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/libm.h"
  23 #include "libavutil/log.h"
  24 #include "libavutil/opt.h"
  25 #include "avcodec.h"
  26 #include "internal.h"
  27 #include "snow_dwt.h"
  28 #include "snow.h"
  29
  30 #include "rangecoder.h"
  31 #include "mathops.h"
  32
  33 #include "mpegvideo.h"
  34 #include "h263.h"
  35
  36 static av_cold int encode_init(AVCodecContext *avctx)
  37 {
  38     SnowContext *s = avctx->priv_data;
  39     int plane_index, ret;
  40     int i;
  41
  42 #if FF_API_PRIVATE_OPT
  43 FF_DISABLE_DEPRECATION_WARNINGS
  44     if (avctx->prediction_method)
  45         s->pred = avctx->prediction_method;
  46 FF_ENABLE_DEPRECATION_WARNINGS
  47 #endif
  48
  49     if(s->pred == DWT_97
  50        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
  51        && avctx->global_quality == 0){
  52         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  53         return AVERROR(EINVAL);
  54     }
  55
  56     s->spatial_decomposition_type= s->pred; //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     s->m.lmin    = avctx->mb_lmin;
  82     s->m.lmax    = avctx->mb_lmax;
  83
  84     s->m.me.temp      =
  85     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  86     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  87     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  88     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  89     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
  90         return AVERROR(ENOMEM);
  91
  92     ff_h263_encode_init(&s->m); //mv_penalty
  93
  94     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
  95
  96     if(avctx->flags&AV_CODEC_FLAG_PASS1){
  97         if(!avctx->stats_out)
  98             avctx->stats_out = av_mallocz(256);
  99
 100         if (!avctx->stats_out)
 101             return AVERROR(ENOMEM);
 102     }
 103     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
 104         ret = ff_rate_control_init(&s->m);
 105         if(ret < 0)
 106             return ret;
 107     }
 108     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
 109
 110     switch(avctx->pix_fmt){
 111     case AV_PIX_FMT_YUV444P:
 112 //    case AV_PIX_FMT_YUV422P:
 113     case AV_PIX_FMT_YUV420P:
 114 //    case AV_PIX_FMT_YUV411P:
 115     case AV_PIX_FMT_YUV410P:
 116         s->nb_planes = 3;
 117         s->colorspace_type= 0;
 118         break;
 119     case AV_PIX_FMT_GRAY8:
 120         s->nb_planes = 1;
 121         s->colorspace_type = 1;
 122         break;
 123 /*    case AV_PIX_FMT_RGB32:
 124         s->colorspace= 1;
 125         break;*/
 126     default:
 127         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 128         return AVERROR_PATCHWELCOME;
 129     }
 130     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 131
 132     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 133     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 134
 135     s->input_picture = av_frame_alloc();
 136     if (!s->input_picture)
 137         return AVERROR(ENOMEM);
 138
 139     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 140         return ret;
 141
 142     if(s->motion_est == FF_ME_ITER){
 143         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 144         for(i=0; i<s->max_ref_frames; i++){
 145             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 146             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 147             if (!s->ref_mvs[i] || !s->ref_scores[i])
 148                 return AVERROR(ENOMEM);
 149         }
 150     }
 151
 152     return 0;
 153 }
 154
 155 //near copy & paste from dsputil, FIXME
 156 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 157 {
 158     int s, i, j;
 159
 160     s = 0;
 161     for (i = 0; i < h; i++) {
 162         for (j = 0; j < w; j++) {
 163             s += pix[0];
 164             pix ++;
 165         }
 166         pix += line_size - w;
 167     }
 168     return s;
 169 }
 170
 171 //near copy & paste from dsputil, FIXME
 172 static int pix_norm1(uint8_t * pix, int line_size, int w)
 173 {
 174     int s, i, j;
 175     uint32_t *sq = ff_square_tab + 256;
 176
 177     s = 0;
 178     for (i = 0; i < w; i++) {
 179         for (j = 0; j < w; j ++) {
 180             s += sq[pix[0]];
 181             pix ++;
 182         }
 183         pix += line_size - w;
 184     }
 185     return s;
 186 }
 187
 188 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 189     switch(type&0xFF){
 190     default:
 191     case FF_CMP_SAD:
 192         return lambda>>FF_LAMBDA_SHIFT;
 193     case FF_CMP_DCT:
 194         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 195     case FF_CMP_W53:
 196         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 197     case FF_CMP_W97:
 198         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 199     case FF_CMP_SATD:
 200     case FF_CMP_DCT264:
 201         return (2*lambda)>>FF_LAMBDA_SHIFT;
 202     case FF_CMP_RD:
 203     case FF_CMP_PSNR:
 204     case FF_CMP_SSE:
 205     case FF_CMP_NSSE:
 206         return lambda2>>FF_LAMBDA_SHIFT;
 207     case FF_CMP_BIT:
 208         return 1;
 209     }
 210 }
 211
 212 //FIXME copy&paste
 213 #define P_LEFT P[1]
 214 #define P_TOP P[2]
 215 #define P_TOPRIGHT P[3]
 216 #define P_MEDIAN P[4]
 217 #define P_MV1 P[9]
 218 #define FLAG_QPEL   1 //must be 1
 219
 220 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 221     uint8_t p_buffer[1024];
 222     uint8_t i_buffer[1024];
 223     uint8_t p_state[sizeof(s->block_state)];
 224     uint8_t i_state[sizeof(s->block_state)];
 225     RangeCoder pc, ic;
 226     uint8_t *pbbak= s->c.bytestream;
 227     uint8_t *pbbak_start= s->c.bytestream_start;
 228     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 229     const int w= s->b_width  << s->block_max_depth;
 230     const int h= s->b_height << s->block_max_depth;
 231     const int rem_depth= s->block_max_depth - level;
 232     const int index= (x + y*w) << rem_depth;
 233     const int block_w= 1<<(LOG2_MB_SIZE - level);
 234     int trx= (x+1)<<rem_depth;
 235     int try= (y+1)<<rem_depth;
 236     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 237     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 238     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 239     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 240     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 241     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 242     int pl = left->color[0];
 243     int pcb= left->color[1];
 244     int pcr= left->color[2];
 245     int pmx, pmy;
 246     int mx=0, my=0;
 247     int l,cr,cb;
 248     const int stride= s->current_picture->linesize[0];
 249     const int uvstride= s->current_picture->linesize[1];
 250     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 251                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 252                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 253     int P[10][2];
 254     int16_t last_mv[3][2];
 255     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
 256     const int shift= 1+qpel;
 257     MotionEstContext *c= &s->m.me;
 258     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 259     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 260     int my_context= av_log2(2*FFABS(left->my - top->my));
 261     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 262     int ref, best_ref, ref_score, ref_mx, ref_my;
 263
 264     av_assert0(sizeof(s->block_state) >= 256);
 265     if(s->keyframe){
 266         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 267         return 0;
 268     }
 269
 270 //    clip predictors / edge ?
