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 "libavutil/pixdesc.h"
  26 #include "avcodec.h"
  27 #include "internal.h"
  28 #include "snow_dwt.h"
  29 #include "snow.h"
  30
  31 #include "rangecoder.h"
  32 #include "mathops.h"
  33
  34 #include "mpegvideo.h"
  35 #include "h263.h"
  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 FF_API_PRIVATE_OPT
  44 FF_DISABLE_DEPRECATION_WARNINGS
  45     if (avctx->prediction_method)
  46         s->pred = avctx->prediction_method;
  47 FF_ENABLE_DEPRECATION_WARNINGS
  48 #endif
  49
  50     if(s->pred == DWT_97
  51        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
  52        && avctx->global_quality == 0){
  53         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  54         return AVERROR(EINVAL);
  55     }
  56
  57     s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
  58
  59     s->mv_scale       = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
  60     s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
  61
  62     for(plane_index=0; plane_index<3; plane_index++){
  63         s->plane[plane_index].diag_mc= 1;
  64         s->plane[plane_index].htaps= 6;
  65         s->plane[plane_index].hcoeff[0]=  40;
  66         s->plane[plane_index].hcoeff[1]= -10;
  67         s->plane[plane_index].hcoeff[2]=   2;
  68         s->plane[plane_index].fast_mc= 1;
  69     }
  70
  71     if ((ret = ff_snow_common_init(avctx)) < 0) {
  72         return ret;
  73     }
  74     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  75
  76     ff_snow_alloc_blocks(s);
  77
  78     s->version=0;
  79
  80     s->m.avctx   = avctx;
  81     s->m.bit_rate= avctx->bit_rate;
  82     s->m.lmin    = avctx->mb_lmin;
  83     s->m.lmax    = avctx->mb_lmax;
  84
  85     s->m.me.temp      =
  86     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  87     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  88     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  89     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  90     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
  91         return AVERROR(ENOMEM);
  92
  93     ff_h263_encode_init(&s->m); //mv_penalty
  94
  95     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
  96
  97     if(avctx->flags&AV_CODEC_FLAG_PASS1){
  98         if(!avctx->stats_out)
  99             avctx->stats_out = av_mallocz(256);
 100
 101         if (!avctx->stats_out)
 102             return AVERROR(ENOMEM);
 103     }
 104     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
 105         ret = ff_rate_control_init(&s->m);
 106         if(ret < 0)
 107             return ret;
 108     }
 109     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
 110
 111     switch(avctx->pix_fmt){
 112     case AV_PIX_FMT_YUV444P:
 113 //    case AV_PIX_FMT_YUV422P:
 114     case AV_PIX_FMT_YUV420P:
 115 //    case AV_PIX_FMT_YUV411P:
 116     case AV_PIX_FMT_YUV410P:
 117         s->nb_planes = 3;
 118         s->colorspace_type= 0;
 119         break;
 120     case AV_PIX_FMT_GRAY8:
 121         s->nb_planes = 1;
 122         s->colorspace_type = 1;
 123         break;
 124 /*    case AV_PIX_FMT_RGB32:
 125         s->colorspace= 1;
 126         break;*/
 127     default:
 128         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 129         return AVERROR_PATCHWELCOME;
 130     }
 131
 132     ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
 133                                            &s->chroma_v_shift);
 134     if (ret) {
 135         av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
 136         return ret;
 137     }
 138
 139     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 140     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 141
 142     s->input_picture = av_frame_alloc();
 143     if (!s->input_picture)
 144         return AVERROR(ENOMEM);
 145
 146     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 147         return ret;
 148
 149     if(s->motion_est == FF_ME_ITER){
 150         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 151         for(i=0; i<s->max_ref_frames; i++){
 152             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 153             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 154             if (!s->ref_mvs[i] || !s->ref_scores[i])
 155                 return AVERROR(ENOMEM);
 156         }
 157     }
 158
 159     return 0;
 160 }
 161
 162 //near copy & paste from dsputil, FIXME
 163 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 164 {
 165     int s, i, j;
 166
 167     s = 0;
 168     for (i = 0; i < h; i++) {
 169         for (j = 0; j < w; j++) {
 170             s += pix[0];
 171             pix ++;
 172         }
 173         pix += line_size - w;
 174     }
 175     return s;
 176 }
 177
 178 //near copy & paste from dsputil, FIXME
 179 static int pix_norm1(uint8_t * pix, int line_size, int w)
 180 {
 181     int s, i, j;
 182     const uint32_t *sq = ff_square_tab + 256;
 183
 184     s = 0;
 185     for (i = 0; i < w; i++) {
 186         for (j = 0; j < w; j ++) {
 187             s += sq[pix[0]];
 188             pix ++;
 189         }
 190         pix += line_size - w;
 191     }
 192     return s;
 193 }
 194
 195 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 196     switch(type&0xFF){
 197     default:
 198     case FF_CMP_SAD:
 199         return lambda>>FF_LAMBDA_SHIFT;
 200     case FF_CMP_DCT:
 201         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 202     case FF_CMP_W53:
 203         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 204     case FF_CMP_W97:
 205         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 206     case FF_CMP_SATD:
 207     case FF_CMP_DCT264:
 208         return (2*lambda)>>FF_LAMBDA_SHIFT;
 209     case FF_CMP_RD:
 210     case FF_CMP_PSNR:
 211     case FF_CMP_SSE:
 212     case FF_CMP_NSSE:
 213         return lambda2>>FF_LAMBDA_SHIFT;
 214     case FF_CMP_BIT:
 215         return 1;
 216     }
 217 }
 218
 219 //FIXME copy&paste
 220 #define P_LEFT P[1]
 221 #define P_TOP P[2]
 222 #define P_TOPRIGHT P[3]
 223 #define P_MEDIAN P[4]
 224 #define P_MV1 P[9]
 225 #define FLAG_QPEL   1 //must be 1
 226
 227 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 228     uint8_t p_buffer[1024];
 229     uint8_t i_buffer[1024];
 230     uint8_t p_state[sizeof(s->block_state)];
 231     uint8_t i_state[sizeof(s->block_state)];
 232     RangeCoder pc, ic;
 233     uint8_t *pbbak= s->c.bytestream;
 234     uint8_t *pbbak_start= s->c.bytestream_start;
 235     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 236     const int w= s->b_width  << s->block_max_depth;
 237     const int h= s->b_height << s->block_max_depth;
 238     const int rem_depth= s->block_max_depth - level;
 239     const int index= (x + y*w) << rem_depth;
 240     const int block_w= 1<<(LOG2_MB_SIZE - level);
 241     int trx= (x+1)<<rem_depth;
 242     int try= (y+1)<<rem_depth;
 243     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 244     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 245     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 246     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 247     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 248     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 249     int pl = left->color[0];
 250     int pcb= left->color[1];
 251     int pcr= left->color[2];
 252     int pmx, pmy;
 253     int mx=0, my=0;
 254     int l,cr,cb;
 255     const int stride= s->current_picture->linesize[0];
 256     const int uvstride= s->current_picture->linesize[1];
 257     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 258                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 259                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 260     int P[10][2];
 261     int16_t last_mv[3][2];
 262     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
 263     const int shift= 1+qpel;
 264     MotionEstContext *c= &s->m.me;
 265     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 266     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 267     int my_context= av_log2(2*FFABS(left->my - top->my));
 268     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 269     int ref, best_ref, ref_score, ref_mx, ref_my;
 270
 271     av_assert0(sizeof(s->block_state) >= 256);
 272     if(s->keyframe){
 273         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 274         return 0;
 275     }
 276
 277 //    clip predictors / edge ?
