git.sesse.net Git - ffmpeg/blob - libavcodec/snowenc.c

   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 "dsputil.h"
  26 #include "internal.h"
  27 #include "snow_dwt.h"
  28 #include "snow.h"
  29
  30 #include "rangecoder.h"
  31 #include "mathops.h"
  32
  33 #include "mpegvideo.h"
  34 #include "h263.h"
  35
  36 static av_cold int encode_init(AVCodecContext *avctx)
  37 {
  38     SnowContext *s = avctx->priv_data;
  39     int plane_index, ret;
  40
  41     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
  42         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
  43                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
  44         return -1;
  45     }
  46
  47     if(avctx->prediction_method == DWT_97
  48        && (avctx->flags & CODEC_FLAG_QSCALE)
  49        && avctx->global_quality == 0){
  50         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  51         return -1;
  52     }
  53
  54     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
  55
  56     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
  57     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
  58
  59     for(plane_index=0; plane_index<3; plane_index++){
  60         s->plane[plane_index].diag_mc= 1;
  61         s->plane[plane_index].htaps= 6;
  62         s->plane[plane_index].hcoeff[0]=  40;
  63         s->plane[plane_index].hcoeff[1]= -10;
  64         s->plane[plane_index].hcoeff[2]=   2;
  65         s->plane[plane_index].fast_mc= 1;
  66     }
  67
  68     if ((ret = ff_snow_common_init(avctx)) < 0) {
  69         ff_snow_common_end(avctx->priv_data);
  70         return ret;
  71     }
  72     ff_snow_alloc_blocks(s);
  73
  74     s->version=0;
  75
  76     s->m.avctx   = avctx;
  77     s->m.flags   = avctx->flags;
  78     s->m.bit_rate= avctx->bit_rate;
  79
  80     s->m.me.temp      =
  81     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
  82     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  83     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  84     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  85     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)
  86         return AVERROR(ENOMEM);
  87
  88     ff_h263_encode_init(&s->m); //mv_penalty
  89
  90     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
  91
  92     if(avctx->flags&CODEC_FLAG_PASS1){
  93         if(!avctx->stats_out)
  94             avctx->stats_out = av_mallocz(256);
  95
  96         if (!avctx->stats_out)
  97             return AVERROR(ENOMEM);
  98     }
  99     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
 100         if(ff_rate_control_init(&s->m) < 0)
 101             return -1;
 102     }
 103     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 104
 105     switch(avctx->pix_fmt){
 106     case AV_PIX_FMT_YUV444P:
 107 //    case AV_PIX_FMT_YUV422P:
 108     case AV_PIX_FMT_YUV420P:
 109 //    case AV_PIX_FMT_YUV411P:
 110     case AV_PIX_FMT_YUV410P:
 111         s->nb_planes = 3;
 112         s->colorspace_type= 0;
 113         break;
 114     case AV_PIX_FMT_GRAY8:
 115         s->nb_planes = 1;
 116         s->colorspace_type = 1;
 117         break;
 118 /*    case AV_PIX_FMT_RGB32:
 119         s->colorspace= 1;
 120         break;*/
 121     default:
 122         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 123         return -1;
 124     }
 125     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 126
 127     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
 128     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
 129
 130     s->input_picture = av_frame_alloc();
 131     if (!s->input_picture)
 132         return AVERROR(ENOMEM);
 133     if ((ret = ff_get_buffer(s->avctx, s->input_picture, AV_GET_BUFFER_FLAG_REF)) < 0)
 134         return ret;
 135
 136     if(s->avctx->me_method == ME_ITER){
 137         int i;
 138         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 139         for(i=0; i<s->max_ref_frames; i++){
 140             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
 141             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
 142             if (!s->ref_mvs[i] || !s->ref_scores[i])
 143                 return AVERROR(ENOMEM);
 144         }
 145     }
 146
 147     return 0;
 148 }
 149
 150 //near copy & paste from dsputil, FIXME
 151 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 152 {
 153     int s, i, j;
 154
 155     s = 0;
 156     for (i = 0; i < h; i++) {
 157         for (j = 0; j < w; j++) {
 158             s += pix[0];
 159             pix ++;
 160         }
 161         pix += line_size - w;
 162     }
 163     return s;
 164 }
 165
 166 //near copy & paste from dsputil, FIXME
 167 static int pix_norm1(uint8_t * pix, int line_size, int w)
 168 {
 169     int s, i, j;
 170     uint32_t *sq = ff_squareTbl + 256;
 171
 172     s = 0;
 173     for (i = 0; i < w; i++) {
 174         for (j = 0; j < w; j ++) {
 175             s += sq[pix[0]];
 176             pix ++;
 177         }
 178         pix += line_size - w;
 179     }
 180     return s;
 181 }
 182
 183 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 184     switch(type&0xFF){
 185     default:
 186     case FF_CMP_SAD:
 187         return lambda>>FF_LAMBDA_SHIFT;
 188     case FF_CMP_DCT:
 189         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 190     case FF_CMP_W53:
 191         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 192     case FF_CMP_W97:
 193         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 194     case FF_CMP_SATD:
 195     case FF_CMP_DCT264:
 196         return (2*lambda)>>FF_LAMBDA_SHIFT;
 197     case FF_CMP_RD:
 198     case FF_CMP_PSNR:
 199     case FF_CMP_SSE:
 200     case FF_CMP_NSSE:
 201         return lambda2>>FF_LAMBDA_SHIFT;
 202     case FF_CMP_BIT:
 203         return 1;
 204     }
 205 }
 206
 207 //FIXME copy&paste
 208 #define P_LEFT P[1]
 209 #define P_TOP P[2]
 210 #define P_TOPRIGHT P[3]
 211 #define P_MEDIAN P[4]
 212 #define P_MV1 P[9]
 213 #define FLAG_QPEL   1 //must be 1
 214
 215 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 216     uint8_t p_buffer[1024];
 217     uint8_t i_buffer[1024];
 218     uint8_t p_state[sizeof(s->block_state)];
 219     uint8_t i_state[sizeof(s->block_state)];
 220     RangeCoder pc, ic;
 221     uint8_t *pbbak= s->c.bytestream;
 222     uint8_t *pbbak_start= s->c.bytestream_start;
 223     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 224     const int w= s->b_width  << s->block_max_depth;
 225     const int h= s->b_height << s->block_max_depth;
 226     const int rem_depth= s->block_max_depth - level;
 227     const int index= (x + y*w) << rem_depth;
 228     const int block_w= 1<<(LOG2_MB_SIZE - level);
 229     int trx= (x+1)<<rem_depth;
 230     int try= (y+1)<<rem_depth;
 231     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 232     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 233     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 234     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 235     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 236     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 237     int pl = left->color[0];
 238     int pcb= left->color[1];
 239     int pcr= left->color[2];
 240     int pmx, pmy;
 241     int mx=0, my=0;
 242     int l,cr,cb;
 243     const int stride= s->current_picture->linesize[0];
 244     const int uvstride= s->current_picture->linesize[1];
 245     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 246                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 247                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 248     int P[10][2];
 249     int16_t last_mv[3][2];
 250     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
 251     const int shift= 1+qpel;
 252     MotionEstContext *c= &s->m.me;
 253     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 254     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 255     int my_context= av_log2(2*FFABS(left->my - top->my));
 256     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 257     int ref, best_ref, ref_score, ref_mx, ref_my;
 258
 259     av_assert0(sizeof(s->block_state) >= 256);
 260     if(s->keyframe){
 261         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 262         return 0;
 263     }
 264
 265 //    clip predictors / edge ?
