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 #define FF_ME_ITER 50
  36
  37 static av_cold int encode_init(AVCodecContext *avctx)
  38 {
  39     SnowContext *s = avctx->priv_data;
  40     int plane_index, ret;
  41     int i;
  42
  43     if(avctx->prediction_method == DWT_97
  44        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
  45        && avctx->global_quality == 0){
  46         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  47         return -1;
  48     }
  49 #if FF_API_MOTION_EST
  50 FF_DISABLE_DEPRECATION_WARNINGS
  51     if (avctx->me_method == ME_ITER)
  52         s->motion_est = FF_ME_ITER;
  53 FF_ENABLE_DEPRECATION_WARNINGS
  54 #endif
  55
  56     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
  57
  58     s->mv_scale       = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
  59     s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
  60
  61     for(plane_index=0; plane_index<3; plane_index++){
  62         s->plane[plane_index].diag_mc= 1;
  63         s->plane[plane_index].htaps= 6;
  64         s->plane[plane_index].hcoeff[0]=  40;
  65         s->plane[plane_index].hcoeff[1]= -10;
  66         s->plane[plane_index].hcoeff[2]=   2;
  67         s->plane[plane_index].fast_mc= 1;
  68     }
  69
  70     if ((ret = ff_snow_common_init(avctx)) < 0) {
  71         return ret;
  72     }
  73     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  74
  75     ff_snow_alloc_blocks(s);
  76
  77     s->version=0;
  78
  79     s->m.avctx   = avctx;
  80     s->m.bit_rate= avctx->bit_rate;
  81
  82     s->m.me.temp      =
  83     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  84     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  85     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  86     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  87     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
  88         return AVERROR(ENOMEM);
  89
  90     ff_h263_encode_init(&s->m); //mv_penalty
  91
  92     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
  93
  94     if(avctx->flags&AV_CODEC_FLAG_PASS1){
  95         if(!avctx->stats_out)
  96             avctx->stats_out = av_mallocz(256);
  97
  98         if (!avctx->stats_out)
  99             return AVERROR(ENOMEM);
 100     }
 101     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
 102         if(ff_rate_control_init(&s->m) < 0)
 103             return -1;
 104     }
 105     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 106
 107     switch(avctx->pix_fmt){
 108     case AV_PIX_FMT_YUV444P:
 109 //    case AV_PIX_FMT_YUV422P:
 110     case AV_PIX_FMT_YUV420P:
 111 //    case AV_PIX_FMT_YUV411P:
 112     case AV_PIX_FMT_YUV410P:
 113         s->nb_planes = 3;
 114         s->colorspace_type= 0;
 115         break;
 116     case AV_PIX_FMT_GRAY8:
 117         s->nb_planes = 1;
 118         s->colorspace_type = 1;
 119         break;
 120 /*    case AV_PIX_FMT_RGB32:
 121         s->colorspace= 1;
 122         break;*/
 123     default:
 124         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 125         return -1;
 126     }
 127     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 128
 129     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 130     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 131
 132     s->input_picture = av_frame_alloc();
 133     if (!s->input_picture)
 134         return AVERROR(ENOMEM);
 135
 136     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 137         return ret;
 138
 139     if(s->motion_est == FF_ME_ITER){
 140         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 141         for(i=0; i<s->max_ref_frames; i++){
 142             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 143             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 144             if (!s->ref_mvs[i] || !s->ref_scores[i])
 145                 return AVERROR(ENOMEM);
 146         }
 147     }
 148
 149     return 0;
 150 }
 151
 152 //near copy & paste from dsputil, FIXME
 153 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 154 {
 155     int s, i, j;
 156
 157     s = 0;
 158     for (i = 0; i < h; i++) {
 159         for (j = 0; j < w; j++) {
 160             s += pix[0];
 161             pix ++;
 162         }
 163         pix += line_size - w;
 164     }
 165     return s;
 166 }
 167
 168 //near copy & paste from dsputil, FIXME
 169 static int pix_norm1(uint8_t * pix, int line_size, int w)
 170 {
 171     int s, i, j;
 172     uint32_t *sq = ff_square_tab + 256;
 173
 174     s = 0;
 175     for (i = 0; i < w; i++) {
 176         for (j = 0; j < w; j ++) {
 177             s += sq[pix[0]];
 178             pix ++;
 179         }
 180         pix += line_size - w;
 181     }
 182     return s;
 183 }
 184
 185 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 186     switch(type&0xFF){
 187     default:
 188     case FF_CMP_SAD:
 189         return lambda>>FF_LAMBDA_SHIFT;
 190     case FF_CMP_DCT:
 191         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 192     case FF_CMP_W53:
 193         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 194     case FF_CMP_W97:
 195         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 196     case FF_CMP_SATD:
 197     case FF_CMP_DCT264:
 198         return (2*lambda)>>FF_LAMBDA_SHIFT;
 199     case FF_CMP_RD:
 200     case FF_CMP_PSNR:
 201     case FF_CMP_SSE:
 202     case FF_CMP_NSSE:
 203         return lambda2>>FF_LAMBDA_SHIFT;
 204     case FF_CMP_BIT:
 205         return 1;
 206     }
 207 }
 208
 209 //FIXME copy&paste
 210 #define P_LEFT P[1]
 211 #define P_TOP P[2]
 212 #define P_TOPRIGHT P[3]
 213 #define P_MEDIAN P[4]
 214 #define P_MV1 P[9]
 215 #define FLAG_QPEL   1 //must be 1
 216
 217 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 218     uint8_t p_buffer[1024];
 219     uint8_t i_buffer[1024];
 220     uint8_t p_state[sizeof(s->block_state)];
 221     uint8_t i_state[sizeof(s->block_state)];
 222     RangeCoder pc, ic;
 223     uint8_t *pbbak= s->c.bytestream;
 224     uint8_t *pbbak_start= s->c.bytestream_start;
 225     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 226     const int w= s->b_width  << s->block_max_depth;
 227     const int h= s->b_height << s->block_max_depth;
 228     const int rem_depth= s->block_max_depth - level;
 229     const int index= (x + y*w) << rem_depth;
 230     const int block_w= 1<<(LOG2_MB_SIZE - level);
 231     int trx= (x+1)<<rem_depth;
 232     int try= (y+1)<<rem_depth;
 233     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 234     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 235     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 236     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 237     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 238     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 239     int pl = left->color[0];
 240     int pcb= left->color[1];
 241     int pcr= left->color[2];
 242     int pmx, pmy;
 243     int mx=0, my=0;
 244     int l,cr,cb;
 245     const int stride= s->current_picture->linesize[0];
 246     const int uvstride= s->current_picture->linesize[1];
 247     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 248                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 249                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 250     int P[10][2];
 251     int16_t last_mv[3][2];
 252     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
 253     const int shift= 1+qpel;
 254     MotionEstContext *c= &s->m.me;
 255     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 256     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 257     int my_context= av_log2(2*FFABS(left->my - top->my));
 258     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 259     int ref, best_ref, ref_score, ref_mx, ref_my;
 260
 261     av_assert0(sizeof(s->block_state) >= 256);
 262     if(s->keyframe){
 263         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 264         return 0;
 265     }
 266
 267 //    clip predictors / edge ?