 271
 272     P_LEFT[0]= left->mx;
 273     P_LEFT[1]= left->my;
 274     P_TOP [0]= top->mx;
 275     P_TOP [1]= top->my;
 276     P_TOPRIGHT[0]= tr->mx;
 277     P_TOPRIGHT[1]= tr->my;
 278
 279     last_mv[0][0]= s->block[index].mx;
 280     last_mv[0][1]= s->block[index].my;
 281     last_mv[1][0]= right->mx;
 282     last_mv[1][1]= right->my;
 283     last_mv[2][0]= bottom->mx;
 284     last_mv[2][1]= bottom->my;
 285
 286     s->m.mb_stride=2;
 287     s->m.mb_x=
 288     s->m.mb_y= 0;
 289     c->skip= 0;
 290
 291     av_assert1(c->  stride ==   stride);
 292     av_assert1(c->uvstride == uvstride);
 293
 294     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 295     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 296     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 297     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
 298
 299     c->xmin = - x*block_w - 16+3;
 300     c->ymin = - y*block_w - 16+3;
 301     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 302     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 303
 304     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 305     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 306     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 307     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 308     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 309     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 310     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 311
 312     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 313     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 314
 315     if (!y) {
 316         c->pred_x= P_LEFT[0];
 317         c->pred_y= P_LEFT[1];
 318     } else {
 319         c->pred_x = P_MEDIAN[0];
 320         c->pred_y = P_MEDIAN[1];
 321     }
 322
 323     score= INT_MAX;
 324     best_ref= 0;
 325     for(ref=0; ref<s->ref_frames; ref++){
 326         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 327
 328         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 329                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 330
 331         av_assert2(ref_mx >= c->xmin);
 332         av_assert2(ref_mx <= c->xmax);
 333         av_assert2(ref_my >= c->ymin);
 334         av_assert2(ref_my <= c->ymax);
 335
 336         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 337         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 338         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 339         if(s->ref_mvs[ref]){
 340             s->ref_mvs[ref][index][0]= ref_mx;
 341             s->ref_mvs[ref][index][1]= ref_my;
 342             s->ref_scores[ref][index]= ref_score;
 343         }
 344         if(score > ref_score){
 345             score= ref_score;
 346             best_ref= ref;
 347             mx= ref_mx;
 348             my= ref_my;
 349         }
 350     }
 351     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 352
 353   //  subpel search
 354     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 355     pc= s->c;
 356     pc.bytestream_start=
 357     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 358     memcpy(p_state, s->block_state, sizeof(s->block_state));
 359
 360     if(level!=s->block_max_depth)
 361         put_rac(&pc, &p_state[4 + s_context], 1);
 362     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 363     if(s->ref_frames > 1)
 364         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 365     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 366     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 367     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 368     p_len= pc.bytestream - pc.bytestream_start;
 369     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 370
 371     block_s= block_w*block_w;
 372     sum = pix_sum(current_data[0], stride, block_w, block_w);
 373     l= (sum + block_s/2)/block_s;
 374     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 375
 376     if (s->nb_planes > 2) {
 377         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 378         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 379         cb= (sum + block_s/2)/block_s;
 380     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 381         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 382         cr= (sum + block_s/2)/block_s;
 383     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 384     }else
 385         cb = cr = 0;
 386
 387     ic= s->c;
 388     ic.bytestream_start=
 389     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 390     memcpy(i_state, s->block_state, sizeof(s->block_state));
 391     if(level!=s->block_max_depth)
 392         put_rac(&ic, &i_state[4 + s_context], 1);
 393     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 394     put_symbol(&ic, &i_state[32],  l-pl , 1);
 395     if (s->nb_planes > 2) {
 396         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 397         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 398     }
 399     i_len= ic.bytestream - ic.bytestream_start;
 400     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 401
 402     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 403     av_assert1(iscore >= 0);
 404     av_assert1(l>=0 && l<=255);
 405     av_assert1(pl>=0 && pl<=255);
 406
 407     if(level==0){
 408         int varc= iscore >> 8;
 409         int vard= score >> 8;
 410         if (vard <= 64 || vard < varc)
 411             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 412         else
 413             c->scene_change_score+= s->m.qscale;
 414     }
 415
 416     if(level!=s->block_max_depth){
 417         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 418         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 419         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 420         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 421         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 422         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 423
 424         if(score2 < score && score2 < iscore)
 425             return score2;
 426     }
 427
 428     if(iscore < score){
 429         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 430         memcpy(pbbak, i_buffer, i_len);
 431         s->c= ic;
 432         s->c.bytestream_start= pbbak_start;
 433         s->c.bytestream= pbbak + i_len;
 434         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 435         memcpy(s->block_state, i_state, sizeof(s->block_state));
 436         return iscore;
 437     }else{
 438         memcpy(pbbak, p_buffer, p_len);
 439         s->c= pc;
 440         s->c.bytestream_start= pbbak_start;
 441         s->c.bytestream= pbbak + p_len;
 442         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 443         memcpy(s->block_state, p_state, sizeof(s->block_state));
 444         return score;
 445     }
 446 }
 447
 448 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 449     const int w= s->b_width  << s->block_max_depth;
 450     const int rem_depth= s->block_max_depth - level;
 451     const int index= (x + y*w) << rem_depth;
 452     int trx= (x+1)<<rem_depth;
 453     BlockNode *b= &s->block[index];
 454     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 455     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 456     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 457     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 458     int pl = left->color[0];
 459     int pcb= left->color[1];
 460     int pcr= left->color[2];
 461     int pmx, pmy;
 462     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 463     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 464     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 465     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 466
 467     if(s->keyframe){
 468         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 469         return;
 470     }
 471
 472     if(level!=s->block_max_depth){
 473         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 474             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 475         }else{
 476             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 477             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 478             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 479             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 480             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 481             return;
 482         }
 483     }
 484     if(b->type & BLOCK_INTRA){
 485         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 486         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 487         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 488         if (s->nb_planes > 2) {
 489             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 490             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 491         }
 492         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 493     }else{
 494         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 495         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 496         if(s->ref_frames > 1)
 497             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 498         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 499         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 500         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 501     }
 502 }
 503
 504 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 505     int i, x2, y2;
 506     Plane *p= &s->plane[plane_index];
 507     const int block_size = MB_SIZE >> s->block_max_depth;
 508     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 509     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 510     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 511     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 512     const int ref_stride= s->current_picture->linesize[plane_index];
