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 "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         return ret;
  64     }
  65     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  66
  67     ff_snow_alloc_blocks(s);
  68
  69     s->version=0;
  70
  71     s->m.avctx   = avctx;
  72     s->m.bit_rate= avctx->bit_rate;
  73
  74     s->m.me.temp      =
  75     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  76     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  77     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  78     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  79     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
  80         return AVERROR(ENOMEM);
  81
  82     ff_h263_encode_init(&s->m); //mv_penalty
  83
  84     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
  85
  86     if(avctx->flags&CODEC_FLAG_PASS1){
  87         if(!avctx->stats_out)
  88             avctx->stats_out = av_mallocz(256);
  89
  90         if (!avctx->stats_out)
  91             return AVERROR(ENOMEM);
  92     }
  93     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
  94         if(ff_rate_control_init(&s->m) < 0)
  95             return -1;
  96     }
  97     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
  98
  99     switch(avctx->pix_fmt){
 100     case AV_PIX_FMT_YUV444P:
 101 //    case AV_PIX_FMT_YUV422P:
 102     case AV_PIX_FMT_YUV420P:
 103 //    case AV_PIX_FMT_YUV411P:
 104     case AV_PIX_FMT_YUV410P:
 105         s->nb_planes = 3;
 106         s->colorspace_type= 0;
 107         break;
 108     case AV_PIX_FMT_GRAY8:
 109         s->nb_planes = 1;
 110         s->colorspace_type = 1;
 111         break;
 112 /*    case AV_PIX_FMT_RGB32:
 113         s->colorspace= 1;
 114         break;*/
 115     default:
 116         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 117         return -1;
 118     }
 119     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 120
 121     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 122     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 123
 124     s->input_picture = av_frame_alloc();
 125     avctx->coded_frame = av_frame_alloc();
 126     if (!s->input_picture || !avctx->coded_frame)
 127         return AVERROR(ENOMEM);
 128
 129     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 130         return ret;
 131
 132     if(s->avctx->me_method == ME_ITER){
 133         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 134         for(i=0; i<s->max_ref_frames; i++){
 135             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 136             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 137             if (!s->ref_mvs[i] || !s->ref_scores[i])
 138                 return AVERROR(ENOMEM);
 139         }
 140     }
 141
 142     return 0;
 143 }
 144
 145 //near copy & paste from dsputil, FIXME
 146 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 147 {
 148     int s, i, j;
 149
 150     s = 0;
 151     for (i = 0; i < h; i++) {
 152         for (j = 0; j < w; j++) {
 153             s += pix[0];
 154             pix ++;
 155         }
 156         pix += line_size - w;
 157     }
 158     return s;
 159 }
 160
 161 //near copy & paste from dsputil, FIXME
 162 static int pix_norm1(uint8_t * pix, int line_size, int w)
 163 {
 164     int s, i, j;
 165     uint32_t *sq = ff_square_tab + 256;
 166
 167     s = 0;
 168     for (i = 0; i < w; i++) {
 169         for (j = 0; j < w; j ++) {
 170             s += sq[pix[0]];
 171             pix ++;
 172         }
 173         pix += line_size - w;
 174     }
 175     return s;
 176 }
 177
 178 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 179     switch(type&0xFF){
 180     default:
 181     case FF_CMP_SAD:
 182         return lambda>>FF_LAMBDA_SHIFT;
 183     case FF_CMP_DCT:
 184         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 185     case FF_CMP_W53:
 186         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 187     case FF_CMP_W97:
 188         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 189     case FF_CMP_SATD:
 190     case FF_CMP_DCT264:
 191         return (2*lambda)>>FF_LAMBDA_SHIFT;
 192     case FF_CMP_RD:
 193     case FF_CMP_PSNR:
 194     case FF_CMP_SSE:
 195     case FF_CMP_NSSE:
 196         return lambda2>>FF_LAMBDA_SHIFT;
 197     case FF_CMP_BIT:
 198         return 1;
 199     }
 200 }
 201
 202 //FIXME copy&paste
 203 #define P_LEFT P[1]
 204 #define P_TOP P[2]
 205 #define P_TOPRIGHT P[3]
 206 #define P_MEDIAN P[4]
 207 #define P_MV1 P[9]
 208 #define FLAG_QPEL   1 //must be 1
 209
 210 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 211     uint8_t p_buffer[1024];
 212     uint8_t i_buffer[1024];
 213     uint8_t p_state[sizeof(s->block_state)];
 214     uint8_t i_state[sizeof(s->block_state)];
 215     RangeCoder pc, ic;
 216     uint8_t *pbbak= s->c.bytestream;
 217     uint8_t *pbbak_start= s->c.bytestream_start;
 218     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 219     const int w= s->b_width  << s->block_max_depth;
 220     const int h= s->b_height << s->block_max_depth;
 221     const int rem_depth= s->block_max_depth - level;
 222     const int index= (x + y*w) << rem_depth;
 223     const int block_w= 1<<(LOG2_MB_SIZE - level);
 224     int trx= (x+1)<<rem_depth;
 225     int try= (y+1)<<rem_depth;
 226     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 227     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 228     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 229     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 230     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 231     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 232     int pl = left->color[0];
 233     int pcb= left->color[1];
 234     int pcr= left->color[2];
 235     int pmx, pmy;
 236     int mx=0, my=0;
 237     int l,cr,cb;
 238     const int stride= s->current_picture->linesize[0];
 239     const int uvstride= s->current_picture->linesize[1];
 240     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 241                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 242                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 243     int P[10][2];
 244     int16_t last_mv[3][2];
 245     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
 246     const int shift= 1+qpel;
 247     MotionEstContext *c= &s->m.me;
 248     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 249     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 250     int my_context= av_log2(2*FFABS(left->my - top->my));
 251     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 252     int ref, best_ref, ref_score, ref_mx, ref_my;
 253
 254     av_assert0(sizeof(s->block_state) >= 256);
 255     if(s->keyframe){
 256         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 257         return 0;
 258     }
 259
 260 //    clip predictors / edge ?
