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.
55886 Commits
32 Branches
366MB
Tree: cf06dee58d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'cf06dee58d'
${ noResults }
FFmpeg/libavcodec/snowenc.c

2004 lines
76KB
Raw Blame History 

  1. /*

  2.  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>

  3.  *

  4.  * This file is part of FFmpeg.

  5.  *

  6.  * FFmpeg is free software; you can redistribute it and/or

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

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

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

  10.  *

  11.  * FFmpeg is distributed in the hope that it will be useful,

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

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

  14.  * Lesser General Public License for more details.

  15.  *

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

  17.  * License along with FFmpeg; if not, write to the Free Software

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

  19.  */

  20. 

  21. #include "libavutil/intmath.h"

  22. #include "libavutil/log.h"

  23. #include "libavutil/opt.h"

  24. #include "avcodec.h"

  25. #include "internal.h"

  26. #include "dsputil.h"

  27. #include "internal.h"

  28. #include "snow_dwt.h"

  29. #include "snow.h"

  30. 

  31. #include "rangecoder.h"

  32. #include "mathops.h"

  33. 

  34. #include "mpegvideo.h"

  35. #include "h263.h"

  36. 

  37. static av_cold int encode_init(AVCodecContext *avctx)

  38. {

  39.     SnowContext *s = avctx->priv_data;

  40.     int plane_index, ret;

  41. 

  42.     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){

  43.         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"

  44.                "Use vstrict=-2 / -strict -2 to use it anyway.\n");

  45.         return -1;

  46.     }

  47. 

  48.     if(avctx->prediction_method == DWT_97

  49.        && (avctx->flags & CODEC_FLAG_QSCALE)

  50.        && avctx->global_quality == 0){

  51.         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");

  52.         return -1;

  53.     }

  54. 

  55.     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type

  56. 

  57.     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;

  58.     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;

  59. 

  60.     for(plane_index=0; plane_index<3; plane_index++){

  61.         s->plane[plane_index].diag_mc= 1;

  62.         s->plane[plane_index].htaps= 6;

  63.         s->plane[plane_index].hcoeff[0]=  40;

  64.         s->plane[plane_index].hcoeff[1]= -10;

  65.         s->plane[plane_index].hcoeff[2]=   2;

  66.         s->plane[plane_index].fast_mc= 1;

  67.     }

  68. 

  69.     if ((ret = ff_snow_common_init(avctx)) < 0) {

  70.         ff_snow_common_end(avctx->priv_data);

  71.         return ret;

  72.     }

  73.     ff_snow_alloc_blocks(s);

  74. 

  75.     s->version=0;

  76. 

  77.     s->m.avctx   = avctx;

  78.     s->m.flags   = avctx->flags;

  79.     s->m.bit_rate= avctx->bit_rate;

  80. 

  81.     s->m.me.temp      =

  82.     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));

  83.     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));

  84.     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));

  85.     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));

  86.     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)

  87.         return AVERROR(ENOMEM);

  88. 

  89.     ff_h263_encode_init(&s->m); //mv_penalty

  90. 

  91.     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);

  92. 

  93.     if(avctx->flags&CODEC_FLAG_PASS1){

  94.         if(!avctx->stats_out)

  95.             avctx->stats_out = av_mallocz(256);

  96. 

  97.         if (!avctx->stats_out)

  98.             return AVERROR(ENOMEM);

  99.     }

  100.     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){

  101.         if(ff_rate_control_init(&s->m) < 0)

  102.             return -1;

  103.     }

  104.     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));

  105. 

  106.     switch(avctx->pix_fmt){

  107.     case AV_PIX_FMT_YUV444P:

  108. //    case AV_PIX_FMT_YUV422P:

  109.     case AV_PIX_FMT_YUV420P:

  110. //     case AV_PIX_FMT_GRAY8:

  111. //    case AV_PIX_FMT_YUV411P:

  112.     case AV_PIX_FMT_YUV410P:

  113.         s->colorspace_type= 0;

  114.         break;

  115. /*    case AV_PIX_FMT_RGB32:

  116.         s->colorspace= 1;

  117.         break;*/

  118.     default:

  119.         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");

  120.         return -1;

  121.     }

  122.     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);

  123. 

  124.     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);

  125.     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);

  126. 

  127.     s->input_picture = av_frame_alloc();

  128.     if ((ret = ff_get_buffer(s->avctx, s->input_picture, AV_GET_BUFFER_FLAG_REF)) < 0)

  129.         return ret;

  130. 

  131.     if(s->avctx->me_method == ME_ITER){

  132.         int i;

  133.         int size= s->b_width * s->b_height << 2*s->block_max_depth;

  134.         for(i=0; i<s->max_ref_frames; i++){

  135.             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));

  136.             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));

  137.             if (!s->ref_mvs[i] || !s->ref_scores[i])

  138.                 return AVERROR(ENOMEM);

  139.         }

  140.     }

  141. 

  142.     return 0;

  143. }

  144. 

  145. //near copy & paste from dsputil, FIXME

  146. static int pix_sum(uint8_t * pix, int line_size, int w, int h)

  147. {

  148.     int s, i, j;

  149. 

  150.     s = 0;

  151.     for (i = 0; i < h; i++) {

  152.         for (j = 0; j < w; j++) {

  153.             s += pix[0];

  154.             pix ++;

  155.         }

  156.         pix += line_size - w;

  157.     }

  158.     return s;

  159. }

  160. 

  161. //near copy & paste from dsputil, FIXME

  162. static int pix_norm1(uint8_t * pix, int line_size, int w)

  163. {

  164.     int s, i, j;

  165.     uint32_t *sq = ff_squareTbl + 256;

  166. 

  167.     s = 0;

  168.     for (i = 0; i < w; i++) {

  169.         for (j = 0; j < w; j ++) {

  170.             s += sq[pix[0]];

  171.             pix ++;

  172.         }

  173.         pix += line_size - w;

  174.     }

  175.     return s;

  176. }

  177. 

  178. static inline int get_penalty_factor(int lambda, int lambda2, int type){

  179.     switch(type&0xFF){

  180.     default:

  181.     case FF_CMP_SAD:

  182.         return lambda>>FF_LAMBDA_SHIFT;

  183.     case FF_CMP_DCT:

  184.         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);

  185.     case FF_CMP_W53:

  186.         return (4*lambda)>>(FF_LAMBDA_SHIFT);

  187.     case FF_CMP_W97:

  188.         return (2*lambda)>>(FF_LAMBDA_SHIFT);

  189.     case FF_CMP_SATD:

  190.     case FF_CMP_DCT264:

  191.         return (2*lambda)>>FF_LAMBDA_SHIFT;

  192.     case FF_CMP_RD:

  193.     case FF_CMP_PSNR:

  194.     case FF_CMP_SSE:

  195.     case FF_CMP_NSSE:

  196.         return lambda2>>FF_LAMBDA_SHIFT;

  197.     case FF_CMP_BIT:

  198.         return 1;

  199.     }

  200. }

  201. 

  202. //FIXME copy&paste

  203. #define P_LEFT P[1]

  204. #define P_TOP P[2]

  205. #define P_TOPRIGHT P[3]

  206. #define P_MEDIAN P[4]

  207. #define P_MV1 P[9]

  208. #define FLAG_QPEL   1 //must be 1

  209. 

  210. static int encode_q_branch(SnowContext *s, int level, int x, int y){

  211.     uint8_t p_buffer[1024];

  212.     uint8_t i_buffer[1024];

  213.     uint8_t p_state[sizeof(s->block_state)];

  214.     uint8_t i_state[sizeof(s->block_state)];

  215.     RangeCoder pc, ic;

  216.     uint8_t *pbbak= s->c.bytestream;

  217.     uint8_t *pbbak_start= s->c.bytestream_start;

  218.     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;

  219.     const int w= s->b_width  << s->block_max_depth;

  220.     const int h= s->b_height << s->block_max_depth;

  221.     const int rem_depth= s->block_max_depth - level;

  222.     const int index= (x + y*w) << rem_depth;

  223.     const int block_w= 1<<(LOG2_MB_SIZE - level);

  224.     int trx= (x+1)<<rem_depth;

  225.     int try= (y+1)<<rem_depth;

  226.     const BlockNode *left  = x ? &s->block[index-1] : &null_block;

  227.     const BlockNode *top   = y ? &s->block[index-w] : &null_block;

  228.     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;

  229.     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;

  230.     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;

  231.     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt

  232.     int pl = left->color[0];

  233.     int pcb= left->color[1];

  234.     int pcr= left->color[2];

  235.     int pmx, pmy;

  236.     int mx=0, my=0;

  237.     int l,cr,cb;

  238.     const int stride= s->current_picture->linesize[0];

  239.     const int uvstride= s->current_picture->linesize[1];

  240.     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,

  241.                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),

  242.                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};

  243.     int P[10][2];

  244.     int16_t last_mv[3][2];

  245.     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused

  246.     const int shift= 1+qpel;

  247.     MotionEstContext *c= &s->m.me;

  248.     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);

  249.     int mx_context= av_log2(2*FFABS(left->mx - top->mx));

  250.     int my_context= av_log2(2*FFABS(left->my - top->my));

  251.     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;

  252.     int ref, best_ref, ref_score, ref_mx, ref_my;

  253. 