 268
 269     P_LEFT[0]= left->mx;
 270     P_LEFT[1]= left->my;
 271     P_TOP [0]= top->mx;
 272     P_TOP [1]= top->my;
 273     P_TOPRIGHT[0]= tr->mx;
 274     P_TOPRIGHT[1]= tr->my;
 275
 276     last_mv[0][0]= s->block[index].mx;
 277     last_mv[0][1]= s->block[index].my;
 278     last_mv[1][0]= right->mx;
 279     last_mv[1][1]= right->my;
 280     last_mv[2][0]= bottom->mx;
 281     last_mv[2][1]= bottom->my;
 282
 283     s->m.mb_stride=2;
 284     s->m.mb_x=
 285     s->m.mb_y= 0;
 286     c->skip= 0;
 287
 288     av_assert1(c->  stride ==   stride);
 289     av_assert1(c->uvstride == uvstride);
 290
 291     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 292     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 293     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 294     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 295
 296     c->xmin = - x*block_w - 16+3;
 297     c->ymin = - y*block_w - 16+3;
 298     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 299     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 300
 301     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 302     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 303     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 304     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 305     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 306     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 307     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 308
 309     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 310     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 311
 312     if (!y) {
 313         c->pred_x= P_LEFT[0];
 314         c->pred_y= P_LEFT[1];
 315     } else {
 316         c->pred_x = P_MEDIAN[0];
 317         c->pred_y = P_MEDIAN[1];
 318     }
 319
 320     score= INT_MAX;
 321     best_ref= 0;
 322     for(ref=0; ref<s->ref_frames; ref++){
 323         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 324
 325         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 326                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 327
 328         av_assert2(ref_mx >= c->xmin);
 329         av_assert2(ref_mx <= c->xmax);
 330         av_assert2(ref_my >= c->ymin);
 331         av_assert2(ref_my <= c->ymax);
 332
 333         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 334         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 335         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 336         if(s->ref_mvs[ref]){
 337             s->ref_mvs[ref][index][0]= ref_mx;
 338             s->ref_mvs[ref][index][1]= ref_my;
 339             s->ref_scores[ref][index]= ref_score;
 340         }
 341         if(score > ref_score){
 342             score= ref_score;
 343             best_ref= ref;
 344             mx= ref_mx;
 345             my= ref_my;
 346         }
 347     }
 348     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 349
 350   //  subpel search
 351     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 352     pc= s->c;
 353     pc.bytestream_start=
 354     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 355     memcpy(p_state, s->block_state, sizeof(s->block_state));
 356
 357     if(level!=s->block_max_depth)
 358         put_rac(&pc, &p_state[4 + s_context], 1);
 359     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 360     if(s->ref_frames > 1)
 361         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 362     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 363     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 364     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 365     p_len= pc.bytestream - pc.bytestream_start;
 366     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 367
 368     block_s= block_w*block_w;
 369     sum = pix_sum(current_data[0], stride, block_w, block_w);
 370     l= (sum + block_s/2)/block_s;
 371     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 372
 373     if (s->nb_planes > 2) {
 374         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 375         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 376         cb= (sum + block_s/2)/block_s;
 377     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 378         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 379         cr= (sum + block_s/2)/block_s;
 380     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 381     }else
 382         cb = cr = 0;
 383
 384     ic= s->c;
 385     ic.bytestream_start=
 386     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 387     memcpy(i_state, s->block_state, sizeof(s->block_state));
 388     if(level!=s->block_max_depth)
 389         put_rac(&ic, &i_state[4 + s_context], 1);
 390     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 391     put_symbol(&ic, &i_state[32],  l-pl , 1);
 392     if (s->nb_planes > 2) {
 393         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 394         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 395     }
 396     i_len= ic.bytestream - ic.bytestream_start;
 397     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 398
 399     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 400     av_assert1(iscore >= 0);
 401     av_assert1(l>=0 && l<=255);
 402     av_assert1(pl>=0 && pl<=255);
 403
 404     if(level==0){
 405         int varc= iscore >> 8;
 406         int vard= score >> 8;
 407         if (vard <= 64 || vard < varc)
 408             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 409         else
 410             c->scene_change_score+= s->m.qscale;
 411     }
 412
 413     if(level!=s->block_max_depth){
 414         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 415         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 416         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 417         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 418         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 419         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 420
 421         if(score2 < score && score2 < iscore)
 422             return score2;
 423     }
 424
 425     if(iscore < score){
 426         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 427         memcpy(pbbak, i_buffer, i_len);
 428         s->c= ic;
 429         s->c.bytestream_start= pbbak_start;
 430         s->c.bytestream= pbbak + i_len;
 431         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 432         memcpy(s->block_state, i_state, sizeof(s->block_state));
 433         return iscore;
 434     }else{
 435         memcpy(pbbak, p_buffer, p_len);
 436         s->c= pc;
 437         s->c.bytestream_start= pbbak_start;
 438         s->c.bytestream= pbbak + p_len;
 439         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 440         memcpy(s->block_state, p_state, sizeof(s->block_state));
 441         return score;
 442     }
 443 }
 444
 445 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 446     const int w= s->b_width  << s->block_max_depth;
 447     const int rem_depth= s->block_max_depth - level;
 448     const int index= (x + y*w) << rem_depth;
 449     int trx= (x+1)<<rem_depth;
 450     BlockNode *b= &s->block[index];
 451     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 452     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 453     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 454     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 455     int pl = left->color[0];
 456     int pcb= left->color[1];
 457     int pcr= left->color[2];
 458     int pmx, pmy;
 459     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 460     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 461     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 462     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 463
 464     if(s->keyframe){
 465         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 466         return;
 467     }
 468
 469     if(level!=s->block_max_depth){
 470         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 471             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 472         }else{
 473             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 474             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 475             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 476             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 477             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 478             return;
 479         }
 480     }
 481     if(b->type & BLOCK_INTRA){
 482         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 483         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 484         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 485         if (s->nb_planes > 2) {
 486             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 487             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 488         }
 489         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 490     }else{
 491         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 492         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 493         if(s->ref_frames > 1)
 494             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 495         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 496         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 497         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 498     }
 499 }
 500
 501 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 502     int i, x2, y2;
 503     Plane *p= &s->plane[plane_index];
 504     const int block_size = MB_SIZE >> s->block_max_depth;
 505     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 506     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 507     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 508     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 509     const int ref_stride= s->current_picture->linesize[plane_index];