 513     uint8_t *src= s-> input_picture->data[plane_index];
 514     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 515     const int b_stride = s->b_width << s->block_max_depth;
 516     const int w= p->width;
 517     const int h= p->height;
 518     int index= mb_x + mb_y*b_stride;
 519     BlockNode *b= &s->block[index];
 520     BlockNode backup= *b;
 521     int ab=0;
 522     int aa=0;
 523
 524     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 525
 526     b->type|= BLOCK_INTRA;
 527     b->color[plane_index]= 0;
 528     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 529
 530     for(i=0; i<4; i++){
 531         int mb_x2= mb_x + (i &1) - 1;
 532         int mb_y2= mb_y + (i>>1) - 1;
 533         int x= block_w*mb_x2 + block_w/2;
 534         int y= block_h*mb_y2 + block_h/2;
 535
 536         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 537                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 538
 539         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 540             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 541                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 542                 int obmc_v= obmc[index];
 543                 int d;
 544                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 545                 if(x<0) obmc_v += obmc[index + block_w];
 546                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 547                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 548                 //FIXME precalculate this or simplify it somehow else
 549
 550                 d = -dst[index] + (1<<(FRAC_BITS-1));
 551                 dst[index] = d;
 552                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 553                 aa += obmc_v * obmc_v; //FIXME precalculate this
 554             }
 555         }
 556     }
 557     *b= backup;
 558
 559     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 560 }
 561
 562 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 563     const int b_stride = s->b_width << s->block_max_depth;
 564     const int b_height = s->b_height<< s->block_max_depth;
 565     int index= x + y*b_stride;
 566     const BlockNode *b     = &s->block[index];
 567     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 568     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 569     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 570     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 571     int dmx, dmy;
 572 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 573 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 574
 575     if(x<0 || x>=b_stride || y>=b_height)
 576         return 0;
 577 /*
 578 1            0      0
 579 01X          1-2    1
 580 001XX        3-6    2-3
 581 0001XXX      7-14   4-7
 582 00001XXXX   15-30   8-15
 583 */
 584 //FIXME try accurate rate
 585 //FIXME intra and inter predictors if surrounding blocks are not the same type
 586     if(b->type & BLOCK_INTRA){
 587         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 588                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 589                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 590     }else{
 591         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 592         dmx-= b->mx;
 593         dmy-= b->my;
 594         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 595                     + av_log2(2*FFABS(dmy))
 596                     + av_log2(2*b->ref));
 597     }
 598 }
 599
 600 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 601     Plane *p= &s->plane[plane_index];
 602     const int block_size = MB_SIZE >> s->block_max_depth;
 603     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 604     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 605     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 606     const int ref_stride= s->current_picture->linesize[plane_index];
 607     uint8_t *dst= s->current_picture->data[plane_index];
 608     uint8_t *src= s->  input_picture->data[plane_index];
 609     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
 610     uint8_t *cur = s->scratchbuf;
 611     uint8_t *tmp = s->emu_edge_buffer;
 612     const int b_stride = s->b_width << s->block_max_depth;
 613     const int b_height = s->b_height<< s->block_max_depth;
 614     const int w= p->width;
 615     const int h= p->height;
 616     int distortion;
 617     int rate= 0;
 618     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 619     int sx= block_w*mb_x - block_w/2;
 620     int sy= block_h*mb_y - block_h/2;
 621     int x0= FFMAX(0,-sx);
 622     int y0= FFMAX(0,-sy);
 623     int x1= FFMIN(block_w*2, w-sx);
 624     int y1= FFMIN(block_h*2, h-sy);
 625     int i,x,y;
 626
 627     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 628
 629     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);
 630
 631     for(y=y0; y<y1; y++){
 632         const uint8_t *obmc1= obmc_edged[y];
 633         const IDWTELEM *pred1 = pred + y*obmc_stride;
 634         uint8_t *cur1 = cur + y*ref_stride;
 635         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 636         for(x=x0; x<x1; x++){
 637 #if FRAC_BITS >= LOG2_OBMC_MAX
 638             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 639 #else
 640             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 641 #endif
 642             v = (v + pred1[x]) >> FRAC_BITS;
 643             if(v&(~255)) v= ~(v>>31);
 644             dst1[x] = v;
 645         }
 646     }
 647
 648     /* copy the regions where obmc[] = (uint8_t)256 */
 649     if(LOG2_OBMC_MAX == 8
 650         && (mb_x == 0 || mb_x == b_stride-1)
 651         && (mb_y == 0 || mb_y == b_height-1)){
 652         if(mb_x == 0)
 653             x1 = block_w;
 654         else
 655             x0 = block_w;
 656         if(mb_y == 0)
 657             y1 = block_h;
 658         else
 659             y0 = block_h;
 660         for(y=y0; y<y1; y++)
 661             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 662     }
 663
 664     if(block_w==16){
 665         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 666         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 667         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 668          * So improving the score of one block is not strictly guaranteed
 669          * to improve the score of the whole frame, thus iterative motion
 670          * estimation does not always converge. */
 671         if(s->avctx->me_cmp == FF_CMP_W97)
 672             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 673         else if(s->avctx->me_cmp == FF_CMP_W53)
 674             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 675         else{
 676             distortion = 0;
 677             for(i=0; i<4; i++){
 678                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 679                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 680             }
 681         }
 682     }else{
 683         av_assert2(block_w==8);
 684         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 685     }
 686
 687     if(plane_index==0){
 688         for(i=0; i<4; i++){
 689 /* ..RRr
 690  * .RXx.
 691  * rxx..
 692  */
 693             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 694         }
 695         if(mb_x == b_stride-2)
 696             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 697     }
 698     return distortion + rate*penalty_factor;
 699 }
 700
 701 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 702     int i, y2;
 703     Plane *p= &s->plane[plane_index];
 704     const int block_size = MB_SIZE >> s->block_max_depth;
 705     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 706     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 707     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 708     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 709     const int ref_stride= s->current_picture->linesize[plane_index];
 710     uint8_t *dst= s->current_picture->data[plane_index];
 711     uint8_t *src= s-> input_picture->data[plane_index];
 712     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 713     // const has only been removed from zero_dst to suppress a warning
 714     static IDWTELEM zero_dst[4096]; //FIXME
 715     const int b_stride = s->b_width << s->block_max_depth;
 716     const int w= p->width;
 717     const int h= p->height;
 718     int distortion= 0;
 719     int rate= 0;
 720     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 721
 722     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 723
 724     for(i=0; i<9; i++){
 725         int mb_x2= mb_x + (i%3) - 1;
 726         int mb_y2= mb_y + (i/3) - 1;
 727         int x= block_w*mb_x2 + block_w/2;
 728         int y= block_h*mb_y2 + block_h/2;
 729
 730         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 731                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 732
 733         //FIXME find a cleaner/simpler way to skip the outside stuff
 734         for(y2= y; y2<0; y2++)
 735             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 736         for(y2= h; y2<y+block_h; y2++)
 737             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 738         if(x<0){
 739             for(y2= y; y2<y+block_h; y2++)
 740                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 741         }
 742         if(x+block_w > w){
 743             for(y2= y; y2<y+block_h; y2++)
 744                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 745         }
 746
 747         av_assert1(block_w== 8 || block_w==16);
 748         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 749     }
 750
 751     if(plane_index==0){
 752         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 753         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 754
 755 /* ..RRRr
 756  * .RXXx.
 757  * .RXXx.
 758  * rxxx.