 278
 279     P_LEFT[0]= left->mx;
 280     P_LEFT[1]= left->my;
 281     P_TOP [0]= top->mx;
 282     P_TOP [1]= top->my;
 283     P_TOPRIGHT[0]= tr->mx;
 284     P_TOPRIGHT[1]= tr->my;
 285
 286     last_mv[0][0]= s->block[index].mx;
 287     last_mv[0][1]= s->block[index].my;
 288     last_mv[1][0]= right->mx;
 289     last_mv[1][1]= right->my;
 290     last_mv[2][0]= bottom->mx;
 291     last_mv[2][1]= bottom->my;
 292
 293     s->m.mb_stride=2;
 294     s->m.mb_x=
 295     s->m.mb_y= 0;
 296     c->skip= 0;
 297
 298     av_assert1(c->  stride ==   stride);
 299     av_assert1(c->uvstride == uvstride);
 300
 301     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 302     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 303     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 304     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
 305
 306     c->xmin = - x*block_w - 16+3;
 307     c->ymin = - y*block_w - 16+3;
 308     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 309     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 310
 311     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 312     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 313     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 314     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 315     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 316     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 317     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 318
 319     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 320     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 321
 322     if (!y) {
 323         c->pred_x= P_LEFT[0];
 324         c->pred_y= P_LEFT[1];
 325     } else {
 326         c->pred_x = P_MEDIAN[0];
 327         c->pred_y = P_MEDIAN[1];
 328     }
 329
 330     score= INT_MAX;
 331     best_ref= 0;
 332     for(ref=0; ref<s->ref_frames; ref++){
 333         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 334
 335         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 336                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 337
 338         av_assert2(ref_mx >= c->xmin);
 339         av_assert2(ref_mx <= c->xmax);
 340         av_assert2(ref_my >= c->ymin);
 341         av_assert2(ref_my <= c->ymax);
 342
 343         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 344         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 345         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 346         if(s->ref_mvs[ref]){
 347             s->ref_mvs[ref][index][0]= ref_mx;
 348             s->ref_mvs[ref][index][1]= ref_my;
 349             s->ref_scores[ref][index]= ref_score;
 350         }
 351         if(score > ref_score){
 352             score= ref_score;
 353             best_ref= ref;
 354             mx= ref_mx;
 355             my= ref_my;
 356         }
 357     }
 358     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 359
 360   //  subpel search
 361     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 362     pc= s->c;
 363     pc.bytestream_start=
 364     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 365     memcpy(p_state, s->block_state, sizeof(s->block_state));
 366
 367     if(level!=s->block_max_depth)
 368         put_rac(&pc, &p_state[4 + s_context], 1);
 369     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 370     if(s->ref_frames > 1)
 371         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 372     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 373     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 374     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 375     p_len= pc.bytestream - pc.bytestream_start;
 376     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 377
 378     block_s= block_w*block_w;
 379     sum = pix_sum(current_data[0], stride, block_w, block_w);
 380     l= (sum + block_s/2)/block_s;
 381     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 382
 383     if (s->nb_planes > 2) {
 384         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 385         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 386         cb= (sum + block_s/2)/block_s;
 387     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 388         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 389         cr= (sum + block_s/2)/block_s;
 390     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 391     }else
 392         cb = cr = 0;
 393
 394     ic= s->c;
 395     ic.bytestream_start=
 396     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 397     memcpy(i_state, s->block_state, sizeof(s->block_state));
 398     if(level!=s->block_max_depth)
 399         put_rac(&ic, &i_state[4 + s_context], 1);
 400     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 401     put_symbol(&ic, &i_state[32],  l-pl , 1);
 402     if (s->nb_planes > 2) {
 403         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 404         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 405     }
 406     i_len= ic.bytestream - ic.bytestream_start;
 407     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 408
 409     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 410     av_assert1(iscore >= 0);
 411     av_assert1(l>=0 && l<=255);
 412     av_assert1(pl>=0 && pl<=255);
 413
 414     if(level==0){
 415         int varc= iscore >> 8;
 416         int vard= score >> 8;
 417         if (vard <= 64 || vard < varc)
 418             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 419         else
 420             c->scene_change_score+= s->m.qscale;
 421     }
 422
 423     if(level!=s->block_max_depth){
 424         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 425         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 426         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 427         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 428         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 429         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 430
 431         if(score2 < score && score2 < iscore)
 432             return score2;
 433     }
 434
 435     if(iscore < score){
 436         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 437         memcpy(pbbak, i_buffer, i_len);
 438         s->c= ic;
 439         s->c.bytestream_start= pbbak_start;
 440         s->c.bytestream= pbbak + i_len;
 441         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 442         memcpy(s->block_state, i_state, sizeof(s->block_state));
 443         return iscore;
 444     }else{
 445         memcpy(pbbak, p_buffer, p_len);
 446         s->c= pc;
 447         s->c.bytestream_start= pbbak_start;
 448         s->c.bytestream= pbbak + p_len;
 449         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 450         memcpy(s->block_state, p_state, sizeof(s->block_state));
 451         return score;
 452     }
 453 }
 454
 455 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 456     const int w= s->b_width  << s->block_max_depth;
 457     const int rem_depth= s->block_max_depth - level;
 458     const int index= (x + y*w) << rem_depth;
 459     int trx= (x+1)<<rem_depth;
 460     BlockNode *b= &s->block[index];
 461     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 462     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 463     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 464     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 465     int pl = left->color[0];
 466     int pcb= left->color[1];
 467     int pcr= left->color[2];
 468     int pmx, pmy;
 469     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 470     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 471     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 472     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 473
 474     if(s->keyframe){
 475         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 476         return;
 477     }
 478
 479     if(level!=s->block_max_depth){
 480         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 481             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 482         }else{
 483             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 484             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 485             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 486             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 487             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 488             return;
 489         }
 490     }
 491     if(b->type & BLOCK_INTRA){
 492         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 493         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 494         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 495         if (s->nb_planes > 2) {
 496             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 497             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 498         }
 499         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 500     }else{
 501         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 502         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 503         if(s->ref_frames > 1)
 504             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 505         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 506         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 507         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 508     }
 509 }
 510
 511 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 512     int i, x2, y2;
 513     Plane *p= &s->plane[plane_index];
 514     const int block_size = MB_SIZE >> s->block_max_depth;
 515     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 516     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 517     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 518     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 519     const int ref_stride= s->current_picture->linesize[plane_index];