 510     uint8_t *src= s-> input_picture->data[plane_index];
 511     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 512     const int b_stride = s->b_width << s->block_max_depth;
 513     const int w= p->width;
 514     const int h= p->height;
 515     int index= mb_x + mb_y*b_stride;
 516     BlockNode *b= &s->block[index];
 517     BlockNode backup= *b;
 518     int ab=0;
 519     int aa=0;
 520
 521     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 522
 523     b->type|= BLOCK_INTRA;
 524     b->color[plane_index]= 0;
 525     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 526
 527     for(i=0; i<4; i++){
 528         int mb_x2= mb_x + (i &1) - 1;
 529         int mb_y2= mb_y + (i>>1) - 1;
 530         int x= block_w*mb_x2 + block_w/2;
 531         int y= block_h*mb_y2 + block_h/2;
 532
 533         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 534                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 535
 536         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 537             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 538                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 539                 int obmc_v= obmc[index];
 540                 int d;
 541                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 542                 if(x<0) obmc_v += obmc[index + block_w];
 543                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 544                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 545                 //FIXME precalculate this or simplify it somehow else
 546
 547                 d = -dst[index] + (1<<(FRAC_BITS-1));
 548                 dst[index] = d;
 549                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 550                 aa += obmc_v * obmc_v; //FIXME precalculate this
 551             }
 552         }
 553     }
 554     *b= backup;
 555
 556     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 557 }
 558
 559 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 560     const int b_stride = s->b_width << s->block_max_depth;
 561     const int b_height = s->b_height<< s->block_max_depth;
 562     int index= x + y*b_stride;
 563     const BlockNode *b     = &s->block[index];
 564     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 565     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 566     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 567     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 568     int dmx, dmy;
 569 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 570 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 571
 572     if(x<0 || x>=b_stride || y>=b_height)
 573         return 0;
 574 /*
 575 1            0      0
 576 01X          1-2    1
 577 001XX        3-6    2-3
 578 0001XXX      7-14   4-7
 579 00001XXXX   15-30   8-15
 580 */
 581 //FIXME try accurate rate
 582 //FIXME intra and inter predictors if surrounding blocks are not the same type
 583     if(b->type & BLOCK_INTRA){
 584         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 585                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 586                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 587     }else{
 588         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 589         dmx-= b->mx;
 590         dmy-= b->my;
 591         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 592                     + av_log2(2*FFABS(dmy))
 593                     + av_log2(2*b->ref));
 594     }
 595 }
 596
 597 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 598     Plane *p= &s->plane[plane_index];
 599     const int block_size = MB_SIZE >> s->block_max_depth;
 600     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 601     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 602     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 603     const int ref_stride= s->current_picture->linesize[plane_index];
 604     uint8_t *dst= s->current_picture->data[plane_index];
 605     uint8_t *src= s->  input_picture->data[plane_index];
 606     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
 607     uint8_t *cur = s->scratchbuf;
 608     uint8_t *tmp = s->emu_edge_buffer;
 609     const int b_stride = s->b_width << s->block_max_depth;
 610     const int b_height = s->b_height<< s->block_max_depth;
 611     const int w= p->width;
 612     const int h= p->height;
 613     int distortion;
 614     int rate= 0;
 615     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 616     int sx= block_w*mb_x - block_w/2;
 617     int sy= block_h*mb_y - block_h/2;
 618     int x0= FFMAX(0,-sx);
 619     int y0= FFMAX(0,-sy);
 620     int x1= FFMIN(block_w*2, w-sx);
 621     int y1= FFMIN(block_h*2, h-sy);
 622     int i,x,y;
 623
 624     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 625
 626     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);
 627
 628     for(y=y0; y<y1; y++){
 629         const uint8_t *obmc1= obmc_edged[y];
 630         const IDWTELEM *pred1 = pred + y*obmc_stride;
 631         uint8_t *cur1 = cur + y*ref_stride;
 632         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 633         for(x=x0; x<x1; x++){
 634 #if FRAC_BITS >= LOG2_OBMC_MAX
 635             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 636 #else
 637             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 638 #endif
 639             v = (v + pred1[x]) >> FRAC_BITS;
 640             if(v&(~255)) v= ~(v>>31);
 641             dst1[x] = v;
 642         }
 643     }
 644
 645     /* copy the regions where obmc[] = (uint8_t)256 */
 646     if(LOG2_OBMC_MAX == 8
 647         && (mb_x == 0 || mb_x == b_stride-1)
 648         && (mb_y == 0 || mb_y == b_height-1)){
 649         if(mb_x == 0)
 650             x1 = block_w;
 651         else
 652             x0 = block_w;
 653         if(mb_y == 0)
 654             y1 = block_h;
 655         else
 656             y0 = block_h;
 657         for(y=y0; y<y1; y++)
 658             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 659     }
 660
 661     if(block_w==16){
 662         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 663         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 664         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 665          * So improving the score of one block is not strictly guaranteed
 666          * to improve the score of the whole frame, thus iterative motion
 667          * estimation does not always converge. */
 668         if(s->avctx->me_cmp == FF_CMP_W97)
 669             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 670         else if(s->avctx->me_cmp == FF_CMP_W53)
 671             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 672         else{
 673             distortion = 0;
 674             for(i=0; i<4; i++){
 675                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 676                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 677             }
 678         }
 679     }else{
 680         av_assert2(block_w==8);
 681         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 682     }
 683
 684     if(plane_index==0){
 685         for(i=0; i<4; i++){
 686 /* ..RRr
 687  * .RXx.
 688  * rxx..
 689  */
 690             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 691         }
 692         if(mb_x == b_stride-2)
 693             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 694     }
 695     return distortion + rate*penalty_factor;
 696 }
 697
 698 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 699     int i, y2;
 700     Plane *p= &s->plane[plane_index];
 701     const int block_size = MB_SIZE >> s->block_max_depth;
 702     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 703     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 704     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 705     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 706     const int ref_stride= s->current_picture->linesize[plane_index];
 707     uint8_t *dst= s->current_picture->data[plane_index];
 708     uint8_t *src= s-> input_picture->data[plane_index];
 709     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 710     // const has only been removed from zero_dst to suppress a warning
 711     static IDWTELEM zero_dst[4096]; //FIXME
 712     const int b_stride = s->b_width << s->block_max_depth;
 713     const int w= p->width;
 714     const int h= p->height;
 715     int distortion= 0;
 716     int rate= 0;
 717     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 718
 719     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 720
 721     for(i=0; i<9; i++){
 722         int mb_x2= mb_x + (i%3) - 1;
 723         int mb_y2= mb_y + (i/3) - 1;
 724         int x= block_w*mb_x2 + block_w/2;
 725         int y= block_h*mb_y2 + block_h/2;
 726
 727         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 728                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 729
 730         //FIXME find a cleaner/simpler way to skip the outside stuff
 731         for(y2= y; y2<0; y2++)
 732             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 733         for(y2= h; y2<y+block_h; y2++)
 734             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 735         if(x<0){
 736             for(y2= y; y2<y+block_h; y2++)
 737                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 738         }
 739         if(x+block_w > w){
 740             for(y2= y; y2<y+block_h; y2++)
 741                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 742         }
 743
 744         av_assert1(block_w== 8 || block_w==16);
 745         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 746     }
 747
 748     if(plane_index==0){
 749         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 750         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 751
 752 /* ..RRRr
 753  * .RXXx.
 754  * .RXXx.
 755  * rxxx.