 266
 267     P_LEFT[0]= left->mx;
 268     P_LEFT[1]= left->my;
 269     P_TOP [0]= top->mx;
 270     P_TOP [1]= top->my;
 271     P_TOPRIGHT[0]= tr->mx;
 272     P_TOPRIGHT[1]= tr->my;
 273
 274     last_mv[0][0]= s->block[index].mx;
 275     last_mv[0][1]= s->block[index].my;
 276     last_mv[1][0]= right->mx;
 277     last_mv[1][1]= right->my;
 278     last_mv[2][0]= bottom->mx;
 279     last_mv[2][1]= bottom->my;
 280
 281     s->m.mb_stride=2;
 282     s->m.mb_x=
 283     s->m.mb_y= 0;
 284     c->skip= 0;
 285
 286     av_assert1(c->  stride ==   stride);
 287     av_assert1(c->uvstride == uvstride);
 288
 289     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 290     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 291     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 292     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 293
 294     c->xmin = - x*block_w - 16+3;
 295     c->ymin = - y*block_w - 16+3;
 296     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 297     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 298
 299     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 300     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 301     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 302     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 303     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 304     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 305     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 306
 307     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 308     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 309
 310     if (!y) {
 311         c->pred_x= P_LEFT[0];
 312         c->pred_y= P_LEFT[1];
 313     } else {
 314         c->pred_x = P_MEDIAN[0];
 315         c->pred_y = P_MEDIAN[1];
 316     }
 317
 318     score= INT_MAX;
 319     best_ref= 0;
 320     for(ref=0; ref<s->ref_frames; ref++){
 321         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 322
 323         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 324                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 325
 326         av_assert2(ref_mx >= c->xmin);
 327         av_assert2(ref_mx <= c->xmax);
 328         av_assert2(ref_my >= c->ymin);
 329         av_assert2(ref_my <= c->ymax);
 330
 331         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 332         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 333         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 334         if(s->ref_mvs[ref]){
 335             s->ref_mvs[ref][index][0]= ref_mx;
 336             s->ref_mvs[ref][index][1]= ref_my;
 337             s->ref_scores[ref][index]= ref_score;
 338         }
 339         if(score > ref_score){
 340             score= ref_score;
 341             best_ref= ref;
 342             mx= ref_mx;
 343             my= ref_my;
 344         }
 345     }
 346     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 347
 348   //  subpel search
 349     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 350     pc= s->c;
 351     pc.bytestream_start=
 352     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 353     memcpy(p_state, s->block_state, sizeof(s->block_state));
 354
 355     if(level!=s->block_max_depth)
 356         put_rac(&pc, &p_state[4 + s_context], 1);
 357     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 358     if(s->ref_frames > 1)
 359         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 360     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 361     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 362     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 363     p_len= pc.bytestream - pc.bytestream_start;
 364     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 365
 366     block_s= block_w*block_w;
 367     sum = pix_sum(current_data[0], stride, block_w, block_w);
 368     l= (sum + block_s/2)/block_s;
 369     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 370
 371     if (s->nb_planes > 2) {
 372         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 373         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 374         cb= (sum + block_s/2)/block_s;
 375     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 376         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 377         cr= (sum + block_s/2)/block_s;
 378     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 379     }else
 380         cb = cr = 0;
 381
 382     ic= s->c;
 383     ic.bytestream_start=
 384     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 385     memcpy(i_state, s->block_state, sizeof(s->block_state));
 386     if(level!=s->block_max_depth)
 387         put_rac(&ic, &i_state[4 + s_context], 1);
 388     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 389     put_symbol(&ic, &i_state[32],  l-pl , 1);
 390     if (s->nb_planes > 2) {
 391         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 392         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 393     }
 394     i_len= ic.bytestream - ic.bytestream_start;
 395     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 396
 397 //    assert(score==256*256*256*64-1);
 398     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 399     av_assert1(iscore >= 0);
 400     av_assert1(l>=0 && l<=255);
 401     av_assert1(pl>=0 && pl<=255);
 402
 403     if(level==0){
 404         int varc= iscore >> 8;
 405         int vard= score >> 8;
 406         if (vard <= 64 || vard < varc)
 407             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 408         else
 409             c->scene_change_score+= s->m.qscale;
 410     }
 411
 412     if(level!=s->block_max_depth){
 413         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 414         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 415         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 416         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 417         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 418         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 419
 420         if(score2 < score && score2 < iscore)
 421             return score2;
 422     }
 423
 424     if(iscore < score){
 425         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 426         memcpy(pbbak, i_buffer, i_len);
 427         s->c= ic;
 428         s->c.bytestream_start= pbbak_start;
 429         s->c.bytestream= pbbak + i_len;
 430         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 431         memcpy(s->block_state, i_state, sizeof(s->block_state));
 432         return iscore;
 433     }else{
 434         memcpy(pbbak, p_buffer, p_len);
 435         s->c= pc;
 436         s->c.bytestream_start= pbbak_start;
 437         s->c.bytestream= pbbak + p_len;
 438         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 439         memcpy(s->block_state, p_state, sizeof(s->block_state));
 440         return score;
 441     }
 442 }
 443
 444 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 445     const int w= s->b_width  << s->block_max_depth;
 446     const int rem_depth= s->block_max_depth - level;
 447     const int index= (x + y*w) << rem_depth;
 448     int trx= (x+1)<<rem_depth;
 449     BlockNode *b= &s->block[index];
 450     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 451     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 452     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 453     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 454     int pl = left->color[0];
 455     int pcb= left->color[1];
 456     int pcr= left->color[2];
 457     int pmx, pmy;
 458     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 459     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 460     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 461     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 462
 463     if(s->keyframe){
 464         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 465         return;
 466     }
 467
 468     if(level!=s->block_max_depth){
 469         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 470             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 471         }else{
 472             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 473             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 474             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 475             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 476             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 477             return;
 478         }
 479     }
 480     if(b->type & BLOCK_INTRA){
 481         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 482         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 483         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 484         if (s->nb_planes > 2) {
 485             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 486             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 487         }
 488         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 489     }else{
 490         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 491         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 492         if(s->ref_frames > 1)
 493             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 494         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 495         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 496         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 497     }
 498 }
 499
 500 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 501     int i, x2, y2;
 502     Plane *p= &s->plane[plane_index];
 503     const int block_size = MB_SIZE >> s->block_max_depth;
 504     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 505     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 506     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 507     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 508     const int ref_stride= s->current_picture->linesize[plane_index];
