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 "internal.h"
  26 #include "snow_dwt.h"
  27 #include "snow.h"
  28
  29 #include "rangecoder.h"
  30 #include "mathops.h"
  31
  32 #include "mpegvideo.h"
  33 #include "h263.h"
  34
  35 static av_cold int encode_init(AVCodecContext *avctx)
  36 {
  37     SnowContext *s = avctx->priv_data;
  38     int plane_index, ret;
  39     int i;
  40
  41     if(avctx->prediction_method == DWT_97
  42        && (avctx->flags & CODEC_FLAG_QSCALE)
  43        && avctx->global_quality == 0){
  44         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  45         return -1;
  46     }
  47
  48     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
  49
  50     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
  51     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
  52
  53     for(plane_index=0; plane_index<3; plane_index++){
  54         s->plane[plane_index].diag_mc= 1;
  55         s->plane[plane_index].htaps= 6;
  56         s->plane[plane_index].hcoeff[0]=  40;
  57         s->plane[plane_index].hcoeff[1]= -10;
  58         s->plane[plane_index].hcoeff[2]=   2;
  59         s->plane[plane_index].fast_mc= 1;
  60     }
  61
  62     if ((ret = ff_snow_common_init(avctx)) < 0) {
  63         ff_snow_common_end(avctx->priv_data);
  64         return ret;
  65     }
  66     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  67
  68     ff_snow_alloc_blocks(s);
  69
  70     s->version=0;
  71
  72     s->m.avctx   = avctx;
  73     s->m.flags   = avctx->flags;
  74     s->m.bit_rate= avctx->bit_rate;
  75
  76     s->m.me.temp      =
  77     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  78     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  79     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  80     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  81     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)
  82         return AVERROR(ENOMEM);
  83
  84     ff_h263_encode_init(&s->m); //mv_penalty
  85
  86     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
  87
  88     if(avctx->flags&CODEC_FLAG_PASS1){
  89         if(!avctx->stats_out)
  90             avctx->stats_out = av_mallocz(256);
  91
  92         if (!avctx->stats_out)
  93             return AVERROR(ENOMEM);
  94     }
  95     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
  96         if(ff_rate_control_init(&s->m) < 0)
  97             return -1;
  98     }
  99     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 100
 101     switch(avctx->pix_fmt){
 102     case AV_PIX_FMT_YUV444P:
 103 //    case AV_PIX_FMT_YUV422P:
 104     case AV_PIX_FMT_YUV420P:
 105 //    case AV_PIX_FMT_YUV411P:
 106     case AV_PIX_FMT_YUV410P:
 107         s->nb_planes = 3;
 108         s->colorspace_type= 0;
 109         break;
 110     case AV_PIX_FMT_GRAY8:
 111         s->nb_planes = 1;
 112         s->colorspace_type = 1;
 113         break;
 114 /*    case AV_PIX_FMT_RGB32:
 115         s->colorspace= 1;
 116         break;*/
 117     default:
 118         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 119         return -1;
 120     }
 121     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 122
 123     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 124     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 125
 126     s->input_picture = av_frame_alloc();
 127     avctx->coded_frame = av_frame_alloc();
 128     if (!s->input_picture || !avctx->coded_frame)
 129         return AVERROR(ENOMEM);
 130
 131     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 132         return ret;
 133
 134     if(s->avctx->me_method == ME_ITER){
 135         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 136         for(i=0; i<s->max_ref_frames; i++){
 137             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 138             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 139             if (!s->ref_mvs[i] || !s->ref_scores[i])
 140                 return AVERROR(ENOMEM);
 141         }
 142     }
 143
 144     return 0;
 145 }
 146
 147 //near copy & paste from dsputil, FIXME
 148 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 149 {
 150     int s, i, j;
 151
 152     s = 0;
 153     for (i = 0; i < h; i++) {
 154         for (j = 0; j < w; j++) {
 155             s += pix[0];
 156             pix ++;
 157         }
 158         pix += line_size - w;
 159     }
 160     return s;
 161 }
 162
 163 //near copy & paste from dsputil, FIXME
 164 static int pix_norm1(uint8_t * pix, int line_size, int w)
 165 {
 166     int s, i, j;
 167     uint32_t *sq = ff_square_tab + 256;
 168
 169     s = 0;
 170     for (i = 0; i < w; i++) {
 171         for (j = 0; j < w; j ++) {
 172             s += sq[pix[0]];
 173             pix ++;
 174         }
 175         pix += line_size - w;
 176     }
 177     return s;
 178 }
 179
 180 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 181     switch(type&0xFF){
 182     default:
 183     case FF_CMP_SAD:
 184         return lambda>>FF_LAMBDA_SHIFT;
 185     case FF_CMP_DCT:
 186         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 187     case FF_CMP_W53:
 188         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 189     case FF_CMP_W97:
 190         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 191     case FF_CMP_SATD:
 192     case FF_CMP_DCT264:
 193         return (2*lambda)>>FF_LAMBDA_SHIFT;
 194     case FF_CMP_RD:
 195     case FF_CMP_PSNR:
 196     case FF_CMP_SSE:
 197     case FF_CMP_NSSE:
 198         return lambda2>>FF_LAMBDA_SHIFT;
 199     case FF_CMP_BIT:
 200         return 1;
 201     }
 202 }
 203
 204 //FIXME copy&paste
 205 #define P_LEFT P[1]
 206 #define P_TOP P[2]
 207 #define P_TOPRIGHT P[3]
 208 #define P_MEDIAN P[4]
 209 #define P_MV1 P[9]
 210 #define FLAG_QPEL   1 //must be 1
 211
 212 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 213     uint8_t p_buffer[1024];
 214     uint8_t i_buffer[1024];
 215     uint8_t p_state[sizeof(s->block_state)];
 216     uint8_t i_state[sizeof(s->block_state)];
 217     RangeCoder pc, ic;
 218     uint8_t *pbbak= s->c.bytestream;
 219     uint8_t *pbbak_start= s->c.bytestream_start;
 220     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 221     const int w= s->b_width  << s->block_max_depth;
 222     const int h= s->b_height << s->block_max_depth;
 223     const int rem_depth= s->block_max_depth - level;
 224     const int index= (x + y*w) << rem_depth;
 225     const int block_w= 1<<(LOG2_MB_SIZE - level);
 226     int trx= (x+1)<<rem_depth;
 227     int try= (y+1)<<rem_depth;
 228     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 229     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 230     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 231     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 232     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 233     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 234     int pl = left->color[0];
 235     int pcb= left->color[1];
 236     int pcr= left->color[2];
 237     int pmx, pmy;
 238     int mx=0, my=0;
 239     int l,cr,cb;
 240     const int stride= s->current_picture->linesize[0];
 241     const int uvstride= s->current_picture->linesize[1];
 242     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 243                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 244                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 245     int P[10][2];
 246     int16_t last_mv[3][2];
 247     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
 248     const int shift= 1+qpel;
 249     MotionEstContext *c= &s->m.me;
 250     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 251     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 252     int my_context= av_log2(2*FFABS(left->my - top->my));
 253     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 254     int ref, best_ref, ref_score, ref_mx, ref_my;
 255
 256     av_assert0(sizeof(s->block_state) >= 256);
 257     if(s->keyframe){
 258         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 259         return 0;
 260     }
 261
 262 //    clip predictors / edge ?