  254.     av_assert0(sizeof(s->block_state) >= 256);

  255.     if(s->keyframe){

  256.         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);

  257.         return 0;

  258.     }

  259. 

  260. //    clip predictors / edge ?

  261. 

  262.     P_LEFT[0]= left->mx;

  263.     P_LEFT[1]= left->my;

  264.     P_TOP [0]= top->mx;

  265.     P_TOP [1]= top->my;

  266.     P_TOPRIGHT[0]= tr->mx;

  267.     P_TOPRIGHT[1]= tr->my;

  268. 

  269.     last_mv[0][0]= s->block[index].mx;

  270.     last_mv[0][1]= s->block[index].my;

  271.     last_mv[1][0]= right->mx;

  272.     last_mv[1][1]= right->my;

  273.     last_mv[2][0]= bottom->mx;

  274.     last_mv[2][1]= bottom->my;

  275. 

  276.     s->m.mb_stride=2;

  277.     s->m.mb_x=

  278.     s->m.mb_y= 0;

  279.     c->skip= 0;

  280. 

  281.     av_assert1(c->  stride ==   stride);

  282.     av_assert1(c->uvstride == uvstride);

  283. 

  284.     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);

  285.     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);

  286.     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);

  287.     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;

  288. 

  289.     c->xmin = - x*block_w - 16+3;

  290.     c->ymin = - y*block_w - 16+3;

  291.     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;

  292.     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;

  293. 

  294.     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);

  295.     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);

  296.     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);

  297.     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);

  298.     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);

  299.     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip

  300.     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);

  301. 

  302.     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  303.     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  304. 

  305.     if (!y) {

  306.         c->pred_x= P_LEFT[0];

  307.         c->pred_y= P_LEFT[1];

  308.     } else {

  309.         c->pred_x = P_MEDIAN[0];

  310.         c->pred_y = P_MEDIAN[1];

  311.     }

  312. 

  313.     score= INT_MAX;

  314.     best_ref= 0;

  315.     for(ref=0; ref<s->ref_frames; ref++){

  316.         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);

  317. 

  318.         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,

  319.                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);

  320. 

  321.         av_assert2(ref_mx >= c->xmin);

  322.         av_assert2(ref_mx <= c->xmax);

  323.         av_assert2(ref_my >= c->ymin);

  324.         av_assert2(ref_my <= c->ymax);

  325. 

  326.         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);

  327.         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);

  328.         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;

  329.         if(s->ref_mvs[ref]){

  330.             s->ref_mvs[ref][index][0]= ref_mx;

  331.             s->ref_mvs[ref][index][1]= ref_my;

  332.             s->ref_scores[ref][index]= ref_score;

  333.         }

  334.         if(score > ref_score){

  335.             score= ref_score;

  336.             best_ref= ref;

  337.             mx= ref_mx;

  338.             my= ref_my;

  339.         }

  340.     }

  341.     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2

  342. 

  343.   //  subpel search

  344.     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);

  345.     pc= s->c;

  346.     pc.bytestream_start=

  347.     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo

  348.     memcpy(p_state, s->block_state, sizeof(s->block_state));

  349. 

  350.     if(level!=s->block_max_depth)

  351.         put_rac(&pc, &p_state[4 + s_context], 1);

  352.     put_rac(&pc, &p_state[1 + left->type + top->type], 0);

  353.     if(s->ref_frames > 1)

  354.         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);

  355.     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);

  356.     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);

  357.     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);

  358.     p_len= pc.bytestream - pc.bytestream_start;

  359.     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;

  360. 

  361.     block_s= block_w*block_w;

  362.     sum = pix_sum(current_data[0], stride, block_w, block_w);

  363.     l= (sum + block_s/2)/block_s;

  364.     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;

  365. 

  366.     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);

  367.     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);

  368.     cb= (sum + block_s/2)/block_s;

  369. //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;

  370.     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);

  371.     cr= (sum + block_s/2)/block_s;

  372. //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;

  373. 

  374.     ic= s->c;

  375.     ic.bytestream_start=

  376.     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo

  377.     memcpy(i_state, s->block_state, sizeof(s->block_state));

  378.     if(level!=s->block_max_depth)

  379.         put_rac(&ic, &i_state[4 + s_context], 1);

  380.     put_rac(&ic, &i_state[1 + left->type + top->type], 1);

  381.     put_symbol(&ic, &i_state[32],  l-pl , 1);

  382.     put_symbol(&ic, &i_state[64], cb-pcb, 1);

  383.     put_symbol(&ic, &i_state[96], cr-pcr, 1);

  384.     i_len= ic.bytestream - ic.bytestream_start;

  385.     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;

  386. 

  387. //    assert(score==256*256*256*64-1);

  388.     av_assert1(iscore < 255*255*256 + s->lambda2*10);

  389.     av_assert1(iscore >= 0);

  390.     av_assert1(l>=0 && l<=255);

  391.     av_assert1(pl>=0 && pl<=255);

  392. 

  393.     if(level==0){

  394.         int varc= iscore >> 8;

  395.         int vard= score >> 8;

  396.         if (vard <= 64 || vard < varc)

  397.             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  398.         else

  399.             c->scene_change_score+= s->m.qscale;

  400.     }

  401. 

  402.     if(level!=s->block_max_depth){

  403.         put_rac(&s->c, &s->block_state[4 + s_context], 0);

  404.         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);

  405.         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);

  406.         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);

  407.         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);

  408.         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead

  409. 

  410.         if(score2 < score && score2 < iscore)

  411.             return score2;

  412.     }

  413. 

  414.     if(iscore < score){

  415.         pred_mv(s, &pmx, &pmy, 0, left, top, tr);

  416.         memcpy(pbbak, i_buffer, i_len);

  417.         s->c= ic;

  418.         s->c.bytestream_start= pbbak_start;

  419.         s->c.bytestream= pbbak + i_len;

  420.         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);

  421.         memcpy(s->block_state, i_state, sizeof(s->block_state));

  422.         return iscore;

  423.     }else{

  424.         memcpy(pbbak, p_buffer, p_len);

  425.         s->c= pc;

  426.         s->c.bytestream_start= pbbak_start;

  427.         s->c.bytestream= pbbak + p_len;

  428.         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);

  429.         memcpy(s->block_state, p_state, sizeof(s->block_state));

  430.         return score;

  431.     }

  432. }

  433. 

  434. static void encode_q_branch2(SnowContext *s, int level, int x, int y){

  435.     const int w= s->b_width  << s->block_max_depth;

  436.     const int rem_depth= s->block_max_depth - level;

  437.     const int index= (x + y*w) << rem_depth;

  438.     int trx= (x+1)<<rem_depth;

  439.     BlockNode *b= &s->block[index];

  440.     const BlockNode *left  = x ? &s->block[index-1] : &null_block;

  441.     const BlockNode *top   = y ? &s->block[index-w] : &null_block;

  442.     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;

  443.     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt

  444.     int pl = left->color[0];

  445.     int pcb= left->color[1];

  446.     int pcr= left->color[2];

  447.     int pmx, pmy;

  448.     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);

  449.     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;

  450.     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;

  451.     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;

  452. 

  453.     if(s->keyframe){

  454.         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);

  455.         return;

  456.     }

  457. 

  458.     if(level!=s->block_max_depth){

  459.         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){

  460.             put_rac(&s->c, &s->block_state[4 + s_context], 1);

  461.         }else{

  462.             put_rac(&s->c, &s->block_state[4 + s_context], 0);

  463.             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);

  464.             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);

  465.             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);

  466.             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);

  467.             return;

  468.         }

  469.     }

  470.     if(b->type & BLOCK_INTRA){

  471.         pred_mv(s, &pmx, &pmy, 0, left, top, tr);

  472.         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);

  473.         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);

  474.         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);

  475.         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);

  476.         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);

  477.     }else{

  478.         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);

  479.         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);

  480.         if(s->ref_frames > 1)

  481.             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);

  482.         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);

  483.         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);

  484.         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);

  485.     }

  486. }

  487. 

  488. static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){

  489.     int i, x2, y2;

  490.     Plane *p= &s->plane[plane_index];

  491.     const int block_size = MB_SIZE >> s->block_max_depth;

  492.     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;

  493.     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;

  494.     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];

  495.     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;

  496.     const int ref_stride= s->current_picture->linesize[plane_index];

  497.     uint8_t *src= s-> input_picture->data[plane_index];

  498.     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned

  499.     const int b_stride = s->b_width << s->block_max_depth;

  500.     const int w= p->width;

  501.     const int h= p->height;

  502.     int index= mb_x + mb_y*b_stride;

  503.     BlockNode *b= &s->block[index];

  504.     BlockNode backup= *b;

  505.     int ab=0;

  506.     int aa=0;

  507. 

  508.     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above

  509. 

  510.     b->type|= BLOCK_INTRA;

  511.     b->color[plane_index]= 0;

  512.     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));

  513. 

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

  515.         int mb_x2= mb_x + (i &1) - 1;

  516.         int mb_y2= mb_y + (i>>1) - 1;

  517.         int x= block_w*mb_x2 + block_w/2;

  518.         int y= block_h*mb_y2 + block_h/2;

  519. 

