fmt;

  937.     s->picture = *((AVFrame *)data);

  938. 

  939.     s->buf = buf;

  940.     c->execute(c, dv_encode_mt, (void**)&dv_anchor[0], NULL, 

  941. 	       s->sys->difseg_size * 27);

  942. 

  943.     emms_c();

  944.     return s->sys->frame_size;

  945. }

  946. 

  947. AVCodec dvvideo_encoder = {

  948.     "dvvideo",

  949.     CODEC_TYPE_VIDEO,

  950.     CODEC_ID_DVVIDEO,

  951.     sizeof(DVVideoContext),

  952.     dvvideo_init,

  953.     dvvideo_encode_frame,

  954.     NULL,

  955.     NULL,

  956.     CODEC_CAP_DR1,

  957.     NULL

  958. };

  959. 

  960. AVCodec dvvideo_decoder = {

  961.     "dvvideo",

  962.     CODEC_TYPE_VIDEO,

  963.     CODEC_ID_DVVIDEO,

  964.     sizeof(DVVideoContext),

  965.     dvvideo_init,

  966.     NULL,

  967.     NULL,

  968.     dvvideo_decode_frame,

  969.     CODEC_CAP_DR1,

  970.     NULL

  971. };