 520     uint8_t *src= s-> input_picture->data[plane_index];
 521     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 522     const int b_stride = s->b_width << s->block_max_depth;
 523     const int w= p->width;
 524     const int h= p->height;
 525     int index= mb_x + mb_y*b_stride;
 526     BlockNode *b= &s->block[index];
 527     BlockNode backup= *b;
 528     int ab=0;
 529     int aa=0;
 530
 531     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 532
 533     b->type|= BLOCK_INTRA;
 534     b->color[plane_index]= 0;
 535     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 536
 537     for(i=0; i<4; i++){
 538         int mb_x2= mb_x + (i &1) - 1;
 539         int mb_y2= mb_y + (i>>1) - 1;
 540         int x= block_w*mb_x2 + block_w/2;
 541         int y= block_h*mb_y2 + block_h/2;
 542
 543         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 544                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 545
 546         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 547             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 548                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 549                 int obmc_v= obmc[index];
 550                 int d;
 551                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 552                 if(x<0) obmc_v += obmc[index + block_w];
 553                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 554                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 555                 //FIXME precalculate this or simplify it somehow else
 556
 557                 d = -dst[index] + (1<<(FRAC_BITS-1));
 558                 dst[index] = d;
 559                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 560                 aa += obmc_v * obmc_v; //FIXME precalculate this
 561             }
 562         }
 563     }
 564     *b= backup;
 565
 566     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 567 }
 568
 569 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 570     const int b_stride = s->b_width << s->block_max_depth;
 571     const int b_height = s->b_height<< s->block_max_depth;
 572     int index= x + y*b_stride;
 573     const BlockNode *b     = &s->block[index];
 574     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 575     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 576     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 577     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 578     int dmx, dmy;
 579 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 580 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 581
 582     if(x<0 || x>=b_stride || y>=b_height)
 583         return 0;
 584 /*
 585 1            0      0
 586 01X          1-2    1
 587 001XX        3-6    2-3
 588 0001XXX      7-14   4-7
 589 00001XXXX   15-30   8-15
 590 */
 591 //FIXME try accurate rate
 592 //FIXME intra and inter predictors if surrounding blocks are not the same type
 593     if(b->type & BLOCK_INTRA){
 594         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 595                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 596                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 597     }else{
 598         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 599         dmx-= b->mx;
 600         dmy-= b->my;
 601         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 602                     + av_log2(2*FFABS(dmy))
 603                     + av_log2(2*b->ref));
 604     }
 605 }
 606
 607 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 608     Plane *p= &s->plane[plane_index];
 609     const int block_size = MB_SIZE >> s->block_max_depth;
 610     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 611     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 612     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 613     const int ref_stride= s->current_picture->linesize[plane_index];
 614     uint8_t *dst= s->current_picture->data[plane_index];
 615     uint8_t *src= s->  input_picture->data[plane_index];
 616     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
 617     uint8_t *cur = s->scratchbuf;
 618     uint8_t *tmp = s->emu_edge_buffer;
 619     const int b_stride = s->b_width << s->block_max_depth;
 620     const int b_height = s->b_height<< s->block_max_depth;
 621     const int w= p->width;
 622     const int h= p->height;
 623     int distortion;
 624     int rate= 0;
 625     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 626     int sx= block_w*mb_x - block_w/2;
 627     int sy= block_h*mb_y - block_h/2;
 628     int x0= FFMAX(0,-sx);
 629     int y0= FFMAX(0,-sy);
 630     int x1= FFMIN(block_w*2, w-sx);
 631     int y1= FFMIN(block_h*2, h-sy);
 632     int i,x,y;
 633
 634     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 635
 636     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);
 637
 638     for(y=y0; y<y1; y++){
 639         const uint8_t *obmc1= obmc_edged[y];
 640         const IDWTELEM *pred1 = pred + y*obmc_stride;
 641         uint8_t *cur1 = cur + y*ref_stride;
 642         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 643         for(x=x0; x<x1; x++){
 644 #if FRAC_BITS >= LOG2_OBMC_MAX
 645             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 646 #else
 647             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 648 #endif
 649             v = (v + pred1[x]) >> FRAC_BITS;
 650             if(v&(~255)) v= ~(v>>31);
 651             dst1[x] = v;
 652         }
 653     }
 654
 655     /* copy the regions where obmc[] = (uint8_t)256 */
 656     if(LOG2_OBMC_MAX == 8
 657         && (mb_x == 0 || mb_x == b_stride-1)
 658         && (mb_y == 0 || mb_y == b_height-1)){
 659         if(mb_x == 0)
 660             x1 = block_w;
 661         else
 662             x0 = block_w;
 663         if(mb_y == 0)
 664             y1 = block_h;
 665         else
 666             y0 = block_h;
 667         for(y=y0; y<y1; y++)
 668             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 669     }
 670
 671     if(block_w==16){
 672         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 673         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 674         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 675          * So improving the score of one block is not strictly guaranteed
 676          * to improve the score of the whole frame, thus iterative motion
 677          * estimation does not always converge. */
 678         if(s->avctx->me_cmp == FF_CMP_W97)
 679             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 680         else if(s->avctx->me_cmp == FF_CMP_W53)
 681             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 682         else{
 683             distortion = 0;
 684             for(i=0; i<4; i++){
 685                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 686                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 687             }
 688         }
 689     }else{
 690         av_assert2(block_w==8);
 691         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 692     }
 693
 694     if(plane_index==0){
 695         for(i=0; i<4; i++){
 696 /* ..RRr
 697  * .RXx.
 698  * rxx..
 699  */
 700             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 701         }
 702         if(mb_x == b_stride-2)
 703             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 704     }
 705     return distortion + rate*penalty_factor;
 706 }
 707
 708 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 709     int i, y2;
 710     Plane *p= &s->plane[plane_index];
 711     const int block_size = MB_SIZE >> s->block_max_depth;
 712     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 713     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 714     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 715     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 716     const int ref_stride= s->current_picture->linesize[plane_index];
 717     uint8_t *dst= s->current_picture->data[plane_index];
 718     uint8_t *src= s-> input_picture->data[plane_index];
 719     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 720     // const has only been removed from zero_dst to suppress a warning
 721     static IDWTELEM zero_dst[4096]; //FIXME
 722     const int b_stride = s->b_width << s->block_max_depth;
 723     const int w= p->width;
 724     const int h= p->height;
 725     int distortion= 0;
 726     int rate= 0;
 727     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 728
 729     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 730
 731     for(i=0; i<9; i++){
 732         int mb_x2= mb_x + (i%3) - 1;
 733         int mb_y2= mb_y + (i/3) - 1;
 734         int x= block_w*mb_x2 + block_w/2;
 735         int y= block_h*mb_y2 + block_h/2;
 736
 737         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 738                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 739
 740         //FIXME find a cleaner/simpler way to skip the outside stuff
 741         for(y2= y; y2<0; y2++)
 742             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 743         for(y2= h; y2<y+block_h; y2++)
 744             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 745         if(x<0){
 746             for(y2= y; y2<y+block_h; y2++)
 747                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 748         }
 749         if(x+block_w > w){
 750             for(y2= y; y2<y+block_h; y2++)
 751                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 752         }
 753
 754         av_assert1(block_w== 8 || block_w==16);
 755         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 756     }
 757
 758     if(plane_index==0){
 759         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 760         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 761
 762 /* ..RRRr
 763  * .RXXx.
 764  * .RXXx.
 765  * rxxx.