  520.         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,

  521.                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);

  522. 

  523.         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){

  524.             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){

  525.                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;

  526.                 int obmc_v= obmc[index];

  527.                 int d;

  528.                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];

  529.                 if(x<0) obmc_v += obmc[index + block_w];

  530.                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];

  531.                 if(x+block_w>w) obmc_v += obmc[index - block_w];

  532.                 //FIXME precalculate this or simplify it somehow else

  533. 

  534.                 d = -dst[index] + (1<<(FRAC_BITS-1));

  535.                 dst[index] = d;

  536.                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;

  537.                 aa += obmc_v * obmc_v; //FIXME precalculate this

  538.             }

  539.         }

  540.     }

  541.     *b= backup;

  542. 

  543.     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping

  544. }

  545. 

  546. static inline int get_block_bits(SnowContext *s, int x, int y, int w){

  547.     const int b_stride = s->b_width << s->block_max_depth;

  548.     const int b_height = s->b_height<< s->block_max_depth;

  549.     int index= x + y*b_stride;

  550.     const BlockNode *b     = &s->block[index];

  551.     const BlockNode *left  = x ? &s->block[index-1] : &null_block;

  552.     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;

  553.     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;

  554.     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;

  555.     int dmx, dmy;

  556. //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));

  557. //  int my_context= av_log2(2*FFABS(left->my - top->my));

  558. 

  559.     if(x<0 || x>=b_stride || y>=b_height)

  560.         return 0;

  561. /*

  562. 1            0      0

  563. 01X          1-2    1

  564. 001XX        3-6    2-3

  565. 0001XXX      7-14   4-7

  566. 00001XXXX   15-30   8-15

  567. */

  568. //FIXME try accurate rate

  569. //FIXME intra and inter predictors if surrounding blocks are not the same type

  570.     if(b->type & BLOCK_INTRA){

  571.         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))

  572.                    + av_log2(2*FFABS(left->color[1] - b->color[1]))

  573.                    + av_log2(2*FFABS(left->color[2] - b->color[2])));

  574.     }else{

  575.         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);

  576.         dmx-= b->mx;

  577.         dmy-= b->my;

  578.         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda

  579.                     + av_log2(2*FFABS(dmy))

  580.                     + av_log2(2*b->ref));

  581.     }

  582. }

  583. 

  584. static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){

  585.     Plane *p= &s->plane[plane_index];

  586.     const int block_size = MB_SIZE >> s->block_max_depth;

  587.     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;

  588.     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;

  589.     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;

  590.     const int ref_stride= s->current_picture->linesize[plane_index];

  591.     uint8_t *dst= s->current_picture->data[plane_index];

  592.     uint8_t *src= s->  input_picture->data[plane_index];

  593.     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;

  594.     uint8_t *cur = s->scratchbuf;

  595.     uint8_t *tmp = s->emu_edge_buffer;

  596.     const int b_stride = s->b_width << s->block_max_depth;

  597.     const int b_height = s->b_height<< s->block_max_depth;

  598.     const int w= p->width;

  599.     const int h= p->height;

  600.     int distortion;

  601.     int rate= 0;

  602.     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);

  603.     int sx= block_w*mb_x - block_w/2;

  604.     int sy= block_h*mb_y - block_h/2;

  605.     int x0= FFMAX(0,-sx);

  606.     int y0= FFMAX(0,-sy);

  607.     int x1= FFMIN(block_w*2, w-sx);

  608.     int y1= FFMIN(block_h*2, h-sy);

  609.     int i,x,y;

  610. 

  611.     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w

  612. 

  613.     ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_h*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);

  614. 

  615.     for(y=y0; y<y1; y++){

  616.         const uint8_t *obmc1= obmc_edged[y];

  617.         const IDWTELEM *pred1 = pred + y*obmc_stride;

  618.         uint8_t *cur1 = cur + y*ref_stride;

  619.         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;

  620.         for(x=x0; x<x1; x++){

  621. #if FRAC_BITS >= LOG2_OBMC_MAX

  622.             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);

  623. #else

  624.             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);

  625. #endif

  626.             v = (v + pred1[x]) >> FRAC_BITS;

  627.             if(v&(~255)) v= ~(v>>31);

  628.             dst1[x] = v;

  629.         }

  630.     }

  631. 

  632.     /* copy the regions where obmc[] = (uint8_t)256 */

  633.     if(LOG2_OBMC_MAX == 8

  634.         && (mb_x == 0 || mb_x == b_stride-1)

  635.         && (mb_y == 0 || mb_y == b_height-1)){

  636.         if(mb_x == 0)

  637.             x1 = block_w;

  638.         else

  639.             x0 = block_w;

  640.         if(mb_y == 0)

  641.             y1 = block_h;

  642.         else

  643.             y0 = block_h;

  644.         for(y=y0; y<y1; y++)

  645.             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);

  646.     }

  647. 

  648.     if(block_w==16){

  649.         /* FIXME rearrange dsputil to fit 32x32 cmp functions */

  650.         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */

  651.         /* FIXME cmps overlap but do not cover the wavelet's whole support.

  652.          * So improving the score of one block is not strictly guaranteed

  653.          * to improve the score of the whole frame, thus iterative motion

  654.          * estimation does not always converge. */

  655.         if(s->avctx->me_cmp == FF_CMP_W97)

  656.             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);

  657.         else if(s->avctx->me_cmp == FF_CMP_W53)

  658.             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);

  659.         else{

  660.             distortion = 0;

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

  662.                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;

  663.                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);

  664.             }

  665.         }

  666.     }else{

  667.         av_assert2(block_w==8);

  668.         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);

  669.     }

  670. 

  671.     if(plane_index==0){

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

  673. /* ..RRr

  674.  * .RXx.

  675.  * rxx..

  676.  */

  677.             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);

  678.         }

  679.         if(mb_x == b_stride-2)

  680.             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);

  681.     }

  682.     return distortion + rate*penalty_factor;

  683. }

  684. 

  685. static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){

  686.     int i, y2;

  687.     Plane *p= &s->plane[plane_index];

  688.     const int block_size = MB_SIZE >> s->block_max_depth;

  689.     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;

  690.     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;

  691.     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];

  692.     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;

  693.     const int ref_stride= s->current_picture->linesize[plane_index];

  694.     uint8_t *dst= s->current_picture->data[plane_index];

  695.     uint8_t *src= s-> input_picture->data[plane_index];

  696.     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst

  697.     // const has only been removed from zero_dst to suppress a warning

  698.     static IDWTELEM zero_dst[4096]; //FIXME

  699.     const int b_stride = s->b_width << s->block_max_depth;

  700.     const int w= p->width;

  701.     const int h= p->height;

  702.     int distortion= 0;

  703.     int rate= 0;

  704.     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);

  705. 

  706.     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below

  707. 

  708.     for(i=0; i<9; i++){

  709.         int mb_x2= mb_x + (i%3) - 1;

  710.         int mb_y2= mb_y + (i/3) - 1;

  711.         int x= block_w*mb_x2 + block_w/2;

  712.         int y= block_h*mb_y2 + block_h/2;

  713. 

  714.         add_yblock(s, 0, NULL, zero_dst, dst, obmc,

  715.                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);

  716. 

  717.         //FIXME find a cleaner/simpler way to skip the outside stuff

  718.         for(y2= y; y2<0; y2++)

  719.             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);

  720.         for(y2= h; y2<y+block_h; y2++)

  721.             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);

  722.         if(x<0){

  723.             for(y2= y; y2<y+block_h; y2++)

  724.                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);

  725.         }

  726.         if(x+block_w > w){

  727.             for(y2= y; y2<y+block_h; y2++)

  728.                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);

  729.         }

  730. 

  731.         av_assert1(block_w== 8 || block_w==16);

  732.         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);

  733.     }

  734. 

  735.     if(plane_index==0){

  736.         BlockNode *b= &s->block[mb_x+mb_y*b_stride];

  737.         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);

  738. 

  739. /* ..RRRr

  740.  * .RXXx.

  741.  * .RXXx.

  742.  * rxxx.

  743.  */

  744.         if(merged)

  745.             rate = get_block_bits(s, mb_x, mb_y, 2);

  746.         for(i=merged?4:0; i<9; i++){

  747.             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};

  748.             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);

  749.         }

  750.     }

  751.     return distortion + rate*penalty_factor;

  752. }

  753. 

  754. static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){

  755.     const int w= b->width;

  756.     const int h= b->height;

  757.     int x, y;

  758. 

  759.     if(1){

  760.         int run=0;

  761.         int *runs = s->run_buffer;

  762.         int run_index=0;

  763.         int max_index;

  764. 

  765.         for(y=0; y<h; y++){

  766.             for(x=0; x<w; x++){

  767.                 int v, p=0;

  768.                 int /*ll=0, */l=0, lt=0, t=0, rt=0;

  769.                 v= src[x + y*stride];

  770. 