 759  */
 760         if(merged)
 761             rate = get_block_bits(s, mb_x, mb_y, 2);
 762         for(i=merged?4:0; i<9; i++){
 763             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 764             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 765         }
 766     }
 767     return distortion + rate*penalty_factor;
 768 }
 769
 770 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 771     const int w= b->width;
 772     const int h= b->height;
 773     int x, y;
 774
 775     if(1){
 776         int run=0;
 777         int *runs = s->run_buffer;
 778         int run_index=0;
 779         int max_index;
 780
 781         for(y=0; y<h; y++){
 782             for(x=0; x<w; x++){
 783                 int v, p=0;
 784                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 785                 v= src[x + y*stride];
 786
 787                 if(y){
 788                     t= src[x + (y-1)*stride];
 789                     if(x){
 790                         lt= src[x - 1 + (y-1)*stride];
 791                     }
 792                     if(x + 1 < w){
 793                         rt= src[x + 1 + (y-1)*stride];
 794                     }
 795                 }
 796                 if(x){
 797                     l= src[x - 1 + y*stride];
 798                     /*if(x > 1){
 799                         if(orientation==1) ll= src[y + (x-2)*stride];
 800                         else               ll= src[x - 2 + y*stride];
 801                     }*/
 802                 }
 803                 if(parent){
 804                     int px= x>>1;
 805                     int py= y>>1;
 806                     if(px<b->parent->width && py<b->parent->height)
 807                         p= parent[px + py*2*stride];
 808                 }
 809                 if(!(/*ll|*/l|lt|t|rt|p)){
 810                     if(v){
 811                         runs[run_index++]= run;
 812                         run=0;
 813                     }else{
 814                         run++;
 815                     }
 816                 }
 817             }
 818         }
 819         max_index= run_index;
 820         runs[run_index++]= run;
 821         run_index=0;
 822         run= runs[run_index++];
 823
 824         put_symbol2(&s->c, b->state[30], max_index, 0);
 825         if(run_index <= max_index)
 826             put_symbol2(&s->c, b->state[1], run, 3);
 827
 828         for(y=0; y<h; y++){
 829             if(s->c.bytestream_end - s->c.bytestream < w*40){
 830                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 831                 return AVERROR(ENOMEM);
 832             }
 833             for(x=0; x<w; x++){
 834                 int v, p=0;
 835                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 836                 v= src[x + y*stride];
 837
 838                 if(y){
 839                     t= src[x + (y-1)*stride];
 840                     if(x){
 841                         lt= src[x - 1 + (y-1)*stride];
 842                     }
 843                     if(x + 1 < w){
 844                         rt= src[x + 1 + (y-1)*stride];
 845                     }
 846                 }
 847                 if(x){
 848                     l= src[x - 1 + y*stride];
 849                     /*if(x > 1){
 850                         if(orientation==1) ll= src[y + (x-2)*stride];
 851                         else               ll= src[x - 2 + y*stride];
 852                     }*/
 853                 }
 854                 if(parent){
 855                     int px= x>>1;
 856                     int py= y>>1;
 857                     if(px<b->parent->width && py<b->parent->height)
 858                         p= parent[px + py*2*stride];
 859                 }
 860                 if(/*ll|*/l|lt|t|rt|p){
 861                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 862
 863                     put_rac(&s->c, &b->state[0][context], !!v);
 864                 }else{
 865                     if(!run){
 866                         run= runs[run_index++];
 867
 868                         if(run_index <= max_index)
 869                             put_symbol2(&s->c, b->state[1], run, 3);
 870                         av_assert2(v);
 871                     }else{
 872                         run--;
 873                         av_assert2(!v);
 874                     }
 875                 }
 876                 if(v){
 877                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 878                     int l2= 2*FFABS(l) + (l<0);
 879                     int t2= 2*FFABS(t) + (t<0);
 880
 881                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 882                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 883                 }
 884             }
 885         }
 886     }
 887     return 0;
 888 }
 889
 890 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 891 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 892 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 893     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 894 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 895 }
 896
 897 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){
 898     const int b_stride= s->b_width << s->block_max_depth;
 899     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 900     BlockNode backup= *block;
 901     unsigned value;
 902     int rd, index;
 903
 904     av_assert2(mb_x>=0 && mb_y>=0);
 905     av_assert2(mb_x<b_stride);
 906
 907     if(intra){
 908         block->color[0] = p[0];
 909         block->color[1] = p[1];
 910         block->color[2] = p[2];
 911         block->type |= BLOCK_INTRA;
 912     }else{
 913         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 914         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 915         if(s->me_cache[index] == value)
 916             return 0;
 917         s->me_cache[index]= value;
 918
 919         block->mx= p[0];
 920         block->my= p[1];
 921         block->type &= ~BLOCK_INTRA;
 922     }
 923
 924     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
 925
 926 //FIXME chroma
 927     if(rd < *best_rd){
 928         *best_rd= rd;
 929         return 1;
 930     }else{
 931         *block= backup;
 932         return 0;
 933     }
 934 }
 935
 936 /* special case for int[2] args we discard afterwards,
 937  * fixes compilation problem with gcc 2.95 */
 938 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){
 939     int p[2] = {p0, p1};
 940     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 941 }
 942
 943 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){
 944     const int b_stride= s->b_width << s->block_max_depth;
 945     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 946     BlockNode backup[4];
 947     unsigned value;
 948     int rd, index;
 949
 950     /* We don't initialize backup[] during variable declaration, because
 951      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 952      * 'int16_t'". */
 953     backup[0] = block[0];
 954     backup[1] = block[1];
 955     backup[2] = block[b_stride];
 956     backup[3] = block[b_stride + 1];
 957
 958     av_assert2(mb_x>=0 && mb_y>=0);
 959     av_assert2(mb_x<b_stride);
 960     av_assert2(((mb_x|mb_y)&1) == 0);
 961
 962     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 963     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 964     if(s->me_cache[index] == value)
 965         return 0;
 966     s->me_cache[index]= value;
 967
 968     block->mx= p0;
 969     block->my= p1;
 970     block->ref= ref;
 971     block->type &= ~BLOCK_INTRA;
 972     block[1]= block[b_stride]= block[b_stride+1]= *block;
 973
 974     rd= get_4block_rd(s, mb_x, mb_y, 0);
 975
 976 //FIXME chroma
 977     if(rd < *best_rd){
 978         *best_rd= rd;
 979         return 1;
 980     }else{
 981         block[0]= backup[0];
 982         block[1]= backup[1];
 983         block[b_stride]= backup[2];
 984         block[b_stride+1]= backup[3];
 985         return 0;
 986     }
 987 }
 988
 989 static void iterative_me(SnowContext *s){
 990     int pass, mb_x, mb_y;
 991     const int b_width = s->b_width  << s->block_max_depth;
 992     const int b_height= s->b_height << s->block_max_depth;
 993     const int b_stride= b_width;
 994     int color[3];
 995
 996     {
 997         RangeCoder r = s->c;
 998         uint8_t state[sizeof(s->block_state)];
 999         memcpy(state, s->block_state, sizeof(s->block_state));
1000         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1001             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1002                 encode_q_branch(s, 0, mb_x, mb_y);
1003         s->c = r;
1004         memcpy(s->block_state, state, sizeof(s->block_state));
1005     }
1006
1007     for(pass=0; pass<25; pass++){
1008         int change= 0;
1009
1010         for(mb_y= 0; mb_y<b_height; mb_y++){
1011             for(mb_x= 0; mb_x<b_width; mb_x++){
1012                 int dia_change, i, j, ref;
1013                 int best_rd= INT_MAX, ref_rd;