 263
 264     P_LEFT[0]= left->mx;
 265     P_LEFT[1]= left->my;
 266     P_TOP [0]= top->mx;
 267     P_TOP [1]= top->my;
 268     P_TOPRIGHT[0]= tr->mx;
 269     P_TOPRIGHT[1]= tr->my;
 270
 271     last_mv[0][0]= s->block[index].mx;
 272     last_mv[0][1]= s->block[index].my;
 273     last_mv[1][0]= right->mx;
 274     last_mv[1][1]= right->my;
 275     last_mv[2][0]= bottom->mx;
 276     last_mv[2][1]= bottom->my;
 277
 278     s->m.mb_stride=2;
 279     s->m.mb_x=
 280     s->m.mb_y= 0;
 281     c->skip= 0;
 282
 283     av_assert1(c->  stride ==   stride);
 284     av_assert1(c->uvstride == uvstride);
 285
 286     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 287     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 288     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 289     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 290
 291     c->xmin = - x*block_w - 16+3;
 292     c->ymin = - y*block_w - 16+3;
 293     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 294     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 295
 296     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 297     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 298     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 299     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 300     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 301     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 302     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 303
 304     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 305     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 306
 307     if (!y) {
 308         c->pred_x= P_LEFT[0];
 309         c->pred_y= P_LEFT[1];
 310     } else {
 311         c->pred_x = P_MEDIAN[0];
 312         c->pred_y = P_MEDIAN[1];
 313     }
 314
 315     score= INT_MAX;
 316     best_ref= 0;
 317     for(ref=0; ref<s->ref_frames; ref++){
 318         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 319
 320         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 321                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 322
 323         av_assert2(ref_mx >= c->xmin);
 324         av_assert2(ref_mx <= c->xmax);
 325         av_assert2(ref_my >= c->ymin);
 326         av_assert2(ref_my <= c->ymax);
 327
 328         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 329         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 330         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 331         if(s->ref_mvs[ref]){
 332             s->ref_mvs[ref][index][0]= ref_mx;
 333             s->ref_mvs[ref][index][1]= ref_my;
 334             s->ref_scores[ref][index]= ref_score;
 335         }
 336         if(score > ref_score){
 337             score= ref_score;
 338             best_ref= ref;
 339             mx= ref_mx;
 340             my= ref_my;
 341         }
 342     }
 343     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 344
 345   //  subpel search
 346     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 347     pc= s->c;
 348     pc.bytestream_start=
 349     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 350     memcpy(p_state, s->block_state, sizeof(s->block_state));
 351
 352     if(level!=s->block_max_depth)
 353         put_rac(&pc, &p_state[4 + s_context], 1);
 354     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 355     if(s->ref_frames > 1)
 356         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 357     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 358     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 359     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 360     p_len= pc.bytestream - pc.bytestream_start;
 361     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 362
 363     block_s= block_w*block_w;
 364     sum = pix_sum(current_data[0], stride, block_w, block_w);
 365     l= (sum + block_s/2)/block_s;
 366     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 367
 368     if (s->nb_planes > 2) {
 369         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 370         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 371         cb= (sum + block_s/2)/block_s;
 372     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 373         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 374         cr= (sum + block_s/2)/block_s;
 375     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 376     }else
 377         cb = cr = 0;
 378
 379     ic= s->c;
 380     ic.bytestream_start=
 381     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 382     memcpy(i_state, s->block_state, sizeof(s->block_state));
 383     if(level!=s->block_max_depth)
 384         put_rac(&ic, &i_state[4 + s_context], 1);
 385     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 386     put_symbol(&ic, &i_state[32],  l-pl , 1);
 387     if (s->nb_planes > 2) {
 388         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 389         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 390     }
 391     i_len= ic.bytestream - ic.bytestream_start;
 392     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 393
 394     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 395     av_assert1(iscore >= 0);
 396     av_assert1(l>=0 && l<=255);
 397     av_assert1(pl>=0 && pl<=255);
 398
 399     if(level==0){
 400         int varc= iscore >> 8;
 401         int vard= score >> 8;
 402         if (vard <= 64 || vard < varc)
 403             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 404         else
 405             c->scene_change_score+= s->m.qscale;
 406     }
 407
 408     if(level!=s->block_max_depth){
 409         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 410         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 411         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 412         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 413         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 414         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 415
 416         if(score2 < score && score2 < iscore)
 417             return score2;
 418     }
 419
 420     if(iscore < score){
 421         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 422         memcpy(pbbak, i_buffer, i_len);
 423         s->c= ic;
 424         s->c.bytestream_start= pbbak_start;
 425         s->c.bytestream= pbbak + i_len;
 426         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 427         memcpy(s->block_state, i_state, sizeof(s->block_state));
 428         return iscore;
 429     }else{
 430         memcpy(pbbak, p_buffer, p_len);
 431         s->c= pc;
 432         s->c.bytestream_start= pbbak_start;
 433         s->c.bytestream= pbbak + p_len;
 434         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 435         memcpy(s->block_state, p_state, sizeof(s->block_state));
 436         return score;
 437     }
 438 }
 439
 440 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 441     const int w= s->b_width  << s->block_max_depth;
 442     const int rem_depth= s->block_max_depth - level;
 443     const int index= (x + y*w) << rem_depth;
 444     int trx= (x+1)<<rem_depth;
 445     BlockNode *b= &s->block[index];
 446     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 447     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 448     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 449     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 450     int pl = left->color[0];
 451     int pcb= left->color[1];
 452     int pcr= left->color[2];
 453     int pmx, pmy;
 454     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 455     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 456     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 457     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 458
 459     if(s->keyframe){
 460         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 461         return;
 462     }
 463
 464     if(level!=s->block_max_depth){
 465         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 466             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 467         }else{
 468             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 469             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 470             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 471             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 472             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 473             return;
 474         }
 475     }
 476     if(b->type & BLOCK_INTRA){
 477         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 478         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 479         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 480         if (s->nb_planes > 2) {
 481             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 482             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 483         }
 484         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 485     }else{
 486         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 487         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 488         if(s->ref_frames > 1)
 489             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 490         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 491         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 492         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 493     }
 494 }
 495
 496 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 497     int i, x2, y2;
 498     Plane *p= &s->plane[plane_index];
 499     const int block_size = MB_SIZE >> s->block_max_depth;
 500     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 501     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 502     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 503     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 504     const int ref_stride= s->current_picture->linesize[plane_index];