  771.                 if(y){

  772.                     t= src[x + (y-1)*stride];

  773.                     if(x){

  774.                         lt= src[x - 1 + (y-1)*stride];

  775.                     }

  776.                     if(x + 1 < w){

  777.                         rt= src[x + 1 + (y-1)*stride];

  778.                     }

  779.                 }

  780.                 if(x){

  781.                     l= src[x - 1 + y*stride];

  782.                     /*if(x > 1){

  783.                         if(orientation==1) ll= src[y + (x-2)*stride];

  784.                         else               ll= src[x - 2 + y*stride];

  785.                     }*/

  786.                 }

  787.                 if(parent){

  788.                     int px= x>>1;

  789.                     int py= y>>1;

  790.                     if(px<b->parent->width && py<b->parent->height)

  791.                         p= parent[px + py*2*stride];

  792.                 }

  793.                 if(!(/*ll|*/l|lt|t|rt|p)){

  794.                     if(v){

  795.                         runs[run_index++]= run;

  796.                         run=0;

  797.                     }else{

  798.                         run++;

  799.                     }

  800.                 }

  801.             }

  802.         }

  803.         max_index= run_index;

  804.         runs[run_index++]= run;

  805.         run_index=0;

  806.         run= runs[run_index++];

  807. 

  808.         put_symbol2(&s->c, b->state[30], max_index, 0);

  809.         if(run_index <= max_index)

  810.             put_symbol2(&s->c, b->state[1], run, 3);

  811. 

  812.         for(y=0; y<h; y++){

  813.             if(s->c.bytestream_end - s->c.bytestream < w*40){

  814.                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");

  815.                 return -1;

  816.             }

  817.             for(x=0; x<w; x++){

  818.                 int v, p=0;

  819.                 int /*ll=0, */l=0, lt=0, t=0, rt=0;

  820.                 v= src[x + y*stride];

  821. 

  822.                 if(y){

  823.                     t= src[x + (y-1)*stride];

  824.                     if(x){

  825.                         lt= src[x - 1 + (y-1)*stride];

  826.                     }

  827.                     if(x + 1 < w){

  828.                         rt= src[x + 1 + (y-1)*stride];

  829.                     }

  830.                 }

  831.                 if(x){

  832.                     l= src[x - 1 + y*stride];

  833.                     /*if(x > 1){

  834.                         if(orientation==1) ll= src[y + (x-2)*stride];

  835.                         else               ll= src[x - 2 + y*stride];

  836.                     }*/

  837.                 }

  838.                 if(parent){

  839.                     int px= x>>1;

  840.                     int py= y>>1;

  841.                     if(px<b->parent->width && py<b->parent->height)

  842.                         p= parent[px + py*2*stride];

  843.                 }

  844.                 if(/*ll|*/l|lt|t|rt|p){

  845.                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));

  846. 

  847.                     put_rac(&s->c, &b->state[0][context], !!v);

  848.                 }else{

  849.                     if(!run){

  850.                         run= runs[run_index++];

  851. 

  852.                         if(run_index <= max_index)

  853.                             put_symbol2(&s->c, b->state[1], run, 3);

  854.                         av_assert2(v);

  855.                     }else{

  856.                         run--;

  857.                         av_assert2(!v);

  858.                     }

  859.                 }

  860.                 if(v){

  861.                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));

  862.                     int l2= 2*FFABS(l) + (l<0);

  863.                     int t2= 2*FFABS(t) + (t<0);

  864. 

  865.                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);

  866.                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);

  867.                 }

  868.             }

  869.         }

  870.     }

  871.     return 0;

  872. }

  873. 

  874. static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){

  875. //    encode_subband_qtree(s, b, src, parent, stride, orientation);

  876. //    encode_subband_z0run(s, b, src, parent, stride, orientation);

  877.     return encode_subband_c0run(s, b, src, parent, stride, orientation);

  878. //    encode_subband_dzr(s, b, src, parent, stride, orientation);

  879. }

  880. 

  881. static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){

  882.     const int b_stride= s->b_width << s->block_max_depth;

  883.     BlockNode *block= &s->block[mb_x + mb_y * b_stride];

  884.     BlockNode backup= *block;

  885.     unsigned value;

  886.     int rd, index;

  887. 

  888.     av_assert2(mb_x>=0 && mb_y>=0);

  889.     av_assert2(mb_x<b_stride);

  890. 

  891.     if(intra){

  892.         block->color[0] = p[0];

  893.         block->color[1] = p[1];

  894.         block->color[2] = p[2];

  895.         block->type |= BLOCK_INTRA;

  896.     }else{

  897.         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);

  898.         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);

  899.         if(s->me_cache[index] == value)

  900.             return 0;

  901.         s->me_cache[index]= value;

  902. 

  903.         block->mx= p[0];

  904.         block->my= p[1];

  905.         block->type &= ~BLOCK_INTRA;

  906.     }

  907. 

  908.     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);

  909. 

  910. //FIXME chroma

  911.     if(rd < *best_rd){

  912.         *best_rd= rd;

  913.         return 1;

  914.     }else{

  915.         *block= backup;

  916.         return 0;

  917.     }

  918. }

  919. 

  920. /* special case for int[2] args we discard afterwards,

  921.  * fixes compilation problem with gcc 2.95 */

  922. static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){

  923.     int p[2] = {p0, p1};

  924.     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);

  925. }

  926. 

  927. static av_always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int ref, int *best_rd){

  928.     const int b_stride= s->b_width << s->block_max_depth;

  929.     BlockNode *block= &s->block[mb_x + mb_y * b_stride];

  930.     BlockNode backup[4];

  931.     unsigned value;

  932.     int rd, index;

  933. 

  934.     /* We don't initialize backup[] during variable declaration, because

  935.      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to

  936.      * 'int16_t'". */

  937.     backup[0] = block[0];

  938.     backup[1] = block[1];

  939.     backup[2] = block[b_stride];

  940.     backup[3] = block[b_stride + 1];

  941. 

  942.     av_assert2(mb_x>=0 && mb_y>=0);

  943.     av_assert2(mb_x<b_stride);

  944.     av_assert2(((mb_x|mb_y)&1) == 0);

  945. 

  946.     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);

  947.     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);

  948.     if(s->me_cache[index] == value)

  949.         return 0;

  950.     s->me_cache[index]= value;

  951. 

  952.     block->mx= p0;

  953.     block->my= p1;

  954.     block->ref= ref;

  955.     block->type &= ~BLOCK_INTRA;

  956.     block[1]= block[b_stride]= block[b_stride+1]= *block;

  957. 

  958.     rd= get_4block_rd(s, mb_x, mb_y, 0);

  959. 

  960. //FIXME chroma

  961.     if(rd < *best_rd){

  962.         *best_rd= rd;

  963.         return 1;

  964.     }else{

  965.         block[0]= backup[0];

  966.         block[1]= backup[1];

  967.         block[b_stride]= backup[2];

  968.         block[b_stride+1]= backup[3];

  969.         return 0;

  970.     }

  971. }

  972. 

  973. static void iterative_me(SnowContext *s){

  974.     int pass, mb_x, mb_y;

  975.     const int b_width = s->b_width  << s->block_max_depth;

  976.     const int b_height= s->b_height << s->block_max_depth;

  977.     const int b_stride= b_width;

  978.     int color[3];

  979. 

  980.     {

  981.         RangeCoder r = s->c;

  982.         uint8_t state[sizeof(s->block_state)];

  983.         memcpy(state, s->block_state, sizeof(s->block_state));

  984.         for(mb_y= 0; mb_y<s->b_height; mb_y++)

  985.             for(mb_x= 0; mb_x<s->b_width; mb_x++)

  986.                 encode_q_branch(s, 0, mb_x, mb_y);

  987.         s->c = r;

  988.         memcpy(s->block_state, state, sizeof(s->block_state));

  989.     }

  990. 

  991.     for(pass=0; pass<25; pass++){

  992.         int change= 0;

  993. 

  994.         for(mb_y= 0; mb_y<b_height; mb_y++){

  995.             for(mb_x= 0; mb_x<b_width; mb_x++){

  996.                 int dia_change, i, j, ref;

  997.                 int best_rd= INT_MAX, ref_rd;

  998.                 BlockNode backup, ref_b;

  999.                 const int index= mb_x + mb_y * b_stride;

  1000.                 BlockNode *block= &s->block[index];

  1001.                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;

  1002.                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;

  1003.                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;

  1004.                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;

  1005.                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;

  1006.                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;

  1007.                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;

  1008.                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;

  1009.                 const int b_w= (MB_SIZE >> s->block_max_depth);

  1010.                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];

  1011. 

  1012.                 if(pass && (block->type & BLOCK_OPT))

  1013.                     continue;

  1014.                 block->type |= BLOCK_OPT;

  1015. 

  1016.                 backup= *block;

  1017. 

  1018.                 if(!s->me_cache_generation)

  1019.                     memset(s->me_cache, 0, sizeof(s->me_cache));

  1020.                 s->me_cache_generation += 1<<22;

  1021. 