 261
 262     P_LEFT[0]= left->mx;
 263     P_LEFT[1]= left->my;
 264     P_TOP [0]= top->mx;
 265     P_TOP [1]= top->my;
 266     P_TOPRIGHT[0]= tr->mx;
 267     P_TOPRIGHT[1]= tr->my;
 268
 269     last_mv[0][0]= s->block[index].mx;
 270     last_mv[0][1]= s->block[index].my;
 271     last_mv[1][0]= right->mx;
 272     last_mv[1][1]= right->my;
 273     last_mv[2][0]= bottom->mx;
 274     last_mv[2][1]= bottom->my;
 275
 276     s->m.mb_stride=2;
 277     s->m.mb_x=
 278     s->m.mb_y= 0;
 279     c->skip= 0;
 280
 281     av_assert1(c->  stride ==   stride);
 282     av_assert1(c->uvstride == uvstride);
 283
 284     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 285     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 286     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 287     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 288
 289     c->xmin = - x*block_w - 16+3;
 290     c->ymin = - y*block_w - 16+3;
 291     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 292     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 293
 294     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 295     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 296     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 297     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 298     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 299     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 300     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 301
 302     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 303     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 304
 305     if (!y) {
 306         c->pred_x= P_LEFT[0];
 307         c->pred_y= P_LEFT[1];
 308     } else {
 309         c->pred_x = P_MEDIAN[0];
 310         c->pred_y = P_MEDIAN[1];
 311     }
 312
 313     score= INT_MAX;
 314     best_ref= 0;
 315     for(ref=0; ref<s->ref_frames; ref++){
 316         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 317
 318         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 319                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 320
 321         av_assert2(ref_mx >= c->xmin);
 322         av_assert2(ref_mx <= c->xmax);
 323         av_assert2(ref_my >= c->ymin);
 324         av_assert2(ref_my <= c->ymax);
 325
 326         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 327         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 328         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 329         if(s->ref_mvs[ref]){
 330             s->ref_mvs[ref][index][0]= ref_mx;
 331             s->ref_mvs[ref][index][1]= ref_my;
 332             s->ref_scores[ref][index]= ref_score;
 333         }
 334         if(score > ref_score){
 335             score= ref_score;
 336             best_ref= ref;
 337             mx= ref_mx;
 338             my= ref_my;
 339         }
 340     }
 341     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 342
 343   //  subpel search
 344     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 345     pc= s->c;
 346     pc.bytestream_start=
 347     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 348     memcpy(p_state, s->block_state, sizeof(s->block_state));
 349
 350     if(level!=s->block_max_depth)
 351         put_rac(&pc, &p_state[4 + s_context], 1);
 352     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 353     if(s->ref_frames > 1)
 354         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 355     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 356     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 357     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 358     p_len= pc.bytestream - pc.bytestream_start;
 359     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 360
 361     block_s= block_w*block_w;
 362     sum = pix_sum(current_data[0], stride, block_w, block_w);
 363     l= (sum + block_s/2)/block_s;
 364     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 365
 366     if (s->nb_planes > 2) {
 367         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 368         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 369         cb= (sum + block_s/2)/block_s;
 370     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 371         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 372         cr= (sum + block_s/2)/block_s;
 373     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 374     }else
 375         cb = cr = 0;
 376
 377     ic= s->c;
 378     ic.bytestream_start=
 379     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 380     memcpy(i_state, s->block_state, sizeof(s->block_state));
 381     if(level!=s->block_max_depth)
 382         put_rac(&ic, &i_state[4 + s_context], 1);
 383     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 384     put_symbol(&ic, &i_state[32],  l-pl , 1);
 385     if (s->nb_planes > 2) {
 386         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 387         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 388     }
 389     i_len= ic.bytestream - ic.bytestream_start;
 390     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 391
 392     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 393     av_assert1(iscore >= 0);
 394     av_assert1(l>=0 && l<=255);
 395     av_assert1(pl>=0 && pl<=255);
 396
 397     if(level==0){
 398         int varc= iscore >> 8;
 399         int vard= score >> 8;
 400         if (vard <= 64 || vard < varc)
 401             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 402         else
 403             c->scene_change_score+= s->m.qscale;
 404     }
 405
 406     if(level!=s->block_max_depth){
 407         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 408         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 409         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 410         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 411         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 412         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 413
 414         if(score2 < score && score2 < iscore)
 415             return score2;
 416     }
 417
 418     if(iscore < score){
 419         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 420         memcpy(pbbak, i_buffer, i_len);
 421         s->c= ic;
 422         s->c.bytestream_start= pbbak_start;
 423         s->c.bytestream= pbbak + i_len;
 424         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 425         memcpy(s->block_state, i_state, sizeof(s->block_state));
 426         return iscore;
 427     }else{
 428         memcpy(pbbak, p_buffer, p_len);
 429         s->c= pc;
 430         s->c.bytestream_start= pbbak_start;
 431         s->c.bytestream= pbbak + p_len;
 432         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 433         memcpy(s->block_state, p_state, sizeof(s->block_state));
 434         return score;
 435     }
 436 }
 437
 438 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 439     const int w= s->b_width  << s->block_max_depth;
 440     const int rem_depth= s->block_max_depth - level;
 441     const int index= (x + y*w) << rem_depth;
 442     int trx= (x+1)<<rem_depth;
 443     BlockNode *b= &s->block[index];
 444     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 445     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 446     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 447     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 448     int pl = left->color[0];
 449     int pcb= left->color[1];
 450     int pcr= left->color[2];
 451     int pmx, pmy;
 452     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 453     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 454     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 455     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 456
 457     if(s->keyframe){
 458         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 459         return;
 460     }
 461
 462     if(level!=s->block_max_depth){
 463         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 464             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 465         }else{
 466             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 467             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 468             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 469             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 470             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 471             return;
 472         }
 473     }
 474     if(b->type & BLOCK_INTRA){