 509     uint8_t *src= s-> input_picture->data[plane_index];
 510     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 511     const int b_stride = s->b_width << s->block_max_depth;
 512     const int w= p->width;
 513     const int h= p->height;
 514     int index= mb_x + mb_y*b_stride;
 515     BlockNode *b= &s->block[index];
 516     BlockNode backup= *b;
 517     int ab=0;
 518     int aa=0;
 519
 520     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 521
 522     b->type|= BLOCK_INTRA;
 523     b->color[plane_index]= 0;
 524     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 525
 526     for(i=0; i<4; i++){
 527         int mb_x2= mb_x + (i &1) - 1;
 528         int mb_y2= mb_y + (i>>1) - 1;
 529         int x= block_w*mb_x2 + block_w/2;
 530         int y= block_h*mb_y2 + block_h/2;
 531
 532         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 533                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 534
 535         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 536             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 537                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 538                 int obmc_v= obmc[index];
 539                 int d;
 540                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 541                 if(x<0) obmc_v += obmc[index + block_w];
 542                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 543                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 544                 //FIXME precalculate this or simplify it somehow else
 545
 546                 d = -dst[index] + (1<<(FRAC_BITS-1));
 547                 dst[index] = d;
 548                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 549                 aa += obmc_v * obmc_v; //FIXME precalculate this
 550             }
 551         }
 552     }
 553     *b= backup;
 554
 555     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
 556 }
 557
 558 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 559     const int b_stride = s->b_width << s->block_max_depth;
 560     const int b_height = s->b_height<< s->block_max_depth;
 561     int index= x + y*b_stride;
 562     const BlockNode *b     = &s->block[index];
 563     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 564     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 565     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 566     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 567     int dmx, dmy;
 568 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 569 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 570
 571     if(x<0 || x>=b_stride || y>=b_height)
 572         return 0;
 573 /*
 574 1            0      0
 575 01X          1-2    1
 576 001XX        3-6    2-3
 577 0001XXX      7-14   4-7
 578 00001XXXX   15-30   8-15
 579 */
 580 //FIXME try accurate rate
 581 //FIXME intra and inter predictors if surrounding blocks are not the same type
 582     if(b->type & BLOCK_INTRA){
 583         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 584                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 585                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 586     }else{
 587         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 588         dmx-= b->mx;
 589         dmy-= b->my;
 590         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 591                     + av_log2(2*FFABS(dmy))
 592                     + av_log2(2*b->ref));
 593     }
 594 }
 595
 596 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 597     Plane *p= &s->plane[plane_index];
 598     const int block_size = MB_SIZE >> s->block_max_depth;
 599     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 600     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 601     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 602     const int ref_stride= s->current_picture->linesize[plane_index];
 603     uint8_t *dst= s->current_picture->data[plane_index];
 604     uint8_t *src= s->  input_picture->data[plane_index];
 605     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
 606     uint8_t *cur = s->scratchbuf;
 607     uint8_t *tmp = s->emu_edge_buffer;
 608     const int b_stride = s->b_width << s->block_max_depth;
 609     const int b_height = s->b_height<< s->block_max_depth;
 610     const int w= p->width;
 611     const int h= p->height;
 612     int distortion;
 613     int rate= 0;
 614     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 615     int sx= block_w*mb_x - block_w/2;
 616     int sy= block_h*mb_y - block_h/2;
 617     int x0= FFMAX(0,-sx);
 618     int y0= FFMAX(0,-sy);
 619     int x1= FFMIN(block_w*2, w-sx);
 620     int y1= FFMIN(block_h*2, h-sy);
 621     int i,x,y;
 622
 623     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 624
 625     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);
 626
 627     for(y=y0; y<y1; y++){
 628         const uint8_t *obmc1= obmc_edged[y];
 629         const IDWTELEM *pred1 = pred + y*obmc_stride;
 630         uint8_t *cur1 = cur + y*ref_stride;
 631         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 632         for(x=x0; x<x1; x++){
 633 #if FRAC_BITS >= LOG2_OBMC_MAX
 634             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 635 #else
 636             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 637 #endif
 638             v = (v + pred1[x]) >> FRAC_BITS;
 639             if(v&(~255)) v= ~(v>>31);
 640             dst1[x] = v;
 641         }
 642     }
 643
 644     /* copy the regions where obmc[] = (uint8_t)256 */
 645     if(LOG2_OBMC_MAX == 8
 646         && (mb_x == 0 || mb_x == b_stride-1)
 647         && (mb_y == 0 || mb_y == b_height-1)){
 648         if(mb_x == 0)
 649             x1 = block_w;
 650         else
 651             x0 = block_w;
 652         if(mb_y == 0)
 653             y1 = block_h;
 654         else
 655             y0 = block_h;
 656         for(y=y0; y<y1; y++)
 657             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 658     }
 659
 660     if(block_w==16){
 661         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 662         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 663         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 664          * So improving the score of one block is not strictly guaranteed
 665          * to improve the score of the whole frame, thus iterative motion
 666          * estimation does not always converge. */
 667         if(s->avctx->me_cmp == FF_CMP_W97)
 668             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 669         else if(s->avctx->me_cmp == FF_CMP_W53)
 670             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 671         else{
 672             distortion = 0;
 673             for(i=0; i<4; i++){
 674                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 675                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 676             }
 677         }
 678     }else{
 679         av_assert2(block_w==8);
 680         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 681     }
 682
 683     if(plane_index==0){
 684         for(i=0; i<4; i++){
 685 /* ..RRr
 686  * .RXx.
 687  * rxx..
 688  */
 689             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 690         }
 691         if(mb_x == b_stride-2)
 692             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 693     }
 694     return distortion + rate*penalty_factor;
 695 }
 696
 697 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 698     int i, y2;
 699     Plane *p= &s->plane[plane_index];
 700     const int block_size = MB_SIZE >> s->block_max_depth;
 701     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 702     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 703     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 704     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 705     const int ref_stride= s->current_picture->linesize[plane_index];
 706     uint8_t *dst= s->current_picture->data[plane_index];
 707     uint8_t *src= s-> input_picture->data[plane_index];
 708     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 709     // const has only been removed from zero_dst to suppress a warning
 710     static IDWTELEM zero_dst[4096]; //FIXME
 711     const int b_stride = s->b_width << s->block_max_depth;
 712     const int w= p->width;
 713     const int h= p->height;
 714     int distortion= 0;
 715     int rate= 0;
 716     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 717
 718     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 719
 720     for(i=0; i<9; i++){
 721         int mb_x2= mb_x + (i%3) - 1;
 722         int mb_y2= mb_y + (i/3) - 1;
 723         int x= block_w*mb_x2 + block_w/2;
 724         int y= block_h*mb_y2 + block_h/2;
 725
 726         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 727                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 728
 729         //FIXME find a cleaner/simpler way to skip the outside stuff
 730         for(y2= y; y2<0; y2++)
 731             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 732         for(y2= h; y2<y+block_h; y2++)
 733             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 734         if(x<0){
 735             for(y2= y; y2<y+block_h; y2++)
 736                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 737         }
 738         if(x+block_w > w){
 739             for(y2= y; y2<y+block_h; y2++)
 740                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 741         }
 742
 743         av_assert1(block_w== 8 || block_w==16);
 744         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 745     }
 746
 747     if(plane_index==0){
 748         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 749         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 750
 751 /* ..RRRr
 752  * .RXXx.
 753  * .RXXx.
 754  * rxxx.