  1022.                 //FIXME precalculate

  1023.                 {

  1024.                     int x, y;

  1025.                     for (y = 0; y < b_w * 2; y++)

  1026.                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);

  1027.                     if(mb_x==0)

  1028.                         for(y=0; y<b_w*2; y++)

  1029.                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);

  1030.                     if(mb_x==b_stride-1)

  1031.                         for(y=0; y<b_w*2; y++)

  1032.                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);

  1033.                     if(mb_y==0){

  1034.                         for(x=0; x<b_w*2; x++)

  1035.                             obmc_edged[0][x] += obmc_edged[b_w-1][x];

  1036.                         for(y=1; y<b_w; y++)

  1037.                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);

  1038.                     }

  1039.                     if(mb_y==b_height-1){

  1040.                         for(x=0; x<b_w*2; x++)

  1041.                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];

  1042.                         for(y=b_w; y<b_w*2-1; y++)

  1043.                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);

  1044.                     }

  1045.                 }

  1046. 

  1047.                 //skip stuff outside the picture

  1048.                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){

  1049.                     uint8_t *src= s->  input_picture->data[0];

  1050.                     uint8_t *dst= s->current_picture->data[0];

  1051.                     const int stride= s->current_picture->linesize[0];

  1052.                     const int block_w= MB_SIZE >> s->block_max_depth;

  1053.                     const int block_h= MB_SIZE >> s->block_max_depth;

  1054.                     const int sx= block_w*mb_x - block_w/2;

  1055.                     const int sy= block_h*mb_y - block_h/2;

  1056.                     const int w= s->plane[0].width;

  1057.                     const int h= s->plane[0].height;

  1058.                     int y;

  1059. 

  1060.                     for(y=sy; y<0; y++)

  1061.                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);

  1062.                     for(y=h; y<sy+block_h*2; y++)

  1063.                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);

  1064.                     if(sx<0){

  1065.                         for(y=sy; y<sy+block_h*2; y++)

  1066.                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);

  1067.                     }

  1068.                     if(sx+block_w*2 > w){

  1069.                         for(y=sy; y<sy+block_h*2; y++)

  1070.                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);

  1071.                     }

  1072.                 }

  1073. 

  1074.                 // intra(black) = neighbors' contribution to the current block

  1075.                 for(i=0; i<3; i++)

  1076.                     color[i]= get_dc(s, mb_x, mb_y, i);

  1077. 

  1078.                 // get previous score (cannot be cached due to OBMC)

  1079.                 if(pass > 0 && (block->type&BLOCK_INTRA)){

  1080.                     int color0[3]= {block->color[0], block->color[1], block->color[2]};

  1081.                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);

  1082.                 }else

  1083.                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);

  1084. 

  1085.                 ref_b= *block;

  1086.                 ref_rd= best_rd;

  1087.                 for(ref=0; ref < s->ref_frames; ref++){

  1088.                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];

  1089.                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold

  1090.                         continue;

  1091.                     block->ref= ref;

  1092.                     best_rd= INT_MAX;

  1093. 

  1094.                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);

  1095.                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);

  1096.                     if(tb)

  1097.                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);

  1098.                     if(lb)

  1099.                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);

  1100.                     if(rb)

  1101.                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);

  1102.                     if(bb)

  1103.                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);

  1104. 

  1105.                     /* fullpel ME */

  1106.                     //FIXME avoid subpel interpolation / round to nearest integer

  1107.                     do{

  1108.                         dia_change=0;

  1109.                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){

  1110.                             for(j=0; j<i; j++){

  1111.                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);

  1112.                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);

  1113.                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);

  1114.                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);

  1115.                             }

  1116.                         }

  1117.                     }while(dia_change);

  1118.                     /* subpel ME */

  1119.                     do{

  1120.                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};

  1121.                         dia_change=0;

  1122.                         for(i=0; i<8; i++)

  1123.                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);

  1124.                     }while(dia_change);

  1125.                     //FIXME or try the standard 2 pass qpel or similar

  1126. 

  1127.                     mvr[0][0]= block->mx;

  1128.                     mvr[0][1]= block->my;

  1129.                     if(ref_rd > best_rd){

  1130.                         ref_rd= best_rd;

  1131.                         ref_b= *block;

  1132.                     }

  1133.                 }

  1134.                 best_rd= ref_rd;

  1135.                 *block= ref_b;

  1136.                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);

  1137.                 //FIXME RD style color selection

  1138.                 if(!same_block(block, &backup)){

  1139.                     if(tb ) tb ->type &= ~BLOCK_OPT;

  1140.                     if(lb ) lb ->type &= ~BLOCK_OPT;

  1141.                     if(rb ) rb ->type &= ~BLOCK_OPT;

  1142.                     if(bb ) bb ->type &= ~BLOCK_OPT;

  1143.                     if(tlb) tlb->type &= ~BLOCK_OPT;

  1144.                     if(trb) trb->type &= ~BLOCK_OPT;

  1145.                     if(blb) blb->type &= ~BLOCK_OPT;

  1146.                     if(brb) brb->type &= ~BLOCK_OPT;

  1147.                     change ++;

  1148.                 }

  1149.             }

  1150.         }

  1151.         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);

  1152.         if(!change)

  1153.             break;

  1154.     }

  1155. 

  1156.     if(s->block_max_depth == 1){

  1157.         int change= 0;

  1158.         for(mb_y= 0; mb_y<b_height; mb_y+=2){

  1159.             for(mb_x= 0; mb_x<b_width; mb_x+=2){

  1160.                 int i;

  1161.                 int best_rd, init_rd;

  1162.                 const int index= mb_x + mb_y * b_stride;

  1163.                 BlockNode *b[4];

  1164. 

  1165.                 b[0]= &s->block[index];

  1166.                 b[1]= b[0]+1;

  1167.                 b[2]= b[0]+b_stride;

  1168.                 b[3]= b[2]+1;

  1169.                 if(same_block(b[0], b[1]) &&

  1170.                    same_block(b[0], b[2]) &&

  1171.                    same_block(b[0], b[3]))

  1172.                     continue;

  1173. 

  1174.                 if(!s->me_cache_generation)

  1175.                     memset(s->me_cache, 0, sizeof(s->me_cache));

  1176.                 s->me_cache_generation += 1<<22;

  1177. 

  1178.                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);

  1179. 

  1180.                 //FIXME more multiref search?

  1181.                 check_4block_inter(s, mb_x, mb_y,

  1182.                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,

  1183.                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);

  1184. 

  1185.                 for(i=0; i<4; i++)

  1186.                     if(!(b[i]->type&BLOCK_INTRA))

  1187.                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);

  1188. 

  1189.                 if(init_rd != best_rd)

  1190.                     change++;

  1191.             }

  1192.         }

  1193.         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);

  1194.     }

  1195. }

  1196. 

  1197. static void encode_blocks(SnowContext *s, int search){

  1198.     int x, y;

  1199.     int w= s->b_width;

  1200.     int h= s->b_height;

  1201. 

  1202.     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)

  1203.         iterative_me(s);

  1204. 

  1205.     for(y=0; y<h; y++){

  1206.         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit

  1207.             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");

  1208.             return;

  1209.         }

  1210.         for(x=0; x<w; x++){

  1211.             if(s->avctx->me_method == ME_ITER || !search)

  1212.                 encode_q_branch2(s, 0, x, y);

  1213.             else

  1214.                 encode_q_branch (s, 0, x, y);

  1215.         }

  1216.     }

  1217. }

  1218. 

  1219. static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){

  1220.     const int w= b->width;

  1221.     const int h= b->height;

  1222.     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);

  1223.     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);

  1224.     int x,y, thres1, thres2;

  1225. 

  1226.     if(s->qlog == LOSSLESS_QLOG){

  1227.         for(y=0; y<h; y++)

  1228.             for(x=0; x<w; x++)

  1229.                 dst[x + y*stride]= src[x + y*stride];

  1230.         return;

  1231.     }

  1232. 

  1233.     bias= bias ? 0 : (3*qmul)>>3;

  1234.     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;

  1235.     thres2= 2*thres1;

  1236. 

  1237.     if(!bias){

  1238.         for(y=0; y<h; y++){

  1239.             for(x=0; x<w; x++){

  1240.                 int i= src[x + y*stride];

  1241. 

  1242.                 if((unsigned)(i+thres1) > thres2){

  1243.                     if(i>=0){

  1244.                         i<<= QEXPSHIFT;

  1245.                         i/= qmul; //FIXME optimize

  1246.                         dst[x + y*stride]=  i;

  1247.                     }else{

  1248.                         i= -i;

  1249.                         i<<= QEXPSHIFT;

  1250.                         i/= qmul; //FIXME optimize

  1251.                         dst[x + y*stride]= -i;

  1252.                     }

  1253.                 }else

  1254.                     dst[x + y*stride]= 0;

  1255.             }

  1256.         }

  1257.     }else{

  1258.         for(y=0; y<h; y++){

  1259.             for(x=0; x<w; x++){

  1260.                 int i= src[x + y*stride];

  1261. 

  1262.                 if((unsigned)(i+thres1) > thres2){

  1263.                     if(i>=0){

  1264.                         i<<= QEXPSHIFT;

  1265.                         i= (i + bias) / qmul; //FIXME optimize

  1266.                         dst[x + y*stride]=  i;

  1267.                     }else{

  1268.                         i= -i;

  1269.                         i<<= QEXPSHIFT;

  1270.                         i= (i + bias) / qmul; //FIXME optimize

  1271.                         dst[x + y*stride]= -i;

  1272.                     }

  1273.                 }else

  1274.                     dst[x + y*stride]= 0;

  1275.             }

  1276.         }

  1277.     }

  1278. }

  1279. 