 766  */
 767         if(merged)
 768             rate = get_block_bits(s, mb_x, mb_y, 2);
 769         for(i=merged?4:0; i<9; i++){
 770             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 771             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 772         }
 773     }
 774     return distortion + rate*penalty_factor;
 775 }
 776
 777 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 778     const int w= b->width;
 779     const int h= b->height;
 780     int x, y;
 781
 782     if(1){
 783         int run=0;
 784         int *runs = s->run_buffer;
 785         int run_index=0;
 786         int max_index;
 787
 788         for(y=0; y<h; y++){
 789             for(x=0; x<w; x++){
 790                 int v, p=0;
 791                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 792                 v= src[x + y*stride];
 793
 794                 if(y){
 795                     t= src[x + (y-1)*stride];
 796                     if(x){
 797                         lt= src[x - 1 + (y-1)*stride];
 798                     }
 799                     if(x + 1 < w){
 800                         rt= src[x + 1 + (y-1)*stride];
 801                     }
 802                 }
 803                 if(x){
 804                     l= src[x - 1 + y*stride];
 805                     /*if(x > 1){
 806                         if(orientation==1) ll= src[y + (x-2)*stride];
 807                         else               ll= src[x - 2 + y*stride];
 808                     }*/
 809                 }
 810                 if(parent){
 811                     int px= x>>1;
 812                     int py= y>>1;
 813                     if(px<b->parent->width && py<b->parent->height)
 814                         p= parent[px + py*2*stride];
 815                 }
 816                 if(!(/*ll|*/l|lt|t|rt|p)){
 817                     if(v){
 818                         runs[run_index++]= run;
 819                         run=0;
 820                     }else{
 821                         run++;
 822                     }
 823                 }
 824             }
 825         }
 826         max_index= run_index;
 827         runs[run_index++]= run;
 828         run_index=0;
 829         run= runs[run_index++];
 830
 831         put_symbol2(&s->c, b->state[30], max_index, 0);
 832         if(run_index <= max_index)
 833             put_symbol2(&s->c, b->state[1], run, 3);
 834
 835         for(y=0; y<h; y++){
 836             if(s->c.bytestream_end - s->c.bytestream < w*40){
 837                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 838                 return AVERROR(ENOMEM);
 839             }
 840             for(x=0; x<w; x++){
 841                 int v, p=0;
 842                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 843                 v= src[x + y*stride];
 844
 845                 if(y){
 846                     t= src[x + (y-1)*stride];
 847                     if(x){
 848                         lt= src[x - 1 + (y-1)*stride];
 849                     }
 850                     if(x + 1 < w){
 851                         rt= src[x + 1 + (y-1)*stride];
 852                     }
 853                 }
 854                 if(x){
 855                     l= src[x - 1 + y*stride];
 856                     /*if(x > 1){
 857                         if(orientation==1) ll= src[y + (x-2)*stride];
 858                         else               ll= src[x - 2 + y*stride];
 859                     }*/
 860                 }
 861                 if(parent){
 862                     int px= x>>1;
 863                     int py= y>>1;
 864                     if(px<b->parent->width && py<b->parent->height)
 865                         p= parent[px + py*2*stride];
 866                 }
 867                 if(/*ll|*/l|lt|t|rt|p){
 868                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 869
 870                     put_rac(&s->c, &b->state[0][context], !!v);
 871                 }else{
 872                     if(!run){
 873                         run= runs[run_index++];
 874
 875                         if(run_index <= max_index)
 876                             put_symbol2(&s->c, b->state[1], run, 3);
 877                         av_assert2(v);
 878                     }else{
 879                         run--;
 880                         av_assert2(!v);
 881                     }
 882                 }
 883                 if(v){
 884                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 885                     int l2= 2*FFABS(l) + (l<0);
 886                     int t2= 2*FFABS(t) + (t<0);
 887
 888                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 889                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 890                 }
 891             }
 892         }
 893     }
 894     return 0;
 895 }
 896
 897 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 898 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 899 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 900     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 901 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 902 }
 903
 904 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){
 905     const int b_stride= s->b_width << s->block_max_depth;
 906     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 907     BlockNode backup= *block;
 908     unsigned value;
 909     int rd, index;
 910
 911     av_assert2(mb_x>=0 && mb_y>=0);
 912     av_assert2(mb_x<b_stride);
 913
 914     if(intra){
 915         block->color[0] = p[0];
 916         block->color[1] = p[1];
 917         block->color[2] = p[2];
 918         block->type |= BLOCK_INTRA;
 919     }else{
 920         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 921         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 922         if(s->me_cache[index] == value)
 923             return 0;
 924         s->me_cache[index]= value;
 925
 926         block->mx= p[0];
 927         block->my= p[1];
 928         block->type &= ~BLOCK_INTRA;
 929     }
 930
 931     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
 932
 933 //FIXME chroma
 934     if(rd < *best_rd){
 935         *best_rd= rd;
 936         return 1;
 937     }else{
 938         *block= backup;
 939         return 0;
 940     }
 941 }
 942
 943 /* special case for int[2] args we discard afterwards,
 944  * fixes compilation problem with gcc 2.95 */
 945 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){
 946     int p[2] = {p0, p1};
 947     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 948 }
 949
 950 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){
 951     const int b_stride= s->b_width << s->block_max_depth;
 952     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 953     BlockNode backup[4];
 954     unsigned value;
 955     int rd, index;
 956
 957     /* We don't initialize backup[] during variable declaration, because
 958      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 959      * 'int16_t'". */
 960     backup[0] = block[0];
 961     backup[1] = block[1];
 962     backup[2] = block[b_stride];
 963     backup[3] = block[b_stride + 1];
 964
 965     av_assert2(mb_x>=0 && mb_y>=0);
 966     av_assert2(mb_x<b_stride);
 967     av_assert2(((mb_x|mb_y)&1) == 0);
 968
 969     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 970     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 971     if(s->me_cache[index] == value)
 972         return 0;
 973     s->me_cache[index]= value;
 974
 975     block->mx= p0;
 976     block->my= p1;
 977     block->ref= ref;
 978     block->type &= ~BLOCK_INTRA;
 979     block[1]= block[b_stride]= block[b_stride+1]= *block;
 980
 981     rd= get_4block_rd(s, mb_x, mb_y, 0);
 982
 983 //FIXME chroma
 984     if(rd < *best_rd){
 985         *best_rd= rd;
 986         return 1;
 987     }else{
 988         block[0]= backup[0];
 989         block[1]= backup[1];
 990         block[b_stride]= backup[2];
 991         block[b_stride+1]= backup[3];
 992         return 0;
 993     }
 994 }
 995
 996 static void iterative_me(SnowContext *s){
 997     int pass, mb_x, mb_y;
 998     const int b_width = s->b_width  << s->block_max_depth;
 999     const int b_height= s->b_height << s->block_max_depth;
1000     const int b_stride= b_width;
1001     int color[3];
1002
1003     {
1004         RangeCoder r = s->c;
1005         uint8_t state[sizeof(s->block_state)];
1006         memcpy(state, s->block_state, sizeof(s->block_state));
1007         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1008             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1009                 encode_q_branch(s, 0, mb_x, mb_y);
1010         s->c = r;
1011         memcpy(s->block_state, state, sizeof(s->block_state));
1012     }
1013
1014     for(pass=0; pass<25; pass++){
1015         int change= 0;
1016
1017         for(mb_y= 0; mb_y<b_height; mb_y++){
1018             for(mb_x= 0; mb_x<b_width; mb_x++){
1019                 int dia_change, i, j, ref;
1020                 int best_rd= INT_MAX, ref_rd;