 756  */
 757         if(merged)
 758             rate = get_block_bits(s, mb_x, mb_y, 2);
 759         for(i=merged?4:0; i<9; i++){
 760             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 761             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 762         }
 763     }
 764     return distortion + rate*penalty_factor;
 765 }
 766
 767 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 768     const int w= b->width;
 769     const int h= b->height;
 770     int x, y;
 771
 772     if(1){
 773         int run=0;
 774         int *runs = s->run_buffer;
 775         int run_index=0;
 776         int max_index;
 777
 778         for(y=0; y<h; y++){
 779             for(x=0; x<w; x++){
 780                 int v, p=0;
 781                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 782                 v= src[x + y*stride];
 783
 784                 if(y){
 785                     t= src[x + (y-1)*stride];
 786                     if(x){
 787                         lt= src[x - 1 + (y-1)*stride];
 788                     }
 789                     if(x + 1 < w){
 790                         rt= src[x + 1 + (y-1)*stride];
 791                     }
 792                 }
 793                 if(x){
 794                     l= src[x - 1 + y*stride];
 795                     /*if(x > 1){
 796                         if(orientation==1) ll= src[y + (x-2)*stride];
 797                         else               ll= src[x - 2 + y*stride];
 798                     }*/
 799                 }
 800                 if(parent){
 801                     int px= x>>1;
 802                     int py= y>>1;
 803                     if(px<b->parent->width && py<b->parent->height)
 804                         p= parent[px + py*2*stride];
 805                 }
 806                 if(!(/*ll|*/l|lt|t|rt|p)){
 807                     if(v){
 808                         runs[run_index++]= run;
 809                         run=0;
 810                     }else{
 811                         run++;
 812                     }
 813                 }
 814             }
 815         }
 816         max_index= run_index;
 817         runs[run_index++]= run;
 818         run_index=0;
 819         run= runs[run_index++];
 820
 821         put_symbol2(&s->c, b->state[30], max_index, 0);
 822         if(run_index <= max_index)
 823             put_symbol2(&s->c, b->state[1], run, 3);
 824
 825         for(y=0; y<h; y++){
 826             if(s->c.bytestream_end - s->c.bytestream < w*40){
 827                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 828                 return -1;
 829             }
 830             for(x=0; x<w; x++){
 831                 int v, p=0;
 832                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 833                 v= src[x + y*stride];
 834
 835                 if(y){
 836                     t= src[x + (y-1)*stride];
 837                     if(x){
 838                         lt= src[x - 1 + (y-1)*stride];
 839                     }
 840                     if(x + 1 < w){
 841                         rt= src[x + 1 + (y-1)*stride];
 842                     }
 843                 }
 844                 if(x){
 845                     l= src[x - 1 + y*stride];
 846                     /*if(x > 1){
 847                         if(orientation==1) ll= src[y + (x-2)*stride];
 848                         else               ll= src[x - 2 + y*stride];
 849                     }*/
 850                 }
 851                 if(parent){
 852                     int px= x>>1;
 853                     int py= y>>1;
 854                     if(px<b->parent->width && py<b->parent->height)
 855                         p= parent[px + py*2*stride];
 856                 }
 857                 if(/*ll|*/l|lt|t|rt|p){
 858                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 859
 860                     put_rac(&s->c, &b->state[0][context], !!v);
 861                 }else{
 862                     if(!run){
 863                         run= runs[run_index++];
 864
 865                         if(run_index <= max_index)
 866                             put_symbol2(&s->c, b->state[1], run, 3);
 867                         av_assert2(v);
 868                     }else{
 869                         run--;
 870                         av_assert2(!v);
 871                     }
 872                 }
 873                 if(v){
 874                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 875                     int l2= 2*FFABS(l) + (l<0);
 876                     int t2= 2*FFABS(t) + (t<0);
 877
 878                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 879                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 880                 }
 881             }
 882         }
 883     }
 884     return 0;
 885 }
 886
 887 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 888 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 889 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 890     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 891 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 892 }
 893
 894 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){
 895     const int b_stride= s->b_width << s->block_max_depth;
 896     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 897     BlockNode backup= *block;
 898     unsigned value;
 899     int rd, index;
 900
 901     av_assert2(mb_x>=0 && mb_y>=0);
 902     av_assert2(mb_x<b_stride);
 903
 904     if(intra){
 905         block->color[0] = p[0];
 906         block->color[1] = p[1];
 907         block->color[2] = p[2];
 908         block->type |= BLOCK_INTRA;
 909     }else{
 910         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 911         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 912         if(s->me_cache[index] == value)
 913             return 0;
 914         s->me_cache[index]= value;
 915
 916         block->mx= p[0];
 917         block->my= p[1];
 918         block->type &= ~BLOCK_INTRA;
 919     }
 920
 921     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
 922
 923 //FIXME chroma
 924     if(rd < *best_rd){
 925         *best_rd= rd;
 926         return 1;
 927     }else{
 928         *block= backup;
 929         return 0;
 930     }
 931 }
 932
 933 /* special case for int[2] args we discard afterwards,
 934  * fixes compilation problem with gcc 2.95 */
 935 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){
 936     int p[2] = {p0, p1};
 937     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 938 }
 939
 940 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){
 941     const int b_stride= s->b_width << s->block_max_depth;
 942     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 943     BlockNode backup[4];
 944     unsigned value;
 945     int rd, index;
 946
 947     /* We don't initialize backup[] during variable declaration, because
 948      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 949      * 'int16_t'". */
 950     backup[0] = block[0];
 951     backup[1] = block[1];
 952     backup[2] = block[b_stride];
 953     backup[3] = block[b_stride + 1];
 954
 955     av_assert2(mb_x>=0 && mb_y>=0);
 956     av_assert2(mb_x<b_stride);
 957     av_assert2(((mb_x|mb_y)&1) == 0);
 958
 959     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 960     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 961     if(s->me_cache[index] == value)
 962         return 0;
 963     s->me_cache[index]= value;
 964
 965     block->mx= p0;
 966     block->my= p1;
 967     block->ref= ref;
 968     block->type &= ~BLOCK_INTRA;
 969     block[1]= block[b_stride]= block[b_stride+1]= *block;
 970
 971     rd= get_4block_rd(s, mb_x, mb_y, 0);
 972
 973 //FIXME chroma
 974     if(rd < *best_rd){
 975         *best_rd= rd;
 976         return 1;
 977     }else{
 978         block[0]= backup[0];
 979         block[1]= backup[1];
 980         block[b_stride]= backup[2];
 981         block[b_stride+1]= backup[3];
 982         return 0;
 983     }
 984 }
 985
 986 static void iterative_me(SnowContext *s){
 987     int pass, mb_x, mb_y;
 988     const int b_width = s->b_width  << s->block_max_depth;
 989     const int b_height= s->b_height << s->block_max_depth;
 990     const int b_stride= b_width;
 991     int color[3];
 992
 993     {
 994         RangeCoder r = s->c;
 995         uint8_t state[sizeof(s->block_state)];
 996         memcpy(state, s->block_state, sizeof(s->block_state));
 997         for(mb_y= 0; mb_y<s->b_height; mb_y++)
 998             for(mb_x= 0; mb_x<s->b_width; mb_x++)
 999                 encode_q_branch(s, 0, mb_x, mb_y);
1000         s->c = r;
1001         memcpy(s->block_state, state, sizeof(s->block_state));
1002     }
1003
1004     for(pass=0; pass<25; pass++){
1005         int change= 0;
1006
1007         for(mb_y= 0; mb_y<b_height; mb_y++){
1008             for(mb_x= 0; mb_x<b_width; mb_x++){
1009                 int dia_change, i, j, ref;
1010                 int best_rd= INT_MAX, ref_rd;
1011                 BlockNode backup, ref_b;
1012                 const int index= mb_x + mb_y * b_stride;
1013                 BlockNode *block= &s->block[index];
1014                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1015                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1016                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1017                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1018                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1019                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1020                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1021                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1022                 const int b_w= (MB_SIZE >> s->block_max_depth);
1023                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1024
1025                 if(pass && (block->type & BLOCK_OPT))
1026                     continue;
1027                 block->type |= BLOCK_OPT;
1028
1029                 backup= *block;
1030
1031                 if(!s->me_cache_generation)
1032                     memset(s->me_cache, 0, sizeof(s->me_cache));
1033                 s->me_cache_generation += 1<<22;