  1280. static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){

  1281.     const int w= b->width;

  1282.     const int h= b->height;

  1283.     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);

  1284.     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);

  1285.     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;

  1286.     int x,y;

  1287. 

  1288.     if(s->qlog == LOSSLESS_QLOG) return;

  1289. 

  1290.     for(y=0; y<h; y++){

  1291.         for(x=0; x<w; x++){

  1292.             int i= src[x + y*stride];

  1293.             if(i<0){

  1294.                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias

  1295.             }else if(i>0){

  1296.                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));

  1297.             }

  1298.         }

  1299.     }

  1300. }

  1301. 

  1302. static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){

  1303.     const int w= b->width;

  1304.     const int h= b->height;

  1305.     int x,y;

  1306. 

  1307.     for(y=h-1; y>=0; y--){

  1308.         for(x=w-1; x>=0; x--){

  1309.             int i= x + y*stride;

  1310. 

  1311.             if(x){

  1312.                 if(use_median){

  1313.                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);

  1314.                     else  src[i] -= src[i - 1];

  1315.                 }else{

  1316.                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);

  1317.                     else  src[i] -= src[i - 1];

  1318.                 }

  1319.             }else{

  1320.                 if(y) src[i] -= src[i - stride];

  1321.             }

  1322.         }

  1323.     }

  1324. }

  1325. 

  1326. static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){

  1327.     const int w= b->width;

  1328.     const int h= b->height;

  1329.     int x,y;

  1330. 

  1331.     for(y=0; y<h; y++){

  1332.         for(x=0; x<w; x++){

  1333.             int i= x + y*stride;

  1334. 

  1335.             if(x){

  1336.                 if(use_median){

  1337.                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);

  1338.                     else  src[i] += src[i - 1];

  1339.                 }else{

  1340.                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);

  1341.                     else  src[i] += src[i - 1];

  1342.                 }

  1343.             }else{

  1344.                 if(y) src[i] += src[i - stride];

  1345.             }

  1346.         }

  1347.     }

  1348. }

  1349. 

  1350. static void encode_qlogs(SnowContext *s){

  1351.     int plane_index, level, orientation;

  1352. 

  1353.     for(plane_index=0; plane_index<2; plane_index++){

  1354.         for(level=0; level<s->spatial_decomposition_count; level++){

  1355.             for(orientation=level ? 1:0; orientation<4; orientation++){

  1356.                 if(orientation==2) continue;

  1357.                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);

  1358.             }

  1359.         }

  1360.     }

  1361. }

  1362. 

  1363. static void encode_header(SnowContext *s){

  1364.     int plane_index, i;

  1365.     uint8_t kstate[32];

  1366. 

  1367.     memset(kstate, MID_STATE, sizeof(kstate));

  1368. 

  1369.     put_rac(&s->c, kstate, s->keyframe);

  1370.     if(s->keyframe || s->always_reset){

  1371.         ff_snow_reset_contexts(s);

  1372.         s->last_spatial_decomposition_type=

  1373.         s->last_qlog=

  1374.         s->last_qbias=

  1375.         s->last_mv_scale=

  1376.         s->last_block_max_depth= 0;

  1377.         for(plane_index=0; plane_index<2; plane_index++){

  1378.             Plane *p= &s->plane[plane_index];

  1379.             p->last_htaps=0;

  1380.             p->last_diag_mc=0;

  1381.             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));

  1382.         }

  1383.     }

  1384.     if(s->keyframe){

  1385.         put_symbol(&s->c, s->header_state, s->version, 0);

  1386.         put_rac(&s->c, s->header_state, s->always_reset);

  1387.         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);

  1388.         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);

  1389.         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);

  1390.         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);

  1391.         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);

  1392.         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);

  1393.         put_rac(&s->c, s->header_state, s->spatial_scalability);

  1394. //        put_rac(&s->c, s->header_state, s->rate_scalability);

  1395.         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);

  1396. 

  1397.         encode_qlogs(s);

  1398.     }

  1399. 

  1400.     if(!s->keyframe){

  1401.         int update_mc=0;

  1402.         for(plane_index=0; plane_index<2; plane_index++){

  1403.             Plane *p= &s->plane[plane_index];

  1404.             update_mc |= p->last_htaps   != p->htaps;

  1405.             update_mc |= p->last_diag_mc != p->diag_mc;

  1406.             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));

  1407.         }

  1408.         put_rac(&s->c, s->header_state, update_mc);

  1409.         if(update_mc){

  1410.             for(plane_index=0; plane_index<2; plane_index++){

  1411.                 Plane *p= &s->plane[plane_index];

  1412.                 put_rac(&s->c, s->header_state, p->diag_mc);

  1413.                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);

  1414.                 for(i= p->htaps/2; i; i--)

  1415.                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);

  1416.             }

  1417.         }

  1418.         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){

  1419.             put_rac(&s->c, s->header_state, 1);

  1420.             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);

  1421.             encode_qlogs(s);

  1422.         }else

  1423.             put_rac(&s->c, s->header_state, 0);

  1424.     }

  1425. 

  1426.     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);

  1427.     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);

  1428.     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);

  1429.     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);

  1430.     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);

  1431. 

  1432. }

  1433. 

  1434. static void update_last_header_values(SnowContext *s){

  1435.     int plane_index;

  1436. 

  1437.     if(!s->keyframe){

  1438.         for(plane_index=0; plane_index<2; plane_index++){

  1439.             Plane *p= &s->plane[plane_index];

  1440.             p->last_diag_mc= p->diag_mc;

  1441.             p->last_htaps  = p->htaps;

  1442.             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));

  1443.         }

  1444.     }

  1445. 

  1446.     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;

  1447.     s->last_qlog                        = s->qlog;

  1448.     s->last_qbias                       = s->qbias;

  1449.     s->last_mv_scale                    = s->mv_scale;

  1450.     s->last_block_max_depth             = s->block_max_depth;

  1451.     s->last_spatial_decomposition_count = s->spatial_decomposition_count;

  1452. }

  1453. 

  1454. static int qscale2qlog(int qscale){

  1455.     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))

  1456.            + 61*QROOT/8; ///< 64 > 60

  1457. }

  1458. 

  1459. static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)

  1460. {

  1461.     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.

  1462.      * FIXME we know exact mv bits at this point,

  1463.      * but ratecontrol isn't set up to include them. */

  1464.     uint32_t coef_sum= 0;

  1465.     int level, orientation, delta_qlog;

  1466. 

  1467.     for(level=0; level<s->spatial_decomposition_count; level++){

  1468.         for(orientation=level ? 1 : 0; orientation<4; orientation++){

  1469.             SubBand *b= &s->plane[0].band[level][orientation];

  1470.             IDWTELEM *buf= b->ibuf;

  1471.             const int w= b->width;

  1472.             const int h= b->height;

  1473.             const int stride= b->stride;

  1474.             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);

  1475.             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);

  1476.             const int qdiv= (1<<16)/qmul;

  1477.             int x, y;

  1478.             //FIXME this is ugly

  1479.             for(y=0; y<h; y++)

  1480.                 for(x=0; x<w; x++)

  1481.                     buf[x+y*stride]= b->buf[x+y*stride];

  1482.             if(orientation==0)

  1483.                 decorrelate(s, b, buf, stride, 1, 0);

  1484.             for(y=0; y<h; y++)

  1485.                 for(x=0; x<w; x++)

  1486.                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;

  1487.         }

  1488.     }

  1489. 

  1490.     /* ugly, ratecontrol just takes a sqrt again */

  1491.     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;

  1492.     av_assert0(coef_sum < INT_MAX);

  1493. 

  1494.     if(pict->pict_type == AV_PICTURE_TYPE_I){

  1495.         s->m.current_picture.mb_var_sum= coef_sum;

  1496.         s->m.current_picture.mc_mb_var_sum= 0;

  1497.     }else{

  1498.         s->m.current_picture.mc_mb_var_sum= coef_sum;

  1499.         s->m.current_picture.mb_var_sum= 0;

  1500.     }

  1501. 

  1502.     pict->quality= ff_rate_estimate_qscale(&s->m, 1);

  1503.     if (pict->quality < 0)

  1504.         return INT_MIN;

  1505.     s->lambda= pict->quality * 3/2;

  1506.     delta_qlog= qscale2qlog(pict->quality) - s->qlog;

  1507.     s->qlog+= delta_qlog;

  1508.     return delta_qlog;

  1509. }

  1510. 

  1511. static void calculate_visual_weight(SnowContext *s, Plane *p){

  1512.     int width = p->width;

  1513.     int height= p->height;

  1514.     int level, orientation, x, y;

  1515. 

  1516.     for(level=0; level<s->spatial_decomposition_count; level++){

  1517.         for(orientation=level ? 1 : 0; orientation<4; orientation++){

  1518.             SubBand *b= &p->band[level][orientation];

  1519.             IDWTELEM *ibuf= b->ibuf;

  1520.             int64_t error=0;

  1521. 

  1522.             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);

  1523.             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;

  1524.             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);

  1525.             for(y=0; y<height; y++){

  1526.                 for(x=0; x<width; x++){

  1527.                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;

  1528.                     error += d*d;

  1529.                 }

  1530.             }

  1531. 