1014                 BlockNode backup, ref_b;
1015                 const int index= mb_x + mb_y * b_stride;
1016                 BlockNode *block= &s->block[index];
1017                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1018                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1019                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1020                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1021                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1022                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1023                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1024                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1025                 const int b_w= (MB_SIZE >> s->block_max_depth);
1026                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1027
1028                 if(pass && (block->type & BLOCK_OPT))
1029                     continue;
1030                 block->type |= BLOCK_OPT;
1031
1032                 backup= *block;
1033
1034                 if(!s->me_cache_generation)
1035                     memset(s->me_cache, 0, sizeof(s->me_cache));
1036                 s->me_cache_generation += 1<<22;
1037
1038                 //FIXME precalculate
1039                 {
1040                     int x, y;
1041                     for (y = 0; y < b_w * 2; y++)
1042                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1043                     if(mb_x==0)
1044                         for(y=0; y<b_w*2; y++)
1045                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1046                     if(mb_x==b_stride-1)
1047                         for(y=0; y<b_w*2; y++)
1048                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1049                     if(mb_y==0){
1050                         for(x=0; x<b_w*2; x++)
1051                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1052                         for(y=1; y<b_w; y++)
1053                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1054                     }
1055                     if(mb_y==b_height-1){
1056                         for(x=0; x<b_w*2; x++)
1057                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1058                         for(y=b_w; y<b_w*2-1; y++)
1059                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1060                     }
1061                 }
1062
1063                 //skip stuff outside the picture
1064                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1065                     uint8_t *src= s->  input_picture->data[0];
1066                     uint8_t *dst= s->current_picture->data[0];
1067                     const int stride= s->current_picture->linesize[0];
1068                     const int block_w= MB_SIZE >> s->block_max_depth;
1069                     const int block_h= MB_SIZE >> s->block_max_depth;
1070                     const int sx= block_w*mb_x - block_w/2;
1071                     const int sy= block_h*mb_y - block_h/2;
1072                     const int w= s->plane[0].width;
1073                     const int h= s->plane[0].height;
1074                     int y;
1075
1076                     for(y=sy; y<0; y++)
1077                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1078                     for(y=h; y<sy+block_h*2; y++)
1079                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1080                     if(sx<0){
1081                         for(y=sy; y<sy+block_h*2; y++)
1082                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1083                     }
1084                     if(sx+block_w*2 > w){
1085                         for(y=sy; y<sy+block_h*2; y++)
1086                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1087                     }
1088                 }
1089
1090                 // intra(black) = neighbors' contribution to the current block
1091                 for(i=0; i < s->nb_planes; i++)
1092                     color[i]= get_dc(s, mb_x, mb_y, i);
1093
1094                 // get previous score (cannot be cached due to OBMC)
1095                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1096                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1097                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1098                 }else
1099                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1100
1101                 ref_b= *block;
1102                 ref_rd= best_rd;
1103                 for(ref=0; ref < s->ref_frames; ref++){
1104                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1105                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1106                         continue;
1107                     block->ref= ref;
1108                     best_rd= INT_MAX;
1109
1110                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1111                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1112                     if(tb)
1113                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1114                     if(lb)
1115                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1116                     if(rb)
1117                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1118                     if(bb)
1119                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1120
1121                     /* fullpel ME */
1122                     //FIXME avoid subpel interpolation / round to nearest integer
1123                     do{
1124                         int newx = block->mx;
1125                         int newy = block->my;
1126                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1127                         dia_change=0;
1128                         for(i=0; i < dia_size; i++){
1129                             for(j=0; j<i; j++){
1130                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1131                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1132                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1133                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1134                             }
1135                         }
1136                     }while(dia_change);
1137                     /* subpel ME */
1138                     do{
1139                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1140                         dia_change=0;
1141                         for(i=0; i<8; i++)
1142                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1143                     }while(dia_change);
1144                     //FIXME or try the standard 2 pass qpel or similar
1145
1146                     mvr[0][0]= block->mx;
1147                     mvr[0][1]= block->my;
1148                     if(ref_rd > best_rd){
1149                         ref_rd= best_rd;
1150                         ref_b= *block;
1151                     }
1152                 }
1153                 best_rd= ref_rd;
1154                 *block= ref_b;
1155                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1156                 //FIXME RD style color selection
1157                 if(!same_block(block, &backup)){
1158                     if(tb ) tb ->type &= ~BLOCK_OPT;
1159                     if(lb ) lb ->type &= ~BLOCK_OPT;
1160                     if(rb ) rb ->type &= ~BLOCK_OPT;
1161                     if(bb ) bb ->type &= ~BLOCK_OPT;
1162                     if(tlb) tlb->type &= ~BLOCK_OPT;
1163                     if(trb) trb->type &= ~BLOCK_OPT;
1164                     if(blb) blb->type &= ~BLOCK_OPT;
1165                     if(brb) brb->type &= ~BLOCK_OPT;
1166                     change ++;
1167                 }
1168             }
1169         }
1170         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1171         if(!change)
1172             break;
1173     }
1174
1175     if(s->block_max_depth == 1){
1176         int change= 0;
1177         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1178             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1179                 int i;
1180                 int best_rd, init_rd;
1181                 const int index= mb_x + mb_y * b_stride;
1182                 BlockNode *b[4];
1183
1184                 b[0]= &s->block[index];
1185                 b[1]= b[0]+1;
1186                 b[2]= b[0]+b_stride;
1187                 b[3]= b[2]+1;
1188                 if(same_block(b[0], b[1]) &&
1189                    same_block(b[0], b[2]) &&
1190                    same_block(b[0], b[3]))
1191                     continue;
1192
1193                 if(!s->me_cache_generation)
1194                     memset(s->me_cache, 0, sizeof(s->me_cache));
1195                 s->me_cache_generation += 1<<22;
1196
1197                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1198
1199                 //FIXME more multiref search?