1034
1035                 //FIXME precalculate
1036                 {
1037                     int x, y;
1038                     for (y = 0; y < b_w * 2; y++)
1039                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1040                     if(mb_x==0)
1041                         for(y=0; y<b_w*2; y++)
1042                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1043                     if(mb_x==b_stride-1)
1044                         for(y=0; y<b_w*2; y++)
1045                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1046                     if(mb_y==0){
1047                         for(x=0; x<b_w*2; x++)
1048                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1049                         for(y=1; y<b_w; y++)
1050                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1051                     }
1052                     if(mb_y==b_height-1){
1053                         for(x=0; x<b_w*2; x++)
1054                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1055                         for(y=b_w; y<b_w*2-1; y++)
1056                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1057                     }
1058                 }
1059
1060                 //skip stuff outside the picture
1061                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1062                     uint8_t *src= s->  input_picture->data[0];
1063                     uint8_t *dst= s->current_picture->data[0];
1064                     const int stride= s->current_picture->linesize[0];
1065                     const int block_w= MB_SIZE >> s->block_max_depth;
1066                     const int block_h= MB_SIZE >> s->block_max_depth;
1067                     const int sx= block_w*mb_x - block_w/2;
1068                     const int sy= block_h*mb_y - block_h/2;
1069                     const int w= s->plane[0].width;
1070                     const int h= s->plane[0].height;
1071                     int y;
1072
1073                     for(y=sy; y<0; y++)
1074                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1075                     for(y=h; y<sy+block_h*2; y++)
1076                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1077                     if(sx<0){
1078                         for(y=sy; y<sy+block_h*2; y++)
1079                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1080                     }
1081                     if(sx+block_w*2 > w){
1082                         for(y=sy; y<sy+block_h*2; y++)
1083                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1084                     }
1085                 }
1086
1087                 // intra(black) = neighbors' contribution to the current block
1088                 for(i=0; i < s->nb_planes; i++)
1089                     color[i]= get_dc(s, mb_x, mb_y, i);
1090
1091                 // get previous score (cannot be cached due to OBMC)
1092                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1093                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1094                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1095                 }else
1096                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1097
1098                 ref_b= *block;
1099                 ref_rd= best_rd;
1100                 for(ref=0; ref < s->ref_frames; ref++){
1101                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1102                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1103                         continue;
1104                     block->ref= ref;
1105                     best_rd= INT_MAX;
1106
1107                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1108                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1109                     if(tb)
1110                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1111                     if(lb)
1112                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1113                     if(rb)
1114                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1115                     if(bb)
1116                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1117
1118                     /* fullpel ME */
1119                     //FIXME avoid subpel interpolation / round to nearest integer
1120                     do{
1121                         int newx = block->mx;
1122                         int newy = block->my;
1123                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1124                         dia_change=0;
1125                         for(i=0; i < dia_size; i++){
1126                             for(j=0; j<i; j++){
1127                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1128                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1129                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1130                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1131                             }
1132                         }
1133                     }while(dia_change);
1134                     /* subpel ME */
1135                     do{
1136                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1137                         dia_change=0;
1138                         for(i=0; i<8; i++)
1139                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1140                     }while(dia_change);
1141                     //FIXME or try the standard 2 pass qpel or similar
1142
1143                     mvr[0][0]= block->mx;
1144                     mvr[0][1]= block->my;
1145                     if(ref_rd > best_rd){
1146                         ref_rd= best_rd;
1147                         ref_b= *block;
1148                     }
1149                 }
1150                 best_rd= ref_rd;
1151                 *block= ref_b;
1152                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1153                 //FIXME RD style color selection
1154                 if(!same_block(block, &backup)){
1155                     if(tb ) tb ->type &= ~BLOCK_OPT;
1156                     if(lb ) lb ->type &= ~BLOCK_OPT;
1157                     if(rb ) rb ->type &= ~BLOCK_OPT;
1158                     if(bb ) bb ->type &= ~BLOCK_OPT;
1159                     if(tlb) tlb->type &= ~BLOCK_OPT;
1160                     if(trb) trb->type &= ~BLOCK_OPT;
1161                     if(blb) blb->type &= ~BLOCK_OPT;
1162                     if(brb) brb->type &= ~BLOCK_OPT;
1163                     change ++;
1164                 }
1165             }
1166         }
1167         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1168         if(!change)
1169             break;
1170     }
1171
1172     if(s->block_max_depth == 1){
1173         int change= 0;
1174         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1175             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1176                 int i;
1177                 int best_rd, init_rd;
1178                 const int index= mb_x + mb_y * b_stride;
1179                 BlockNode *b[4];
1180
1181                 b[0]= &s->block[index];
1182                 b[1]= b[0]+1;
1183                 b[2]= b[0]+b_stride;
1184                 b[3]= b[2]+1;
1185                 if(same_block(b[0], b[1]) &&
1186                    same_block(b[0], b[2]) &&
1187                    same_block(b[0], b[3]))
1188                     continue;
1189
1190                 if(!s->me_cache_generation)
1191                     memset(s->me_cache, 0, sizeof(s->me_cache));
1192                 s->me_cache_generation += 1<<22;
1193
1194                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1195
1196                 //FIXME more multiref search?
1197                 check_4block_inter(s, mb_x, mb_y,
1198                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1199                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1200
1201                 for(i=0; i<4; i++)
1202                     if(!(b[i]->type&BLOCK_INTRA))
1203                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1204
1205                 if(init_rd != best_rd)
1206                     change++;
1207             }
1208         }
1209         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1210     }
1211 }
1212
1213 static void encode_blocks(SnowContext *s, int search){
1214     int x, y;
1215     int w= s->b_width;
1216     int h= s->b_height;
1217
1218     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1219         iterative_me(s);
1220
1221     for(y=0; y<h; y++){
1222         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1223             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1224             return;
1225         }
1226         for(x=0; x<w; x++){
1227             if(s->motion_est == FF_ME_ITER || !search)
1228                 encode_q_branch2(s, 0, x, y);
1229             else
1230                 encode_q_branch (s, 0, x, y);
1231         }
1232     }
1233 }
1234
1235 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1236     const int w= b->width;
1237     const int h= b->height;
1238     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1239     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1240     int x,y, thres1, thres2;
1241
1242     if(s->qlog == LOSSLESS_QLOG){
1243         for(y=0; y<h; y++)
1244             for(x=0; x<w; x++)
1245                 dst[x + y*stride]= src[x + y*stride];
1246         return;
1247     }
1248
1249     bias= bias ? 0 : (3*qmul)>>3;
1250     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1251     thres2= 2*thres1;
1252
1253     if(!bias){
1254         for(y=0; y<h; y++){
1255             for(x=0; x<w; x++){
1256                 int i= src[x + y*stride];
1257
1258                 if((unsigned)(i+thres1) > thres2){
1259                     if(i>=0){
1260                         i<<= QEXPSHIFT;
1261                         i/= qmul; //FIXME optimize
1262                         dst[x + y*stride]=  i;
1263                     }else{
1264                         i= -i;
1265                         i<<= QEXPSHIFT;
1266                         i/= qmul; //FIXME optimize
1267                         dst[x + y*stride]= -i;