 505     uint8_t *src= s-> input_picture->data[plane_index];
 506     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 507     const int b_stride = s->b_width << s->block_max_depth;
 508     const int w= p->width;
 509     const int h= p->height;
 510     int index= mb_x + mb_y*b_stride;
 511     BlockNode *b= &s->block[index];
 512     BlockNode backup= *b;
 513     int ab=0;
 514     int aa=0;
 515
 516     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 517
 518     b->type|= BLOCK_INTRA;
 519     b->color[plane_index]= 0;
 520     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 521
 522     for(i=0; i<4; i++){
 523         int mb_x2= mb_x + (i &1) - 1;
 524         int mb_y2= mb_y + (i>>1) - 1;
 525         int x= block_w*mb_x2 + block_w/2;
 526         int y= block_h*mb_y2 + block_h/2;
 527
 528         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 529                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 530
 531         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 532             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 533                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 534                 int obmc_v= obmc[index];
 535                 int d;
 536                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 537                 if(x<0) obmc_v += obmc[index + block_w];
 538                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 539                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 540                 //FIXME precalculate this or simplify it somehow else
 541
 542                 d = -dst[index] + (1<<(FRAC_BITS-1));
 543                 dst[index] = d;
 544                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 545                 aa += obmc_v * obmc_v; //FIXME precalculate this
 546             }
 547         }
 548     }
 549     *b= backup;
 550
 551     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 552 }
 553
 554 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 555     const int b_stride = s->b_width << s->block_max_depth;
 556     const int b_height = s->b_height<< s->block_max_depth;
 557     int index= x + y*b_stride;
 558     const BlockNode *b     = &s->block[index];
 559     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 560     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 561     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 562     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 563     int dmx, dmy;
 564 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 565 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 566
 567     if(x<0 || x>=b_stride || y>=b_height)
 568         return 0;
 569 /*
 570 1            0      0
 571 01X          1-2    1
 572 001XX        3-6    2-3
 573 0001XXX      7-14   4-7
 574 00001XXXX   15-30   8-15
 575 */
 576 //FIXME try accurate rate
 577 //FIXME intra and inter predictors if surrounding blocks are not the same type
 578     if(b->type & BLOCK_INTRA){
 579         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 580                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 581                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 582     }else{
 583         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 584         dmx-= b->mx;
 585         dmy-= b->my;
 586         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 587                     + av_log2(2*FFABS(dmy))
 588                     + av_log2(2*b->ref));
 589     }
 590 }
 591
 592 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 593     Plane *p= &s->plane[plane_index];
 594     const int block_size = MB_SIZE >> s->block_max_depth;
 595     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 596     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 597     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 598     const int ref_stride= s->current_picture->linesize[plane_index];
 599     uint8_t *dst= s->current_picture->data[plane_index];
 600     uint8_t *src= s->  input_picture->data[plane_index];
 601     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
 602     uint8_t *cur = s->scratchbuf;
 603     uint8_t *tmp = s->emu_edge_buffer;
 604     const int b_stride = s->b_width << s->block_max_depth;
 605     const int b_height = s->b_height<< s->block_max_depth;
 606     const int w= p->width;
 607     const int h= p->height;
 608     int distortion;
 609     int rate= 0;
 610     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 611     int sx= block_w*mb_x - block_w/2;
 612     int sy= block_h*mb_y - block_h/2;
 613     int x0= FFMAX(0,-sx);
 614     int y0= FFMAX(0,-sy);
 615     int x1= FFMIN(block_w*2, w-sx);
 616     int y1= FFMIN(block_h*2, h-sy);
 617     int i,x,y;
 618
 619     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 620
 621     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);
 622
 623     for(y=y0; y<y1; y++){
 624         const uint8_t *obmc1= obmc_edged[y];
 625         const IDWTELEM *pred1 = pred + y*obmc_stride;
 626         uint8_t *cur1 = cur + y*ref_stride;
 627         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 628         for(x=x0; x<x1; x++){
 629 #if FRAC_BITS >= LOG2_OBMC_MAX
 630             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 631 #else
 632             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 633 #endif
 634             v = (v + pred1[x]) >> FRAC_BITS;
 635             if(v&(~255)) v= ~(v>>31);
 636             dst1[x] = v;
 637         }
 638     }
 639
 640     /* copy the regions where obmc[] = (uint8_t)256 */
 641     if(LOG2_OBMC_MAX == 8
 642         && (mb_x == 0 || mb_x == b_stride-1)
 643         && (mb_y == 0 || mb_y == b_height-1)){
 644         if(mb_x == 0)
 645             x1 = block_w;
 646         else
 647             x0 = block_w;
 648         if(mb_y == 0)
 649             y1 = block_h;
 650         else
 651             y0 = block_h;
 652         for(y=y0; y<y1; y++)
 653             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 654     }
 655
 656     if(block_w==16){
 657         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 658         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 659         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 660          * So improving the score of one block is not strictly guaranteed
 661          * to improve the score of the whole frame, thus iterative motion
 662          * estimation does not always converge. */
 663         if(s->avctx->me_cmp == FF_CMP_W97)
 664             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 665         else if(s->avctx->me_cmp == FF_CMP_W53)
 666             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 667         else{
 668             distortion = 0;
 669             for(i=0; i<4; i++){
 670                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 671                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 672             }
 673         }
 674     }else{
 675         av_assert2(block_w==8);
 676         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 677     }
 678
 679     if(plane_index==0){
 680         for(i=0; i<4; i++){
 681 /* ..RRr
 682  * .RXx.
 683  * rxx..
 684  */
 685             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 686         }
 687         if(mb_x == b_stride-2)
 688             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 689     }
 690     return distortion + rate*penalty_factor;
 691 }
 692
 693 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 694     int i, y2;
 695     Plane *p= &s->plane[plane_index];
 696     const int block_size = MB_SIZE >> s->block_max_depth;
 697     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 698     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 699     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 700     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 701     const int ref_stride= s->current_picture->linesize[plane_index];
 702     uint8_t *dst= s->current_picture->data[plane_index];
 703     uint8_t *src= s-> input_picture->data[plane_index];
 704     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 705     // const has only been removed from zero_dst to suppress a warning
 706     static IDWTELEM zero_dst[4096]; //FIXME
 707     const int b_stride = s->b_width << s->block_max_depth;
 708     const int w= p->width;
 709     const int h= p->height;
 710     int distortion= 0;
 711     int rate= 0;
 712     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 713
 714     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 715
 716     for(i=0; i<9; i++){
 717         int mb_x2= mb_x + (i%3) - 1;
 718         int mb_y2= mb_y + (i/3) - 1;
 719         int x= block_w*mb_x2 + block_w/2;
 720         int y= block_h*mb_y2 + block_h/2;
 721
 722         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 723                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 724
 725         //FIXME find a cleaner/simpler way to skip the outside stuff
 726         for(y2= y; y2<0; y2++)
 727             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 728         for(y2= h; y2<y+block_h; y2++)
 729             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 730         if(x<0){
 731             for(y2= y; y2<y+block_h; y2++)
 732                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 733         }
 734         if(x+block_w > w){
 735             for(y2= y; y2<y+block_h; y2++)
 736                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 737         }
 738
 739         av_assert1(block_w== 8 || block_w==16);
 740         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 741     }
 742
 743     if(plane_index==0){
 744         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 745         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 746
 747 /* ..RRRr
 748  * .RXXx.
 749  * .RXXx.
 750  * rxxx.