1200                 check_4block_inter(s, mb_x, mb_y,
1201                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1202                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1203
1204                 for(i=0; i<4; i++)
1205                     if(!(b[i]->type&BLOCK_INTRA))
1206                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1207
1208                 if(init_rd != best_rd)
1209                     change++;
1210             }
1211         }
1212         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1213     }
1214 }
1215
1216 static void encode_blocks(SnowContext *s, int search){
1217     int x, y;
1218     int w= s->b_width;
1219     int h= s->b_height;
1220
1221     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1222         iterative_me(s);
1223
1224     for(y=0; y<h; y++){
1225         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1226             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1227             return;
1228         }
1229         for(x=0; x<w; x++){
1230             if(s->motion_est == FF_ME_ITER || !search)
1231                 encode_q_branch2(s, 0, x, y);
1232             else
1233                 encode_q_branch (s, 0, x, y);
1234         }
1235     }
1236 }
1237
1238 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1239     const int w= b->width;
1240     const int h= b->height;
1241     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1242     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1243     int x,y, thres1, thres2;
1244
1245     if(s->qlog == LOSSLESS_QLOG){
1246         for(y=0; y<h; y++)
1247             for(x=0; x<w; x++)
1248                 dst[x + y*stride]= src[x + y*stride];
1249         return;
1250     }
1251
1252     bias= bias ? 0 : (3*qmul)>>3;
1253     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1254     thres2= 2*thres1;
1255
1256     if(!bias){
1257         for(y=0; y<h; y++){
1258             for(x=0; x<w; x++){
1259                 int i= src[x + y*stride];
1260
1261                 if((unsigned)(i+thres1) > thres2){
1262                     if(i>=0){
1263                         i<<= QEXPSHIFT;
1264                         i/= qmul; //FIXME optimize
1265                         dst[x + y*stride]=  i;
1266                     }else{
1267                         i= -i;
1268                         i<<= QEXPSHIFT;
1269                         i/= qmul; //FIXME optimize
1270                         dst[x + y*stride]= -i;
1271                     }
1272                 }else
1273                     dst[x + y*stride]= 0;
1274             }
1275         }
1276     }else{
1277         for(y=0; y<h; y++){
1278             for(x=0; x<w; x++){
1279                 int i= src[x + y*stride];
1280
1281                 if((unsigned)(i+thres1) > thres2){
1282                     if(i>=0){
1283                         i<<= QEXPSHIFT;
1284                         i= (i + bias) / qmul; //FIXME optimize
1285                         dst[x + y*stride]=  i;
1286                     }else{
1287                         i= -i;
1288                         i<<= QEXPSHIFT;
1289                         i= (i + bias) / qmul; //FIXME optimize
1290                         dst[x + y*stride]= -i;
1291                     }
1292                 }else
1293                     dst[x + y*stride]= 0;
1294             }
1295         }
1296     }
1297 }
1298
1299 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1300     const int w= b->width;
1301     const int h= b->height;
1302     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1303     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1304     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1305     int x,y;
1306
1307     if(s->qlog == LOSSLESS_QLOG) return;
1308
1309     for(y=0; y<h; y++){
1310         for(x=0; x<w; x++){
1311             int i= src[x + y*stride];
1312             if(i<0){
1313                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1314             }else if(i>0){
1315                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1316             }
1317         }
1318     }
1319 }
1320
1321 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1322     const int w= b->width;
1323     const int h= b->height;
1324     int x,y;
1325
1326     for(y=h-1; y>=0; y--){
1327         for(x=w-1; x>=0; x--){
1328             int i= x + y*stride;
1329
1330             if(x){
1331                 if(use_median){
1332                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1333                     else  src[i] -= src[i - 1];
1334                 }else{
1335                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1336                     else  src[i] -= src[i - 1];
1337                 }
1338             }else{
1339                 if(y) src[i] -= src[i - stride];
1340             }
1341         }
1342     }
1343 }
1344
1345 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1346     const int w= b->width;
1347     const int h= b->height;
1348     int x,y;
1349
1350     for(y=0; y<h; y++){
1351         for(x=0; x<w; x++){
1352             int i= x + y*stride;
1353
1354             if(x){
1355                 if(use_median){
1356                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1357                     else  src[i] += src[i - 1];
1358                 }else{
1359                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1360                     else  src[i] += src[i - 1];
1361                 }
1362             }else{
1363                 if(y) src[i] += src[i - stride];
1364             }
1365         }
1366     }
1367 }
1368
1369 static void encode_qlogs(SnowContext *s){
1370     int plane_index, level, orientation;
1371
1372     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1373         for(level=0; level<s->spatial_decomposition_count; level++){
1374             for(orientation=level ? 1:0; orientation<4; orientation++){
1375                 if(orientation==2) continue;
1376                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1377             }
1378         }
1379     }
1380 }
1381
1382 static void encode_header(SnowContext *s){
1383     int plane_index, i;
1384     uint8_t kstate[32];
1385
1386     memset(kstate, MID_STATE, sizeof(kstate));
1387
1388     put_rac(&s->c, kstate, s->keyframe);
1389     if(s->keyframe || s->always_reset){
1390         ff_snow_reset_contexts(s);
1391         s->last_spatial_decomposition_type=
1392         s->last_qlog=
1393         s->last_qbias=
1394         s->last_mv_scale=
1395         s->last_block_max_depth= 0;
1396         for(plane_index=0; plane_index<2; plane_index++){
1397             Plane *p= &s->plane[plane_index];
1398             p->last_htaps=0;
1399             p->last_diag_mc=0;
1400             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1401         }
1402     }
1403     if(s->keyframe){
1404         put_symbol(&s->c, s->header_state, s->version, 0);
1405         put_rac(&s->c, s->header_state, s->always_reset);
1406         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1407         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1408         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1409         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1410         if (s->nb_planes > 2) {
1411             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1412             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1413         }
1414         put_rac(&s->c, s->header_state, s->spatial_scalability);
1415 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1416         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1417
1418         encode_qlogs(s);
1419     }
1420
1421     if(!s->keyframe){
1422         int update_mc=0;
1423         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1424             Plane *p= &s->plane[plane_index];
1425             update_mc |= p->last_htaps   != p->htaps;
1426             update_mc |= p->last_diag_mc != p->diag_mc;
1427             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1428         }
1429         put_rac(&s->c, s->header_state, update_mc);
1430         if(update_mc){
1431             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1432                 Plane *p= &s->plane[plane_index];
1433                 put_rac(&s->c, s->header_state, p->diag_mc);
1434                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1435                 for(i= p->htaps/2; i; i--)
1436                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1437             }
1438         }
1439         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1440             put_rac(&s->c, s->header_state, 1);
1441             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1442             encode_qlogs(s);
1443         }else
1444             put_rac(&s->c, s->header_state, 0);
1445     }
1446
1447     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1448     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1449     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1450     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1451     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1452
1453 }
1454
1455 static void update_last_header_values(SnowContext *s){
1456     int plane_index;
1457
1458     if(!s->keyframe){
1459         for(plane_index=0; plane_index<2; plane_index++){
1460             Plane *p= &s->plane[plane_index];
1461             p->last_diag_mc= p->diag_mc;
1462             p->last_htaps  = p->htaps;
1463             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1464         }
1465     }
1466
1467     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1468     s->last_qlog                        = s->qlog;
1469     s->last_qbias                       = s->qbias;
1470     s->last_mv_scale                    = s->mv_scale;
1471     s->last_block_max_depth             = s->block_max_depth;