  1532.             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);

  1533.         }

  1534.     }

  1535. }

  1536. 

  1537. static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,

  1538.                         const AVFrame *pict, int *got_packet)

  1539. {

  1540.     SnowContext *s = avctx->priv_data;

  1541.     RangeCoder * const c= &s->c;

  1542.     AVFrame *pic = pict;

  1543.     const int width= s->avctx->width;

  1544.     const int height= s->avctx->height;

  1545.     int level, orientation, plane_index, i, y, ret;

  1546.     uint8_t rc_header_bak[sizeof(s->header_state)];

  1547.     uint8_t rc_block_bak[sizeof(s->block_state)];

  1548. 

  1549.     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)

  1550.         return ret;

  1551. 

  1552.     ff_init_range_encoder(c, pkt->data, pkt->size);

  1553.     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);

  1554. 

  1555.     for(i=0; i<3; i++){

  1556.         int hshift= i ? s->chroma_h_shift : 0;

  1557.         int vshift= i ? s->chroma_v_shift : 0;

  1558.         for(y=0; y<(height>>vshift); y++)

  1559.             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],

  1560.                    &pict->data[i][y * pict->linesize[i]],

  1561.                    width>>hshift);

  1562.         s->dsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],

  1563.                             width >> hshift, height >> vshift,

  1564.                             EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,

  1565.                             EDGE_TOP | EDGE_BOTTOM);

  1566. 

  1567.     }

  1568.     emms_c();

  1569.     s->new_picture = pict;

  1570. 

  1571.     s->m.picture_number= avctx->frame_number;

  1572.     if(avctx->flags&CODEC_FLAG_PASS2){

  1573.         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;

  1574.         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;

  1575.         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {

  1576.             pic->quality = ff_rate_estimate_qscale(&s->m, 0);

  1577.             if (pic->quality < 0)

  1578.                 return -1;

  1579.         }

  1580.     }else{

  1581.         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;

  1582.         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  1583.     }

  1584. 

  1585.     if(s->pass1_rc && avctx->frame_number == 0)

  1586.         pic->quality = 2*FF_QP2LAMBDA;

  1587.     if (pic->quality) {

  1588.         s->qlog   = qscale2qlog(pic->quality);

  1589.         s->lambda = pic->quality * 3/2;

  1590.     }

  1591.     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {

  1592.         s->qlog= LOSSLESS_QLOG;

  1593.         s->lambda = 0;

  1594.     }//else keep previous frame's qlog until after motion estimation

  1595. 

  1596.     ff_snow_frame_start(s);

  1597.     avctx->coded_frame= s->current_picture;

  1598. 

  1599.     s->m.current_picture_ptr= &s->m.current_picture;

  1600.     s->m.last_picture.f.pts = s->m.current_picture.f.pts;

  1601.     s->m.current_picture.f.pts = pict->pts;

  1602.     if(pic->pict_type == AV_PICTURE_TYPE_P){

  1603.         int block_width = (width +15)>>4;

  1604.         int block_height= (height+15)>>4;

  1605.         int stride= s->current_picture->linesize[0];

  1606. 

  1607.         av_assert0(s->current_picture->data[0]);

  1608.         av_assert0(s->last_picture[0]->data[0]);

  1609. 

  1610.         s->m.avctx= s->avctx;

  1611.         s->m.current_picture.f.data[0] = s->current_picture->data[0];

  1612.         s->m.   last_picture.f.data[0] = s->last_picture[0]->data[0];

  1613.         s->m.    new_picture.f.data[0] = s->  input_picture->data[0];

  1614.         s->m.   last_picture_ptr= &s->m.   last_picture;

  1615.         s->m.linesize=

  1616.         s->m.   last_picture.f.linesize[0] =

  1617.         s->m.    new_picture.f.linesize[0] =

  1618.         s->m.current_picture.f.linesize[0] = stride;

  1619.         s->m.uvlinesize= s->current_picture->linesize[1];

  1620.         s->m.width = width;

  1621.         s->m.height= height;

  1622.         s->m.mb_width = block_width;

  1623.         s->m.mb_height= block_height;

  1624.         s->m.mb_stride=   s->m.mb_width+1;

  1625.         s->m.b8_stride= 2*s->m.mb_width+1;

  1626.         s->m.f_code=1;

  1627.         s->m.pict_type = pic->pict_type;

  1628.         s->m.me_method= s->avctx->me_method;

  1629.         s->m.me.scene_change_score=0;

  1630.         s->m.flags= s->avctx->flags;

  1631.         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;

  1632.         s->m.out_format= FMT_H263;

  1633.         s->m.unrestricted_mv= 1;

  1634. 

  1635.         s->m.lambda = s->lambda;

  1636.         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);

  1637.         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;

  1638. 

  1639.         s->m.dsp= s->dsp; //move

  1640.         s->m.hdsp = s->hdsp;

  1641.         ff_init_me(&s->m);

  1642.         s->hdsp = s->m.hdsp;

  1643.         s->dsp= s->m.dsp;

  1644.     }

  1645. 

  1646.     if(s->pass1_rc){

  1647.         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));

  1648.         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));

  1649.     }

  1650. 

  1651. redo_frame:

  1652. 

  1653.     if (pic->pict_type == AV_PICTURE_TYPE_I)

  1654.         s->spatial_decomposition_count= 5;

  1655.     else

  1656.         s->spatial_decomposition_count= 5;

  1657. 

  1658.     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))

  1659.           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))

  1660.         s->spatial_decomposition_count--;

  1661. 

  1662.     if (s->spatial_decomposition_count <= 0) {

  1663.         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");

  1664.         return AVERROR(EINVAL);

  1665.     }

  1666. 

  1667.     s->m.pict_type = pic->pict_type;

  1668.     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;

  1669. 

  1670.     ff_snow_common_init_after_header(avctx);

  1671. 

  1672.     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){

  1673.         for(plane_index=0; plane_index<3; plane_index++){

  1674.             calculate_visual_weight(s, &s->plane[plane_index]);

  1675.         }

  1676.     }

  1677. 

  1678.     encode_header(s);

  1679.     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);

  1680.     encode_blocks(s, 1);

  1681.     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;

  1682. 

  1683.     for(plane_index=0; plane_index<3; plane_index++){

  1684.         Plane *p= &s->plane[plane_index];

  1685.         int w= p->width;

  1686.         int h= p->height;

  1687.         int x, y;

  1688. //        int bits= put_bits_count(&s->c.pb);

  1689. 

  1690.         if (!s->memc_only) {

  1691.             //FIXME optimize

  1692.             if(pict->data[plane_index]) //FIXME gray hack

  1693.                 for(y=0; y<h; y++){

  1694.                     for(x=0; x<w; x++){

  1695.                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;

  1696.                     }

  1697.                 }

  1698.             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);

  1699. 

  1700.             if(   plane_index==0

  1701.                && pic->pict_type == AV_PICTURE_TYPE_P

  1702.                && !(avctx->flags&CODEC_FLAG_PASS2)

  1703.                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){

  1704.                 ff_init_range_encoder(c, pkt->data, pkt->size);

  1705.                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);

  1706.                 pic->pict_type= AV_PICTURE_TYPE_I;

  1707.                 s->keyframe=1;

  1708.                 s->current_picture->key_frame=1;

  1709.                 goto redo_frame;

  1710.             }

  1711. 

  1712.             if(s->qlog == LOSSLESS_QLOG){

  1713.                 for(y=0; y<h; y++){

  1714.                     for(x=0; x<w; x++){

  1715.                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;

  1716.                     }

  1717.                 }

  1718.             }else{

  1719.                 for(y=0; y<h; y++){

  1720.                     for(x=0; x<w; x++){

  1721.                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;

  1722.                     }

  1723.                 }

  1724.             }

  1725. 

  1726.             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);

  1727. 

  1728.             if(s->pass1_rc && plane_index==0){

  1729.                 int delta_qlog = ratecontrol_1pass(s, pic);

  1730.                 if (delta_qlog <= INT_MIN)

  1731.                     return -1;

  1732.                 if(delta_qlog){

  1733.                     //reordering qlog in the bitstream would eliminate this reset

  1734.                     ff_init_range_encoder(c, pkt->data, pkt->size);

  1735.                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));

  1736.                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));

  1737.                     encode_header(s);

  1738.                     encode_blocks(s, 0);

  1739.                 }

  1740.             }

  1741. 

  1742.             for(level=0; level<s->spatial_decomposition_count; level++){

  1743.                 for(orientation=level ? 1 : 0; orientation<4; orientation++){

  1744.                     SubBand *b= &p->band[level][orientation];

  1745. 

  1746.                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);

  1747.                     if(orientation==0)

  1748.                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);

  1749.                     if (!s->no_bitstream)

  1750.                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);

  1751.                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);

  1752.                     if(orientation==0)

  1753.                         correlate(s, b, b->ibuf, b->stride, 1, 0);

  1754.                 }

  1755.             }

  1756. 

  1757.             for(level=0; level<s->spatial_decomposition_count; level++){

  1758.                 for(orientation=level ? 1 : 0; orientation<4; orientation++){

  1759.                     SubBand *b= &p->band[level][orientation];

  1760. 