 755  */
 756         if(merged)
 757             rate = get_block_bits(s, mb_x, mb_y, 2);
 758         for(i=merged?4:0; i<9; i++){
 759             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 760             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 761         }
 762     }
 763     return distortion + rate*penalty_factor;
 764 }
 765
 766 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 767     const int w= b->width;
 768     const int h= b->height;
 769     int x, y;
 770
 771     if(1){
 772         int run=0;
 773         int *runs = s->run_buffer;
 774         int run_index=0;
 775         int max_index;
 776
 777         for(y=0; y<h; y++){
 778             for(x=0; x<w; x++){
 779                 int v, p=0;
 780                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 781                 v= src[x + y*stride];
 782
 783                 if(y){
 784                     t= src[x + (y-1)*stride];
 785                     if(x){
 786                         lt= src[x - 1 + (y-1)*stride];
 787                     }
 788                     if(x + 1 < w){
 789                         rt= src[x + 1 + (y-1)*stride];
 790                     }
 791                 }
 792                 if(x){
 793                     l= src[x - 1 + y*stride];
 794                     /*if(x > 1){
 795                         if(orientation==1) ll= src[y + (x-2)*stride];
 796                         else               ll= src[x - 2 + y*stride];
 797                     }*/
 798                 }
 799                 if(parent){
 800                     int px= x>>1;
 801                     int py= y>>1;
 802                     if(px<b->parent->width && py<b->parent->height)
 803                         p= parent[px + py*2*stride];
 804                 }
 805                 if(!(/*ll|*/l|lt|t|rt|p)){
 806                     if(v){
 807                         runs[run_index++]= run;
 808                         run=0;
 809                     }else{
 810                         run++;
 811                     }
 812                 }
 813             }
 814         }
 815         max_index= run_index;
 816         runs[run_index++]= run;
 817         run_index=0;
 818         run= runs[run_index++];
 819
 820         put_symbol2(&s->c, b->state[30], max_index, 0);
 821         if(run_index <= max_index)
 822             put_symbol2(&s->c, b->state[1], run, 3);
 823
 824         for(y=0; y<h; y++){
 825             if(s->c.bytestream_end - s->c.bytestream < w*40){
 826                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 827                 return -1;
 828             }
 829             for(x=0; x<w; x++){
 830                 int v, p=0;
 831                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 832                 v= src[x + y*stride];
 833
 834                 if(y){
 835                     t= src[x + (y-1)*stride];
 836                     if(x){
 837                         lt= src[x - 1 + (y-1)*stride];
 838                     }
 839                     if(x + 1 < w){
 840                         rt= src[x + 1 + (y-1)*stride];
 841                     }
 842                 }
 843                 if(x){
 844                     l= src[x - 1 + y*stride];
 845                     /*if(x > 1){
 846                         if(orientation==1) ll= src[y + (x-2)*stride];
 847                         else               ll= src[x - 2 + y*stride];
 848                     }*/
 849                 }
 850                 if(parent){
 851                     int px= x>>1;
 852                     int py= y>>1;
 853                     if(px<b->parent->width && py<b->parent->height)
 854                         p= parent[px + py*2*stride];
 855                 }
 856                 if(/*ll|*/l|lt|t|rt|p){
 857                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 858
 859                     put_rac(&s->c, &b->state[0][context], !!v);
 860                 }else{
 861                     if(!run){
 862                         run= runs[run_index++];
 863
 864                         if(run_index <= max_index)
 865                             put_symbol2(&s->c, b->state[1], run, 3);
 866                         av_assert2(v);
 867                     }else{
 868                         run--;
 869                         av_assert2(!v);
 870                     }
 871                 }
 872                 if(v){
 873                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 874                     int l2= 2*FFABS(l) + (l<0);
 875                     int t2= 2*FFABS(t) + (t<0);
 876
 877                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 878                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 879                 }
 880             }
 881         }
 882     }
 883     return 0;
 884 }
 885
 886 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 887 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 888 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 889     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 890 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 891 }
 892
 893 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){
 894     const int b_stride= s->b_width << s->block_max_depth;
 895     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 896     BlockNode backup= *block;
 897     unsigned value;
 898     int rd, index;
 899
 900     av_assert2(mb_x>=0 && mb_y>=0);
 901     av_assert2(mb_x<b_stride);
 902
 903     if(intra){
 904         block->color[0] = p[0];
 905         block->color[1] = p[1];
 906         block->color[2] = p[2];
 907         block->type |= BLOCK_INTRA;
 908     }else{
 909         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 910         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 911         if(s->me_cache[index] == value)
 912             return 0;
 913         s->me_cache[index]= value;
 914
 915         block->mx= p[0];
 916         block->my= p[1];
 917         block->type &= ~BLOCK_INTRA;
 918     }
 919
 920     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 921
 922 //FIXME chroma
 923     if(rd < *best_rd){
 924         *best_rd= rd;
 925         return 1;
 926     }else{
 927         *block= backup;
 928         return 0;
 929     }
 930 }
 931
 932 /* special case for int[2] args we discard afterwards,
 933  * fixes compilation problem with gcc 2.95 */
 934 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){
 935     int p[2] = {p0, p1};
 936     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 937 }
 938
 939 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){
 940     const int b_stride= s->b_width << s->block_max_depth;
 941     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 942     BlockNode backup[4];
 943     unsigned value;
 944     int rd, index;
 945
 946     /* We don't initialize backup[] during variable declaration, because
 947      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 948      * 'int16_t'". */
 949     backup[0] = block[0];
 950     backup[1] = block[1];
 951     backup[2] = block[b_stride];
 952     backup[3] = block[b_stride + 1];
 953
 954     av_assert2(mb_x>=0 && mb_y>=0);
 955     av_assert2(mb_x<b_stride);
 956     av_assert2(((mb_x|mb_y)&1) == 0);
 957
 958     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 959     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 960     if(s->me_cache[index] == value)
 961         return 0;
 962     s->me_cache[index]= value;
 963
 964     block->mx= p0;
 965     block->my= p1;
 966     block->ref= ref;
 967     block->type &= ~BLOCK_INTRA;
 968     block[1]= block[b_stride]= block[b_stride+1]= *block;
 969
 970     rd= get_4block_rd(s, mb_x, mb_y, 0);
 971
 972 //FIXME chroma
 973     if(rd < *best_rd){
 974         *best_rd= rd;
 975         return 1;
 976     }else{
 977         block[0]= backup[0];
 978         block[1]= backup[1];
 979         block[b_stride]= backup[2];
 980         block[b_stride+1]= backup[3];
 981         return 0;
 982     }
 983 }
 984
 985 static void iterative_me(SnowContext *s){
 986     int pass, mb_x, mb_y;
 987     const int b_width = s->b_width  << s->block_max_depth;
 988     const int b_height= s->b_height << s->block_max_depth;
 989     const int b_stride= b_width;
 990     int color[3];
 991
 992     {
 993         RangeCoder r = s->c;
 994         uint8_t state[sizeof(s->block_state)];
 995         memcpy(state, s->block_state, sizeof(s->block_state));
 996         for(mb_y= 0; mb_y<s->b_height; mb_y++)
 997             for(mb_x= 0; mb_x<s->b_width; mb_x++)
 998                 encode_q_branch(s, 0, mb_x, mb_y);
 999         s->c = r;
1000         memcpy(s->block_state, state, sizeof(s->block_state));
1001     }
1002
1003     for(pass=0; pass<25; pass++){
1004         int change= 0;
1005
1006         for(mb_y= 0; mb_y<b_height; mb_y++){
1007             for(mb_x= 0; mb_x<b_width; mb_x++){
1008                 int dia_change, i, j, ref;
1009                 int best_rd= INT_MAX, ref_rd;
1010                 BlockNode backup, ref_b;
1011                 const int index= mb_x + mb_y * b_stride;
1012                 BlockNode *block= &s->block[index];
1013                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1014                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1015                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1016                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1017                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1018                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1019                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1020                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1021                 const int b_w= (MB_SIZE >> s->block_max_depth);
1022                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1023