1268                     }
1269                 }else
1270                     dst[x + y*stride]= 0;
1271             }
1272         }
1273     }else{
1274         for(y=0; y<h; y++){
1275             for(x=0; x<w; x++){
1276                 int i= src[x + y*stride];
1277
1278                 if((unsigned)(i+thres1) > thres2){
1279                     if(i>=0){
1280                         i<<= QEXPSHIFT;
1281                         i= (i + bias) / qmul; //FIXME optimize
1282                         dst[x + y*stride]=  i;
1283                     }else{
1284                         i= -i;
1285                         i<<= QEXPSHIFT;
1286                         i= (i + bias) / qmul; //FIXME optimize
1287                         dst[x + y*stride]= -i;
1288                     }
1289                 }else
1290                     dst[x + y*stride]= 0;
1291             }
1292         }
1293     }
1294 }
1295
1296 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1297     const int w= b->width;
1298     const int h= b->height;
1299     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1300     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1301     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1302     int x,y;
1303
1304     if(s->qlog == LOSSLESS_QLOG) return;
1305
1306     for(y=0; y<h; y++){
1307         for(x=0; x<w; x++){
1308             int i= src[x + y*stride];
1309             if(i<0){
1310                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1311             }else if(i>0){
1312                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1313             }
1314         }
1315     }
1316 }
1317
1318 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1319     const int w= b->width;
1320     const int h= b->height;
1321     int x,y;
1322
1323     for(y=h-1; y>=0; y--){
1324         for(x=w-1; x>=0; x--){
1325             int i= x + y*stride;
1326
1327             if(x){
1328                 if(use_median){
1329                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1330                     else  src[i] -= src[i - 1];
1331                 }else{
1332                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1333                     else  src[i] -= src[i - 1];
1334                 }
1335             }else{
1336                 if(y) src[i] -= src[i - stride];
1337             }
1338         }
1339     }
1340 }
1341
1342 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1343     const int w= b->width;
1344     const int h= b->height;
1345     int x,y;
1346
1347     for(y=0; y<h; y++){
1348         for(x=0; x<w; x++){
1349             int i= x + y*stride;
1350
1351             if(x){
1352                 if(use_median){
1353                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1354                     else  src[i] += src[i - 1];
1355                 }else{
1356                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1357                     else  src[i] += src[i - 1];
1358                 }
1359             }else{
1360                 if(y) src[i] += src[i - stride];
1361             }
1362         }
1363     }
1364 }
1365
1366 static void encode_qlogs(SnowContext *s){
1367     int plane_index, level, orientation;
1368
1369     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1370         for(level=0; level<s->spatial_decomposition_count; level++){
1371             for(orientation=level ? 1:0; orientation<4; orientation++){
1372                 if(orientation==2) continue;
1373                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1374             }
1375         }
1376     }
1377 }
1378
1379 static void encode_header(SnowContext *s){
1380     int plane_index, i;
1381     uint8_t kstate[32];
1382
1383     memset(kstate, MID_STATE, sizeof(kstate));
1384
1385     put_rac(&s->c, kstate, s->keyframe);
1386     if(s->keyframe || s->always_reset){
1387         ff_snow_reset_contexts(s);
1388         s->last_spatial_decomposition_type=
1389         s->last_qlog=
1390         s->last_qbias=
1391         s->last_mv_scale=
1392         s->last_block_max_depth= 0;
1393         for(plane_index=0; plane_index<2; plane_index++){
1394             Plane *p= &s->plane[plane_index];
1395             p->last_htaps=0;
1396             p->last_diag_mc=0;
1397             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1398         }
1399     }
1400     if(s->keyframe){
1401         put_symbol(&s->c, s->header_state, s->version, 0);
1402         put_rac(&s->c, s->header_state, s->always_reset);
1403         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1404         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1405         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1406         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1407         if (s->nb_planes > 2) {
1408             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1409             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1410         }
1411         put_rac(&s->c, s->header_state, s->spatial_scalability);
1412 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1413         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1414
1415         encode_qlogs(s);
1416     }
1417
1418     if(!s->keyframe){
1419         int update_mc=0;
1420         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1421             Plane *p= &s->plane[plane_index];
1422             update_mc |= p->last_htaps   != p->htaps;
1423             update_mc |= p->last_diag_mc != p->diag_mc;
1424             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1425         }
1426         put_rac(&s->c, s->header_state, update_mc);
1427         if(update_mc){
1428             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1429                 Plane *p= &s->plane[plane_index];
1430                 put_rac(&s->c, s->header_state, p->diag_mc);
1431                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1432                 for(i= p->htaps/2; i; i--)
1433                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1434             }
1435         }
1436         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1437             put_rac(&s->c, s->header_state, 1);
1438             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1439             encode_qlogs(s);
1440         }else
1441             put_rac(&s->c, s->header_state, 0);
1442     }
1443
1444     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1445     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1446     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1447     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1448     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1449
1450 }
1451
1452 static void update_last_header_values(SnowContext *s){
1453     int plane_index;
1454
1455     if(!s->keyframe){
1456         for(plane_index=0; plane_index<2; plane_index++){
1457             Plane *p= &s->plane[plane_index];
1458             p->last_diag_mc= p->diag_mc;
1459             p->last_htaps  = p->htaps;
1460             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1461         }
1462     }
1463
1464     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1465     s->last_qlog                        = s->qlog;
1466     s->last_qbias                       = s->qbias;
1467     s->last_mv_scale                    = s->mv_scale;
1468     s->last_block_max_depth             = s->block_max_depth;
1469     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1470 }
1471
1472 static int qscale2qlog(int qscale){
1473     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1474            + 61*QROOT/8; ///< 64 > 60
1475 }
1476
1477 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1478 {
1479     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1480      * FIXME we know exact mv bits at this point,
1481      * but ratecontrol isn't set up to include them. */
1482     uint32_t coef_sum= 0;
1483     int level, orientation, delta_qlog;
1484
1485     for(level=0; level<s->spatial_decomposition_count; level++){
1486         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1487             SubBand *b= &s->plane[0].band[level][orientation];
1488             IDWTELEM *buf= b->ibuf;
1489             const int w= b->width;
1490             const int h= b->height;
1491             const int stride= b->stride;
1492             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1493             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1494             const int qdiv= (1<<16)/qmul;
1495             int x, y;
1496             //FIXME this is ugly
1497             for(y=0; y<h; y++)
1498                 for(x=0; x<w; x++)
1499                     buf[x+y*stride]= b->buf[x+y*stride];
1500             if(orientation==0)
1501                 decorrelate(s, b, buf, stride, 1, 0);
1502             for(y=0; y<h; y++)
1503                 for(x=0; x<w; x++)
1504                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1505         }
1506     }
1507
1508     /* ugly, ratecontrol just takes a sqrt again */
1509     av_assert0(coef_sum < INT_MAX);
1510     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1511
1512     if(pict->pict_type == AV_PICTURE_TYPE_I){
1513         s->m.current_picture.mb_var_sum= coef_sum;
1514         s->m.current_picture.mc_mb_var_sum= 0;
1515     }else{
1516         s->m.current_picture.mc_mb_var_sum= coef_sum;
1517         s->m.current_picture.mb_var_sum= 0;
1518     }
1519
1520     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1521     if (pict->quality < 0)
1522         return INT_MIN;
1523     s->lambda= pict->quality * 3/2;
1524     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1525     s->qlog+= delta_qlog;
1526     return delta_qlog;
1527 }
1528
1529 static void calculate_visual_weight(SnowContext *s, Plane *p){
1530     int width = p->width;
1531     int height= p->height;
1532     int level, orientation, x, y;
1533
1534     for(level=0; level<s->spatial_decomposition_count; level++){
1535         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1536             SubBand *b= &p->band[level][orientation];
1537             IDWTELEM *ibuf= b->ibuf;
1538             int64_t error=0;
1539