 475         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 476         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 477         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 478         if (s->nb_planes > 2) {
 479             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 480             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 481         }
 482         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 483     }else{
 484         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 485         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 486         if(s->ref_frames > 1)
 487             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 488         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 489         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 490         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 491     }
 492 }
 493
 494 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 495     int i, x2, y2;
 496     Plane *p= &s->plane[plane_index];
 497     const int block_size = MB_SIZE >> s->block_max_depth;
 498     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 499     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 500     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 501     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 502     const int ref_stride= s->current_picture->linesize[plane_index];
 503     uint8_t *src= s-> input_picture->data[plane_index];
 504     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 505     const int b_stride = s->b_width << s->block_max_depth;
 506     const int w= p->width;
 507     const int h= p->height;
 508     int index= mb_x + mb_y*b_stride;
 509     BlockNode *b= &s->block[index];
 510     BlockNode backup= *b;
 511     int ab=0;
 512     int aa=0;
 513
 514     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 515
 516     b->type|= BLOCK_INTRA;
 517     b->color[plane_index]= 0;
 518     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 519
 520     for(i=0; i<4; i++){
 521         int mb_x2= mb_x + (i &1) - 1;
 522         int mb_y2= mb_y + (i>>1) - 1;
 523         int x= block_w*mb_x2 + block_w/2;
 524         int y= block_h*mb_y2 + block_h/2;
 525
 526         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 527                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 528
 529         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 530             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 531                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 532                 int obmc_v= obmc[index];
 533                 int d;
 534                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 535                 if(x<0) obmc_v += obmc[index + block_w];
 536                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 537                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 538                 //FIXME precalculate this or simplify it somehow else
 539
 540                 d = -dst[index] + (1<<(FRAC_BITS-1));
 541                 dst[index] = d;
 542                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 543                 aa += obmc_v * obmc_v; //FIXME precalculate this
 544             }
 545         }
 546     }
 547     *b= backup;
 548
 549     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 550 }
 551
 552 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 553     const int b_stride = s->b_width << s->block_max_depth;
 554     const int b_height = s->b_height<< s->block_max_depth;
 555     int index= x + y*b_stride;
 556     const BlockNode *b     = &s->block[index];
 557     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 558     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 559     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 560     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 561     int dmx, dmy;
 562 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 563 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 564
 565     if(x<0 || x>=b_stride || y>=b_height)
 566         return 0;
 567 /*
 568 1            0      0
 569 01X          1-2    1
 570 001XX        3-6    2-3
 571 0001XXX      7-14   4-7
 572 00001XXXX   15-30   8-15
 573 */
 574 //FIXME try accurate rate
 575 //FIXME intra and inter predictors if surrounding blocks are not the same type
 576     if(b->type & BLOCK_INTRA){
 577         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 578                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 579                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 580     }else{
 581         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 582         dmx-= b->mx;
 583         dmy-= b->my;
 584         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 585                     + av_log2(2*FFABS(dmy))
 586                     + av_log2(2*b->ref));
 587     }
 588 }
 589
 590 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 591     Plane *p= &s->plane[plane_index];
 592     const int block_size = MB_SIZE >> s->block_max_depth;
 593     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 594     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 595     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 596     const int ref_stride= s->current_picture->linesize[plane_index];
 597     uint8_t *dst= s->current_picture->data[plane_index];
 598     uint8_t *src= s->  input_picture->data[plane_index];
 599     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
 600     uint8_t *cur = s->scratchbuf;
 601     uint8_t *tmp = s->emu_edge_buffer;
 602     const int b_stride = s->b_width << s->block_max_depth;
 603     const int b_height = s->b_height<< s->block_max_depth;
 604     const int w= p->width;
 605     const int h= p->height;
 606     int distortion;
 607     int rate= 0;
 608     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 609     int sx= block_w*mb_x - block_w/2;
 610     int sy= block_h*mb_y - block_h/2;
 611     int x0= FFMAX(0,-sx);
 612     int y0= FFMAX(0,-sy);
 613     int x1= FFMIN(block_w*2, w-sx);
 614     int y1= FFMIN(block_h*2, h-sy);
 615     int i,x,y;
 616
 617     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 618
 619     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);
 620
 621     for(y=y0; y<y1; y++){
 622         const uint8_t *obmc1= obmc_edged[y];
 623         const IDWTELEM *pred1 = pred + y*obmc_stride;
 624         uint8_t *cur1 = cur + y*ref_stride;
 625         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 626         for(x=x0; x<x1; x++){
 627 #if FRAC_BITS >= LOG2_OBMC_MAX
 628             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 629 #else
 630             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 631 #endif
 632             v = (v + pred1[x]) >> FRAC_BITS;
 633             if(v&(~255)) v= ~(v>>31);
 634             dst1[x] = v;
 635         }
 636     }
 637
 638     /* copy the regions where obmc[] = (uint8_t)256 */
 639     if(LOG2_OBMC_MAX == 8
 640         && (mb_x == 0 || mb_x == b_stride-1)
 641         && (mb_y == 0 || mb_y == b_height-1)){
 642         if(mb_x == 0)
 643             x1 = block_w;
 644         else
 645             x0 = block_w;
 646         if(mb_y == 0)
 647             y1 = block_h;
 648         else
 649             y0 = block_h;
 650         for(y=y0; y<y1; y++)
 651             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 652     }
 653
 654     if(block_w==16){
 655         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 656         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 657         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 658          * So improving the score of one block is not strictly guaranteed
 659          * to improve the score of the whole frame, thus iterative motion
 660          * estimation does not always converge. */
 661         if(s->avctx->me_cmp == FF_CMP_W97)
 662             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 663         else if(s->avctx->me_cmp == FF_CMP_W53)
 664             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 665         else{
 666             distortion = 0;
 667             for(i=0; i<4; i++){
 668                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 669                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 670             }
 671         }
 672     }else{
 673         av_assert2(block_w==8);
 674         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 675     }
 676
 677     if(plane_index==0){
 678         for(i=0; i<4; i++){
 679 /* ..RRr
 680  * .RXx.