1021                 BlockNode backup, ref_b;
1022                 const int index= mb_x + mb_y * b_stride;
1023                 BlockNode *block= &s->block[index];
1024                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1025                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1026                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1027                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1028                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1029                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1030                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1031                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1032                 const int b_w= (MB_SIZE >> s->block_max_depth);
1033                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1034
1035                 if(pass && (block->type & BLOCK_OPT))
1036                     continue;
1037                 block->type |= BLOCK_OPT;
1038
1039                 backup= *block;
1040
1041                 if(!s->me_cache_generation)
1042                     memset(s->me_cache, 0, sizeof(s->me_cache));
1043                 s->me_cache_generation += 1<<22;
1044
1045                 //FIXME precalculate
1046                 {
1047                     int x, y;
1048                     for (y = 0; y < b_w * 2; y++)
1049                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1050                     if(mb_x==0)
1051                         for(y=0; y<b_w*2; y++)
1052                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1053                     if(mb_x==b_stride-1)
1054                         for(y=0; y<b_w*2; y++)
1055                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1056                     if(mb_y==0){
1057                         for(x=0; x<b_w*2; x++)
1058                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1059                         for(y=1; y<b_w; y++)
1060                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1061                     }
1062                     if(mb_y==b_height-1){
1063                         for(x=0; x<b_w*2; x++)
1064                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1065                         for(y=b_w; y<b_w*2-1; y++)
1066                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1067                     }
1068                 }
1069
1070                 //skip stuff outside the picture
1071                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1072                     uint8_t *src= s->  input_picture->data[0];
1073                     uint8_t *dst= s->current_picture->data[0];
1074                     const int stride= s->current_picture->linesize[0];
1075                     const int block_w= MB_SIZE >> s->block_max_depth;
1076                     const int block_h= MB_SIZE >> s->block_max_depth;
1077                     const int sx= block_w*mb_x - block_w/2;
1078                     const int sy= block_h*mb_y - block_h/2;
1079                     const int w= s->plane[0].width;
1080                     const int h= s->plane[0].height;
1081                     int y;
1082
1083                     for(y=sy; y<0; y++)
1084                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1085                     for(y=h; y<sy+block_h*2; y++)
1086                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1087                     if(sx<0){
1088                         for(y=sy; y<sy+block_h*2; y++)
1089                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1090                     }
1091                     if(sx+block_w*2 > w){
1092                         for(y=sy; y<sy+block_h*2; y++)
1093                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1094                     }
1095                 }
1096
1097                 // intra(black) = neighbors' contribution to the current block
1098                 for(i=0; i < s->nb_planes; i++)
1099                     color[i]= get_dc(s, mb_x, mb_y, i);
1100
1101                 // get previous score (cannot be cached due to OBMC)
1102                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1103                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1104                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1105                 }else
1106                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1107
1108                 ref_b= *block;
1109                 ref_rd= best_rd;
1110                 for(ref=0; ref < s->ref_frames; ref++){
1111                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1112                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1113                         continue;
1114                     block->ref= ref;
1115                     best_rd= INT_MAX;
1116
1117                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1118                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1119                     if(tb)
1120                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1121                     if(lb)
1122                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1123                     if(rb)
1124                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1125                     if(bb)
1126                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1127
1128                     /* fullpel ME */
1129                     //FIXME avoid subpel interpolation / round to nearest integer
1130                     do{
1131                         int newx = block->mx;
1132                         int newy = block->my;
1133                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1134                         dia_change=0;
1135                         for(i=0; i < dia_size; i++){
1136                             for(j=0; j<i; j++){
1137                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1138                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1139                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1140                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1141                             }
1142                         }
1143                     }while(dia_change);
1144                     /* subpel ME */
1145                     do{
1146                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1147                         dia_change=0;
1148                         for(i=0; i<8; i++)
1149                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1150                     }while(dia_change);
1151                     //FIXME or try the standard 2 pass qpel or similar
1152
1153                     mvr[0][0]= block->mx;
1154                     mvr[0][1]= block->my;
1155                     if(ref_rd > best_rd){
1156                         ref_rd= best_rd;
1157                         ref_b= *block;
1158                     }
1159                 }
1160                 best_rd= ref_rd;
1161                 *block= ref_b;
1162                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1163                 //FIXME RD style color selection
1164                 if(!same_block(block, &backup)){
1165                     if(tb ) tb ->type &= ~BLOCK_OPT;
1166                     if(lb ) lb ->type &= ~BLOCK_OPT;
1167                     if(rb ) rb ->type &= ~BLOCK_OPT;
1168                     if(bb ) bb ->type &= ~BLOCK_OPT;
1169                     if(tlb) tlb->type &= ~BLOCK_OPT;
1170                     if(trb) trb->type &= ~BLOCK_OPT;
1171                     if(blb) blb->type &= ~BLOCK_OPT;
1172                     if(brb) brb->type &= ~BLOCK_OPT;
1173                     change ++;
1174                 }
1175             }
1176         }
1177         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1178         if(!change)
1179             break;
1180     }
1181
1182     if(s->block_max_depth == 1){
1183         int change= 0;
1184         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1185             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1186                 int i;
1187                 int best_rd, init_rd;
1188                 const int index= mb_x + mb_y * b_stride;
1189                 BlockNode *b[4];
1190
1191                 b[0]= &s->block[index];
1192                 b[1]= b[0]+1;
1193                 b[2]= b[0]+b_stride;
1194                 b[3]= b[2]+1;
1195                 if(same_block(b[0], b[1]) &&
1196                    same_block(b[0], b[2]) &&
1197                    same_block(b[0], b[3]))
1198                     continue;
1199
1200                 if(!s->me_cache_generation)
1201                     memset(s->me_cache, 0, sizeof(s->me_cache));
1202                 s->me_cache_generation += 1<<22;
1203
1204                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1205
1206                 //FIXME more multiref search?