1540             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1541             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1542             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1543             for(y=0; y<height; y++){
1544                 for(x=0; x<width; x++){
1545                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1546                     error += d*d;
1547                 }
1548             }
1549
1550             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1551         }
1552     }
1553 }
1554
1555 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1556                         const AVFrame *pict, int *got_packet)
1557 {
1558     SnowContext *s = avctx->priv_data;
1559     RangeCoder * const c= &s->c;
1560     AVFrame *pic = pict;
1561     const int width= s->avctx->width;
1562     const int height= s->avctx->height;
1563     int level, orientation, plane_index, i, y, ret;
1564     uint8_t rc_header_bak[sizeof(s->header_state)];
1565     uint8_t rc_block_bak[sizeof(s->block_state)];
1566
1567     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + AV_INPUT_BUFFER_MIN_SIZE, 0)) < 0)
1568         return ret;
1569
1570     ff_init_range_encoder(c, pkt->data, pkt->size);
1571     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1572
1573     for(i=0; i < s->nb_planes; i++){
1574         int hshift= i ? s->chroma_h_shift : 0;
1575         int vshift= i ? s->chroma_v_shift : 0;
1576         for(y=0; y<FF_CEIL_RSHIFT(height, vshift); y++)
1577             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1578                    &pict->data[i][y * pict->linesize[i]],
1579                    FF_CEIL_RSHIFT(width, hshift));
1580         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1581                                 FF_CEIL_RSHIFT(width, hshift), FF_CEIL_RSHIFT(height, vshift),
1582                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1583                                 EDGE_TOP | EDGE_BOTTOM);
1584
1585     }
1586     emms_c();
1587     s->new_picture = pict;
1588
1589     s->m.picture_number= avctx->frame_number;
1590     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1591         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1592         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1593         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1594             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1595             if (pic->quality < 0)
1596                 return -1;
1597         }
1598     }else{
1599         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1600         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1601     }
1602
1603     if(s->pass1_rc && avctx->frame_number == 0)
1604         pic->quality = 2*FF_QP2LAMBDA;
1605     if (pic->quality) {
1606         s->qlog   = qscale2qlog(pic->quality);
1607         s->lambda = pic->quality * 3/2;
1608     }
1609     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1610         s->qlog= LOSSLESS_QLOG;
1611         s->lambda = 0;
1612     }//else keep previous frame's qlog until after motion estimation
1613
1614     if (s->current_picture->data[0]
1615 #if FF_API_EMU_EDGE
1616         && !(s->avctx->flags&CODEC_FLAG_EMU_EDGE)
1617 #endif
1618         ) {
1619         int w = s->avctx->width;
1620         int h = s->avctx->height;
1621
1622         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1623                                 s->current_picture->linesize[0], w   , h   ,
1624                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1625         if (s->current_picture->data[2]) {
1626             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1627                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1628                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1629             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1630                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1631                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1632         }
1633     }
1634
1635     ff_snow_frame_start(s);
1636     av_frame_unref(avctx->coded_frame);
1637     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1638     if (ret < 0)
1639         return ret;
1640
1641     s->m.current_picture_ptr= &s->m.current_picture;
1642     s->m.current_picture.f = s->current_picture;
1643     s->m.current_picture.f->pts = pict->pts;
1644     if(pic->pict_type == AV_PICTURE_TYPE_P){
1645         int block_width = (width +15)>>4;
1646         int block_height= (height+15)>>4;
1647         int stride= s->current_picture->linesize[0];
1648
1649         av_assert0(s->current_picture->data[0]);
1650         av_assert0(s->last_picture[0]->data[0]);
1651
1652         s->m.avctx= s->avctx;
1653         s->m.   last_picture.f = s->last_picture[0];
1654         s->m.    new_picture.f = s->input_picture;
1655         s->m.   last_picture_ptr= &s->m.   last_picture;
1656         s->m.linesize = stride;
1657         s->m.uvlinesize= s->current_picture->linesize[1];
1658         s->m.width = width;
1659         s->m.height= height;
1660         s->m.mb_width = block_width;
1661         s->m.mb_height= block_height;
1662         s->m.mb_stride=   s->m.mb_width+1;
1663         s->m.b8_stride= 2*s->m.mb_width+1;
1664         s->m.f_code=1;
1665         s->m.pict_type = pic->pict_type;
1666 #if FF_API_MOTION_EST
1667         s->m.me_method= s->avctx->me_method;
1668 #endif
1669         s->m.motion_est= s->motion_est;
1670         s->m.me.scene_change_score=0;
1671         s->m.me.dia_size = avctx->dia_size;
1672         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1673         s->m.out_format= FMT_H263;
1674         s->m.unrestricted_mv= 1;
1675
1676         s->m.lambda = s->lambda;
1677         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1678         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1679
1680         s->m.mecc= s->mecc; //move
1681         s->m.qdsp= s->qdsp; //move
1682         s->m.hdsp = s->hdsp;
1683         ff_init_me(&s->m);
1684         s->hdsp = s->m.hdsp;
1685         s->mecc= s->m.mecc;
1686     }
1687
1688     if(s->pass1_rc){
1689         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1690         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1691     }
1692
1693 redo_frame:
1694
1695     s->spatial_decomposition_count= 5;
1696
1697     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1698           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1699         s->spatial_decomposition_count--;
1700
1701     if (s->spatial_decomposition_count <= 0) {
1702         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1703         return AVERROR(EINVAL);
1704     }
1705
1706     s->m.pict_type = pic->pict_type;
1707     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1708
1709     ff_snow_common_init_after_header(avctx);
1710
1711     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1712         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1713             calculate_visual_weight(s, &s->plane[plane_index]);
1714         }
1715     }
1716
1717     encode_header(s);
1718     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1719     encode_blocks(s, 1);
1720     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1721
1722     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1723         Plane *p= &s->plane[plane_index];
1724         int w= p->width;
1725         int h= p->height;
1726         int x, y;
1727 //        int bits= put_bits_count(&s->c.pb);
1728
1729         if (!s->memc_only) {
1730             //FIXME optimize
1731             if(pict->data[plane_index]) //FIXME gray hack
1732                 for(y=0; y<h; y++){
1733                     for(x=0; x<w; x++){
1734                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1735                     }
1736                 }
1737             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1738
1739             if(   plane_index==0
1740                && pic->pict_type == AV_PICTURE_TYPE_P
1741                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1742                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1743                 ff_init_range_encoder(c, pkt->data, pkt->size);
1744                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1745                 pic->pict_type= AV_PICTURE_TYPE_I;
1746                 s->keyframe=1;
1747                 s->current_picture->key_frame=1;
1748                 goto redo_frame;
1749             }
1750
1751             if(s->qlog == LOSSLESS_QLOG){
1752                 for(y=0; y<h; y++){
1753                     for(x=0; x<w; x++){
1754                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1755                     }
1756                 }
1757             }else{
1758                 for(y=0; y<h; y++){
1759                     for(x=0; x<w; x++){
1760                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1761                     }
1762                 }
1763             }
1764
1765             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1766
1767             if(s->pass1_rc && plane_index==0){
1768                 int delta_qlog = ratecontrol_1pass(s, pic);
1769                 if (delta_qlog <= INT_MIN)
1770                     return -1;
1771                 if(delta_qlog){
1772                     //reordering qlog in the bitstream would eliminate this reset
1773                     ff_init_range_encoder(c, pkt->data, pkt->size);
1774                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1775                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1776                     encode_header(s);
1777                     encode_blocks(s, 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                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1786                     if(orientation==0)
1787                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1788                     if (!s->no_bitstream)
1789                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1790                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1791                     if(orientation==0)