 681  * rxx..
 682  */
 683             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 684         }
 685         if(mb_x == b_stride-2)
 686             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 687     }
 688     return distortion + rate*penalty_factor;
 689 }
 690
 691 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 692     int i, y2;
 693     Plane *p= &s->plane[plane_index];
 694     const int block_size = MB_SIZE >> s->block_max_depth;
 695     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 696     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 697     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 698     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 699     const int ref_stride= s->current_picture->linesize[plane_index];
 700     uint8_t *dst= s->current_picture->data[plane_index];
 701     uint8_t *src= s-> input_picture->data[plane_index];
 702     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 703     // const has only been removed from zero_dst to suppress a warning
 704     static IDWTELEM zero_dst[4096]; //FIXME
 705     const int b_stride = s->b_width << s->block_max_depth;
 706     const int w= p->width;
 707     const int h= p->height;
 708     int distortion= 0;
 709     int rate= 0;
 710     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 711
 712     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 713
 714     for(i=0; i<9; i++){
 715         int mb_x2= mb_x + (i%3) - 1;
 716         int mb_y2= mb_y + (i/3) - 1;
 717         int x= block_w*mb_x2 + block_w/2;
 718         int y= block_h*mb_y2 + block_h/2;
 719
 720         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 721                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 722
 723         //FIXME find a cleaner/simpler way to skip the outside stuff
 724         for(y2= y; y2<0; y2++)
 725             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 726         for(y2= h; y2<y+block_h; y2++)
 727             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 728         if(x<0){
 729             for(y2= y; y2<y+block_h; y2++)
 730                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 731         }
 732         if(x+block_w > w){
 733             for(y2= y; y2<y+block_h; y2++)
 734                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 735         }
 736
 737         av_assert1(block_w== 8 || block_w==16);
 738         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 739     }
 740
 741     if(plane_index==0){
 742         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 743         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 744
 745 /* ..RRRr
 746  * .RXXx.
 747  * .RXXx.
 748  * rxxx.
 749  */
 750         if(merged)
 751             rate = get_block_bits(s, mb_x, mb_y, 2);
 752         for(i=merged?4:0; i<9; i++){
 753             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 754             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 755         }
 756     }
 757     return distortion + rate*penalty_factor;
 758 }
 759
 760 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 761     const int w= b->width;
 762     const int h= b->height;
 763     int x, y;
 764
 765     if(1){
 766         int run=0;
 767         int *runs = s->run_buffer;
 768         int run_index=0;
 769         int max_index;
 770
 771         for(y=0; y<h; y++){
 772             for(x=0; x<w; x++){
 773                 int v, p=0;
 774                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 775                 v= src[x + y*stride];
 776
 777                 if(y){
 778                     t= src[x + (y-1)*stride];
 779                     if(x){
 780                         lt= src[x - 1 + (y-1)*stride];
 781                     }
 782                     if(x + 1 < w){
 783                         rt= src[x + 1 + (y-1)*stride];
 784                     }
 785                 }
 786                 if(x){
 787                     l= src[x - 1 + y*stride];
 788                     /*if(x > 1){
 789                         if(orientation==1) ll= src[y + (x-2)*stride];
 790                         else               ll= src[x - 2 + y*stride];
 791                     }*/
 792                 }
 793                 if(parent){
 794                     int px= x>>1;
 795                     int py= y>>1;
 796                     if(px<b->parent->width && py<b->parent->height)
 797                         p= parent[px + py*2*stride];
 798                 }
 799                 if(!(/*ll|*/l|lt|t|rt|p)){
 800                     if(v){
 801                         runs[run_index++]= run;
 802                         run=0;
 803                     }else{
 804                         run++;
 805                     }
 806                 }
 807             }
 808         }
 809         max_index= run_index;
 810         runs[run_index++]= run;
 811         run_index=0;
 812         run= runs[run_index++];
 813
 814         put_symbol2(&s->c, b->state[30], max_index, 0);
 815         if(run_index <= max_index)
 816             put_symbol2(&s->c, b->state[1], run, 3);
 817
 818         for(y=0; y<h; y++){
 819             if(s->c.bytestream_end - s->c.bytestream < w*40){
 820                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 821                 return -1;
 822             }
 823             for(x=0; x<w; x++){
 824                 int v, p=0;
 825                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 826                 v= src[x + y*stride];
 827
 828                 if(y){
 829                     t= src[x + (y-1)*stride];
 830                     if(x){
 831                         lt= src[x - 1 + (y-1)*stride];
 832                     }
 833                     if(x + 1 < w){
 834                         rt= src[x + 1 + (y-1)*stride];
 835                     }
 836                 }
 837                 if(x){
 838                     l= src[x - 1 + y*stride];
 839                     /*if(x > 1){
 840                         if(orientation==1) ll= src[y + (x-2)*stride];
 841                         else               ll= src[x - 2 + y*stride];
 842                     }*/
 843                 }
 844                 if(parent){
 845                     int px= x>>1;
 846                     int py= y>>1;
 847                     if(px<b->parent->width && py<b->parent->height)
 848                         p= parent[px + py*2*stride];
 849                 }
 850                 if(/*ll|*/l|lt|t|rt|p){
 851                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 852
 853                     put_rac(&s->c, &b->state[0][context], !!v);
 854                 }else{
 855                     if(!run){
 856                         run= runs[run_index++];
 857
 858                         if(run_index <= max_index)