1472     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1473 }
1474
1475 static int qscale2qlog(int qscale){
1476     return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1477            + 61*QROOT/8; ///< 64 > 60
1478 }
1479
1480 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1481 {
1482     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1483      * FIXME we know exact mv bits at this point,
1484      * but ratecontrol isn't set up to include them. */
1485     uint32_t coef_sum= 0;
1486     int level, orientation, delta_qlog;
1487
1488     for(level=0; level<s->spatial_decomposition_count; level++){
1489         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1490             SubBand *b= &s->plane[0].band[level][orientation];
1491             IDWTELEM *buf= b->ibuf;
1492             const int w= b->width;
1493             const int h= b->height;
1494             const int stride= b->stride;
1495             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1496             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1497             const int qdiv= (1<<16)/qmul;
1498             int x, y;
1499             //FIXME this is ugly
1500             for(y=0; y<h; y++)
1501                 for(x=0; x<w; x++)
1502                     buf[x+y*stride]= b->buf[x+y*stride];
1503             if(orientation==0)
1504                 decorrelate(s, b, buf, stride, 1, 0);
1505             for(y=0; y<h; y++)
1506                 for(x=0; x<w; x++)
1507                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1508         }
1509     }
1510
1511     /* ugly, ratecontrol just takes a sqrt again */
1512     av_assert0(coef_sum < INT_MAX);
1513     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1514
1515     if(pict->pict_type == AV_PICTURE_TYPE_I){
1516         s->m.current_picture.mb_var_sum= coef_sum;
1517         s->m.current_picture.mc_mb_var_sum= 0;
1518     }else{
1519         s->m.current_picture.mc_mb_var_sum= coef_sum;
1520         s->m.current_picture.mb_var_sum= 0;
1521     }
1522
1523     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1524     if (pict->quality < 0)
1525         return INT_MIN;
1526     s->lambda= pict->quality * 3/2;
1527     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1528     s->qlog+= delta_qlog;
1529     return delta_qlog;
1530 }
1531
1532 static void calculate_visual_weight(SnowContext *s, Plane *p){
1533     int width = p->width;
1534     int height= p->height;
1535     int level, orientation, x, y;
1536
1537     for(level=0; level<s->spatial_decomposition_count; level++){
1538         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1539             SubBand *b= &p->band[level][orientation];
1540             IDWTELEM *ibuf= b->ibuf;
1541             int64_t error=0;
1542
1543             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1544             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1545             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1546             for(y=0; y<height; y++){
1547                 for(x=0; x<width; x++){
1548                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1549                     error += d*d;
1550                 }
1551             }
1552
1553             b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1554         }
1555     }
1556 }
1557
1558 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1559                         const AVFrame *pict, int *got_packet)
1560 {
1561     SnowContext *s = avctx->priv_data;
1562     RangeCoder * const c= &s->c;
1563     AVFrame *pic;
1564     const int width= s->avctx->width;
1565     const int height= s->avctx->height;
1566     int level, orientation, plane_index, i, y, ret;
1567     uint8_t rc_header_bak[sizeof(s->header_state)];
1568     uint8_t rc_block_bak[sizeof(s->block_state)];
1569
1570     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)
1571         return ret;
1572
1573     ff_init_range_encoder(c, pkt->data, pkt->size);
1574     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1575
1576     for(i=0; i < s->nb_planes; i++){
1577         int hshift= i ? s->chroma_h_shift : 0;
1578         int vshift= i ? s->chroma_v_shift : 0;
1579         for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1580             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1581                    &pict->data[i][y * pict->linesize[i]],
1582                    AV_CEIL_RSHIFT(width, hshift));
1583         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1584                                 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1585                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1586                                 EDGE_TOP | EDGE_BOTTOM);
1587
1588     }
1589     emms_c();
1590     pic = s->input_picture;
1591     pic->pict_type = pict->pict_type;
1592     pic->quality = pict->quality;
1593
1594     s->m.picture_number= avctx->frame_number;
1595     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1596         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1597         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1598         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1599             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1600             if (pic->quality < 0)
1601                 return -1;
1602         }
1603     }else{
1604         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1605         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1606     }
1607
1608     if(s->pass1_rc && avctx->frame_number == 0)
1609         pic->quality = 2*FF_QP2LAMBDA;
1610     if (pic->quality) {
1611         s->qlog   = qscale2qlog(pic->quality);
1612         s->lambda = pic->quality * 3/2;
1613     }
1614     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1615         s->qlog= LOSSLESS_QLOG;
1616         s->lambda = 0;
1617     }//else keep previous frame's qlog until after motion estimation
1618
1619     if (s->current_picture->data[0]) {
1620         int w = s->avctx->width;
1621         int h = s->avctx->height;
1622
1623         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1624                                 s->current_picture->linesize[0], w   , h   ,
1625                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1626         if (s->current_picture->data[2]) {
1627             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1628                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1629                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1630             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1631                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1632                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1633         }
1634         emms_c();
1635     }
1636
1637     ff_snow_frame_start(s);
1638 #if FF_API_CODED_FRAME
1639 FF_DISABLE_DEPRECATION_WARNINGS
1640     av_frame_unref(avctx->coded_frame);
1641     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1642 FF_ENABLE_DEPRECATION_WARNINGS
1643 #endif
1644     if (ret < 0)
1645         return ret;
1646
1647     s->m.current_picture_ptr= &s->m.current_picture;
1648     s->m.current_picture.f = s->current_picture;
1649     s->m.current_picture.f->pts = pict->pts;
1650     if(pic->pict_type == AV_PICTURE_TYPE_P){
1651         int block_width = (width +15)>>4;
1652         int block_height= (height+15)>>4;
1653         int stride= s->current_picture->linesize[0];
1654
1655         av_assert0(s->current_picture->data[0]);
1656         av_assert0(s->last_picture[0]->data[0]);
1657
1658         s->m.avctx= s->avctx;
1659         s->m.   last_picture.f = s->last_picture[0];
1660         s->m.    new_picture.f = s->input_picture;
1661         s->m.   last_picture_ptr= &s->m.   last_picture;
1662         s->m.linesize = stride;
1663         s->m.uvlinesize= s->current_picture->linesize[1];
1664         s->m.width = width;
1665         s->m.height= height;
1666         s->m.mb_width = block_width;
1667         s->m.mb_height= block_height;
1668         s->m.mb_stride=   s->m.mb_width+1;
1669         s->m.b8_stride= 2*s->m.mb_width+1;
1670         s->m.f_code=1;
1671         s->m.pict_type = pic->pict_type;
1672         s->m.motion_est= s->motion_est;
1673         s->m.me.scene_change_score=0;
1674         s->m.me.dia_size = avctx->dia_size;
1675         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1676         s->m.out_format= FMT_H263;
1677         s->m.unrestricted_mv= 1;
1678
1679         s->m.lambda = s->lambda;
1680         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1681         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1682
1683         s->m.mecc= s->mecc; //move
1684         s->m.qdsp= s->qdsp; //move
1685         s->m.hdsp = s->hdsp;
1686         ff_init_me(&s->m);
1687         s->hdsp = s->m.hdsp;
1688         s->mecc= s->m.mecc;
1689     }
1690
1691     if(s->pass1_rc){
1692         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1693         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1694     }
1695
1696 redo_frame:
1697
1698     s->spatial_decomposition_count= 5;
1699
1700     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1701           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1702         s->spatial_decomposition_count--;
1703
1704     if (s->spatial_decomposition_count <= 0) {
1705         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1706         return AVERROR(EINVAL);
1707     }
1708
1709     s->m.pict_type = pic->pict_type;
1710     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1711
1712     ff_snow_common_init_after_header(avctx);
1713
1714     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1715         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1716             calculate_visual_weight(s, &s->plane[plane_index]);
1717         }
1718     }
1719
1720     encode_header(s);