1024                 if(pass && (block->type & BLOCK_OPT))
1025                     continue;
1026                 block->type |= BLOCK_OPT;
1027
1028                 backup= *block;
1029
1030                 if(!s->me_cache_generation)
1031                     memset(s->me_cache, 0, sizeof(s->me_cache));
1032                 s->me_cache_generation += 1<<22;
1033
1034                 //FIXME precalculate
1035                 {
1036                     int x, y;
1037                     for (y = 0; y < b_w * 2; y++)
1038                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1039                     if(mb_x==0)
1040                         for(y=0; y<b_w*2; y++)
1041                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1042                     if(mb_x==b_stride-1)
1043                         for(y=0; y<b_w*2; y++)
1044                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1045                     if(mb_y==0){
1046                         for(x=0; x<b_w*2; x++)
1047                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1048                         for(y=1; y<b_w; y++)
1049                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1050                     }
1051                     if(mb_y==b_height-1){
1052                         for(x=0; x<b_w*2; x++)
1053                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1054                         for(y=b_w; y<b_w*2-1; y++)
1055                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1056                     }
1057                 }
1058
1059                 //skip stuff outside the picture
1060                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1061                     uint8_t *src= s->  input_picture->data[0];
1062                     uint8_t *dst= s->current_picture->data[0];
1063                     const int stride= s->current_picture->linesize[0];
1064                     const int block_w= MB_SIZE >> s->block_max_depth;
1065                     const int block_h= MB_SIZE >> s->block_max_depth;
1066                     const int sx= block_w*mb_x - block_w/2;
1067                     const int sy= block_h*mb_y - block_h/2;
1068                     const int w= s->plane[0].width;
1069                     const int h= s->plane[0].height;
1070                     int y;
1071
1072                     for(y=sy; y<0; y++)
1073                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1074                     for(y=h; y<sy+block_h*2; y++)
1075                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1076                     if(sx<0){
1077                         for(y=sy; y<sy+block_h*2; y++)
1078                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1079                     }
1080                     if(sx+block_w*2 > w){
1081                         for(y=sy; y<sy+block_h*2; y++)
1082                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1083                     }
1084                 }
1085
1086                 // intra(black) = neighbors' contribution to the current block
1087                 for(i=0; i < s->nb_planes; i++)
1088                     color[i]= get_dc(s, mb_x, mb_y, i);
1089
1090                 // get previous score (cannot be cached due to OBMC)
1091                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1092                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1093                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1094                 }else
1095                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1096
1097                 ref_b= *block;
1098                 ref_rd= best_rd;
1099                 for(ref=0; ref < s->ref_frames; ref++){
1100                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1101                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1102                         continue;
1103                     block->ref= ref;
1104                     best_rd= INT_MAX;
1105
1106                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1107                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1108                     if(tb)
1109                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1110                     if(lb)
1111                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1112                     if(rb)
1113                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1114                     if(bb)
1115                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1116
1117                     /* fullpel ME */
1118                     //FIXME avoid subpel interpolation / round to nearest integer
1119                     do{
1120                         dia_change=0;
1121                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1122                             for(j=0; j<i; j++){
1123                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1124                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1125                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1126                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1127                             }
1128                         }
1129                     }while(dia_change);
1130                     /* subpel ME */
1131                     do{
1132                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1133                         dia_change=0;
1134                         for(i=0; i<8; i++)
1135                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1136                     }while(dia_change);
1137                     //FIXME or try the standard 2 pass qpel or similar
1138
1139                     mvr[0][0]= block->mx;
1140                     mvr[0][1]= block->my;
1141                     if(ref_rd > best_rd){
1142                         ref_rd= best_rd;
1143                         ref_b= *block;
1144                     }
1145                 }
1146                 best_rd= ref_rd;
1147                 *block= ref_b;
1148                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1149                 //FIXME RD style color selection
1150                 if(!same_block(block, &backup)){
1151                     if(tb ) tb ->type &= ~BLOCK_OPT;
1152                     if(lb ) lb ->type &= ~BLOCK_OPT;
1153                     if(rb ) rb ->type &= ~BLOCK_OPT;
1154                     if(bb ) bb ->type &= ~BLOCK_OPT;
1155                     if(tlb) tlb->type &= ~BLOCK_OPT;
1156                     if(trb) trb->type &= ~BLOCK_OPT;
1157                     if(blb) blb->type &= ~BLOCK_OPT;
1158                     if(brb) brb->type &= ~BLOCK_OPT;
1159                     change ++;
1160                 }
1161             }
1162         }
1163         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1164         if(!change)
1165             break;
1166     }
1167
1168     if(s->block_max_depth == 1){
1169         int change= 0;
1170         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1171             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1172                 int i;
1173                 int best_rd, init_rd;
1174                 const int index= mb_x + mb_y * b_stride;
1175                 BlockNode *b[4];
1176
1177                 b[0]= &s->block[index];
1178                 b[1]= b[0]+1;
1179                 b[2]= b[0]+b_stride;
1180                 b[3]= b[2]+1;
1181                 if(same_block(b[0], b[1]) &&
1182                    same_block(b[0], b[2]) &&
1183                    same_block(b[0], b[3]))
1184                     continue;
1185
1186                 if(!s->me_cache_generation)
1187                     memset(s->me_cache, 0, sizeof(s->me_cache));
1188                 s->me_cache_generation += 1<<22;
1189
1190                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1191
1192                 //FIXME more multiref search?
1193                 check_4block_inter(s, mb_x, mb_y,
1194                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1195                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1196
1197                 for(i=0; i<4; i++)
1198                     if(!(b[i]->type&BLOCK_INTRA))
1199                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1200
1201                 if(init_rd != best_rd)
1202                     change++;
1203             }
1204         }
1205         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1206     }
1207 }
1208
1209 static void encode_blocks(SnowContext *s, int search){
1210     int x, y;
1211     int w= s->b_width;
1212     int h= s->b_height;
1213
1214     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1215         iterative_me(s);
1216
1217     for(y=0; y<h; y++){
1218         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1219             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1220             return;
1221         }
1222         for(x=0; x<w; x++){
1223             if(s->avctx->me_method == ME_ITER || !search)
1224                 encode_q_branch2(s, 0, x, y);
1225             else
1226                 encode_q_branch (s, 0, x, y);
1227         }
1228     }
1229 }
1230
1231 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1232     const int w= b->width;
1233     const int h= b->height;
1234     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1235     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1236     int x,y, thres1, thres2;
1237
1238     if(s->qlog == LOSSLESS_QLOG){
1239         for(y=0; y<h; y++)
1240             for(x=0; x<w; x++)
1241                 dst[x + y*stride]= src[x + y*stride];
1242         return;
1243     }
1244
1245     bias= bias ? 0 : (3*qmul)>>3;
1246     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1247     thres2= 2*thres1;
1248
1249     if(!bias){
1250         for(y=0; y<h; y++){