 859                             put_symbol2(&s->c, b->state[1], run, 3);
 860                         av_assert2(v);
 861                     }else{
 862                         run--;
 863                         av_assert2(!v);
 864                     }
 865                 }
 866                 if(v){
 867                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 868                     int l2= 2*FFABS(l) + (l<0);
 869                     int t2= 2*FFABS(t) + (t<0);
 870
 871                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 872                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 873                 }
 874             }
 875         }
 876     }
 877     return 0;
 878 }
 879
 880 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 881 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 882 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 883     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 884 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 885 }
 886
 887 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){
 888     const int b_stride= s->b_width << s->block_max_depth;
 889     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 890     BlockNode backup= *block;
 891     unsigned value;
 892     int rd, index;
 893
 894     av_assert2(mb_x>=0 && mb_y>=0);
 895     av_assert2(mb_x<b_stride);
 896
 897     if(intra){
 898         block->color[0] = p[0];
 899         block->color[1] = p[1];
 900         block->color[2] = p[2];
 901         block->type |= BLOCK_INTRA;
 902     }else{
 903         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 904         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 905         if(s->me_cache[index] == value)
 906             return 0;
 907         s->me_cache[index]= value;
 908
 909         block->mx= p[0];
 910         block->my= p[1];
 911         block->type &= ~BLOCK_INTRA;
 912     }
 913
 914     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
 915
 916 //FIXME chroma
 917     if(rd < *best_rd){
 918         *best_rd= rd;
 919         return 1;
 920     }else{
 921         *block= backup;
 922         return 0;
 923     }
 924 }
 925
 926 /* special case for int[2] args we discard afterwards,
 927  * fixes compilation problem with gcc 2.95 */
 928 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){
 929     int p[2] = {p0, p1};
 930     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 931 }
 932
 933 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){
 934     const int b_stride= s->b_width << s->block_max_depth;
 935     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 936     BlockNode backup[4];
 937     unsigned value;
 938     int rd, index;
 939
 940     /* We don't initialize backup[] during variable declaration, because
 941      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 942      * 'int16_t'". */
 943     backup[0] = block[0];
 944     backup[1] = block[1];
 945     backup[2] = block[b_stride];
 946     backup[3] = block[b_stride + 1];
 947
 948     av_assert2(mb_x>=0 && mb_y>=0);
 949     av_assert2(mb_x<b_stride);
 950     av_assert2(((mb_x|mb_y)&1) == 0);
 951
 952     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 953     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 954     if(s->me_cache[index] == value)
 955         return 0;
 956     s->me_cache[index]= value;
 957
 958     block->mx= p0;
 959     block->my= p1;
 960     block->ref= ref;
 961     block->type &= ~BLOCK_INTRA;
 962     block[1]= block[b_stride]= block[b_stride+1]= *block;
 963
 964     rd= get_4block_rd(s, mb_x, mb_y, 0);
 965
 966 //FIXME chroma
 967     if(rd < *best_rd){
 968         *best_rd= rd;
 969         return 1;
 970     }else{
 971         block[0]= backup[0];
 972         block[1]= backup[1];
 973         block[b_stride]= backup[2];
 974         block[b_stride+1]= backup[3];
 975         return 0;
 976     }
 977 }
 978
 979 static void iterative_me(SnowContext *s){
 980     int pass, mb_x, mb_y;
 981     const int b_width = s->b_width  << s->block_max_depth;
 982     const int b_height= s->b_height << s->block_max_depth;
 983     const int b_stride= b_width;
 984     int color[3];
 985
 986     {
 987         RangeCoder r = s->c;
 988         uint8_t state[sizeof(s->block_state)];
 989         memcpy(state, s->block_state, sizeof(s->block_state));
 990         for(mb_y= 0; mb_y<s->b_height; mb_y++)
 991             for(mb_x= 0; mb_x<s->b_width; mb_x++)
 992                 encode_q_branch(s, 0, mb_x, mb_y);
 993         s->c = r;
 994         memcpy(s->block_state, state, sizeof(s->block_state));
 995     }
 996
 997     for(pass=0; pass<25; pass++){
 998         int change= 0;
 999
1000         for(mb_y= 0; mb_y<b_height; mb_y++){
1001             for(mb_x= 0; mb_x<b_width; mb_x++){
1002                 int dia_change, i, j, ref;
1003                 int best_rd= INT_MAX, ref_rd;
1004                 BlockNode backup, ref_b;
1005                 const int index= mb_x + mb_y * b_stride;
1006                 BlockNode *block= &s->block[index];
1007                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1008                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1009                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1010                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1011                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1012                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1013                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1014                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1015                 const int b_w= (MB_SIZE >> s->block_max_depth);
1016                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1017
1018                 if(pass && (block->type & BLOCK_OPT))
1019                     continue;
1020                 block->type |= BLOCK_OPT;
1021
1022                 backup= *block;
1023
1024                 if(!s->me_cache_generation)
1025                     memset(s->me_cache, 0, sizeof(s->me_cache));
1026                 s->me_cache_generation += 1<<22;
1027
1028                 //FIXME precalculate
1029                 {
1030                     int x, y;
1031                     for (y = 0; y < b_w * 2; y++)
1032                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1033                     if(mb_x==0)
1034                         for(y=0; y<b_w*2; y++)
1035                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1036                     if(mb_x==b_stride-1)
1037                         for(y=0; y<b_w*2; y++)
1038                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1039                     if(mb_y==0){
1040                         for(x=0; x<b_w*2; x++)
1041                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1042                         for(y=1; y<b_w; y++)