1792                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1793                 }
1794             }
1795
1796             for(level=0; level<s->spatial_decomposition_count; level++){
1797                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1798                     SubBand *b= &p->band[level][orientation];
1799
1800                     dequantize(s, b, b->ibuf, b->stride);
1801                 }
1802             }
1803
1804             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1805             if(s->qlog == LOSSLESS_QLOG){
1806                 for(y=0; y<h; y++){
1807                     for(x=0; x<w; x++){
1808                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1809                     }
1810                 }
1811             }
1812             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1813         }else{
1814             //ME/MC only
1815             if(pic->pict_type == AV_PICTURE_TYPE_I){
1816                 for(y=0; y<h; y++){
1817                     for(x=0; x<w; x++){
1818                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1819                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1820                     }
1821                 }
1822             }else{
1823                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1824                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1825             }
1826         }
1827         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1828             int64_t error= 0;
1829
1830             if(pict->data[plane_index]) //FIXME gray hack
1831                 for(y=0; y<h; y++){
1832                     for(x=0; x<w; x++){
1833                         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];
1834                         error += d*d;
1835                     }
1836                 }
1837             s->avctx->error[plane_index] += error;
1838             s->current_picture->error[plane_index] = error;
1839         }
1840
1841     }
1842
1843     update_last_header_values(s);
1844
1845     ff_snow_release_buffer(avctx);
1846
1847     s->current_picture->coded_picture_number = avctx->frame_number;
1848     s->current_picture->pict_type = pict->pict_type;
1849     s->current_picture->quality = pict->quality;
1850     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1851     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1852     s->m.current_picture.f->display_picture_number =
1853     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1854     s->m.current_picture.f->quality                = pic->quality;
1855     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1856     if(s->pass1_rc)
1857         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1858             return -1;
1859     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1860         ff_write_pass1_stats(&s->m);
1861     s->m.last_pict_type = s->m.pict_type;
1862     avctx->frame_bits = s->m.frame_bits;
1863     avctx->mv_bits = s->m.mv_bits;
1864     avctx->misc_bits = s->m.misc_bits;
1865     avctx->p_tex_bits = s->m.p_tex_bits;
1866
1867     emms_c();
1868
1869     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1870                                    s->current_picture->error,
1871                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1872                                    s->current_picture->pict_type);
1873
1874     pkt->size = ff_rac_terminate(c);
1875     if (s->current_picture->key_frame)
1876         pkt->flags |= AV_PKT_FLAG_KEY;
1877     *got_packet = 1;
1878
1879     return 0;
1880 }
1881
1882 static av_cold int encode_end(AVCodecContext *avctx)
1883 {
1884     SnowContext *s = avctx->priv_data;
1885
1886     ff_snow_common_end(s);
1887     ff_rate_control_uninit(&s->m);
1888     av_frame_free(&s->input_picture);
1889     av_freep(&avctx->stats_out);
1890
1891     return 0;
1892 }
1893
1894 #define OFFSET(x) offsetof(SnowContext, x)
1895 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1896 static const AVOption options[] = {
1897     FF_MPV_COMMON_OPTS
1898     { "iter",           NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, FF_MPV_OPT_FLAGS, "motion_est" },
1899     { "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 },
1900     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1901     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1902     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1903     { NULL },
1904 };
1905
1906 static const AVClass snowenc_class = {
1907     .class_name = "snow encoder",
1908     .item_name  = av_default_item_name,
1909     .option     = options,
1910     .version    = LIBAVUTIL_VERSION_INT,
1911 };
1912
1913 AVCodec ff_snow_encoder = {
1914     .name           = "snow",
1915     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1916     .type           = AVMEDIA_TYPE_VIDEO,
1917     .id             = AV_CODEC_ID_SNOW,
1918     .priv_data_size = sizeof(SnowContext),
1919     .init           = encode_init,
1920     .encode2        = encode_frame,
1921     .close          = encode_end,
1922     .pix_fmts       = (const enum AVPixelFormat[]){
1923         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1924         AV_PIX_FMT_GRAY8,
1925         AV_PIX_FMT_NONE
1926     },
1927     .priv_class     = &snowenc_class,
1928     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1929                       FF_CODEC_CAP_INIT_CLEANUP,
1930 };
1931
1932
1933 #ifdef TEST
1934 #undef malloc
1935 #undef free
1936 #undef printf
1937
1938 #include "libavutil/lfg.h"
1939 #include "libavutil/mathematics.h"
1940
1941 int main(void){
1942 #define width  256
1943 #define height 256
1944     int buffer[2][width*height];
1945     SnowContext s;
1946     int i;
1947     AVLFG prng;
1948     s.spatial_decomposition_count=6;
1949     s.spatial_decomposition_type=1;
1950
1951     s.temp_dwt_buffer  = av_mallocz_array(width, sizeof(DWTELEM));
1952     s.temp_idwt_buffer = av_mallocz_array(width, sizeof(IDWTELEM));
1953
1954     if (!s.temp_dwt_buffer || !s.temp_idwt_buffer) {
1955         fprintf(stderr, "Failed to allocate memory\n");
1956         return 1;
1957     }
1958
1959     av_lfg_init(&prng, 1);
1960
1961     printf("testing 5/3 DWT\n");
1962     for(i=0; i<width*height; i++)
1963         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1964
1965     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1966     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1967
1968     for(i=0; i<width*height; i++)
1969         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1970
1971     printf("testing 9/7 DWT\n");
1972     s.spatial_decomposition_type=0;
1973     for(i=0; i<width*height; i++)
1974         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1975
1976     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1977     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1978
1979     for(i=0; i<width*height; i++)
1980         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1981
1982     {
1983     int level, orientation, x, y;
1984     int64_t errors[8][4];
1985     int64_t g=0;
1986
1987         memset(errors, 0, sizeof(errors));
1988         s.spatial_decomposition_count=3;
1989         s.spatial_decomposition_type=0;
1990         for(level=0; level<s.spatial_decomposition_count; level++){
1991             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1992                 int w= width  >> (s.spatial_decomposition_count-level);
1993                 int h= height >> (s.spatial_decomposition_count-level);
1994                 int stride= width  << (s.spatial_decomposition_count-level);
1995                 DWTELEM *buf= buffer[0];
1996                 int64_t error=0;
1997
1998                 if(orientation&1) buf+=w;
1999                 if(orientation>1) buf+=stride>>1;
2000
2001                 memset(buffer[0], 0, sizeof(int)*width*height);
2002                 buf[w/2 + h/2*stride]= 256*256;
2003                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2004                 for(y=0; y<height; y++){
2005                     for(x=0; x<width; x++){
2006                         int64_t d= buffer[0][x + y*width];
2007                         error += d*d;
2008                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2009                     }
2010                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2011                 }
2012                 error= (int)(sqrt(error)+0.5);
2013                 errors[level][orientation]= error;
2014                 if(g) g=av_gcd(g, error);
2015                 else g= error;
2016             }
2017         }
2018         printf("static int const visual_weight[][4]={\n");
2019         for(level=0; level<s.spatial_decomposition_count; level++){
2020             printf("  {");
2021             for(orientation=0; orientation<4; orientation++){
2022                 printf("%8"PRId64",", errors[level][orientation]/g);
2023             }
2024             printf("},\n");
2025         }
2026         printf("};\n");
2027         {
2028             int level=2;
2029             int w= width  >> (s.spatial_decomposition_count-level);
2030             //int h= height >> (s.spatial_decomposition_count-level);
2031             int stride= width  << (s.spatial_decomposition_count-level);
2032             DWTELEM *buf= buffer[0];
2033             int64_t error=0;
2034
2035             buf+=w;
2036             buf+=stride>>1;
2037
2038             memset(buffer[0], 0, sizeof(int)*width*height);
2039             for(y=0; y<height; y++){
2040                 for(x=0; x<width; x++){
2041                     int tab[4]={0,2,3,1};
2042                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2043                 }
2044             }
2045             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2046             for(y=0; y<height; y++){
2047                 for(x=0; x<width; x++){
2048                     int64_t d= buffer[0][x + y*width];
2049                     error += d*d;
2050                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2051                 }
2052                 if(FFABS(height/2-y)<9) printf("\n");
2053             }
2054         }
2055
2056     }
2057     return 0;
2058 }
2059 #endif /* TEST */