 751  */
 752         if(merged)
 753             rate = get_block_bits(s, mb_x, mb_y, 2);
 754         for(i=merged?4:0; i<9; i++){
 755             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 756             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 757         }
 758     }
 759     return distortion + rate*penalty_factor;
 760 }
 761
 762 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 763     const int w= b->width;
 764     const int h= b->height;
 765     int x, y;
 766
 767     if(1){
 768         int run=0;
 769         int *runs = s->run_buffer;
 770         int run_index=0;
 771         int max_index;
 772
 773         for(y=0; y<h; y++){
 774             for(x=0; x<w; x++){
 775                 int v, p=0;
 776                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 777                 v= src[x + y*stride];
 778
 779                 if(y){
 780                     t= src[x + (y-1)*stride];
 781                     if(x){
 782                         lt= src[x - 1 + (y-1)*stride];
 783                     }
 784                     if(x + 1 < w){
 785                         rt= src[x + 1 + (y-1)*stride];
 786                     }
 787                 }
 788                 if(x){
 789                     l= src[x - 1 + y*stride];
 790                     /*if(x > 1){
 791                         if(orientation==1) ll= src[y + (x-2)*stride];
 792                         else               ll= src[x - 2 + y*stride];
 793                     }*/
 794                 }
 795                 if(parent){
 796                     int px= x>>1;
 797                     int py= y>>1;
 798                     if(px<b->parent->width && py<b->parent->height)
 799                         p= parent[px + py*2*stride];
 800                 }
 801                 if(!(/*ll|*/l|lt|t|rt|p)){
 802                     if(v){
 803                         runs[run_index++]= run;
 804                         run=0;
 805                     }else{
 806                         run++;
 807                     }
 808                 }
 809             }
 810         }
 811         max_index= run_index;
 812         runs[run_index++]= run;
 813         run_index=0;
 814         run= runs[run_index++];
 815
 816         put_symbol2(&s->c, b->state[30], max_index, 0);
 817         if(run_index <= max_index)
 818             put_symbol2(&s->c, b->state[1], run, 3);
 819
 820         for(y=0; y<h; y++){
 821             if(s->c.bytestream_end - s->c.bytestream < w*40){
 822                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 823                 return -1;
 824             }
 825             for(x=0; x<w; x++){
 826                 int v, p=0;
 827                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 828                 v= src[x + y*stride];
 829
 830                 if(y){
 831                     t= src[x + (y-1)*stride];
 832                     if(x){
 833                         lt= src[x - 1 + (y-1)*stride];
 834                     }
 835                     if(x + 1 < w){
 836                         rt= src[x + 1 + (y-1)*stride];
 837                     }
 838                 }
 839                 if(x){
 840                     l= src[x - 1 + y*stride];
 841                     /*if(x > 1){
 842                         if(orientation==1) ll= src[y + (x-2)*stride];
 843                         else               ll= src[x - 2 + y*stride];
 844                     }*/
 845                 }
 846                 if(parent){
 847                     int px= x>>1;
 848                     int py= y>>1;
 849                     if(px<b->parent->width && py<b->parent->height)
 850                         p= parent[px + py*2*stride];
 851                 }
 852                 if(/*ll|*/l|lt|t|rt|p){
 853                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 854
 855                     put_rac(&s->c, &b->state[0][context], !!v);
 856                 }else{
 857                     if(!run){
 858                         run= runs[run_index++];
 859
 860                         if(run_index <= max_index)
 861                             put_symbol2(&s->c, b->state[1], run, 3);
 862                         av_assert2(v);
 863                     }else{
 864                         run--;
 865                         av_assert2(!v);
 866                     }
 867                 }
 868                 if(v){
 869                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 870                     int l2= 2*FFABS(l) + (l<0);
 871                     int t2= 2*FFABS(t) + (t<0);
 872
 873                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 874                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 875                 }
 876             }
 877         }
 878     }
 879     return 0;
 880 }
 881
 882 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 883 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 884 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 885     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 886 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 887 }
 888
 889 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){
 890     const int b_stride= s->b_width << s->block_max_depth;
 891     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 892     BlockNode backup= *block;
 893     unsigned value;
 894     int rd, index;
 895
 896     av_assert2(mb_x>=0 && mb_y>=0);
 897     av_assert2(mb_x<b_stride);
 898
 899     if(intra){
 900         block->color[0] = p[0];
 901         block->color[1] = p[1];
 902         block->color[2] = p[2];
 903         block->type |= BLOCK_INTRA;
 904     }else{
 905         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 906         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 907         if(s->me_cache[index] == value)
 908             return 0;
 909         s->me_cache[index]= value;
 910
 911         block->mx= p[0];
 912         block->my= p[1];
 913         block->type &= ~BLOCK_INTRA;
 914     }
 915
 916     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 917
 918 //FIXME chroma
 919     if(rd < *best_rd){
 920         *best_rd= rd;
 921         return 1;
 922     }else{
 923         *block= backup;
 924         return 0;
 925     }
 926 }
 927
 928 /* special case for int[2] args we discard afterwards,
 929  * fixes compilation problem with gcc 2.95 */
 930 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){
 931     int p[2] = {p0, p1};
 932     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 933 }
 934
 935 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){
 936     const int b_stride= s->b_width << s->block_max_depth;
 937     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 938     BlockNode backup[4];
 939     unsigned value;
 940     int rd, index;
 941
 942     /* We don't initialize backup[] during variable declaration, because
 943      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 944      * 'int16_t'". */
 945     backup[0] = block[0];
 946     backup[1] = block[1];
 947     backup[2] = block[b_stride];
 948     backup[3] = block[b_stride + 1];
 949
 950     av_assert2(mb_x>=0 && mb_y>=0);
 951     av_assert2(mb_x<b_stride);
 952     av_assert2(((mb_x|mb_y)&1) == 0);
 953
 954     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 955     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 956     if(s->me_cache[index] == value)
 957         return 0;
 958     s->me_cache[index]= value;
 959
 960     block->mx= p0;
 961     block->my= p1;
 962     block->ref= ref;
 963     block->type &= ~BLOCK_INTRA;
 964     block[1]= block[b_stride]= block[b_stride+1]= *block;
 965
 966     rd= get_4block_rd(s, mb_x, mb_y, 0);
 967
 968 //FIXME chroma
 969     if(rd < *best_rd){
 970         *best_rd= rd;
 971         return 1;
 972     }else{
 973         block[0]= backup[0];
 974         block[1]= backup[1];
 975         block[b_stride]= backup[2];
 976         block[b_stride+1]= backup[3];
 977         return 0;
 978     }
 979 }
 980
 981 static void iterative_me(SnowContext *s){
 982     int pass, mb_x, mb_y;
 983     const int b_width = s->b_width  << s->block_max_depth;
 984     const int b_height= s->b_height << s->block_max_depth;
 985     const int b_stride= b_width;
 986     int color[3];
 987
 988     {
 989         RangeCoder r = s->c;
 990         uint8_t state[sizeof(s->block_state)];
 991         memcpy(state, s->block_state, sizeof(s->block_state));
 992         for(mb_y= 0; mb_y<s->b_height; mb_y++)
 993             for(mb_x= 0; mb_x<s->b_width; mb_x++)
 994                 encode_q_branch(s, 0, mb_x, mb_y);
 995         s->c = r;
 996         memcpy(s->block_state, state, sizeof(s->block_state));
 997     }
 998
 999     for(pass=0; pass<25; pass++){
1000         int change= 0;
1001
1002         for(mb_y= 0; mb_y<b_height; mb_y++){
1003             for(mb_x= 0; mb_x<b_width; mb_x++){
1004                 int dia_change, i, j, ref;
1005                 int best_rd= INT_MAX, ref_rd;
1006                 BlockNode backup, ref_b;
1007                 const int index= mb_x + mb_y * b_stride;
1008                 BlockNode *block= &s->block[index];
1009                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1010                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1011                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1012                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1013                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1014                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1015                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1016                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1017                 const int b_w= (MB_SIZE >> s->block_max_depth);
1018                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1019
1020                 if(pass && (block->type & BLOCK_OPT))
1021                     continue;
1022                 block->type |= BLOCK_OPT;
1023
1024                 backup= *block;