1043                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1044                     }
1045                     if(mb_y==b_height-1){
1046                         for(x=0; x<b_w*2; x++)
1047                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1048                         for(y=b_w; y<b_w*2-1; y++)
1049                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1050                     }
1051                 }
1052
1053                 //skip stuff outside the picture
1054                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1055                     uint8_t *src= s->  input_picture->data[0];
1056                     uint8_t *dst= s->current_picture->data[0];
1057                     const int stride= s->current_picture->linesize[0];
1058                     const int block_w= MB_SIZE >> s->block_max_depth;
1059                     const int block_h= MB_SIZE >> s->block_max_depth;
1060                     const int sx= block_w*mb_x - block_w/2;
1061                     const int sy= block_h*mb_y - block_h/2;
1062                     const int w= s->plane[0].width;
1063                     const int h= s->plane[0].height;
1064                     int y;
1065
1066                     for(y=sy; y<0; y++)
1067                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1068                     for(y=h; y<sy+block_h*2; y++)
1069                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1070                     if(sx<0){
1071                         for(y=sy; y<sy+block_h*2; y++)
1072                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1073                     }
1074                     if(sx+block_w*2 > w){
1075                         for(y=sy; y<sy+block_h*2; y++)
1076                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1077                     }
1078                 }
1079
1080                 // intra(black) = neighbors' contribution to the current block
1081                 for(i=0; i < s->nb_planes; i++)
1082                     color[i]= get_dc(s, mb_x, mb_y, i);
1083
1084                 // get previous score (cannot be cached due to OBMC)
1085                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1086                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1087                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1088                 }else
1089                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1090
1091                 ref_b= *block;
1092                 ref_rd= best_rd;
1093                 for(ref=0; ref < s->ref_frames; ref++){
1094                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1095                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1096                         continue;
1097                     block->ref= ref;
1098                     best_rd= INT_MAX;
1099
1100                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1101                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1102                     if(tb)
1103                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1104                     if(lb)
1105                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1106                     if(rb)
1107                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1108                     if(bb)
1109                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1110
1111                     /* fullpel ME */
1112                     //FIXME avoid subpel interpolation / round to nearest integer
1113                     do{
1114                         int newx = block->mx;
1115                         int newy = block->my;
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, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1120                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1121                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1122                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-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_DEBUG, "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, (1LL<<32)/20, 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<FF_CEIL_RSHIFT(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                    FF_CEIL_RSHIFT(width, hshift));
1572         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1573                                 FF_CEIL_RSHIFT(width, hshift), FF_CEIL_RSHIFT(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     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1626     if (ret < 0)
1627         return ret;
1628
1629     s->m.current_picture_ptr= &s->m.current_picture;
1630     s->m.current_picture.f = s->current_picture;
1631     s->m.current_picture.f->pts = pict->pts;
1632     if(pic->pict_type == AV_PICTURE_TYPE_P){
1633         int block_width = (width +15)>>4;
1634         int block_height= (height+15)>>4;
1635         int stride= s->current_picture->linesize[0];
1636
1637         av_assert0(s->current_picture->data[0]);
1638         av_assert0(s->last_picture[0]->data[0]);
1639
1640         s->m.avctx= s->avctx;
1641         s->m.   last_picture.f = s->last_picture[0];
1642         s->m.    new_picture.f = s->input_picture;
1643         s->m.   last_picture_ptr= &s->m.   last_picture;
1644         s->m.linesize = stride;
1645         s->m.uvlinesize= s->current_picture->linesize[1];
1646         s->m.width = width;
1647         s->m.height= height;
1648         s->m.mb_width = block_width;
1649         s->m.mb_height= block_height;
1650         s->m.mb_stride=   s->m.mb_width+1;
1651         s->m.b8_stride= 2*s->m.mb_width+1;
1652         s->m.f_code=1;
1653         s->m.pict_type = pic->pict_type;
1654         s->m.me_method= s->avctx->me_method;
1655         s->m.me.scene_change_score=0;
1656         s->m.me.dia_size = avctx->dia_size;
1657         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1658         s->m.out_format= FMT_H263;
1659         s->m.unrestricted_mv= 1;
1660
1661         s->m.lambda = s->lambda;
1662         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1663         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1664
1665         s->m.mecc= s->mecc; //move
1666         s->m.qdsp= s->qdsp; //move
1667         s->m.hdsp = s->hdsp;
1668         ff_init_me(&s->m);
1669         s->hdsp = s->m.hdsp;
1670         s->mecc= s->m.mecc;
1671     }
1672
1673     if(s->pass1_rc){
1674         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1675         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1676     }
1677
1678 redo_frame:
1679
1680     s->spatial_decomposition_count= 5;
1681
1682     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1683           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1684         s->spatial_decomposition_count--;