1251             for(x=0; x<w; x++){
1252                 int i= src[x + y*stride];
1253
1254                 if((unsigned)(i+thres1) > thres2){
1255                     if(i>=0){
1256                         i<<= QEXPSHIFT;
1257                         i/= qmul; //FIXME optimize
1258                         dst[x + y*stride]=  i;
1259                     }else{
1260                         i= -i;
1261                         i<<= QEXPSHIFT;
1262                         i/= qmul; //FIXME optimize
1263                         dst[x + y*stride]= -i;
1264                     }
1265                 }else
1266                     dst[x + y*stride]= 0;
1267             }
1268         }
1269     }else{
1270         for(y=0; y<h; y++){
1271             for(x=0; x<w; x++){
1272                 int i= src[x + y*stride];
1273
1274                 if((unsigned)(i+thres1) > thres2){
1275                     if(i>=0){
1276                         i<<= QEXPSHIFT;
1277                         i= (i + bias) / qmul; //FIXME optimize
1278                         dst[x + y*stride]=  i;
1279                     }else{
1280                         i= -i;
1281                         i<<= QEXPSHIFT;
1282                         i= (i + bias) / qmul; //FIXME optimize
1283                         dst[x + y*stride]= -i;
1284                     }
1285                 }else
1286                     dst[x + y*stride]= 0;
1287             }
1288         }
1289     }
1290 }
1291
1292 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1293     const int w= b->width;
1294     const int h= b->height;
1295     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1296     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1297     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1298     int x,y;
1299
1300     if(s->qlog == LOSSLESS_QLOG) return;
1301
1302     for(y=0; y<h; y++){
1303         for(x=0; x<w; x++){
1304             int i= src[x + y*stride];
1305             if(i<0){
1306                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1307             }else if(i>0){
1308                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1309             }
1310         }
1311     }
1312 }
1313
1314 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1315     const int w= b->width;
1316     const int h= b->height;
1317     int x,y;
1318
1319     for(y=h-1; y>=0; y--){
1320         for(x=w-1; x>=0; x--){
1321             int i= x + y*stride;
1322
1323             if(x){
1324                 if(use_median){
1325                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1326                     else  src[i] -= src[i - 1];
1327                 }else{
1328                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1329                     else  src[i] -= src[i - 1];
1330                 }
1331             }else{
1332                 if(y) src[i] -= src[i - stride];
1333             }
1334         }
1335     }
1336 }
1337
1338 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1339     const int w= b->width;
1340     const int h= b->height;
1341     int x,y;
1342
1343     for(y=0; y<h; y++){
1344         for(x=0; x<w; x++){
1345             int i= x + y*stride;
1346
1347             if(x){
1348                 if(use_median){
1349                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1350                     else  src[i] += src[i - 1];
1351                 }else{
1352                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1353                     else  src[i] += src[i - 1];
1354                 }
1355             }else{
1356                 if(y) src[i] += src[i - stride];
1357             }
1358         }
1359     }
1360 }
1361
1362 static void encode_qlogs(SnowContext *s){
1363     int plane_index, level, orientation;
1364
1365     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1366         for(level=0; level<s->spatial_decomposition_count; level++){
1367             for(orientation=level ? 1:0; orientation<4; orientation++){
1368                 if(orientation==2) continue;
1369                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1370             }
1371         }
1372     }
1373 }
1374
1375 static void encode_header(SnowContext *s){
1376     int plane_index, i;
1377     uint8_t kstate[32];
1378
1379     memset(kstate, MID_STATE, sizeof(kstate));
1380
1381     put_rac(&s->c, kstate, s->keyframe);
1382     if(s->keyframe || s->always_reset){
1383         ff_snow_reset_contexts(s);
1384         s->last_spatial_decomposition_type=
1385         s->last_qlog=
1386         s->last_qbias=
1387         s->last_mv_scale=
1388         s->last_block_max_depth= 0;
1389         for(plane_index=0; plane_index<2; plane_index++){
1390             Plane *p= &s->plane[plane_index];
1391             p->last_htaps=0;
1392             p->last_diag_mc=0;
1393             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1394         }
1395     }
1396     if(s->keyframe){
1397         put_symbol(&s->c, s->header_state, s->version, 0);
1398         put_rac(&s->c, s->header_state, s->always_reset);
1399         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1400         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1401         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1402         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1403         if (s->nb_planes > 2) {
1404             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1405             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1406         }
1407         put_rac(&s->c, s->header_state, s->spatial_scalability);
1408 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1409         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1410
1411         encode_qlogs(s);
1412     }
1413
1414     if(!s->keyframe){
1415         int update_mc=0;
1416         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1417             Plane *p= &s->plane[plane_index];
1418             update_mc |= p->last_htaps   != p->htaps;
1419             update_mc |= p->last_diag_mc != p->diag_mc;
1420             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1421         }
1422         put_rac(&s->c, s->header_state, update_mc);
1423         if(update_mc){
1424             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1425                 Plane *p= &s->plane[plane_index];
1426                 put_rac(&s->c, s->header_state, p->diag_mc);
1427                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1428                 for(i= p->htaps/2; i; i--)
1429                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1430             }
1431         }
1432         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1433             put_rac(&s->c, s->header_state, 1);
1434             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1435             encode_qlogs(s);
1436         }else
1437             put_rac(&s->c, s->header_state, 0);
1438     }
1439
1440     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1441     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1442     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1443     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1444     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1445
1446 }
1447
1448 static void update_last_header_values(SnowContext *s){
1449     int plane_index;
1450
1451     if(!s->keyframe){
1452         for(plane_index=0; plane_index<2; plane_index++){
1453             Plane *p= &s->plane[plane_index];
1454             p->last_diag_mc= p->diag_mc;
1455             p->last_htaps  = p->htaps;
1456             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1457         }
1458     }
1459
1460     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1461     s->last_qlog                        = s->qlog;
1462     s->last_qbias                       = s->qbias;
1463     s->last_mv_scale                    = s->mv_scale;
1464     s->last_block_max_depth             = s->block_max_depth;
1465     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1466 }
1467
1468 static int qscale2qlog(int qscale){
1469     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1470            + 61*QROOT/8; ///< 64 > 60
1471 }
1472
1473 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1474 {
1475     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1476      * FIXME we know exact mv bits at this point,
1477      * but ratecontrol isn't set up to include them. */
1478     uint32_t coef_sum= 0;
1479     int level, orientation, delta_qlog;
1480
1481     for(level=0; level<s->spatial_decomposition_count; level++){
1482         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1483             SubBand *b= &s->plane[0].band[level][orientation];
1484             IDWTELEM *buf= b->ibuf;
1485             const int w= b->width;
1486             const int h= b->height;
1487             const int stride= b->stride;
1488             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1489             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1490             const int qdiv= (1<<16)/qmul;
1491             int x, y;
1492             //FIXME this is ugly
1493             for(y=0; y<h; y++)
1494                 for(x=0; x<w; x++)
1495                     buf[x+y*stride]= b->buf[x+y*stride];
1496             if(orientation==0)
1497                 decorrelate(s, b, buf, stride, 1, 0);
1498             for(y=0; y<h; y++)
1499                 for(x=0; x<w; x++)
1500                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1501         }
1502     }
1503
1504     /* ugly, ratecontrol just takes a sqrt again */
1505     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1506     av_assert0(coef_sum < INT_MAX);
1507
1508     if(pict->pict_type == AV_PICTURE_TYPE_I){
1509         s->m.current_picture.mb_var_sum= coef_sum;
1510         s->m.current_picture.mc_mb_var_sum= 0;