1025
1026                 if(!s->me_cache_generation)
1027                     memset(s->me_cache, 0, sizeof(s->me_cache));
1028                 s->me_cache_generation += 1<<22;
1029
1030                 //FIXME precalculate
1031                 {
1032                     int x, y;
1033                     for (y = 0; y < b_w * 2; y++)
1034                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1035                     if(mb_x==0)
1036                         for(y=0; y<b_w*2; y++)
1037                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1038                     if(mb_x==b_stride-1)
1039                         for(y=0; y<b_w*2; y++)
1040                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1041                     if(mb_y==0){
1042                         for(x=0; x<b_w*2; x++)
1043                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1044                         for(y=1; y<b_w; y++)
1045                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1046                     }
1047                     if(mb_y==b_height-1){
1048                         for(x=0; x<b_w*2; x++)
1049                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1050                         for(y=b_w; y<b_w*2-1; y++)
1051                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1052                     }
1053                 }
1054
1055                 //skip stuff outside the picture
1056                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1057                     uint8_t *src= s->  input_picture->data[0];
1058                     uint8_t *dst= s->current_picture->data[0];
1059                     const int stride= s->current_picture->linesize[0];
1060                     const int block_w= MB_SIZE >> s->block_max_depth;
1061                     const int block_h= MB_SIZE >> s->block_max_depth;
1062                     const int sx= block_w*mb_x - block_w/2;
1063                     const int sy= block_h*mb_y - block_h/2;
1064                     const int w= s->plane[0].width;
1065                     const int h= s->plane[0].height;
1066                     int y;
1067
1068                     for(y=sy; y<0; y++)
1069                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1070                     for(y=h; y<sy+block_h*2; y++)
1071                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1072                     if(sx<0){
1073                         for(y=sy; y<sy+block_h*2; y++)
1074                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1075                     }
1076                     if(sx+block_w*2 > w){
1077                         for(y=sy; y<sy+block_h*2; y++)
1078                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1079                     }
1080                 }
1081
1082                 // intra(black) = neighbors' contribution to the current block
1083                 for(i=0; i < s->nb_planes; i++)
1084                     color[i]= get_dc(s, mb_x, mb_y, i);
1085
1086                 // get previous score (cannot be cached due to OBMC)
1087                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1088                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1089                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1090                 }else
1091                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1092
1093                 ref_b= *block;
1094                 ref_rd= best_rd;
1095                 for(ref=0; ref < s->ref_frames; ref++){
1096                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1097                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1098                         continue;
1099                     block->ref= ref;
1100                     best_rd= INT_MAX;
1101
1102                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1103                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1104                     if(tb)
1105                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1106                     if(lb)
1107                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1108                     if(rb)
1109                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1110                     if(bb)
1111                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1112
1113                     /* fullpel ME */
1114                     //FIXME avoid subpel interpolation / round to nearest integer
1115                     do{
1116                         dia_change=0;
1117                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1118                             for(j=0; j<i; j++){
1119                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1120                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1121                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1122                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1123                             }
1124                         }
1125                     }while(dia_change);
1126                     /* subpel ME */
1127                     do{
1128                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1129                         dia_change=0;
1130                         for(i=0; i<8; i++)
1131                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1132                     }while(dia_change);
1133                     //FIXME or try the standard 2 pass qpel or similar
1134
1135                     mvr[0][0]= block->mx;
1136                     mvr[0][1]= block->my;
1137                     if(ref_rd > best_rd){
1138                         ref_rd= best_rd;
1139                         ref_b= *block;
1140                     }
1141                 }
1142                 best_rd= ref_rd;
1143                 *block= ref_b;
1144                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1145                 //FIXME RD style color selection
1146                 if(!same_block(block, &backup)){
1147                     if(tb ) tb ->type &= ~BLOCK_OPT;
1148                     if(lb ) lb ->type &= ~BLOCK_OPT;
1149                     if(rb ) rb ->type &= ~BLOCK_OPT;
1150                     if(bb ) bb ->type &= ~BLOCK_OPT;
1151                     if(tlb) tlb->type &= ~BLOCK_OPT;
1152                     if(trb) trb->type &= ~BLOCK_OPT;
1153                     if(blb) blb->type &= ~BLOCK_OPT;
1154                     if(brb) brb->type &= ~BLOCK_OPT;
1155                     change ++;
1156                 }
1157             }
1158         }
1159         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1160         if(!change)
1161             break;
1162     }
1163
1164     if(s->block_max_depth == 1){
1165         int change= 0;
1166         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1167             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1168                 int i;
1169                 int best_rd, init_rd;
1170                 const int index= mb_x + mb_y * b_stride;
1171                 BlockNode *b[4];
1172
1173                 b[0]= &s->block[index];
1174                 b[1]= b[0]+1;
1175                 b[2]= b[0]+b_stride;
1176                 b[3]= b[2]+1;
1177                 if(same_block(b[0], b[1]) &&
1178                    same_block(b[0], b[2]) &&
1179                    same_block(b[0], b[3]))
1180                     continue;
1181
1182                 if(!s->me_cache_generation)
1183                     memset(s->me_cache, 0, sizeof(s->me_cache));
1184                 s->me_cache_generation += 1<<22;
1185
1186                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1187
1188                 //FIXME more multiref search?
1189                 check_4block_inter(s, mb_x, mb_y,
1190                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1191                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1192
1193                 for(i=0; i<4; i++)
1194                     if(!(b[i]->type&BLOCK_INTRA))
1195                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1196
1197                 if(init_rd != best_rd)
1198                     change++;
1199             }
1200         }
1201         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1202     }
1203 }
1204
1205 static void encode_blocks(SnowContext *s, int search){
1206     int x, y;
1207     int w= s->b_width;
1208     int h= s->b_height;
1209
1210     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1211         iterative_me(s);
1212
1213     for(y=0; y<h; y++){
1214         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1215             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1216             return;
1217         }
1218         for(x=0; x<w; x++){
1219             if(s->avctx->me_method == ME_ITER || !search)
1220                 encode_q_branch2(s, 0, x, y);
1221             else
1222                 encode_q_branch (s, 0, x, y);
1223         }
1224     }
1225 }
1226
1227 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1228     const int w= b->width;
1229     const int h= b->height;
1230     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1231     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1232     int x,y, thres1, thres2;
1233
1234     if(s->qlog == LOSSLESS_QLOG){
1235         for(y=0; y<h; y++)
1236             for(x=0; x<w; x++)
1237                 dst[x + y*stride]= src[x + y*stride];
1238         return;
1239     }
1240
1241     bias= bias ? 0 : (3*qmul)>>3;
1242     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1243     thres2= 2*thres1;
1244
1245     if(!bias){
1246         for(y=0; y<h; y++){
1247             for(x=0; x<w; x++){
1248                 int i= src[x + y*stride];
1249
1250                 if((unsigned)(i+thres1) > thres2){
1251                     if(i>=0){
1252                         i<<= QEXPSHIFT;
1253                         i/= qmul; //FIXME optimize
1254                         dst[x + y*stride]=  i;