1685
1686     if (s->spatial_decomposition_count <= 0) {
1687         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1688         return AVERROR(EINVAL);
1689     }
1690
1691     s->m.pict_type = pic->pict_type;
1692     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1693
1694     ff_snow_common_init_after_header(avctx);
1695
1696     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1697         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1698             calculate_visual_weight(s, &s->plane[plane_index]);
1699         }
1700     }
1701
1702     encode_header(s);
1703     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1704     encode_blocks(s, 1);
1705     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1706
1707     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1708         Plane *p= &s->plane[plane_index];
1709         int w= p->width;
1710         int h= p->height;
1711         int x, y;
1712 //        int bits= put_bits_count(&s->c.pb);
1713
1714         if (!s->memc_only) {
1715             //FIXME optimize
1716             if(pict->data[plane_index]) //FIXME gray hack
1717                 for(y=0; y<h; y++){
1718                     for(x=0; x<w; x++){
1719                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1720                     }
1721                 }
1722             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1723
1724             if(   plane_index==0
1725                && pic->pict_type == AV_PICTURE_TYPE_P
1726                && !(avctx->flags&CODEC_FLAG_PASS2)
1727                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1728                 ff_init_range_encoder(c, pkt->data, pkt->size);
1729                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1730                 pic->pict_type= AV_PICTURE_TYPE_I;
1731                 s->keyframe=1;
1732                 s->current_picture->key_frame=1;
1733                 goto redo_frame;
1734             }
1735
1736             if(s->qlog == LOSSLESS_QLOG){
1737                 for(y=0; y<h; y++){
1738                     for(x=0; x<w; x++){
1739                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1740                     }
1741                 }
1742             }else{
1743                 for(y=0; y<h; y++){
1744                     for(x=0; x<w; x++){
1745                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1746                     }
1747                 }
1748             }
1749
1750             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1751
1752             if(s->pass1_rc && plane_index==0){
1753                 int delta_qlog = ratecontrol_1pass(s, pic);
1754                 if (delta_qlog <= INT_MIN)
1755                     return -1;
1756                 if(delta_qlog){
1757                     //reordering qlog in the bitstream would eliminate this reset
1758                     ff_init_range_encoder(c, pkt->data, pkt->size);
1759                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1760                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1761                     encode_header(s);
1762                     encode_blocks(s, 0);
1763                 }
1764             }
1765
1766             for(level=0; level<s->spatial_decomposition_count; level++){
1767                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1768                     SubBand *b= &p->band[level][orientation];
1769
1770                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1771                     if(orientation==0)
1772                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1773                     if (!s->no_bitstream)
1774                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1775                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1776                     if(orientation==0)
1777                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1778                 }
1779             }
1780
1781             for(level=0; level<s->spatial_decomposition_count; level++){
1782                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1783                     SubBand *b= &p->band[level][orientation];
1784
1785                     dequantize(s, b, b->ibuf, b->stride);
1786                 }
1787             }
1788
1789             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1790             if(s->qlog == LOSSLESS_QLOG){
1791                 for(y=0; y<h; y++){
1792                     for(x=0; x<w; x++){
1793                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1794                     }
1795                 }
1796             }
1797             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1798         }else{
1799             //ME/MC only
1800             if(pic->pict_type == AV_PICTURE_TYPE_I){
1801                 for(y=0; y<h; y++){
1802                     for(x=0; x<w; x++){
1803                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1804                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1805                     }
1806                 }
1807             }else{
1808                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1809                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1810             }
1811         }
1812         if(s->avctx->flags&CODEC_FLAG_PSNR){
1813             int64_t error= 0;
1814
1815             if(pict->data[plane_index]) //FIXME gray hack
1816                 for(y=0; y<h; y++){
1817                     for(x=0; x<w; x++){
1818                         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];
1819                         error += d*d;
1820                     }
1821                 }
1822             s->avctx->error[plane_index] += error;
1823             s->current_picture->error[plane_index] = error;
1824         }
1825
1826     }
1827
1828     update_last_header_values(s);
1829
1830     ff_snow_release_buffer(avctx);
1831
1832     s->current_picture->coded_picture_number = avctx->frame_number;
1833     s->current_picture->pict_type = pict->pict_type;
1834     s->current_picture->quality = pict->quality;
1835     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1836     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1837     s->m.current_picture.f->display_picture_number =
1838     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1839     s->m.current_picture.f->quality                = pic->quality;
1840     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1841     if(s->pass1_rc)
1842         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1843             return -1;
1844     if(avctx->flags&CODEC_FLAG_PASS1)
1845         ff_write_pass1_stats(&s->m);
1846     s->m.last_pict_type = s->m.pict_type;
1847     avctx->frame_bits = s->m.frame_bits;