1511     }else{
1512         s->m.current_picture.mc_mb_var_sum= coef_sum;
1513         s->m.current_picture.mb_var_sum= 0;
1514     }
1515
1516     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1517     if (pict->quality < 0)
1518         return INT_MIN;
1519     s->lambda= pict->quality * 3/2;
1520     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1521     s->qlog+= delta_qlog;
1522     return delta_qlog;
1523 }
1524
1525 static void calculate_visual_weight(SnowContext *s, Plane *p){
1526     int width = p->width;
1527     int height= p->height;
1528     int level, orientation, x, y;
1529
1530     for(level=0; level<s->spatial_decomposition_count; level++){
1531         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1532             SubBand *b= &p->band[level][orientation];
1533             IDWTELEM *ibuf= b->ibuf;
1534             int64_t error=0;
1535
1536             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1537             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1538             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1539             for(y=0; y<height; y++){
1540                 for(x=0; x<width; x++){
1541                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1542                     error += d*d;
1543                 }
1544             }
1545
1546             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1547         }
1548     }
1549 }
1550
1551 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1552                         AVFrame *pict, int *got_packet)
1553 {
1554     SnowContext *s = avctx->priv_data;
1555     RangeCoder * const c= &s->c;
1556     AVFrame *pic = pict;
1557     const int width= s->avctx->width;
1558     const int height= s->avctx->height;
1559     int level, orientation, plane_index, i, y, ret;
1560     uint8_t rc_header_bak[sizeof(s->header_state)];
1561     uint8_t rc_block_bak[sizeof(s->block_state)];
1562
1563     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1564         return ret;
1565
1566     ff_init_range_encoder(c, pkt->data, pkt->size);
1567     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1568
1569     for(i=0; i < s->nb_planes; i++){
1570         int hshift= i ? s->chroma_h_shift : 0;
1571         int vshift= i ? s->chroma_v_shift : 0;
1572         for(y=0; y<(height>>vshift); y++)
1573             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1574                    &pict->data[i][y * pict->linesize[i]],
1575                    width>>hshift);
1576         s->dsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1577                             width >> hshift, height >> vshift,
1578                             EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1579                             EDGE_TOP | EDGE_BOTTOM);
1580
1581     }
1582     emms_c();
1583     s->new_picture = pict;
1584
1585     s->m.picture_number= avctx->frame_number;
1586     if(avctx->flags&CODEC_FLAG_PASS2){
1587         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1588         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1589         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1590             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1591             if (pic->quality < 0)
1592                 return -1;
1593         }
1594     }else{
1595         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1596         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1597     }
1598
1599     if(s->pass1_rc && avctx->frame_number == 0)
1600         pic->quality = 2*FF_QP2LAMBDA;
1601     if (pic->quality) {
1602         s->qlog   = qscale2qlog(pic->quality);
1603         s->lambda = pic->quality * 3/2;
1604     }
1605     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1606         s->qlog= LOSSLESS_QLOG;
1607         s->lambda = 0;
1608     }//else keep previous frame's qlog until after motion estimation
1609
1610     ff_snow_frame_start(s);
1611     avctx->coded_frame= s->current_picture;
1612
1613     s->m.current_picture_ptr= &s->m.current_picture;
1614     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1615     s->m.current_picture.f.pts = pict->pts;
1616     if(pic->pict_type == AV_PICTURE_TYPE_P){
1617         int block_width = (width +15)>>4;
1618         int block_height= (height+15)>>4;
1619         int stride= s->current_picture->linesize[0];
1620
1621         av_assert0(s->current_picture->data[0]);
1622         av_assert0(s->last_picture[0]->data[0]);
1623
1624         s->m.avctx= s->avctx;
1625         s->m.current_picture.f.data[0] = s->current_picture->data[0];
1626         s->m.   last_picture.f.data[0] = s->last_picture[0]->data[0];
1627         s->m.    new_picture.f.data[0] = s->  input_picture->data[0];
1628         s->m.   last_picture_ptr= &s->m.   last_picture;
1629         s->m.linesize=
1630         s->m.   last_picture.f.linesize[0] =
1631         s->m.    new_picture.f.linesize[0] =
1632         s->m.current_picture.f.linesize[0] = stride;
1633         s->m.uvlinesize= s->current_picture->linesize[1];
1634         s->m.width = width;
1635         s->m.height= height;
1636         s->m.mb_width = block_width;
1637         s->m.mb_height= block_height;
1638         s->m.mb_stride=   s->m.mb_width+1;
1639         s->m.b8_stride= 2*s->m.mb_width+1;
1640         s->m.f_code=1;
1641         s->m.pict_type = pic->pict_type;
1642         s->m.me_method= s->avctx->me_method;
1643         s->m.me.scene_change_score=0;
1644         s->m.flags= s->avctx->flags;
1645         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1646         s->m.out_format= FMT_H263;
1647         s->m.unrestricted_mv= 1;
1648
1649         s->m.lambda = s->lambda;
1650         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1651         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1652
1653         s->m.dsp= s->dsp; //move
1654         s->m.hdsp = s->hdsp;
1655         ff_init_me(&s->m);
1656         s->hdsp = s->m.hdsp;
1657         s->dsp= s->m.dsp;
1658     }
1659
1660     if(s->pass1_rc){
1661         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1662         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1663     }
1664
1665 redo_frame:
1666
1667     if (pic->pict_type == AV_PICTURE_TYPE_I)
1668         s->spatial_decomposition_count= 5;
1669     else
1670         s->spatial_decomposition_count= 5;
1671
1672     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1673           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1674         s->spatial_decomposition_count--;
1675
1676     if (s->spatial_decomposition_count <= 0) {
1677         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1678         return AVERROR(EINVAL);
1679     }
1680
1681     s->m.pict_type = pic->pict_type;
1682     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1683
1684     ff_snow_common_init_after_header(avctx);
1685
1686     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1687         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1688             calculate_visual_weight(s, &s->plane[plane_index]);
1689         }
1690     }
1691
1692     encode_header(s);
1693     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1694     encode_blocks(s, 1);
1695     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1696
1697     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1698         Plane *p= &s->plane[plane_index];
1699         int w= p->width;
1700         int h= p->height;
1701         int x, y;
1702 //        int bits= put_bits_count(&s->c.pb);
1703
1704         if (!s->memc_only) {
1705             //FIXME optimize
1706             if(pict->data[plane_index]) //FIXME gray hack
1707                 for(y=0; y<h; y++){
1708                     for(x=0; x<w; x++){
1709                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1710                     }
1711                 }
1712             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1713
1714             if(   plane_index==0
1715                && pic->pict_type == AV_PICTURE_TYPE_P
1716                && !(avctx->flags&CODEC_FLAG_PASS2)
1717                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1718                 ff_init_range_encoder(c, pkt->data, pkt->size);
1719                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1720                 pic->pict_type= AV_PICTURE_TYPE_I;
1721                 s->keyframe=1;
1722                 s->current_picture->key_frame=1;
1723                 goto redo_frame;
1724             }
1725
1726             if(s->qlog == LOSSLESS_QLOG){
1727                 for(y=0; y<h; y++){
1728                     for(x=0; x<w; x++){
1729                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1730                     }
1731                 }
1732             }else{
1733                 for(y=0; y<h; y++){
1734                     for(x=0; x<w; x++){
1735                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1736                     }
1737                 }
1738             }
1739
1740             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1741
1742             if(s->pass1_rc && plane_index==0){
1743                 int delta_qlog = ratecontrol_1pass(s, pic);
1744                 if (delta_qlog <= INT_MIN)
1745                     return -1;
1746                 if(delta_qlog){
1747                     //reordering qlog in the bitstream would eliminate this reset
1748                     ff_init_range_encoder(c, pkt->data, pkt->size);
1749                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1750                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1751                     encode_header(s);
1752                     encode_blocks(s, 0);
1753                 }
1754             }
1755
1756             for(level=0; level<s->spatial_decomposition_count; level++){