1721     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1722     encode_blocks(s, 1);
1723     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1724
1725     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1726         Plane *p= &s->plane[plane_index];
1727         int w= p->width;
1728         int h= p->height;
1729         int x, y;
1730 //        int bits= put_bits_count(&s->c.pb);
1731
1732         if (!s->memc_only) {
1733             //FIXME optimize
1734             if(pict->data[plane_index]) //FIXME gray hack
1735                 for(y=0; y<h; y++){
1736                     for(x=0; x<w; x++){
1737                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1738                     }
1739                 }
1740             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1741
1742 #if FF_API_PRIVATE_OPT
1743 FF_DISABLE_DEPRECATION_WARNINGS
1744             if(s->avctx->scenechange_threshold)
1745                 s->scenechange_threshold = s->avctx->scenechange_threshold;
1746 FF_ENABLE_DEPRECATION_WARNINGS
1747 #endif
1748
1749             if(   plane_index==0
1750                && pic->pict_type == AV_PICTURE_TYPE_P
1751                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1752                && s->m.me.scene_change_score > s->scenechange_threshold){
1753                 ff_init_range_encoder(c, pkt->data, pkt->size);
1754                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1755                 pic->pict_type= AV_PICTURE_TYPE_I;
1756                 s->keyframe=1;
1757                 s->current_picture->key_frame=1;
1758                 goto redo_frame;
1759             }
1760
1761             if(s->qlog == LOSSLESS_QLOG){
1762                 for(y=0; y<h; y++){
1763                     for(x=0; x<w; x++){
1764                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1765                     }
1766                 }
1767             }else{
1768                 for(y=0; y<h; y++){
1769                     for(x=0; x<w; x++){
1770                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1771                     }
1772                 }
1773             }
1774
1775             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1776
1777             if(s->pass1_rc && plane_index==0){
1778                 int delta_qlog = ratecontrol_1pass(s, pic);
1779                 if (delta_qlog <= INT_MIN)
1780                     return -1;
1781                 if(delta_qlog){
1782                     //reordering qlog in the bitstream would eliminate this reset
1783                     ff_init_range_encoder(c, pkt->data, pkt->size);
1784                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1785                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1786                     encode_header(s);
1787                     encode_blocks(s, 0);
1788                 }
1789             }
1790
1791             for(level=0; level<s->spatial_decomposition_count; level++){
1792                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1793                     SubBand *b= &p->band[level][orientation];
1794
1795                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1796                     if(orientation==0)
1797                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1798                     if (!s->no_bitstream)
1799                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1800                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1801                     if(orientation==0)
1802                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1803                 }
1804             }
1805
1806             for(level=0; level<s->spatial_decomposition_count; level++){
1807                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1808                     SubBand *b= &p->band[level][orientation];
1809
1810                     dequantize(s, b, b->ibuf, b->stride);
1811                 }
1812             }
1813
1814             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1815             if(s->qlog == LOSSLESS_QLOG){
1816                 for(y=0; y<h; y++){
1817                     for(x=0; x<w; x++){
1818                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1819                     }
1820                 }
1821             }
1822             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1823         }else{
1824             //ME/MC only
1825             if(pic->pict_type == AV_PICTURE_TYPE_I){
1826                 for(y=0; y<h; y++){
1827                     for(x=0; x<w; x++){
1828                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1829                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1830                     }
1831                 }
1832             }else{
1833                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1834                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1835             }
1836         }
1837         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1838             int64_t error= 0;
1839
1840             if(pict->data[plane_index]) //FIXME gray hack
1841                 for(y=0; y<h; y++){
1842                     for(x=0; x<w; x++){
1843                         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];
1844                         error += d*d;
1845                     }
1846                 }
1847             s->avctx->error[plane_index] += error;
1848             s->encoding_error[plane_index] = error;
1849         }
1850
1851     }
1852     emms_c();
1853
1854     update_last_header_values(s);
1855
1856     ff_snow_release_buffer(avctx);
1857
1858     s->current_picture->coded_picture_number = avctx->frame_number;
1859     s->current_picture->pict_type = pic->pict_type;
1860     s->current_picture->quality = pic->quality;
1861     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1862     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1863     s->m.current_picture.f->display_picture_number =
1864     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1865     s->m.current_picture.f->quality                = pic->quality;
1866     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1867     if(s->pass1_rc)
1868         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1869             return -1;
1870     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1871         ff_write_pass1_stats(&s->m);
1872     s->m.last_pict_type = s->m.pict_type;
1873 #if FF_API_STAT_BITS
1874 FF_DISABLE_DEPRECATION_WARNINGS
1875     avctx->frame_bits = s->m.frame_bits;
1876     avctx->mv_bits = s->m.mv_bits;
1877     avctx->misc_bits = s->m.misc_bits;
1878     avctx->p_tex_bits = s->m.p_tex_bits;
1879 FF_ENABLE_DEPRECATION_WARNINGS
1880 #endif
1881
1882     emms_c();
1883
1884     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1885                                    s->encoding_error,
1886                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1887                                    s->current_picture->pict_type);
1888
1889 #if FF_API_ERROR_FRAME
1890 FF_DISABLE_DEPRECATION_WARNINGS
1891     memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1892 FF_ENABLE_DEPRECATION_WARNINGS
1893 #endif
1894
1895     pkt->size = ff_rac_terminate(c);
1896     if (s->current_picture->key_frame)
1897         pkt->flags |= AV_PKT_FLAG_KEY;
1898     *got_packet = 1;
1899
1900     return 0;
1901 }
1902
1903 static av_cold int encode_end(AVCodecContext *avctx)
1904 {
1905     SnowContext *s = avctx->priv_data;
1906
1907     ff_snow_common_end(s);
1908     ff_rate_control_uninit(&s->m);
1909     av_frame_free(&s->input_picture);
1910     av_freep(&avctx->stats_out);
1911
1912     return 0;
1913 }
1914
1915 #define OFFSET(x) offsetof(SnowContext, x)
1916 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1917 static const AVOption options[] = {
1918     {"motion_est", "motion estimation algorithm", OFFSET(motion_est), AV_OPT_TYPE_INT, {.i64 = FF_ME_EPZS }, FF_ME_ZERO, FF_ME_ITER, VE, "motion_est" },
1919     { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1920     { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1921     { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1922     { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1923     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1924     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1925     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1926     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1927     { "sc_threshold",   "Scene change threshold",                   OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1928     { "pred",           "Spatial decomposition type",                                OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1929         { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1930         { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1931     { NULL },
1932 };
1933
1934 static const AVClass snowenc_class = {
1935     .class_name = "snow encoder",
1936     .item_name  = av_default_item_name,
1937     .option     = options,
1938     .version    = LIBAVUTIL_VERSION_INT,
1939 };
1940
1941 AVCodec ff_snow_encoder = {
1942     .name           = "snow",
1943     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1944     .type           = AVMEDIA_TYPE_VIDEO,
1945     .id             = AV_CODEC_ID_SNOW,
1946     .priv_data_size = sizeof(SnowContext),
1947     .init           = encode_init,
1948     .encode2        = encode_frame,
1949     .close          = encode_end,
1950     .pix_fmts       = (const enum AVPixelFormat[]){
1951         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1952         AV_PIX_FMT_GRAY8,
1953         AV_PIX_FMT_NONE
1954     },
1955     .priv_class     = &snowenc_class,
1956     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1957                       FF_CODEC_CAP_INIT_CLEANUP,
1958 };