1255                     }else{
1256                         i= -i;
1257                         i<<= QEXPSHIFT;
1258                         i/= qmul; //FIXME optimize
1259                         dst[x + y*stride]= -i;
1260                     }
1261                 }else
1262                     dst[x + y*stride]= 0;
1263             }
1264         }
1265     }else{
1266         for(y=0; y<h; y++){
1267             for(x=0; x<w; x++){
1268                 int i= src[x + y*stride];
1269
1270                 if((unsigned)(i+thres1) > thres2){
1271                     if(i>=0){
1272                         i<<= QEXPSHIFT;
1273                         i= (i + bias) / qmul; //FIXME optimize
1274                         dst[x + y*stride]=  i;
1275                     }else{
1276                         i= -i;
1277                         i<<= QEXPSHIFT;
1278                         i= (i + bias) / qmul; //FIXME optimize
1279                         dst[x + y*stride]= -i;
1280                     }
1281                 }else
1282                     dst[x + y*stride]= 0;
1283             }
1284         }
1285     }
1286 }
1287
1288 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1289     const int w= b->width;
1290     const int h= b->height;
1291     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1292     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1293     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1294     int x,y;
1295
1296     if(s->qlog == LOSSLESS_QLOG) return;
1297
1298     for(y=0; y<h; y++){
1299         for(x=0; x<w; x++){
1300             int i= src[x + y*stride];
1301             if(i<0){
1302                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1303             }else if(i>0){
1304                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1305             }
1306         }
1307     }
1308 }
1309
1310 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1311     const int w= b->width;
1312     const int h= b->height;
1313     int x,y;
1314
1315     for(y=h-1; y>=0; y--){
1316         for(x=w-1; x>=0; x--){
1317             int i= x + y*stride;
1318
1319             if(x){
1320                 if(use_median){
1321                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1322                     else  src[i] -= src[i - 1];
1323                 }else{
1324                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1325                     else  src[i] -= src[i - 1];
1326                 }
1327             }else{
1328                 if(y) src[i] -= src[i - stride];
1329             }
1330         }
1331     }
1332 }
1333
1334 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1335     const int w= b->width;
1336     const int h= b->height;
1337     int x,y;
1338
1339     for(y=0; y<h; y++){
1340         for(x=0; x<w; x++){
1341             int i= x + y*stride;
1342
1343             if(x){
1344                 if(use_median){
1345                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1346                     else  src[i] += src[i - 1];
1347                 }else{
1348                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1349                     else  src[i] += src[i - 1];
1350                 }
1351             }else{
1352                 if(y) src[i] += src[i - stride];
1353             }
1354         }
1355     }
1356 }
1357
1358 static void encode_qlogs(SnowContext *s){
1359     int plane_index, level, orientation;
1360
1361     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1362         for(level=0; level<s->spatial_decomposition_count; level++){
1363             for(orientation=level ? 1:0; orientation<4; orientation++){
1364                 if(orientation==2) continue;
1365                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1366             }
1367         }
1368     }
1369 }
1370
1371 static void encode_header(SnowContext *s){
1372     int plane_index, i;
1373     uint8_t kstate[32];
1374
1375     memset(kstate, MID_STATE, sizeof(kstate));
1376
1377     put_rac(&s->c, kstate, s->keyframe);
1378     if(s->keyframe || s->always_reset){
1379         ff_snow_reset_contexts(s);
1380         s->last_spatial_decomposition_type=
1381         s->last_qlog=
1382         s->last_qbias=
1383         s->last_mv_scale=
1384         s->last_block_max_depth= 0;
1385         for(plane_index=0; plane_index<2; plane_index++){
1386             Plane *p= &s->plane[plane_index];
1387             p->last_htaps=0;
1388             p->last_diag_mc=0;
1389             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1390         }
1391     }
1392     if(s->keyframe){
1393         put_symbol(&s->c, s->header_state, s->version, 0);
1394         put_rac(&s->c, s->header_state, s->always_reset);
1395         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1396         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1397         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1398         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1399         if (s->nb_planes > 2) {
1400             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1401             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1402         }
1403         put_rac(&s->c, s->header_state, s->spatial_scalability);
1404 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1405         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1406
1407         encode_qlogs(s);
1408     }
1409
1410     if(!s->keyframe){
1411         int update_mc=0;
1412         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1413             Plane *p= &s->plane[plane_index];
1414             update_mc |= p->last_htaps   != p->htaps;
1415             update_mc |= p->last_diag_mc != p->diag_mc;
1416             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1417         }
1418         put_rac(&s->c, s->header_state, update_mc);
1419         if(update_mc){
1420             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1421                 Plane *p= &s->plane[plane_index];
1422                 put_rac(&s->c, s->header_state, p->diag_mc);
1423                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1424                 for(i= p->htaps/2; i; i--)
1425                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1426             }
1427         }
1428         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1429             put_rac(&s->c, s->header_state, 1);
1430             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1431             encode_qlogs(s);
1432         }else
1433             put_rac(&s->c, s->header_state, 0);
1434     }
1435
1436     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1437     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1438     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1439     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1440     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1441
1442 }
1443
1444 static void update_last_header_values(SnowContext *s){
1445     int plane_index;
1446
1447     if(!s->keyframe){
1448         for(plane_index=0; plane_index<2; plane_index++){
1449             Plane *p= &s->plane[plane_index];
1450             p->last_diag_mc= p->diag_mc;
1451             p->last_htaps  = p->htaps;
1452             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1453         }
1454     }
1455
1456     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1457     s->last_qlog                        = s->qlog;
1458     s->last_qbias                       = s->qbias;
1459     s->last_mv_scale                    = s->mv_scale;
1460     s->last_block_max_depth             = s->block_max_depth;
1461     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1462 }
1463
1464 static int qscale2qlog(int qscale){
1465     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1466            + 61*QROOT/8; ///< 64 > 60
1467 }
1468
1469 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1470 {
1471     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1472      * FIXME we know exact mv bits at this point,
1473      * but ratecontrol isn't set up to include them. */
1474     uint32_t coef_sum= 0;
1475     int level, orientation, delta_qlog;
1476
1477     for(level=0; level<s->spatial_decomposition_count; level++){
1478         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1479             SubBand *b= &s->plane[0].band[level][orientation];
1480             IDWTELEM *buf= b->ibuf;
1481             const int w= b->width;
1482             const int h= b->height;
1483             const int stride= b->stride;
1484             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1485             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1486             const int qdiv= (1<<16)/qmul;
1487             int x, y;
1488             //FIXME this is ugly
1489             for(y=0; y<h; y++)
1490                 for(x=0; x<w; x++)
1491                     buf[x+y*stride]= b->buf[x+y*stride];
1492             if(orientation==0)
1493                 decorrelate(s, b, buf, stride, 1, 0);
1494             for(y=0; y<h; y++)
1495                 for(x=0; x<w; x++)
1496                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1497         }
1498     }
1499
1500     /* ugly, ratecontrol just takes a sqrt again */
1501     av_assert0(coef_sum < INT_MAX);
1502     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1503
1504     if(pict->pict_type == AV_PICTURE_TYPE_I){
1505         s->m.current_picture.mb_var_sum= coef_sum;
1506         s->m.current_picture.mc_mb_var_sum= 0;
1507     }else{
1508         s->m.current_picture.mc_mb_var_sum= coef_sum;
1509         s->m.current_picture.mb_var_sum= 0;
1510     }
1511
1512     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1513     if (pict->quality < 0)
1514         return INT_MIN;
1515     s->lambda= pict->quality * 3/2;
1516     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1517     s->qlog+= delta_qlog;
1518     return delta_qlog;
1519 }
1520
1521 static void calculate_visual_weight(SnowContext *s, Plane *p){
1522     int width = p->width;