1757                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1758                     SubBand *b= &p->band[level][orientation];
1759
1760                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1761                     if(orientation==0)
1762                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1763                     if (!s->no_bitstream)
1764                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1765                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1766                     if(orientation==0)
1767                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1768                 }
1769             }
1770
1771             for(level=0; level<s->spatial_decomposition_count; level++){
1772                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1773                     SubBand *b= &p->band[level][orientation];
1774
1775                     dequantize(s, b, b->ibuf, b->stride);
1776                 }
1777             }
1778
1779             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1780             if(s->qlog == LOSSLESS_QLOG){
1781                 for(y=0; y<h; y++){
1782                     for(x=0; x<w; x++){
1783                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1784                     }
1785                 }
1786             }
1787             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1788         }else{
1789             //ME/MC only
1790             if(pic->pict_type == AV_PICTURE_TYPE_I){
1791                 for(y=0; y<h; y++){
1792                     for(x=0; x<w; x++){
1793                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1794                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1795                     }
1796                 }
1797             }else{
1798                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1799                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1800             }
1801         }
1802         if(s->avctx->flags&CODEC_FLAG_PSNR){
1803             int64_t error= 0;
1804
1805             if(pict->data[plane_index]) //FIXME gray hack
1806                 for(y=0; y<h; y++){
1807                     for(x=0; x<w; x++){
1808                         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];
1809                         error += d*d;
1810                     }
1811                 }
1812             s->avctx->error[plane_index] += error;
1813             s->current_picture->error[plane_index] = error;
1814         }
1815
1816     }
1817
1818     update_last_header_values(s);
1819
1820     ff_snow_release_buffer(avctx);
1821
1822     s->current_picture->coded_picture_number = avctx->frame_number;
1823     s->current_picture->pict_type = pict->pict_type;
1824     s->current_picture->quality = pict->quality;
1825     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1826     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1827     s->m.current_picture.f.display_picture_number =
1828     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1829     s->m.current_picture.f.quality                = pic->quality;
1830     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1831     if(s->pass1_rc)
1832         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1833             return -1;
1834     if(avctx->flags&CODEC_FLAG_PASS1)
1835         ff_write_pass1_stats(&s->m);
1836     s->m.last_pict_type = s->m.pict_type;
1837     avctx->frame_bits = s->m.frame_bits;
1838     avctx->mv_bits = s->m.mv_bits;
1839     avctx->misc_bits = s->m.misc_bits;
1840     avctx->p_tex_bits = s->m.p_tex_bits;
1841
1842     emms_c();
1843
1844     pkt->size = ff_rac_terminate(c);
1845     if (avctx->coded_frame->key_frame)
1846         pkt->flags |= AV_PKT_FLAG_KEY;
1847     *got_packet = 1;
1848
1849     return 0;
1850 }
1851
1852 static av_cold int encode_end(AVCodecContext *avctx)
1853 {
1854     SnowContext *s = avctx->priv_data;
1855
1856     ff_snow_common_end(s);
1857     ff_rate_control_uninit(&s->m);
1858     av_frame_free(&s->input_picture);
1859     av_free(avctx->stats_out);
1860
1861     return 0;
1862 }
1863
1864 #define OFFSET(x) offsetof(SnowContext, x)
1865 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1866 static const AVOption options[] = {
1867     { "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 },
1868     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1869     { NULL },
1870 };
1871
1872 static const AVClass snowenc_class = {
1873     .class_name = "snow encoder",
1874     .item_name  = av_default_item_name,
1875     .option     = options,
1876     .version    = LIBAVUTIL_VERSION_INT,
1877 };
1878
1879 AVCodec ff_snow_encoder = {
1880     .name           = "snow",
1881     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1882     .type           = AVMEDIA_TYPE_VIDEO,
1883     .id             = AV_CODEC_ID_SNOW,
1884     .priv_data_size = sizeof(SnowContext),
1885     .init           = encode_init,
1886     .encode2        = encode_frame,
1887     .close          = encode_end,
1888     .pix_fmts       = (const enum AVPixelFormat[]){
1889         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1890         AV_PIX_FMT_GRAY8,
1891         AV_PIX_FMT_NONE
1892     },
1893     .priv_class     = &snowenc_class,
1894 };
1895
1896
1897 #ifdef TEST
1898 #undef malloc
1899 #undef free
1900 #undef printf
1901
1902 #include "libavutil/lfg.h"
1903 #include "libavutil/mathematics.h"
1904
1905 int main(void){
1906 #define width  256
1907 #define height 256
1908     int buffer[2][width*height];
1909     SnowContext s;
1910     int i;
1911     AVLFG prng;
1912     s.spatial_decomposition_count=6;
1913     s.spatial_decomposition_type=1;
1914
1915     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1916     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1917
1918     av_lfg_init(&prng, 1);
1919
1920     printf("testing 5/3 DWT\n");
1921     for(i=0; i<width*height; i++)
1922         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1923
1924     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1925     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1926
1927     for(i=0; i<width*height; i++)
1928         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1929
1930     printf("testing 9/7 DWT\n");
1931     s.spatial_decomposition_type=0;
1932     for(i=0; i<width*height; i++)
1933         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1934
1935     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1936     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1937
1938     for(i=0; i<width*height; i++)
1939         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1940
1941     {
1942     int level, orientation, x, y;
1943     int64_t errors[8][4];
1944     int64_t g=0;
1945
1946         memset(errors, 0, sizeof(errors));
1947         s.spatial_decomposition_count=3;
1948         s.spatial_decomposition_type=0;
1949         for(level=0; level<s.spatial_decomposition_count; level++){
1950             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1951                 int w= width  >> (s.spatial_decomposition_count-level);
1952                 int h= height >> (s.spatial_decomposition_count-level);
1953                 int stride= width  << (s.spatial_decomposition_count-level);
1954                 DWTELEM *buf= buffer[0];
1955                 int64_t error=0;
1956
1957                 if(orientation&1) buf+=w;
1958                 if(orientation>1) buf+=stride>>1;
1959
1960                 memset(buffer[0], 0, sizeof(int)*width*height);
1961                 buf[w/2 + h/2*stride]= 256*256;
1962                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1963                 for(y=0; y<height; y++){
1964                     for(x=0; x<width; x++){
1965                         int64_t d= buffer[0][x + y*width];
1966                         error += d*d;
1967                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1968                     }
1969                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1970                 }
1971                 error= (int)(sqrt(error)+0.5);
1972                 errors[level][orientation]= error;
1973                 if(g) g=av_gcd(g, error);
1974                 else g= error;
1975             }
1976         }
1977         printf("static int const visual_weight[][4]={\n");
1978         for(level=0; level<s.spatial_decomposition_count; level++){
1979             printf("  {");
1980             for(orientation=0; orientation<4; orientation++){
1981                 printf("%8"PRId64",", errors[level][orientation]/g);
1982             }
1983             printf("},\n");
1984         }
1985         printf("};\n");
1986         {
1987             int level=2;
1988             int w= width  >> (s.spatial_decomposition_count-level);
1989             //int h= height >> (s.spatial_decomposition_count-level);
1990             int stride= width  << (s.spatial_decomposition_count-level);
1991             DWTELEM *buf= buffer[0];
1992             int64_t error=0;
1993
1994             buf+=w;
1995             buf+=stride>>1;
1996
1997             memset(buffer[0], 0, sizeof(int)*width*height);
1998             for(y=0; y<height; y++){
1999                 for(x=0; x<width; x++){
2000                     int tab[4]={0,2,3,1};
2001                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2002                 }
2003             }
2004             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2005             for(y=0; y<height; y++){
2006                 for(x=0; x<width; x++){
2007                     int64_t d= buffer[0][x + y*width];
2008                     error += d*d;
2009                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2010                 }
2011                 if(FFABS(height/2-y)<9) printf("\n");
2012             }
2013         }
2014
2015     }
2016     return 0;
2017 }
2018 #endif /* TEST */