1848     avctx->mv_bits = s->m.mv_bits;
1849     avctx->misc_bits = s->m.misc_bits;
1850     avctx->p_tex_bits = s->m.p_tex_bits;
1851
1852     emms_c();
1853
1854     pkt->size = ff_rac_terminate(c);
1855     if (avctx->coded_frame->key_frame)
1856         pkt->flags |= AV_PKT_FLAG_KEY;
1857     *got_packet = 1;
1858
1859     return 0;
1860 }
1861
1862 static av_cold int encode_end(AVCodecContext *avctx)
1863 {
1864     SnowContext *s = avctx->priv_data;
1865
1866     ff_snow_common_end(s);
1867     ff_rate_control_uninit(&s->m);
1868     av_frame_free(&s->input_picture);
1869     av_frame_free(&avctx->coded_frame);
1870     av_freep(&avctx->stats_out);
1871
1872     return 0;
1873 }
1874
1875 #define OFFSET(x) offsetof(SnowContext, x)
1876 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1877 static const AVOption options[] = {
1878     FF_MPV_COMMON_OPTS
1879     { "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 },
1880     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1881     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1882     { NULL },
1883 };
1884
1885 static const AVClass snowenc_class = {
1886     .class_name = "snow encoder",
1887     .item_name  = av_default_item_name,
1888     .option     = options,
1889     .version    = LIBAVUTIL_VERSION_INT,
1890 };
1891
1892 AVCodec ff_snow_encoder = {
1893     .name           = "snow",
1894     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1895     .type           = AVMEDIA_TYPE_VIDEO,
1896     .id             = AV_CODEC_ID_SNOW,
1897     .priv_data_size = sizeof(SnowContext),
1898     .init           = encode_init,
1899     .encode2        = encode_frame,
1900     .close          = encode_end,
1901     .pix_fmts       = (const enum AVPixelFormat[]){
1902         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1903         AV_PIX_FMT_GRAY8,
1904         AV_PIX_FMT_NONE
1905     },
1906     .priv_class     = &snowenc_class,
1907     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1908                       FF_CODEC_CAP_INIT_CLEANUP,
1909 };
1910
1911
1912 #ifdef TEST
1913 #undef malloc
1914 #undef free
1915 #undef printf
1916
1917 #include "libavutil/lfg.h"
1918 #include "libavutil/mathematics.h"
1919
1920 int main(void){
1921 #define width  256
1922 #define height 256
1923     int buffer[2][width*height];
1924     SnowContext s;
1925     int i;
1926     AVLFG prng;
1927     s.spatial_decomposition_count=6;
1928     s.spatial_decomposition_type=1;
1929
1930     s.temp_dwt_buffer  = av_mallocz_array(width, sizeof(DWTELEM));
1931     s.temp_idwt_buffer = av_mallocz_array(width, sizeof(IDWTELEM));
1932
1933     if (!s.temp_dwt_buffer || !s.temp_idwt_buffer) {
1934         fprintf(stderr, "Failed to allocate memory\n");
1935         return 1;
1936     }
1937
1938     av_lfg_init(&prng, 1);
1939
1940     printf("testing 5/3 DWT\n");
1941     for(i=0; i<width*height; i++)
1942         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1943
1944     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1945     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1946
1947     for(i=0; i<width*height; i++)
1948         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1949
1950     printf("testing 9/7 DWT\n");
1951     s.spatial_decomposition_type=0;
1952     for(i=0; i<width*height; i++)
1953         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1954
1955     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1956     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1957
1958     for(i=0; i<width*height; i++)
1959         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1960
1961     {
1962     int level, orientation, x, y;
1963     int64_t errors[8][4];
1964     int64_t g=0;
1965
1966         memset(errors, 0, sizeof(errors));
1967         s.spatial_decomposition_count=3;
1968         s.spatial_decomposition_type=0;
1969         for(level=0; level<s.spatial_decomposition_count; level++){
1970             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1971                 int w= width  >> (s.spatial_decomposition_count-level);
1972                 int h= height >> (s.spatial_decomposition_count-level);
1973                 int stride= width  << (s.spatial_decomposition_count-level);
1974                 DWTELEM *buf= buffer[0];
1975                 int64_t error=0;
1976
1977                 if(orientation&1) buf+=w;
1978                 if(orientation>1) buf+=stride>>1;
1979
1980                 memset(buffer[0], 0, sizeof(int)*width*height);
1981                 buf[w/2 + h/2*stride]= 256*256;
1982                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1983                 for(y=0; y<height; y++){
1984                     for(x=0; x<width; x++){
1985                         int64_t d= buffer[0][x + y*width];
1986                         error += d*d;
1987                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1988                     }
1989                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1990                 }
1991                 error= (int)(sqrt(error)+0.5);
1992                 errors[level][orientation]= error;
1993                 if(g) g=av_gcd(g, error);
1994                 else g= error;
1995             }
1996         }
1997         printf("static int const visual_weight[][4]={\n");
1998         for(level=0; level<s.spatial_decomposition_count; level++){
1999             printf("  {");
2000             for(orientation=0; orientation<4; orientation++){
2001                 printf("%8"PRId64",", errors[level][orientation]/g);
2002             }
2003             printf("},\n");
2004         }
2005         printf("};\n");
2006         {
2007             int level=2;
2008             int w= width  >> (s.spatial_decomposition_count-level);
2009             //int h= height >> (s.spatial_decomposition_count-level);
2010             int stride= width  << (s.spatial_decomposition_count-level);
2011             DWTELEM *buf= buffer[0];
2012             int64_t error=0;
2013
2014             buf+=w;
2015             buf+=stride>>1;
2016
2017             memset(buffer[0], 0, sizeof(int)*width*height);
2018             for(y=0; y<height; y++){
2019                 for(x=0; x<width; x++){
2020                     int tab[4]={0,2,3,1};
2021                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2022                 }
2023             }
2024             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2025             for(y=0; y<height; y++){
2026                 for(x=0; x<width; x++){
2027                     int64_t d= buffer[0][x + y*width];
2028                     error += d*d;
2029                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2030                 }
2031                 if(FFABS(height/2-y)<9) printf("\n");
2032             }
2033         }
2034
2035     }
2036     return 0;
2037 }
2038 #endif /* TEST */