1523     int height= p->height;
1524     int level, orientation, x, y;
1525
1526     for(level=0; level<s->spatial_decomposition_count; level++){
1527         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1528             SubBand *b= &p->band[level][orientation];
1529             IDWTELEM *ibuf= b->ibuf;
1530             int64_t error=0;
1531
1532             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1533             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1534             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1535             for(y=0; y<height; y++){
1536                 for(x=0; x<width; x++){
1537                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1538                     error += d*d;
1539                 }
1540             }
1541
1542             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1543         }
1544     }
1545 }
1546
1547 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1548                         const AVFrame *pict, int *got_packet)
1549 {
1550     SnowContext *s = avctx->priv_data;
1551     RangeCoder * const c= &s->c;
1552     AVFrame *pic = pict;
1553     const int width= s->avctx->width;
1554     const int height= s->avctx->height;
1555     int level, orientation, plane_index, i, y, ret;
1556     uint8_t rc_header_bak[sizeof(s->header_state)];
1557     uint8_t rc_block_bak[sizeof(s->block_state)];
1558
1559     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1560         return ret;
1561
1562     ff_init_range_encoder(c, pkt->data, pkt->size);
1563     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1564
1565     for(i=0; i < s->nb_planes; i++){
1566         int hshift= i ? s->chroma_h_shift : 0;
1567         int vshift= i ? s->chroma_v_shift : 0;
1568         for(y=0; y<(height>>vshift); y++)
1569             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1570                    &pict->data[i][y * pict->linesize[i]],
1571                    width>>hshift);
1572         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1573                                 width >> hshift, height >> vshift,
1574                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1575                                 EDGE_TOP | EDGE_BOTTOM);
1576
1577     }
1578     emms_c();
1579     s->new_picture = pict;
1580
1581     s->m.picture_number= avctx->frame_number;
1582     if(avctx->flags&CODEC_FLAG_PASS2){
1583         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1584         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1585         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1586             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1587             if (pic->quality < 0)
1588                 return -1;
1589         }
1590     }else{
1591         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1592         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1593     }
1594
1595     if(s->pass1_rc && avctx->frame_number == 0)
1596         pic->quality = 2*FF_QP2LAMBDA;
1597     if (pic->quality) {
1598         s->qlog   = qscale2qlog(pic->quality);
1599         s->lambda = pic->quality * 3/2;
1600     }
1601     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1602         s->qlog= LOSSLESS_QLOG;
1603         s->lambda = 0;
1604     }//else keep previous frame's qlog until after motion estimation
1605
1606     if (s->current_picture->data[0] && !(s->avctx->flags&CODEC_FLAG_EMU_EDGE)) {
1607         int w = s->avctx->width;
1608         int h = s->avctx->height;
1609
1610         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1611                                 s->current_picture->linesize[0], w   , h   ,
1612                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1613         if (s->current_picture->data[2]) {
1614             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1615                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1616                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1617             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1618                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1619                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1620         }
1621     }
1622
1623     ff_snow_frame_start(s);
1624     av_frame_unref(avctx->coded_frame);
1625     av_frame_ref(avctx->coded_frame, s->current_picture);
1626
1627     s->m.current_picture_ptr= &s->m.current_picture;
1628     s->m.current_picture.f = s->current_picture;
1629     s->m.current_picture.f->pts = pict->pts;
1630     if(pic->pict_type == AV_PICTURE_TYPE_P){
1631         int block_width = (width +15)>>4;
1632         int block_height= (height+15)>>4;
1633         int stride= s->current_picture->linesize[0];
1634
1635         av_assert0(s->current_picture->data[0]);
1636         av_assert0(s->last_picture[0]->data[0]);
1637
1638         s->m.avctx= s->avctx;
1639         s->m.   last_picture.f = s->last_picture[0];
1640         s->m.    new_picture.f = s->input_picture;
1641         s->m.   last_picture_ptr= &s->m.   last_picture;
1642         s->m.linesize = stride;
1643         s->m.uvlinesize= s->current_picture->linesize[1];
1644         s->m.width = width;
1645         s->m.height= height;
1646         s->m.mb_width = block_width;
1647         s->m.mb_height= block_height;
1648         s->m.mb_stride=   s->m.mb_width+1;
1649         s->m.b8_stride= 2*s->m.mb_width+1;
1650         s->m.f_code=1;
1651         s->m.pict_type = pic->pict_type;
1652         s->m.me_method= s->avctx->me_method;
1653         s->m.me.scene_change_score=0;
1654         s->m.flags= s->avctx->flags;
1655         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1656         s->m.out_format= FMT_H263;
1657         s->m.unrestricted_mv= 1;
1658
1659         s->m.lambda = s->lambda;
1660         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1661         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1662
1663         s->m.mecc= s->mecc; //move
1664         s->m.qdsp= s->qdsp; //move
1665         s->m.hdsp = s->hdsp;
1666         ff_init_me(&s->m);
1667         s->hdsp = s->m.hdsp;
1668         s->mecc= s->m.mecc;
1669     }
1670
1671     if(s->pass1_rc){
1672         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1673         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1674     }
1675
1676 redo_frame:
1677
1678     s->spatial_decomposition_count= 5;
1679
1680     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1681           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1682         s->spatial_decomposition_count--;
1683
1684     if (s->spatial_decomposition_count <= 0) {
1685         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1686         return AVERROR(EINVAL);
1687     }
1688
1689     s->m.pict_type = pic->pict_type;
1690     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1691
1692     ff_snow_common_init_after_header(avctx);
1693
1694     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1695         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1696             calculate_visual_weight(s, &s->plane[plane_index]);
1697         }
1698     }
1699
1700     encode_header(s);
1701     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1702     encode_blocks(s, 1);
1703     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1704
1705     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1706         Plane *p= &s->plane[plane_index];
1707         int w= p->width;
1708         int h= p->height;
1709         int x, y;
1710 //        int bits= put_bits_count(&s->c.pb);
1711
1712         if (!s->memc_only) {
1713             //FIXME optimize
1714             if(pict->data[plane_index]) //FIXME gray hack
1715                 for(y=0; y<h; y++){
1716                     for(x=0; x<w; x++){
1717                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1718                     }
1719                 }
1720             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1721
1722             if(   plane_index==0
1723                && pic->pict_type == AV_PICTURE_TYPE_P
1724                && !(avctx->flags&CODEC_FLAG_PASS2)
1725                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1726                 ff_init_range_encoder(c, pkt->data, pkt->size);
1727                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1728                 pic->pict_type= AV_PICTURE_TYPE_I;
1729                 s->keyframe=1;
1730                 s->current_picture->key_frame=1;
1731                 goto redo_frame;
1732             }
1733
1734             if(s->qlog == LOSSLESS_QLOG){
1735                 for(y=0; y<h; y++){
1736                     for(x=0; x<w; x++){
1737                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1738                     }
1739                 }
1740             }else{
1741                 for(y=0; y<h; y++){
1742                     for(x=0; x<w; x++){
1743                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1744                     }
1745                 }
1746             }
1747
1748             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1749
1750             if(s->pass1_rc && plane_index==0){
1751                 int delta_qlog = ratecontrol_1pass(s, pic);
1752                 if (delta_qlog <= INT_MIN)
1753                     return -1;
1754                 if(delta_qlog){
1755                     //reordering qlog in the bitstream would eliminate this reset
1756                     ff_init_range_encoder(c, pkt->data, pkt->size);
1757                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1758                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1759                     encode_header(s);
1760                     encode_blocks(s, 0);
1761                 }
1762             }
1763
1764             for(level=0; level<s->spatial_decomposition_count; level++){
1765                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1766                     SubBand *b= &p->band[level][orientation];
1767