1207                 check_4block_inter(s, mb_x, mb_y,
1208                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1209                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1210
1211                 for(i=0; i<4; i++)
1212                     if(!(b[i]->type&BLOCK_INTRA))
1213                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1214
1215                 if(init_rd != best_rd)
1216                     change++;
1217             }
1218         }
1219         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1220     }
1221 }
1222
1223 static void encode_blocks(SnowContext *s, int search){
1224     int x, y;
1225     int w= s->b_width;
1226     int h= s->b_height;
1227
1228     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1229         iterative_me(s);
1230
1231     for(y=0; y<h; y++){
1232         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1233             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1234             return;
1235         }
1236         for(x=0; x<w; x++){
1237             if(s->motion_est == FF_ME_ITER || !search)
1238                 encode_q_branch2(s, 0, x, y);
1239             else
1240                 encode_q_branch (s, 0, x, y);
1241         }
1242     }
1243 }
1244
1245 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1246     const int w= b->width;
1247     const int h= b->height;
1248     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1249     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1250     int x,y, thres1, thres2;
1251
1252     if(s->qlog == LOSSLESS_QLOG){
1253         for(y=0; y<h; y++)
1254             for(x=0; x<w; x++)
1255                 dst[x + y*stride]= src[x + y*stride];
1256         return;
1257     }
1258
1259     bias= bias ? 0 : (3*qmul)>>3;
1260     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1261     thres2= 2*thres1;
1262
1263     if(!bias){
1264         for(y=0; y<h; y++){
1265             for(x=0; x<w; x++){
1266                 int i= src[x + y*stride];
1267
1268                 if((unsigned)(i+thres1) > thres2){
1269                     if(i>=0){
1270                         i<<= QEXPSHIFT;
1271                         i/= qmul; //FIXME optimize
1272                         dst[x + y*stride]=  i;
1273                     }else{
1274                         i= -i;
1275                         i<<= QEXPSHIFT;
1276                         i/= qmul; //FIXME optimize
1277                         dst[x + y*stride]= -i;
1278                     }
1279                 }else
1280                     dst[x + y*stride]= 0;
1281             }
1282         }
1283     }else{
1284         for(y=0; y<h; y++){
1285             for(x=0; x<w; x++){
1286                 int i= src[x + y*stride];
1287
1288                 if((unsigned)(i+thres1) > thres2){
1289                     if(i>=0){
1290                         i<<= QEXPSHIFT;
1291                         i= (i + bias) / qmul; //FIXME optimize
1292                         dst[x + y*stride]=  i;
1293                     }else{
1294                         i= -i;
1295                         i<<= QEXPSHIFT;
1296                         i= (i + bias) / qmul; //FIXME optimize
1297                         dst[x + y*stride]= -i;
1298                     }
1299                 }else
1300                     dst[x + y*stride]= 0;
1301             }
1302         }
1303     }
1304 }
1305
1306 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1307     const int w= b->width;
1308     const int h= b->height;
1309     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1310     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1311     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1312     int x,y;
1313
1314     if(s->qlog == LOSSLESS_QLOG) return;
1315
1316     for(y=0; y<h; y++){
1317         for(x=0; x<w; x++){
1318             int i= src[x + y*stride];
1319             if(i<0){
1320                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1321             }else if(i>0){
1322                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1323             }
1324         }
1325     }
1326 }
1327
1328 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1329     const int w= b->width;
1330     const int h= b->height;
1331     int x,y;
1332
1333     for(y=h-1; y>=0; y--){
1334         for(x=w-1; x>=0; x--){
1335             int i= x + y*stride;
1336
1337             if(x){
1338                 if(use_median){
1339                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1340                     else  src[i] -= src[i - 1];
1341                 }else{
1342                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1343                     else  src[i] -= src[i - 1];
1344                 }
1345             }else{
1346                 if(y) src[i] -= src[i - stride];
1347             }
1348         }
1349     }
1350 }
1351
1352 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1353     const int w= b->width;
1354     const int h= b->height;
1355     int x,y;
1356
1357     for(y=0; y<h; y++){
1358         for(x=0; x<w; x++){
1359             int i= x + y*stride;
1360
1361             if(x){
1362                 if(use_median){
1363                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1364                     else  src[i] += src[i - 1];
1365                 }else{
1366                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1367                     else  src[i] += src[i - 1];
1368                 }
1369             }else{
1370                 if(y) src[i] += src[i - stride];
1371             }
1372         }
1373     }
1374 }
1375
1376 static void encode_qlogs(SnowContext *s){
1377     int plane_index, level, orientation;
1378
1379     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1380         for(level=0; level<s->spatial_decomposition_count; level++){
1381             for(orientation=level ? 1:0; orientation<4; orientation++){
1382                 if(orientation==2) continue;
1383                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1384             }
1385         }
1386     }
1387 }
1388
1389 static void encode_header(SnowContext *s){
1390     int plane_index, i;
1391     uint8_t kstate[32];
1392
1393     memset(kstate, MID_STATE, sizeof(kstate));
1394
1395     put_rac(&s->c, kstate, s->keyframe);
1396     if(s->keyframe || s->always_reset){
1397         ff_snow_reset_contexts(s);
1398         s->last_spatial_decomposition_type=
1399         s->last_qlog=
1400         s->last_qbias=
1401         s->last_mv_scale=
1402         s->last_block_max_depth= 0;
1403         for(plane_index=0; plane_index<2; plane_index++){
1404             Plane *p= &s->plane[plane_index];
1405             p->last_htaps=0;
1406             p->last_diag_mc=0;
1407             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1408         }
1409     }
1410     if(s->keyframe){
1411         put_symbol(&s->c, s->header_state, s->version, 0);
1412         put_rac(&s->c, s->header_state, s->always_reset);
1413         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1414         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1415         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1416         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1417         if (s->nb_planes > 2) {
1418             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1419             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1420         }
1421         put_rac(&s->c, s->header_state, s->spatial_scalability);
1422 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1423         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1424
1425         encode_qlogs(s);
1426     }
1427
1428     if(!s->keyframe){
1429         int update_mc=0;
1430         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1431             Plane *p= &s->plane[plane_index];
1432             update_mc |= p->last_htaps   != p->htaps;
1433             update_mc |= p->last_diag_mc != p->diag_mc;
1434             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1435         }
1436         put_rac(&s->c, s->header_state, update_mc);
1437         if(update_mc){
1438             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1439                 Plane *p= &s->plane[plane_index];
1440                 put_rac(&s->c, s->header_state, p->diag_mc);
1441                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1442                 for(i= p->htaps/2; i; i--)
1443                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1444             }
1445         }
1446         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1447             put_rac(&s->c, s->header_state, 1);
1448             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1449             encode_qlogs(s);
1450         }else
1451             put_rac(&s->c, s->header_state, 0);
1452     }
1453
1454     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1455     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1456     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1457     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1458     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1459
1460 }
1461
1462 static void update_last_header_values(SnowContext *s){
1463     int plane_index;
1464
1465     if(!s->keyframe){
1466         for(plane_index=0; plane_index<2; plane_index++){
1467             Plane *p= &s->plane[plane_index];
1468             p->last_diag_mc= p->diag_mc;
1469             p->last_htaps  = p->htaps;
1470             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1471         }
1472     }
1473
1474     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1475     s->last_qlog                        = s->qlog;
1476     s->last_qbias                       = s->qbias;
1477     s->last_mv_scale                    = s->mv_scale;
1478     s->last_block_max_depth             = s->block_max_depth;