  1761.                     dequantize(s, b, b->ibuf, b->stride);

  1762.                 }

  1763.             }

  1764. 

  1765.             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);

  1766.             if(s->qlog == LOSSLESS_QLOG){

  1767.                 for(y=0; y<h; y++){

  1768.                     for(x=0; x<w; x++){

  1769.                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;

  1770.                     }

  1771.                 }

  1772.             }

  1773.             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);

  1774.         }else{

  1775.             //ME/MC only

  1776.             if(pic->pict_type == AV_PICTURE_TYPE_I){

  1777.                 for(y=0; y<h; y++){

  1778.                     for(x=0; x<w; x++){

  1779.                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=

  1780.                             pict->data[plane_index][y*pict->linesize[plane_index] + x];

  1781.                     }

  1782.                 }

  1783.             }else{

  1784.                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);

  1785.                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);

  1786.             }

  1787.         }

  1788.         if(s->avctx->flags&CODEC_FLAG_PSNR){

  1789.             int64_t error= 0;

  1790. 

  1791.             if(pict->data[plane_index]) //FIXME gray hack

  1792.                 for(y=0; y<h; y++){

  1793.                     for(x=0; x<w; x++){

  1794.                         int d= s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x] - pict->data[plane_index][y*pict->linesize[plane_index] + x];

  1795.                         error += d*d;

  1796.                     }

  1797.                 }

  1798.             s->avctx->error[plane_index] += error;

  1799.             s->current_picture->error[plane_index] = error;

  1800.         }

  1801. 

  1802.     }

  1803. 

  1804.     update_last_header_values(s);

  1805. 

  1806.     ff_snow_release_buffer(avctx);

  1807. 

  1808.     s->current_picture->coded_picture_number = avctx->frame_number;

  1809.     s->current_picture->pict_type = pict->pict_type;

  1810.     s->current_picture->quality = pict->quality;

  1811.     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);

  1812.     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;

  1813.     s->m.current_picture.f.display_picture_number =

  1814.     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;

  1815.     s->m.current_picture.f.quality                = pic->quality;

  1816.     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);

  1817.     if(s->pass1_rc)

  1818.         if (ff_rate_estimate_qscale(&s->m, 0) < 0)

  1819.             return -1;

  1820.     if(avctx->flags&CODEC_FLAG_PASS1)

  1821.         ff_write_pass1_stats(&s->m);

  1822.     s->m.last_pict_type = s->m.pict_type;

  1823.     avctx->frame_bits = s->m.frame_bits;

  1824.     avctx->mv_bits = s->m.mv_bits;

  1825.     avctx->misc_bits = s->m.misc_bits;

  1826.     avctx->p_tex_bits = s->m.p_tex_bits;

  1827. 

  1828.     emms_c();

  1829. 

  1830.     pkt->size = ff_rac_terminate(c);

  1831.     if (avctx->coded_frame->key_frame)

  1832.         pkt->flags |= AV_PKT_FLAG_KEY;

  1833.     *got_packet = 1;

  1834. 

  1835.     return 0;

  1836. }

  1837. 

  1838. static av_cold int encode_end(AVCodecContext *avctx)

  1839. {

  1840.     SnowContext *s = avctx->priv_data;

  1841. 

  1842.     ff_snow_common_end(s);

  1843.     ff_rate_control_uninit(&s->m);

  1844.     av_frame_free(&s->input_picture);

  1845.     av_free(avctx->stats_out);

  1846. 

  1847.     return 0;

  1848. }

  1849. 

  1850. #define OFFSET(x) offsetof(SnowContext, x)

  1851. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

  1852. static const AVOption options[] = {

  1853.     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },

  1854.     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },

  1855.     { NULL },

  1856. };

  1857. 

  1858. static const AVClass snowenc_class = {

  1859.     .class_name = "snow encoder",

  1860.     .item_name  = av_default_item_name,

  1861.     .option     = options,

  1862.     .version    = LIBAVUTIL_VERSION_INT,

  1863. };

  1864. 

  1865. AVCodec ff_snow_encoder = {

  1866.     .name           = "snow",

  1867.     .type           = AVMEDIA_TYPE_VIDEO,

  1868.     .id             = AV_CODEC_ID_SNOW,

  1869.     .priv_data_size = sizeof(SnowContext),

  1870.     .init           = encode_init,

  1871.     .encode2        = encode_frame,

  1872.     .close          = encode_end,

  1873.     .pix_fmts       = (const enum AVPixelFormat[]){

  1874.         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,

  1875.         AV_PIX_FMT_NONE

  1876.     },

  1877.     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),

  1878.     .priv_class     = &snowenc_class,

  1879. };

  1880. 

  1881. 

  1882. #ifdef TEST

  1883. #undef malloc

  1884. #undef free

  1885. #undef printf

  1886. 

  1887. #include "libavutil/lfg.h"

  1888. #include "libavutil/mathematics.h"

  1889. 

  1890. int main(void){

  1891. #define width  256

  1892. #define height 256

  1893.     int buffer[2][width*height];

  1894.     SnowContext s;

  1895.     int i;

  1896.     AVLFG prng;

  1897.     s.spatial_decomposition_count=6;

  1898.     s.spatial_decomposition_type=1;

  1899. 

  1900.     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));

  1901.     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));

  1902. 

  1903.     av_lfg_init(&prng, 1);

  1904. 

  1905.     printf("testing 5/3 DWT\n");

  1906.     for(i=0; i<width*height; i++)

  1907.         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;

  1908. 

  1909.     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1910.     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1911. 

  1912.     for(i=0; i<width*height; i++)

  1913.         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);

  1914. 

  1915.     printf("testing 9/7 DWT\n");

  1916.     s.spatial_decomposition_type=0;

  1917.     for(i=0; i<width*height; i++)

  1918.         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;

  1919. 

  1920.     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1921.     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1922. 

  1923.     for(i=0; i<width*height; i++)

  1924.         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);

  1925. 

  1926.     {

  1927.     int level, orientation, x, y;

  1928.     int64_t errors[8][4];

  1929.     int64_t g=0;

  1930. 

  1931.         memset(errors, 0, sizeof(errors));

  1932.         s.spatial_decomposition_count=3;

  1933.         s.spatial_decomposition_type=0;

  1934.         for(level=0; level<s.spatial_decomposition_count; level++){

  1935.             for(orientation=level ? 1 : 0; orientation<4; orientation++){

  1936.                 int w= width  >> (s.spatial_decomposition_count-level);

  1937.                 int h= height >> (s.spatial_decomposition_count-level);

  1938.                 int stride= width  << (s.spatial_decomposition_count-level);

  1939.                 DWTELEM *buf= buffer[0];

  1940.                 int64_t error=0;

  1941. 

  1942.                 if(orientation&1) buf+=w;

  1943.                 if(orientation>1) buf+=stride>>1;

  1944. 

  1945.                 memset(buffer[0], 0, sizeof(int)*width*height);

  1946.                 buf[w/2 + h/2*stride]= 256*256;

  1947.                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1948.                 for(y=0; y<height; y++){

  1949.                     for(x=0; x<width; x++){

  1950.                         int64_t d= buffer[0][x + y*width];

  1951.                         error += d*d;

  1952.                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);

  1953.                     }

  1954.                     if(FFABS(height/2-y)<9 && level==2) printf("\n");

  1955.                 }

  1956.                 error= (int)(sqrt(error)+0.5);

  1957.                 errors[level][orientation]= error;

  1958.                 if(g) g=av_gcd(g, error);

  1959.                 else g= error;

  1960.             }

  1961.         }

  1962.         printf("static int const visual_weight[][4]={\n");

  1963.         for(level=0; level<s.spatial_decomposition_count; level++){

  1964.             printf("  {");

  1965.             for(orientation=0; orientation<4; orientation++){

  1966.                 printf("%8"PRId64",", errors[level][orientation]/g);

  1967.             }

  1968.             printf("},\n");

  1969.         }

  1970.         printf("};\n");

  1971.         {

  1972.             int level=2;

  1973.             int w= width  >> (s.spatial_decomposition_count-level);

  1974.             //int h= height >> (s.spatial_decomposition_count-level);

  1975.             int stride= width  << (s.spatial_decomposition_count-level);

  1976.             DWTELEM *buf= buffer[0];

  1977.             int64_t error=0;

  1978. 

  1979.             buf+=w;

  1980.             buf+=stride>>1;

  1981. 

  1982.             memset(buffer[0], 0, sizeof(int)*width*height);

  1983.             for(y=0; y<height; y++){

  1984.                 for(x=0; x<width; x++){

  1985.                     int tab[4]={0,2,3,1};

  1986.                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];

  1987.                 }

  1988.             }

  1989.             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);

  1990.             for(y=0; y<height; y++){

  1991.                 for(x=0; x<width; x++){

  1992.                     int64_t d= buffer[0][x + y*width];

  1993.                     error += d*d;

  1994.                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);

  1995.                 }

  1996.                 if(FFABS(height/2-y)<9) printf("\n");

  1997.             }

  1998.         }

  1999. 

  2000.     }

  2001.     return 0;

  2002. }

  2003. #endif /* TEST */