1768                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1769                     if(orientation==0)
1770                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1771                     if (!s->no_bitstream)
1772                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1773                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1774                     if(orientation==0)
1775                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1776                 }
1777             }
1778
1779             for(level=0; level<s->spatial_decomposition_count; level++){
1780                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1781                     SubBand *b= &p->band[level][orientation];
1782
1783                     dequantize(s, b, b->ibuf, b->stride);
1784                 }
1785             }
1786
1787             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1788             if(s->qlog == LOSSLESS_QLOG){
1789                 for(y=0; y<h; y++){
1790                     for(x=0; x<w; x++){
1791                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1792                     }
1793                 }
1794             }
1795             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1796         }else{
1797             //ME/MC only
1798             if(pic->pict_type == AV_PICTURE_TYPE_I){
1799                 for(y=0; y<h; y++){
1800                     for(x=0; x<w; x++){
1801                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1802                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1803                     }
1804                 }
1805             }else{
1806                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1807                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1808             }
1809         }
1810         if(s->avctx->flags&CODEC_FLAG_PSNR){
1811             int64_t error= 0;
1812
1813             if(pict->data[plane_index]) //FIXME gray hack
1814                 for(y=0; y<h; y++){
1815                     for(x=0; x<w; x++){
1816                         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];
1817                         error += d*d;
1818                     }
1819                 }
1820             s->avctx->error[plane_index] += error;
1821             s->current_picture->error[plane_index] = error;
1822         }
1823
1824     }
1825
1826     update_last_header_values(s);
1827
1828     ff_snow_release_buffer(avctx);
1829
1830     s->current_picture->coded_picture_number = avctx->frame_number;
1831     s->current_picture->pict_type = pict->pict_type;
1832     s->current_picture->quality = pict->quality;
1833     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1834     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1835     s->m.current_picture.f->display_picture_number =
1836     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1837     s->m.current_picture.f->quality                = pic->quality;
1838     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1839     if(s->pass1_rc)
1840         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1841             return -1;
1842     if(avctx->flags&CODEC_FLAG_PASS1)
1843         ff_write_pass1_stats(&s->m);
1844     s->m.last_pict_type = s->m.pict_type;
1845     avctx->frame_bits = s->m.frame_bits;
1846     avctx->mv_bits = s->m.mv_bits;
1847     avctx->misc_bits = s->m.misc_bits;
1848     avctx->p_tex_bits = s->m.p_tex_bits;
1849
1850     emms_c();
1851
1852     pkt->size = ff_rac_terminate(c);
1853     if (avctx->coded_frame->key_frame)
1854         pkt->flags |= AV_PKT_FLAG_KEY;
1855     *got_packet = 1;
1856
1857     return 0;
1858 }
1859
1860 static av_cold int encode_end(AVCodecContext *avctx)
1861 {
1862     SnowContext *s = avctx->priv_data;
1863
1864     ff_snow_common_end(s);
1865     ff_rate_control_uninit(&s->m);
1866     av_frame_free(&s->input_picture);
1867     av_frame_free(&avctx->coded_frame);
1868     av_freep(&avctx->stats_out);
1869
1870     return 0;
1871 }
1872
1873 #define OFFSET(x) offsetof(SnowContext, x)
1874 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1875 static const AVOption options[] = {
1876     FF_MPV_COMMON_OPTS
1877     { "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 },
1878     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1879     { NULL },
1880 };
1881
1882 static const AVClass snowenc_class = {
1883     .class_name = "snow encoder",
1884     .item_name  = av_default_item_name,
1885     .option     = options,
1886     .version    = LIBAVUTIL_VERSION_INT,
1887 };
1888
1889 AVCodec ff_snow_encoder = {
1890     .name           = "snow",
1891     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1892     .type           = AVMEDIA_TYPE_VIDEO,
1893     .id             = AV_CODEC_ID_SNOW,
1894     .priv_data_size = sizeof(SnowContext),
1895     .init           = encode_init,
1896     .encode2        = encode_frame,
1897     .close          = encode_end,
1898     .pix_fmts       = (const enum AVPixelFormat[]){
1899         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1900         AV_PIX_FMT_GRAY8,
1901         AV_PIX_FMT_NONE
1902     },
1903     .priv_class     = &snowenc_class,
1904 };
1905
1906
1907 #ifdef TEST
1908 #undef malloc
1909 #undef free
1910 #undef printf
1911
1912 #include "libavutil/lfg.h"
1913 #include "libavutil/mathematics.h"
1914
1915 int main(void){
1916 #define width  256
1917 #define height 256
1918     int buffer[2][width*height];
1919     SnowContext s;
1920     int i;
1921     AVLFG prng;
1922     s.spatial_decomposition_count=6;
1923     s.spatial_decomposition_type=1;
1924
1925     s.temp_dwt_buffer  = av_mallocz_array(width, sizeof(DWTELEM));
1926     s.temp_idwt_buffer = av_mallocz_array(width, sizeof(IDWTELEM));
1927
1928     if (!s.temp_dwt_buffer || !s.temp_idwt_buffer) {
1929         fprintf(stderr, "Failed to allocate memory\n");
1930         return 1;
1931     }
1932
1933     av_lfg_init(&prng, 1);
1934
1935     printf("testing 5/3 DWT\n");
1936     for(i=0; i<width*height; i++)
1937         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1938
1939     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1940     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1941
1942     for(i=0; i<width*height; i++)
1943         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1944
1945     printf("testing 9/7 DWT\n");
1946     s.spatial_decomposition_type=0;
1947     for(i=0; i<width*height; i++)
1948         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1949
1950     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1951     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1952
1953     for(i=0; i<width*height; i++)
1954         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1955
1956     {
1957     int level, orientation, x, y;
1958     int64_t errors[8][4];
1959     int64_t g=0;
1960
1961         memset(errors, 0, sizeof(errors));
1962         s.spatial_decomposition_count=3;
1963         s.spatial_decomposition_type=0;
1964         for(level=0; level<s.spatial_decomposition_count; level++){
1965             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1966                 int w= width  >> (s.spatial_decomposition_count-level);
1967                 int h= height >> (s.spatial_decomposition_count-level);
1968                 int stride= width  << (s.spatial_decomposition_count-level);
1969                 DWTELEM *buf= buffer[0];
1970                 int64_t error=0;
1971
1972                 if(orientation&1) buf+=w;
1973                 if(orientation>1) buf+=stride>>1;
1974
1975                 memset(buffer[0], 0, sizeof(int)*width*height);
1976                 buf[w/2 + h/2*stride]= 256*256;
1977                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1978                 for(y=0; y<height; y++){
1979                     for(x=0; x<width; x++){
1980                         int64_t d= buffer[0][x + y*width];
1981                         error += d*d;
1982                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1983                     }
1984                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1985                 }
1986                 error= (int)(sqrt(error)+0.5);
1987                 errors[level][orientation]= error;
1988                 if(g) g=av_gcd(g, error);
1989                 else g= error;
1990             }
1991         }
1992         printf("static int const visual_weight[][4]={\n");
1993         for(level=0; level<s.spatial_decomposition_count; level++){
1994             printf("  {");
1995             for(orientation=0; orientation<4; orientation++){
1996                 printf("%8"PRId64",", errors[level][orientation]/g);
1997             }
1998             printf("},\n");
1999         }
2000         printf("};\n");
2001         {
2002             int level=2;
2003             int w= width  >> (s.spatial_decomposition_count-level);
2004             //int h= height >> (s.spatial_decomposition_count-level);
2005             int stride= width  << (s.spatial_decomposition_count-level);
2006             DWTELEM *buf= buffer[0];
2007             int64_t error=0;
2008
2009             buf+=w;
2010             buf+=stride>>1;
2011
2012             memset(buffer[0], 0, sizeof(int)*width*height);
2013             for(y=0; y<height; y++){
2014                 for(x=0; x<width; x++){
2015                     int tab[4]={0,2,3,1};
2016                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2017                 }
2018             }
2019             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2020             for(y=0; y<height; y++){
2021                 for(x=0; x<width; x++){
2022                     int64_t d= buffer[0][x + y*width];
2023                     error += d*d;
2024                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2025                 }
2026                 if(FFABS(height/2-y)<9) printf("\n");
2027             }
2028         }
2029
2030     }
2031     return 0;
2032 }
2033 #endif /* TEST */