1479     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1480 }
1481
1482 static int qscale2qlog(int qscale){
1483     return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1484            + 61*QROOT/8; ///< 64 > 60
1485 }
1486
1487 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1488 {
1489     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1490      * FIXME we know exact mv bits at this point,
1491      * but ratecontrol isn't set up to include them. */
1492     uint32_t coef_sum= 0;
1493     int level, orientation, delta_qlog;
1494
1495     for(level=0; level<s->spatial_decomposition_count; level++){
1496         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1497             SubBand *b= &s->plane[0].band[level][orientation];
1498             IDWTELEM *buf= b->ibuf;
1499             const int w= b->width;
1500             const int h= b->height;
1501             const int stride= b->stride;
1502             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1503             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1504             const int qdiv= (1<<16)/qmul;
1505             int x, y;
1506             //FIXME this is ugly
1507             for(y=0; y<h; y++)
1508                 for(x=0; x<w; x++)
1509                     buf[x+y*stride]= b->buf[x+y*stride];
1510             if(orientation==0)
1511                 decorrelate(s, b, buf, stride, 1, 0);
1512             for(y=0; y<h; y++)
1513                 for(x=0; x<w; x++)
1514                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1515         }
1516     }
1517
1518     /* ugly, ratecontrol just takes a sqrt again */
1519     av_assert0(coef_sum < INT_MAX);
1520     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1521
1522     if(pict->pict_type == AV_PICTURE_TYPE_I){
1523         s->m.current_picture.mb_var_sum= coef_sum;
1524         s->m.current_picture.mc_mb_var_sum= 0;
1525     }else{
1526         s->m.current_picture.mc_mb_var_sum= coef_sum;
1527         s->m.current_picture.mb_var_sum= 0;
1528     }
1529
1530     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1531     if (pict->quality < 0)
1532         return INT_MIN;
1533     s->lambda= pict->quality * 3/2;
1534     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1535     s->qlog+= delta_qlog;
1536     return delta_qlog;
1537 }
1538
1539 static void calculate_visual_weight(SnowContext *s, Plane *p){
1540     int width = p->width;
1541     int height= p->height;
1542     int level, orientation, x, y;
1543
1544     for(level=0; level<s->spatial_decomposition_count; level++){
1545         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1546             SubBand *b= &p->band[level][orientation];
1547             IDWTELEM *ibuf= b->ibuf;
1548             int64_t error=0;
1549
1550             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1551             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1552             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1553             for(y=0; y<height; y++){
1554                 for(x=0; x<width; x++){
1555                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1556                     error += d*d;
1557                 }
1558             }
1559
1560             b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1561         }
1562     }
1563 }
1564
1565 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1566                         const AVFrame *pict, int *got_packet)
1567 {
1568     SnowContext *s = avctx->priv_data;
1569     RangeCoder * const c= &s->c;
1570     AVFrame *pic;
1571     const int width= s->avctx->width;
1572     const int height= s->avctx->height;
1573     int level, orientation, plane_index, i, y, ret;
1574     uint8_t rc_header_bak[sizeof(s->header_state)];
1575     uint8_t rc_block_bak[sizeof(s->block_state)];
1576
1577     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)
1578         return ret;
1579
1580     ff_init_range_encoder(c, pkt->data, pkt->size);
1581     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1582
1583     for(i=0; i < s->nb_planes; i++){
1584         int hshift= i ? s->chroma_h_shift : 0;
1585         int vshift= i ? s->chroma_v_shift : 0;
1586         for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1587             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1588                    &pict->data[i][y * pict->linesize[i]],
1589                    AV_CEIL_RSHIFT(width, hshift));
1590         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1591                                 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1592                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1593                                 EDGE_TOP | EDGE_BOTTOM);
1594
1595     }
1596     emms_c();
1597     pic = s->input_picture;
1598     pic->pict_type = pict->pict_type;
1599     pic->quality = pict->quality;
1600
1601     s->m.picture_number= avctx->frame_number;
1602     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1603         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1604         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1605         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1606             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1607             if (pic->quality < 0)
1608                 return -1;
1609         }
1610     }else{
1611         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1612         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1613     }
1614
1615     if(s->pass1_rc && avctx->frame_number == 0)
1616         pic->quality = 2*FF_QP2LAMBDA;
1617     if (pic->quality) {
1618         s->qlog   = qscale2qlog(pic->quality);
1619         s->lambda = pic->quality * 3/2;
1620     }
1621     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1622         s->qlog= LOSSLESS_QLOG;
1623         s->lambda = 0;
1624     }//else keep previous frame's qlog until after motion estimation
1625
1626     if (s->current_picture->data[0]) {
1627         int w = s->avctx->width;
1628         int h = s->avctx->height;
1629
1630         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1631                                 s->current_picture->linesize[0], w   , h   ,
1632                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1633         if (s->current_picture->data[2]) {
1634             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1635                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1636                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1637             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1638                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1639                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1640         }
1641         emms_c();
1642     }
1643
1644     ff_snow_frame_start(s);
1645 #if FF_API_CODED_FRAME
1646 FF_DISABLE_DEPRECATION_WARNINGS
1647     av_frame_unref(avctx->coded_frame);
1648     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1649 FF_ENABLE_DEPRECATION_WARNINGS
1650 #endif
1651     if (ret < 0)
1652         return ret;
1653
1654     s->m.current_picture_ptr= &s->m.current_picture;
1655     s->m.current_picture.f = s->current_picture;
1656     s->m.current_picture.f->pts = pict->pts;
1657     if(pic->pict_type == AV_PICTURE_TYPE_P){
1658         int block_width = (width +15)>>4;
1659         int block_height= (height+15)>>4;
1660         int stride= s->current_picture->linesize[0];
1661
1662         av_assert0(s->current_picture->data[0]);
1663         av_assert0(s->last_picture[0]->data[0]);
1664
1665         s->m.avctx= s->avctx;
1666         s->m.   last_picture.f = s->last_picture[0];
1667         s->m.    new_picture.f = s->input_picture;
1668         s->m.   last_picture_ptr= &s->m.   last_picture;
1669         s->m.linesize = stride;
1670         s->m.uvlinesize= s->current_picture->linesize[1];
1671         s->m.width = width;
1672         s->m.height= height;
1673         s->m.mb_width = block_width;
1674         s->m.mb_height= block_height;
1675         s->m.mb_stride=   s->m.mb_width+1;
1676         s->m.b8_stride= 2*s->m.mb_width+1;
1677         s->m.f_code=1;
1678         s->m.pict_type = pic->pict_type;
1679         s->m.motion_est= s->motion_est;
1680         s->m.me.scene_change_score=0;
1681         s->m.me.dia_size = avctx->dia_size;
1682         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1683         s->m.out_format= FMT_H263;
1684         s->m.unrestricted_mv= 1;
1685
1686         s->m.lambda = s->lambda;
1687         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1688         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1689
1690         s->m.mecc= s->mecc; //move
1691         s->m.qdsp= s->qdsp; //move
1692         s->m.hdsp = s->hdsp;
1693         ff_init_me(&s->m);
1694         s->hdsp = s->m.hdsp;
1695         s->mecc= s->m.mecc;
1696     }
1697
1698     if(s->pass1_rc){
1699         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1700         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1701     }
1702
1703 redo_frame:
1704
1705     s->spatial_decomposition_count= 5;
1706
1707     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1708           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1709         s->spatial_decomposition_count--;
1710
1711     if (s->spatial_decomposition_count <= 0) {
1712         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1713         return AVERROR(EINVAL);
1714     }
1715
1716     s->m.pict_type = pic->pict_type;
1717     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1718
1719     ff_snow_common_init_after_header(avctx);
1720
1721     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1722         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1723             calculate_visual_weight(s, &s->plane[plane_index]);
1724         }
1725     }
1726
1727     encode_header(s);
1728     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1729     encode_blocks(s, 1);
1730     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1731
1732     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1733         Plane *p= &s->plane[plane_index];
1734         int w= p->width;
1735         int h= p->height;
1736         int x, y;
1737 //        int bits= put_bits_count(&s->c.pb);
1738
1739         if (!s->memc_only) {
1740             //FIXME optimize
1741             if(pict->data[plane_index]) //FIXME gray hack
1742                 for(y=0; y<h; y++){
1743                     for(x=0; x<w; x++){
1744                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1745                     }
1746                 }
1747             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1748
1749 #if FF_API_PRIVATE_OPT
1750 FF_DISABLE_DEPRECATION_WARNINGS
1751             if(s->avctx->scenechange_threshold)
1752                 s->scenechange_threshold = s->avctx->scenechange_threshold;
1753 FF_ENABLE_DEPRECATION_WARNINGS
1754 #endif
1755
1756             if(   plane_index==0
1757                && pic->pict_type == AV_PICTURE_TYPE_P
1758                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1759                && s->m.me.scene_change_score > s->scenechange_threshold){
1760                 ff_init_range_encoder(c, pkt->data, pkt->size);
1761                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1762                 pic->pict_type= AV_PICTURE_TYPE_I;
1763                 s->keyframe=1;
1764                 s->current_picture->key_frame=1;
1765                 goto redo_frame;
1766             }
1767
1768             if(s->qlog == LOSSLESS_QLOG){
1769                 for(y=0; y<h; y++){
1770                     for(x=0; x<w; x++){
1771                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1772                     }
1773                 }
1774             }else{
1775                 for(y=0; y<h; y++){
1776                     for(x=0; x<w; x++){
1777                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1778                     }
1779                 }
1780             }
1781
1782             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1783
1784             if(s->pass1_rc && plane_index==0){
1785                 int delta_qlog = ratecontrol_1pass(s, pic);
1786                 if (delta_qlog <= INT_MIN)
1787                     return -1;
1788                 if(delta_qlog){
1789                     //reordering qlog in the bitstream would eliminate this reset
1790                     ff_init_range_encoder(c, pkt->data, pkt->size);
1791                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1792                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1793                     encode_header(s);
1794                     encode_blocks(s, 0);
1795                 }
1796             }
1797
1798             for(level=0; level<s->spatial_decomposition_count; level++){
1799                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1800                     SubBand *b= &p->band[level][orientation];
1801
1802                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1803                     if(orientation==0)
1804                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1805                     if (!s->no_bitstream)
1806                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1807                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1808                     if(orientation==0)
1809                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1810                 }
1811             }
1812
1813             for(level=0; level<s->spatial_decomposition_count; level++){
1814                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1815                     SubBand *b= &p->band[level][orientation];
1816
1817                     dequantize(s, b, b->ibuf, b->stride);
1818                 }
1819             }
1820
1821             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1822             if(s->qlog == LOSSLESS_QLOG){
1823                 for(y=0; y<h; y++){
1824                     for(x=0; x<w; x++){
1825                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1826                     }
1827                 }
1828             }
1829             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1830         }else{
1831             //ME/MC only
1832             if(pic->pict_type == AV_PICTURE_TYPE_I){
1833                 for(y=0; y<h; y++){
1834                     for(x=0; x<w; x++){
1835                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1836                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1837                     }
1838                 }
1839             }else{
1840                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1841                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1842             }
1843         }
1844         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1845             int64_t error= 0;
1846
1847             if(pict->data[plane_index]) //FIXME gray hack
1848                 for(y=0; y<h; y++){
1849                     for(x=0; x<w; x++){
1850                         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];
1851                         error += d*d;
1852                     }
1853                 }
1854             s->avctx->error[plane_index] += error;
1855             s->encoding_error[plane_index] = error;
1856         }
1857
1858     }
1859     emms_c();
1860
1861     update_last_header_values(s);
1862
1863     ff_snow_release_buffer(avctx);
1864
1865     s->current_picture->coded_picture_number = avctx->frame_number;
1866     s->current_picture->pict_type = pic->pict_type;
1867     s->current_picture->quality = pic->quality;
1868     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1869     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1870     s->m.current_picture.f->display_picture_number =
1871     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1872     s->m.current_picture.f->quality                = pic->quality;
1873     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1874     if(s->pass1_rc)
1875         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1876             return -1;
1877     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1878         ff_write_pass1_stats(&s->m);
1879     s->m.last_pict_type = s->m.pict_type;
1880 #if FF_API_STAT_BITS
1881 FF_DISABLE_DEPRECATION_WARNINGS
1882     avctx->frame_bits = s->m.frame_bits;
1883     avctx->mv_bits = s->m.mv_bits;
1884     avctx->misc_bits = s->m.misc_bits;
1885     avctx->p_tex_bits = s->m.p_tex_bits;
1886 FF_ENABLE_DEPRECATION_WARNINGS
1887 #endif
1888
1889     emms_c();
1890
1891     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1892                                    s->encoding_error,
1893                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1894                                    s->current_picture->pict_type);
1895
1896 #if FF_API_ERROR_FRAME
1897 FF_DISABLE_DEPRECATION_WARNINGS
1898     memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1899 FF_ENABLE_DEPRECATION_WARNINGS
1900 #endif
1901
1902     pkt->size = ff_rac_terminate(c);
1903     if (s->current_picture->key_frame)
1904         pkt->flags |= AV_PKT_FLAG_KEY;
1905     *got_packet = 1;
1906
1907     return 0;
1908 }
1909
1910 static av_cold int encode_end(AVCodecContext *avctx)
1911 {
1912     SnowContext *s = avctx->priv_data;
1913
1914     ff_snow_common_end(s);
1915     ff_rate_control_uninit(&s->m);
1916     av_frame_free(&s->input_picture);
1917     av_freep(&avctx->stats_out);
1918
1919     return 0;
1920 }
1921
1922 #define OFFSET(x) offsetof(SnowContext, x)
1923 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1924 static const AVOption options[] = {
1925     {"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" },
1926     { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1927     { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1928     { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1929     { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1930     { "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 },
1931     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1932     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1933     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1934     { "sc_threshold",   "Scene change threshold",                   OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1935     { "pred",           "Spatial decomposition type",                                OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1936         { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1937         { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1938     { NULL },
1939 };
1940
1941 static const AVClass snowenc_class = {
1942     .class_name = "snow encoder",
1943     .item_name  = av_default_item_name,
1944     .option     = options,
1945     .version    = LIBAVUTIL_VERSION_INT,
1946 };
1947
1948 AVCodec ff_snow_encoder = {
1949     .name           = "snow",
1950     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1951     .type           = AVMEDIA_TYPE_VIDEO,
1952     .id             = AV_CODEC_ID_SNOW,
1953     .priv_data_size = sizeof(SnowContext),
1954     .init           = encode_init,
1955     .encode2        = encode_frame,
1956     .close          = encode_end,
1957     .pix_fmts       = (const enum AVPixelFormat[]){
1958         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1959         AV_PIX_FMT_GRAY8,
1960         AV_PIX_FMT_NONE
1961     },
1962     .priv_class     = &snowenc_class,
1963     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1964                       FF_CODEC_CAP_INIT_CLEANUP,
1965 };